From: Taras <tg...@mo...> - 2008-10-20 19:10:08
|
Hi, What's the status of oprofile on atom? Are there any up to date patches I can try? I tried http://thread.gmane.org/gmane.linux.kernel/723941 patch which seems like something that would enable the NMI interrupt, but force_arch_perfmon does not work with it on my Asus eee 901. Reason why it doesn't work is that cpu_has_arch_perfmon is false in my 2.6.27 kernel. Commenting out checks for that puts oprofile module into NMI interrupt mode, but unsurprisingly things do not work and oprofile userspace hangs. Thanks, Taras |
From: Andi K. <an...@fi...> - 2008-10-21 17:11:23
|
Taras <tg...@mo...> writes: > Hi, > What's the status of oprofile on atom? Are there any up to date patches > I can try? It should work with the architectural perfmon support on Atom. You need an 2.6.28 git kernel (or 2.6.27 + ftp://firstfloor.org/pub/ak/oprofile/) and a post 0.9.3 oprofile userland (or the userland patches from the site above) > I tried http://thread.gmane.org/gmane.linux.kernel/723941 patch which > seems like something that would enable the NMI interrupt, but > force_arch_perfmon does not work with it on my Asus eee 901. > > Reason why it doesn't work is that cpu_has_arch_perfmon is false in my Hmm? My eeepc 901 has arch perfmon. -Andi -- ak...@li... |
From: Robert R. <rob...@am...> - 2008-10-21 22:23:47
|
On 21.10.08 19:11:07, Andi Kleen wrote: > Taras <tg...@mo...> writes: > > > Hi, > > What's the status of oprofile on atom? Are there any up to date patches > > I can try? > > It should work with the architectural perfmon support on Atom. > > You need an 2.6.28 git kernel (or 2.6.27 + ftp://firstfloor.org/pub/ak/oprofile/) > and a post 0.9.3 oprofile userland (or the userland patches from the site > above) Taras, the arch perfmon patches are scheduled for v2.6.28 but not yet upstream. You can use also the latest oprofile kernel in the 'next' branch of git://git.kernel.org/pub/scm/linux/kernel/git/rric/oprofile.git -Robert -- Advanced Micro Devices, Inc. Operating System Research Center email: rob...@am... |
From: Taras G. <tg...@mo...> - 2008-10-22 21:04:20
|
Andi Kleen wrote: > Taras <tg...@mo...> writes: > > >> Hi, >> What's the status of oprofile on atom? Are there any up to date patches >> I can try? >> > > It should work with the architectural perfmon support on Atom. > > You need an 2.6.28 git kernel (or 2.6.27 + ftp://firstfloor.org/pub/ak/oprofile/) > and a post 0.9.3 oprofile userland (or the userland patches from the site > above) > > I retried with 2.6.27 + patches from above: modprobe oprofile force_arch_perfmon=1 yields oprofile: using timer interrupt. in dmesg. Are there any .config options for enabling cpu_has_arch_perfmon in the kernel? > > Hmm? My eeepc 901 has arch perfmon. > Glad to hear, I wonder why it doesn't work here. Taras |
From: Andi K. <an...@fi...> - 2008-10-22 21:09:15
|
On Wed, Oct 22, 2008 at 02:04:10PM -0700, Taras Glek wrote: > Andi Kleen wrote: > >Taras <tg...@mo...> writes: > > > > > >>Hi, > >>What's the status of oprofile on atom? Are there any up to date patches > >>I can try? > >> > > > >It should work with the architectural perfmon support on Atom. > > > >You need an 2.6.28 git kernel (or 2.6.27 + > >ftp://firstfloor.org/pub/ak/oprofile/) > >and a post 0.9.3 oprofile userland (or the userland patches from the site > >above) > > > > > I retried with 2.6.27 + patches from above: > modprobe oprofile force_arch_perfmon=1 > yields > oprofile: using timer interrupt. > in dmesg. Do you see arch_perfmon in /proc/cpuinfo? > > Are there any .config options for enabling cpu_has_arch_perfmon in the > kernel? No. -Andi -- ak...@li... |
From: Taras G. <tg...@mo...> - 2008-10-22 21:16:42
|
Andi Kleen wrote: > On Wed, Oct 22, 2008 at 02:04:10PM -0700, Taras Glek wrote: > >> Andi Kleen wrote: >> >>> Taras <tg...@mo...> writes: >>> >>> >>> >>>> Hi, >>>> What's the status of oprofile on atom? Are there any up to date patches >>>> I can try? >>>> >>>> >>> It should work with the architectural perfmon support on Atom. >>> >>> You need an 2.6.28 git kernel (or 2.6.27 + >>> ftp://firstfloor.org/pub/ak/oprofile/) >>> and a post 0.9.3 oprofile userland (or the userland patches from the site >>> above) >>> >>> >>> >> I retried with 2.6.27 + patches from above: >> modprobe oprofile force_arch_perfmon=1 >> yields >> oprofile: using timer interrupt. >> in dmesg. >> > > Do you see arch_perfmon in /proc/cpuinfo? > > No Taras |
From: Taras G. <tg...@mo...> - 2008-10-22 22:55:08
|
Andi Kleen wrote: > On Wed, Oct 22, 2008 at 02:04:10PM -0700, Taras Glek wrote: > >> Andi Kleen wrote: >> >>> Taras <tg...@mo...> writes: >>> >>> >>> >>>> Hi, >>>> What's the status of oprofile on atom? Are there any up to date patches >>>> I can try? >>>> >>>> >>> It should work with the architectural perfmon support on Atom. >>> >>> You need an 2.6.28 git kernel (or 2.6.27 + >>> ftp://firstfloor.org/pub/ak/oprofile/) >>> and a post 0.9.3 oprofile userland (or the userland patches from the site >>> above) >>> >>> >>> >> I retried with 2.6.27 + patches from above: >> modprobe oprofile force_arch_perfmon=1 >> yields >> oprofile: using timer interrupt. >> in dmesg. >> > > Do you see arch_perfmon in /proc/cpuinfo? > Ok, got it to appear by disabling hyperthreading in the bios. Not the most obvious thing :) Thanks, Taras |
From: Andi K. <an...@fi...> - 2008-10-22 22:40:55
|
> Ok, got it to appear by disabling hyperthreading in the bios. Not the > most obvious thing :) I see it with hyperthreading. Maybe it's some BIOS problem. This machine has a 06/12/2008 BIOS. Anyways I assume oprofile works now? -Andi -- ak...@li... |
From: Taras G. <tg...@mo...> - 2008-10-22 22:42:43
|
Andi Kleen wrote: >> Ok, got it to appear by disabling hyperthreading in the bios. Not the >> most obvious thing :) >> > > I see it with hyperthreading. Maybe it's some BIOS problem. > > This machine has a 06/12/2008 BIOS. > > Anyways I assume oprofile works now? > It sort of works. I'm using the cvs oprofile and it fails to parse the events file. opcontrol --init opcontrol --vmlinux=/boot/vmlinux-2.6.27.2-oprofile --start which gave: oprofile: parse error in /usr/local/share/oprofile//i386/arch_perfmon/events, line 6 invalid event name Using default event: CPU_CLK_UNHALTED:100000:0:1:1 oprofile: parse error in /usr/local/share/oprofile//i386/arch_perfmon/events, line 6 invalid event name So I deleted offending line and tried again. This time it started: Using default event: CPU_CLK_UNHALTED:100000:0:1:1 /usr/local/bin/opcontrol: 1: let: not found test: 1: Illegal number: /usr/local/bin/opcontrol: 1: let: not found Using 2.6+ OProfile kernel interface. Reading module info. Using log file /var/lib/oprofile/samples/oprofiled.log Daemon started. Profiler running. /usr/local/bin/oprofiled --session-dir=/var/lib/oprofile --separate-lib=0 --separate-kernel=0 --separate-thread=0 --separate-cpu=0 --events=CPU_CLK_UNHALTED:60:0:100000:0:1:1, --vmlinux=/boot/vmlinux-2.6.27.2-oprofile --kernel-range=c0101000,c04b5d32 However opcontrol --dump only seems to record samples the first time it is called. It also doesn't seem to record anything at all if I do --separate=kernel,library Taras |
From: Andi K. <an...@fi...> - 2008-10-23 00:03:45
|
On Wed, Oct 22, 2008 at 03:42:34PM -0700, Taras Glek wrote: > Andi Kleen wrote: > >>Ok, got it to appear by disabling hyperthreading in the bios. Not the > >>most obvious thing :) > >> > > > >I see it with hyperthreading. Maybe it's some BIOS problem. > > > >This machine has a 06/12/2008 BIOS. > > > >Anyways I assume oprofile works now? > > > It sort of works. I'm using the cvs oprofile and it fails to parse the > events file. Hmm my version worked. Must be some problem in CVS. I already got an earlier report, but I hadn't had time to investigate that yet. Anyways using 0.9.3 + the userland patch on firstfloor.org should work. -Andi -- ak...@li... |