[PATCH] x86,perf: Disable intel_bts when PTI

From: Peter Zijlstra
Date: Sun Jan 14 2018 - 05:27:34 EST



The intel_bts driver does not use the 'normal' BTS buffer we allocated
space for in the cpu_entry_area but instead uses the memory allocated
for the perf AUX buffer.

This obviously comes apart when using PTI because then the kernel
mapping; which includes that AUX buffer memory; disappears.

For now disable this driver when PTI is enabled, we'll try and sort
something out later.

Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Reported-by: Vince Weaver <vincent.weaver@xxxxxxxxx>
Reported-by: Robert ÅwiÄcki <robert@xxxxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
arch/x86/events/intel/bts.c | 18 ++++++++++++++++++
1 file changed, 18 insertions(+)

diff --git a/arch/x86/events/intel/bts.c b/arch/x86/events/intel/bts.c
index 141e07b06216..24ffa1e88cf9 100644
--- a/arch/x86/events/intel/bts.c
+++ b/arch/x86/events/intel/bts.c
@@ -582,6 +582,24 @@ static __init int bts_init(void)
if (!boot_cpu_has(X86_FEATURE_DTES64) || !x86_pmu.bts)
return -ENODEV;

+ if (boot_cpu_has(X86_FEATURE_PTI)) {
+ /*
+ * BTS hardware writes through a virtual memory map we must
+ * either use the kernel physical map, or the user mapping of
+ * the AUX buffer.
+ *
+ * However, since this driver supports per-CPU and per-task inherit
+ * we cannot use the user mapping since it will not be availble
+ * if we're not running the owning process.
+ *
+ * With PTI we can't use the kernal map either, because its not
+ * there when we run userspace.
+ *
+ * For now, disable this driver when using PTI.
+ */
+ return -ENODEV;
+ }
+
bts_pmu.capabilities = PERF_PMU_CAP_AUX_NO_SG | PERF_PMU_CAP_ITRACE |
PERF_PMU_CAP_EXCLUSIVE;
bts_pmu.task_ctx_nr = perf_sw_context;