When Architectural Performance Monitoring is available, the PERF_GLOBAL_CTRL
MSR contains per-counter enable bits that is ANDed with the enable bit in the
counter EVNTSEL MSR in order for a PMC counter to be enabled.
So far the watchdog code seems to have relied on the PERF_GLOBAL_CTRL enable
bits being set by default, but at least on some Intel Sapphire and Emerald
Rapids this is no longer the case, and Xen reports:
Testing NMI watchdog on all CPUs: 0 40 stuck
The first CPU on each package is started with PERF_GLOBAL_CTRL zeroed, so PMC0
doesn't start counting when the enable bit in EVNTSEL0 is set, due to the
relevant enable bit in PERF_GLOBAL_CTRL not being set.
Check and adjust PERF_GLOBAL_CTRL during CPU initialization so that all the
general-purpose PMCs are enabled. Doing so brings the state of the package-BSP
PERF_GLOBAL_CTRL in line with the rest of the CPUs on the system.
Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
---
Changes since v3:
- Revert previous change and require 2 counters.
Changes since v2:
- Print message on invalid PERF_GLOBAL_CTRL.
- Use a mask instead of type truncation.
- Adjust the check to require a single counter instead of 2.
Changes since v1:
- Do the adjustment of PERF_GLOBAL_CTRL even if the watchdog is not used, and
enable all counters.
---
xen/arch/x86/cpu/intel.c | 23 ++++++++++++++++++++++-
1 file changed, 22 insertions(+), 1 deletion(-)
diff --git a/xen/arch/x86/cpu/intel.c b/xen/arch/x86/cpu/intel.c
index dfee64689ffe..b5490eb06e00 100644
--- a/xen/arch/x86/cpu/intel.c
+++ b/xen/arch/x86/cpu/intel.c
@@ -533,9 +533,30 @@ static void cf_check init_intel(struct cpuinfo_x86 *c)
init_intel_cacheinfo(c);
if (c->cpuid_level > 9) {
unsigned eax = cpuid_eax(10);
+ unsigned int cnt = (eax >> 8) & 0xff;
+
/* Check for version and the number of counters */
- if ((eax & 0xff) && (((eax>>8) & 0xff) > 1))
+ if ((eax & 0xff) && (cnt > 1) && (cnt <= 32)) {
+ uint64_t global_ctrl;
+ unsigned int cnt_mask = (1UL << cnt) - 1;
+
+ /*
+ * On (some?) Sapphire/Emerald Rapids platforms each
+ * package-BSP starts with all the enable bits for the
+ * general-purpose PMCs cleared. Adjust so counters
+ * can be enabled from EVNTSEL.
+ */
+ rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, global_ctrl);
+ if ((global_ctrl & cnt_mask) != cnt_mask) {
+ printk("CPU%u: invalid PERF_GLOBAL_CTRL: %#"
+ PRIx64 " adjusting to %#" PRIx64 "\n",
+ smp_processor_id(), global_ctrl,
+ global_ctrl | cnt_mask);
+ wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
+ global_ctrl | cnt_mask);
+ }
__set_bit(X86_FEATURE_ARCH_PERFMON, c->x86_capability);
+ }
}
if ( !cpu_has(c, X86_FEATURE_XTOPOLOGY) )
--
2.43.0
On 12.01.2024 12:09, Roger Pau Monne wrote: > When Architectural Performance Monitoring is available, the PERF_GLOBAL_CTRL > MSR contains per-counter enable bits that is ANDed with the enable bit in the > counter EVNTSEL MSR in order for a PMC counter to be enabled. > > So far the watchdog code seems to have relied on the PERF_GLOBAL_CTRL enable > bits being set by default, but at least on some Intel Sapphire and Emerald > Rapids this is no longer the case, and Xen reports: > > Testing NMI watchdog on all CPUs: 0 40 stuck > > The first CPU on each package is started with PERF_GLOBAL_CTRL zeroed, so PMC0 > doesn't start counting when the enable bit in EVNTSEL0 is set, due to the > relevant enable bit in PERF_GLOBAL_CTRL not being set. > > Check and adjust PERF_GLOBAL_CTRL during CPU initialization so that all the > general-purpose PMCs are enabled. Doing so brings the state of the package-BSP > PERF_GLOBAL_CTRL in line with the rest of the CPUs on the system. > > Signed-off-by: Roger Pau Monné <roger.pau@citrix.com> Acked-by: Jan Beulich <jbeulich@suse.com>
© 2016 - 2024 Red Hat, Inc.