Hello,
On Intel X86 processors, the DEBUGCTL msr contains lots of bits
to control various unrelated features, such as LBR, BTS, PMU.
Nowadays, several subsystems need access to this MSR to set or
clear the bits they care about. In 2.6.27, the new DS/BTS interface
was introduced and DEBUGCTL can now be cleared and restored
on context switches.
The KPROBE interface also uses DEBUGCTL.
In the context of perfmon, it can be interesting to also modify bits
in DEBUGCTL, such as FREEZE_PMU_ON_PMI but also the
bits controlling the LBR (bit 0 and 11). With perfmon, however,
there are situations, e.g. LBR, where the DEBUGCTL bits are
managed on a per-CPU level and not per-task level. Simply
overwriting DEBUGCTL on context switch may conflict with
per-cpu usage.
It appears that modifications to DEBUGCTL must be done using
a read-modify-write approach rather than pure write. To that extent,
I would like to propose the attached patch which modifies the
update_debugctlmsr() function to implement a read-modify-write
access. The caller passes the value of debugctl and a mask
of bits of interest. For instance, if you want to set bit 1 and clear bit 0,
then you call update_debugctlmsr(0x2, 0x3).
Please let me know what you think about this patch.
Thanks.
On Fri, Nov 07, 2008 at 03:44:09PM +0100, stephane eranian wrote:
> Hello,
>
> On Intel X86 processors, the DEBUGCTL msr contains lots of bits
> to control various unrelated features, such as LBR, BTS, PMU.
>
> Nowadays, several subsystems need access to this MSR to set or
> clear the bits they care about. In 2.6.27, the new DS/BTS interface
> was introduced and DEBUGCTL can now be cleared and restored
> on context switches.
>
> The KPROBE interface also uses DEBUGCTL.
Stephane,
I see no problems with the kprobes part of the patch.
Ananth