Re: [Xen-devel] [PATCH] x86/vpmu_intel: Handle SMT consistently for programmable and fixed counters

[Mohit Gambhir <mohit.gambhir@xxxxxxxxxx>]

On 04/03/2017 06:36 AM, Jan Beulich wrote:

On 31.03.17 at 16:46, <mohit.gambhir@xxxxxxxxxx> wrote:

This patch masks .AnyThread bits in IA32_FIXED_CTR_CTRL MSR for all
versions of Intel Arhcitectural Performance Monitoring. Note that
.AnyThread bit (21) is already masked in IA32_PERFEVTSELx MSRs since
hyperthreading is not exposed to guests and Intel SDM discourages the use of
.AnyThread bit in virtualized environments (per section 18.2.3.1
AnyThread Counting and Software Evolution)

All nice and presumably correct, but the main thing is missing: The
bits aren't defined prior to version 3 afaics, so ...

--- a/xen/arch/x86/cpu/vpmu_intel.c
+++ b/xen/arch/x86/cpu/vpmu_intel.c
@@ -979,8 +979,7 @@ int __init core2_vpmu_init(void)
     full_width_write = (caps >> 13) & 1;
 
     fixed_ctrl_mask = ~((1ull << (fixed_pmc_cnt * FIXED_CTR_CTRL_BITS)) - 1);
-    if ( version == 2 )
-        fixed_ctrl_mask |= 0x444;
+    fixed_ctrl_mask |= 0x444;

... the main thing to explain is why removing the conditional is
(a) correct and (b) necessary (going through the uses of the
variable I can see (a) to be true, but not (b)). And of course it
would be quite helpful if the literal number changed to a
manifest constant at once, or a comment was attached to
clarify what the number represents.

I do agree that replacing the hard coded constant with a macro would be nice and I will update the patch with that.

The answer to why this change is (b) necessary is two folds -

1. We need to be consistent in the implementation. As said in the commit log - we disable .Anythread bit in
programmable counters (regardless of the version) by masking bit 21 in IA32_PERFEVTSELx.  (See code snippet
below from vpmu_intel.c)

 /* Masks used for testing whether and MSR is valid */                          
 #define ARCH_CTRL_MASK  (~((1ull << 32) - 1) | (1ull << 21))

But we leave it enabled in fixed function counters for version 3. Removing the condition disables the bit in fixed function
counters regardless of the version,  which is consistent with what is done for programmable counters.

2. We don't want to expose event counts from another guest (or hypervisor) which can happen if .AnyThread bit is not masked and
a VCPU is only scheduled to run on one of the hardware threads in a hyper-threaded CPU.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel