Hello everyone,
I have released an updated version of the perfmon new code base package.
This release is relative to 2.6.14-rc4-mm1. I have also updated the library,
libpfm-3.2, to match the kernel level changes.
Kernel-package features:
------------------------
- preliminary support for MIPS R5000 by Phil Mucci
- on X86-64, P6, P4 32-bits, the PMC enable bits are now
under the control of the users. Pfm_start/pfm_stop do not
touch them anymore. That means applications must set them.
- simplified arch-specific interface (merged calls)
- simplified PMU description tables (removed dep_pmc[])
I now have a compilation environment for PPC64 and MIPS64, as such
I have verified that the patches for those architectures compile,
no actual testing has been done, though.
For MIPS, the patch is relative to the http://www.linux-mips.org GIT
tree as they still maintain a separate tree. The common perfmon
patch does apply cleanly even though the MIPS tree is sligthly
behind (see README.mips).
For libpfm-3.2, the updates is to reflect the changes for the
enable bits for P6, X86-64. The P4 standalone programs, and
PEBS examples have also been updated for enable bits. Note
that the PEBS support does not seem to work when Hyperthreading
is enabled. I have not yet tracked this one down, any volunteer?
You can grab both packages at our SourceForge web site:
http://www.sf.net/projects/perfmon2
You must download:
- 2.6.14-rc4-mm1-051017
- libpfm-3.2-051018
Enjoy,
--
-Stephane
Stephane Eranian wrote:
> Hello everyone,
>
> I have released an updated version of the perfmon new code base package.
> This release is relative to 2.6.14-rc4-mm1. I have also updated the library,
> libpfm-3.2, to match the kernel level changes.
>
> Kernel-package features:
> ------------------------
> - preliminary support for MIPS R5000 by Phil Mucci
> - on X86-64, P6, P4 32-bits, the PMC enable bits are now
> under the control of the users. Pfm_start/pfm_stop do not
> touch them anymore. That means applications must set them.
> - simplified arch-specific interface (merged calls)
> - simplified PMU description tables (removed dep_pmc[])
>
> I now have a compilation environment for PPC64 and MIPS64, as such
> I have verified that the patches for those architectures compile,
> no actual testing has been done, though.
>
> For MIPS, the patch is relative to the http://www.linux-mips.org GIT
> tree as they still maintain a separate tree. The common perfmon
> patch does apply cleanly even though the MIPS tree is sligthly
> behind (see README.mips).
>
> For libpfm-3.2, the updates is to reflect the changes for the
> enable bits for P6, X86-64. The P4 standalone programs, and
> PEBS examples have also been updated for enable bits. Note
> that the PEBS support does not seem to work when Hyperthreading
> is enabled. I have not yet tracked this one down, any volunteer?
>
> You can grab both packages at our SourceForge web site:
>
> http://www.sf.net/projects/perfmon2
>
> You must download:
> - 2.6.14-rc4-mm1-051017
> - libpfm-3.2-051018
>
> Enjoy,
>
Hi Stephane,
I have been looking at what changes are required to get oprofile to be
able to use the custom sampling format in perfmon2. It looks like there
have been some changes between the perfmon and perfmon2. The ia64
oprofile support uses the older interface. I don't have easy access to
an ia64 machine, so I have been making similar support available on the
x86 version.
I noticed that the older interface passed in "struct pt_regs *regs", but
the newer interface does not. The oprofile code extracted the program
counter and whether the interrupted process was in kernel or user mode
from regs. The newer perfmon interface passes in the instruction
pointer, but the information about user/kernel mode is lacking.
Reading through the perform2 documentation the last argument passed into
fmt_handler is a void pointer. The Perfmon2 specification
(HPL-2004-200R1.pdf) says:
data : a pointer to an implementation-specific data structure which may
be needed by a handler. For instance, this could point to the
interrupted machine state and/or the thread to which the overflow is
attributed.
However, the actual call in linux/perfmon/perfmon.c just passes NULL for
it. Would it be possible to pass the regs instead? Why not pass regs to
the handler? Was there some thought to allow other data to be passed?
-Will
Will,
On Fri, Oct 28, 2005 at 03:11:40PM -0400, William Cohen wrote:
>
> I have been looking at what changes are required to get oprofile to be
> able to use the custom sampling format in perfmon2. It looks like there
> have been some changes between the perfmon and perfmon2. The ia64
> oprofile support uses the older interface. I don't have easy access to
> an ia64 machine, so I have been making similar support available on the
> x86 version.
>
> I noticed that the older interface passed in "struct pt_regs *regs", but
> the newer interface does not. The oprofile code extracted the program
> counter and whether the interrupted process was in kernel or user mode
> from regs. The newer perfmon interface passes in the instruction
> pointer, but the information about user/kernel mode is lacking.
>
> Reading through the perform2 documentation the last argument passed into
> fmt_handler is a void pointer. The Perfmon2 specification
> (HPL-2004-200R1.pdf) says:
>
> data : a pointer to an implementation-specific data structure which may
> be needed by a handler. For instance, this could point to the
> interrupted machine state and/or the thread to which the overflow is
> attributed.
>
> However, the actual call in linux/perfmon/perfmon.c just passes NULL for
> it. Would it be possible to pass the regs instead? Why not pass regs to
> the handler? Was there some thought to allow other data to be passed?
>
I remember removing it, I just don't remeber why. may I thought that there
was enough information carried by the ip argument (i.e. where you were
interrupted), but I don't see a problmen in passing pt_regs again. The
documentation clearly hints at that. I'll do that in the next patch
for all platforms.
Thanks.
--
-Stephane