Re: Tracing over a suspend/resume cycle

["Bruno Prémont" <bonbons@linux-vserver.org>]

[-- Attachment #1: Type: text/plain, Size: 4148 bytes --]

Hello,

Attached is a tarball with script and log of such a suspend cycle.

The tarball contains:
- config: kernel config
- dmesg.log: kernel log
- mmiotrace.sh script used
  (/sbin/tinyinit is a shell script that runs a few things to
  mount virtual filesystems like proc, sys, tmp; pupulate /dev;
  run sshd; start netconsole; root partition remains RO)
- mmiotrace.log is output of mmiotrace.sh
  (note, I inserted the comment regarding "rmmod nvidia" for the
  sequence as to when I did it)
- trace.log is what came out of /sys/kernel/debug/tracing/trace_pipe


System details:
00:00.0 Host bridge [0600]: nVidia Corporation nForce CPU bridge [10de:01a4] (rev b2)
00:00.1 RAM memory [0500]: nVidia Corporation nForce 220/420 Memory Controller [10de:01ac] (rev b2)
00:00.2 RAM memory [0500]: nVidia Corporation nForce 220/420 Memory Controller [10de:01ad] (rev b2)
00:00.3 RAM memory [0500]: nVidia Corporation nForce 420 Memory Controller (DDR) [10de:01ab] (rev b2)
00:01.0 ISA bridge [0601]: nVidia Corporation nForce ISA Bridge [10de:01b2] (rev c3)
00:01.1 SMBus [0c05]: nVidia Corporation nForce PCI System Management [10de:01b4] (rev c1)
00:02.0 USB Controller [0c03]: nVidia Corporation nForce USB Controller [10de:01c2] (rev c3)
00:03.0 USB Controller [0c03]: nVidia Corporation nForce USB Controller [10de:01c2] (rev c3)
00:04.0 Ethernet controller [0200]: nVidia Corporation nForce Ethernet Controller [10de:01c3] (rev c2)
00:05.0 Multimedia audio controller [0401]: nVidia Corporation nForce Audio Processing Unit [10de:01b0] (rev c2)
00:06.0 Multimedia audio controller [0401]: nVidia Corporation nForce AC'97 Audio Controller [10de:01b1] (rev c2)
00:08.0 PCI bridge [0604]: nVidia Corporation nForce PCI-to-PCI bridge [10de:01b8] (rev c2)
00:09.0 IDE interface [0101]: nVidia Corporation nForce IDE [10de:01bc] (rev c3)
00:1e.0 PCI bridge [0604]: nVidia Corporation nForce AGP to PCI Bridge [10de:01b7] (rev b2)
01:04.0 RAID bus controller [0104]: HighPoint Technologies, Inc. HPT374 [1103:0008] (rev 07)
01:04.1 RAID bus controller [0104]: HighPoint Technologies, Inc. HPT374 [1103:0008] (rev 07)
02:00.0 VGA compatible controller [0300]: nVidia Corporation NVCrush11 [GeForce2 MX Integrated Graphics] [10de:01a0] (rev b1)

Kernel used:
Nearly vanilla (just one patch of mine for pata_hpt37x so it does most
init operations during resume - I will post that patch later on).

Distribution is Gentoo

Bruno


On Sun, 07 June 2009 Pekka Paalanen <pq@iki.fi> wrote:
> Hello,
> 
> a long time ago I had reports, that mmiotracing over a suspend/resume
> cycle worked fine. Perhaps by accident, but there were successful
> runs.
> 
> Right now, it does not work for the user "bonbons". Nothing after
> resume gets to the log, even though 'cat .../trace_pipe' is still
> running. Kernel log shows that mmiotrace is still noticing ioremaps
> and iounmaps, but either they don't get into the ring buffer, or the
> ring buffer is somehow disabled. The mmiotrace session is practically
> dead. Switching to the "nop" tracer and then back to "mmiotrace"
> makes things work again.
> 
> If you need further information about this case, I hope bonbons can
> provide that. And also correct me if I got something wrong. His kernel
> version is 2.6.30-rc7-git4 and suspending here means
> echo mem > /sys/power/state
> 
> Steven, you probably know the tracing and ring buffer framework very
> well. Are there any known issues with tracing over a suspend/resume
> cycle? Is it suspend/resume aware at all? Should it be to make this
> work?
> 
> A traced driver may do a lot of things while processing suspend and
> resume events, but here mmiotrace does not say it's losing events.
> 
> Now that I think of it, it sounds like a very difficult problem to
> make tracing over suspend reliable. The user space should be kept
> running, so that the ring buffer may drain, and at the same time
> the drivers should be doing their suspending. Is this even possible
> by other means than pure luck (of not running out of free ring buffer
> space)?
> 
> Tracing over a suspend/resume cycle is useful for e.g. spying on
> nvidia how they suspend their GPU.
> 

[-- Attachment #2: mmiotrace.tar.gz --]
[-- Type: application/x-gzip, Size: 28971 bytes --]