perf/x86/intel: Save/restore LBR stack during context switch
authorYan, Zheng <zheng.z.yan@intel.com>
Wed, 5 Nov 2014 02:56:05 +0000 (21:56 -0500)
committerIngo Molnar <mingo@kernel.org>
Wed, 18 Feb 2015 16:16:10 +0000 (17:16 +0100)
When the LBR call stack is enabled, it is necessary to save/restore
the LBR stack on context switch. The solution is saving/restoring
the LBR stack to/from task's perf event context.

The LBR stack is saved/restored only when there are events that use
the LBR call stack. If no event uses LBR call stack, the LBR stack
is reset when task is scheduled in.

Signed-off-by: Yan, Zheng <zheng.z.yan@intel.com>
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: eranian@google.com
Cc: jolsa@redhat.com
Link: http://lkml.kernel.org/r/1415156173-10035-10-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/kernel/cpu/perf_event_intel_lbr.c

index ac8e54200934d6298a89fe43bd5212034ce1004b..a8b3f236cf373868b6cb51175fdc4ccf29fc7d1b 100644 (file)
@@ -180,13 +180,89 @@ void intel_pmu_lbr_reset(void)
                intel_pmu_lbr_reset_64();
 }
 
+/*
+ * TOS = most recently recorded branch
+ */
+static inline u64 intel_pmu_lbr_tos(void)
+{
+       u64 tos;
+
+       rdmsrl(x86_pmu.lbr_tos, tos);
+       return tos;
+}
+
+enum {
+       LBR_NONE,
+       LBR_VALID,
+};
+
+static void __intel_pmu_lbr_restore(struct x86_perf_task_context *task_ctx)
+{
+       int i;
+       unsigned lbr_idx, mask;
+       u64 tos;
+
+       if (task_ctx->lbr_callstack_users == 0 ||
+           task_ctx->lbr_stack_state == LBR_NONE) {
+               intel_pmu_lbr_reset();
+               return;
+       }
+
+       mask = x86_pmu.lbr_nr - 1;
+       tos = intel_pmu_lbr_tos();
+       for (i = 0; i < x86_pmu.lbr_nr; i++) {
+               lbr_idx = (tos - i) & mask;
+               wrmsrl(x86_pmu.lbr_from + lbr_idx, task_ctx->lbr_from[i]);
+               wrmsrl(x86_pmu.lbr_to + lbr_idx, task_ctx->lbr_to[i]);
+       }
+       task_ctx->lbr_stack_state = LBR_NONE;
+}
+
+static void __intel_pmu_lbr_save(struct x86_perf_task_context *task_ctx)
+{
+       int i;
+       unsigned lbr_idx, mask;
+       u64 tos;
+
+       if (task_ctx->lbr_callstack_users == 0) {
+               task_ctx->lbr_stack_state = LBR_NONE;
+               return;
+       }
+
+       mask = x86_pmu.lbr_nr - 1;
+       tos = intel_pmu_lbr_tos();
+       for (i = 0; i < x86_pmu.lbr_nr; i++) {
+               lbr_idx = (tos - i) & mask;
+               rdmsrl(x86_pmu.lbr_from + lbr_idx, task_ctx->lbr_from[i]);
+               rdmsrl(x86_pmu.lbr_to + lbr_idx, task_ctx->lbr_to[i]);
+       }
+       task_ctx->lbr_stack_state = LBR_VALID;
+}
+
 void intel_pmu_lbr_sched_task(struct perf_event_context *ctx, bool sched_in)
 {
        struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+       struct x86_perf_task_context *task_ctx;
 
        if (!x86_pmu.lbr_nr)
                return;
 
+       /*
+        * If LBR callstack feature is enabled and the stack was saved when
+        * the task was scheduled out, restore the stack. Otherwise flush
+        * the LBR stack.
+        */
+       task_ctx = ctx ? ctx->task_ctx_data : NULL;
+       if (task_ctx) {
+               if (sched_in) {
+                       __intel_pmu_lbr_restore(task_ctx);
+                       cpuc->lbr_context = ctx;
+               } else {
+                       __intel_pmu_lbr_save(task_ctx);
+               }
+               return;
+       }
+
        /*
         * When sampling the branck stack in system-wide, it may be
         * necessary to flush the stack on context switch. This happens
@@ -279,18 +355,6 @@ void intel_pmu_lbr_disable_all(void)
                __intel_pmu_lbr_disable();
 }
 
-/*
- * TOS = most recently recorded branch
- */
-static inline u64 intel_pmu_lbr_tos(void)
-{
-       u64 tos;
-
-       rdmsrl(x86_pmu.lbr_tos, tos);
-
-       return tos;
-}
-
 static void intel_pmu_lbr_read_32(struct cpu_hw_events *cpuc)
 {
        unsigned long mask = x86_pmu.lbr_nr - 1;