Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1399586 > unrolled thread

[PATCH v2 26/32] perf/x86/intel/cqm: make one write of PQR_ASSOC per ctx switch

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  [PATCH v2 26/32] perf/x86/intel/cqm: make one write of PQR_ASSOC per ctx switch David Carrillo-Cisneros <davidcc@google.com> - 2016-05-12 01:10 +0200

#1399586 — [PATCH v2 26/32] perf/x86/intel/cqm: make one write of PQR_ASSOC per ctx switch

For perf_events, PQR_ASSOC is written twice: in ctx_in and ctx_out.
Since writes to PQR_ASSOC are slow, this has a high impact in context
swich cost.

Modify pqr_common so pqr_update_rmid do not write the msr
right away, but until a call to __pqr_ctx_switch that is called from
the recently introduced finish_arch_pre_lock_switch hook.

Reviewed-by: Stephane Eranian <eranian@google.com>
Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
---
 arch/x86/events/intel/cqm.c       | 12 +++++++-----
 arch/x86/include/asm/pqr_common.h | 21 ++++++++++++++-------
 arch/x86/kernel/cpu/pqr_common.c  | 11 ++++++++++-
 3 files changed, 31 insertions(+), 13 deletions(-)

diff --git a/arch/x86/events/intel/cqm.c b/arch/x86/events/intel/cqm.c
index 5928bdb..83b041a 100644
--- a/arch/x86/events/intel/cqm.c
+++ b/arch/x86/events/intel/cqm.c
@@ -2515,7 +2515,7 @@ static inline void __intel_cqm_event_start(
 	if (!(event->hw.state & PERF_HES_STOPPED))
 		return;
 	event->hw.state &= ~PERF_HES_STOPPED;
-	pqr_update_rmid(summary.sched_rmid, PQR_RMID_MODE_EVENT);
+	pqr_cache_update_rmid(summary.sched_rmid, PQR_RMID_MODE_EVENT);
 }
 
 static void intel_cqm_event_start(struct perf_event *event, int mode)
@@ -2545,7 +2545,7 @@ static void intel_cqm_event_stop(struct perf_event *event, int mode)
 	/* Occupancy of CQM events is obtained at read. No need to read
 	 * when event is stopped since read on inactive cpus succeed.
 	 */
-	pqr_update_rmid(summary.sched_rmid, PQR_RMID_MODE_NOEVENT);
+	pqr_cache_update_rmid(summary.sched_rmid, PQR_RMID_MODE_NOEVENT);
 }
 
 static int intel_cqm_event_add(struct perf_event *event, int mode)
@@ -2962,8 +2962,10 @@ static void intel_cqm_cpu_starting(unsigned int cpu)
 	u16 pkg_id = topology_physical_package_id(cpu);
 
 	state->rmid = 0;
-	state->rmid_mode = PQR_RMID_MODE_NOEVENT;
+	state->next_rmid = 0;
+	state->next_rmid_mode = PQR_RMID_MODE_NOEVENT;
 	state->closid = 0;
+	state->next_closid = 0;
 
 	/* XXX: lock */
 	/* XXX: Make sure this case is handled when hotplug happens. */
@@ -3190,12 +3192,12 @@ inline void __intel_cqm_no_event_sched_in(void)
 	if (WARN_ON_ONCE(!__valid_rmid(pkg_id, summary.sched_rmid)))
 		goto no_rmid;
 
-	pqr_update_rmid(summary.sched_rmid, PQR_RMID_MODE_NOEVENT);
+	pqr_cache_update_rmid(summary.sched_rmid, PQR_RMID_MODE_NOEVENT);
 	return;
 
 no_rmid:
 	summary.value = atomic64_read(&root_pmonr->prmid_summary_atomic);
-	pqr_update_rmid(summary.sched_rmid, PQR_RMID_MODE_NOEVENT);
+	pqr_cache_update_rmid(summary.sched_rmid, PQR_RMID_MODE_NOEVENT);
 #endif
 }
 
diff --git a/arch/x86/include/asm/pqr_common.h b/arch/x86/include/asm/pqr_common.h
index 0af04d2..bfaa7d8 100644
--- a/arch/x86/include/asm/pqr_common.h
+++ b/arch/x86/include/asm/pqr_common.h
@@ -26,8 +26,10 @@ enum intel_pqr_rmid_mode {
 /**
  * struct intel_pqr_state - State cache for the PQR MSR
  * @rmid:		Last RMID written to hw.
+ * @next_rmid:		Next RMID to write to hw.
  * @rmid_mode:		Next RMID's mode.
  * @closid:		The current Class Of Service ID
+ * @next_closid:	The Class Of Service ID to use.
  *
  * The upper 32 bits of MSR_IA32_PQR_ASSOC contain closid and the
  * lower 10 bits rmid. The update to MSR_IA32_PQR_ASSOC always
@@ -39,22 +41,27 @@ enum intel_pqr_rmid_mode {
  */
 struct intel_pqr_state {
 	u32				rmid;
-	enum intel_pqr_rmid_mode	rmid_mode;
+	u32				next_rmid;
+	enum intel_pqr_rmid_mode	next_rmid_mode;
 	u32				closid;
+	u32				next_closid;
 };
 
 DECLARE_PER_CPU(struct intel_pqr_state, pqr_state);
 
-static inline void pqr_update_rmid(u32 rmid, enum intel_pqr_rmid_mode mode)
+static inline void pqr_cache_update_rmid(u32 rmid, enum intel_pqr_rmid_mode mode)
 {
 	struct intel_pqr_state *state = this_cpu_ptr(&pqr_state);
 
-	state->rmid_mode = mode;
+	state->next_rmid_mode = mode;
+	state->next_rmid = rmid;
+}
+
+static inline void pqr_cache_update_closid(u32 closid)
+{
+	struct intel_pqr_state *state = this_cpu_ptr(&pqr_state);
 
-	if (state->rmid == rmid)
-		return;
-	state->rmid = rmid;
-	wrmsr(MSR_IA32_PQR_ASSOC, rmid, state->closid);
+	state->next_closid = closid;
 }
 
 void __pqr_ctx_switch(void);
diff --git a/arch/x86/kernel/cpu/pqr_common.c b/arch/x86/kernel/cpu/pqr_common.c
index e36702f..0ae2b2c 100644
--- a/arch/x86/kernel/cpu/pqr_common.c
+++ b/arch/x86/kernel/cpu/pqr_common.c
@@ -19,6 +19,15 @@ inline void __pqr_ctx_switch(void)
 	/* If perf_event did set rmid that is used, do not try
 	 * to obtain another one from current task.
 	 */
-	if (state->rmid_mode == PQR_RMID_MODE_NOEVENT)
+	if (state->next_rmid_mode == PQR_RMID_MODE_NOEVENT)
 		__intel_cqm_no_event_sched_in();
+
+	/* __intel_cqm_no_event_sched_in might have changed next_rmid. */
+	if (state->rmid == state->next_rmid &&
+	    state->closid == state->next_closid)
+		return;
+
+	state->rmid = state->next_rmid;
+	state->closid = state->next_closid;
+	wrmsr(MSR_IA32_PQR_ASSOC, state->rmid, state->closid);
 }
-- 
2.8.0.rc3.226.g39d4020

[toc] | [standalone]

Back to top | Article view | linux.kernel

csiph-web