ACPI: processor: Take over RCU-idle for C3-BM idle
authorPeter Zijlstra <peterz@infradead.org>
Tue, 15 Sep 2020 10:32:01 +0000 (12:32 +0200)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Wed, 16 Sep 2020 17:36:26 +0000 (19:36 +0200)
The C3 BusMaster idle code takes lock in a number of places, some deep
inside the ACPI code. Instead of wrapping it all in RCU_NONIDLE, have
the driver take over RCU-idle duty and avoid flipping RCU state back
and forth a lot.

( by marking 'C3 && bm_check' as RCU_IDLE, we _must_ call enter_bm() for
  that combination, otherwise we'll loose RCU-idle, this requires
  shuffling some code around )

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/acpi/processor_idle.c

index 9ec504067f9843e0f5251daf7c54686dbdcf7436..7ecb90e90afdee578cceec676f3c1caae6bf1333 100644 (file)
@@ -558,22 +558,43 @@ static DEFINE_RAW_SPINLOCK(c3_lock);
 
 /**
  * acpi_idle_enter_bm - enters C3 with proper BM handling
+ * @drv: cpuidle driver
  * @pr: Target processor
  * @cx: Target state context
+ * @index: index of target state
  */
-static void acpi_idle_enter_bm(struct acpi_processor *pr,
-                              struct acpi_processor_cx *cx)
+static int acpi_idle_enter_bm(struct cpuidle_driver *drv,
+                              struct acpi_processor *pr,
+                              struct acpi_processor_cx *cx,
+                              int index)
 {
+       static struct acpi_processor_cx safe_cx = {
+               .entry_method = ACPI_CSTATE_HALT,
+       };
+
        /*
         * disable bus master
         * bm_check implies we need ARB_DIS
         * bm_control implies whether we can do ARB_DIS
         *
-        * That leaves a case where bm_check is set and bm_control is
-        * not set. In that case we cannot do much, we enter C3
-        * without doing anything.
+        * That leaves a case where bm_check is set and bm_control is not set.
+        * In that case we cannot do much, we enter C3 without doing anything.
         */
-       if (pr->flags.bm_control) {
+       bool dis_bm = pr->flags.bm_control;
+
+       /* If we can skip BM, demote to a safe state. */
+       if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
+               dis_bm = false;
+               index = drv->safe_state_index;
+               if (index >= 0) {
+                       cx = this_cpu_read(acpi_cstate[index]);
+               } else {
+                       cx = &safe_cx;
+                       index = -EBUSY;
+               }
+       }
+
+       if (dis_bm) {
                raw_spin_lock(&c3_lock);
                c3_cpu_count++;
                /* Disable bus master arbitration when all CPUs are in C3 */
@@ -582,15 +603,21 @@ static void acpi_idle_enter_bm(struct acpi_processor *pr,
                raw_spin_unlock(&c3_lock);
        }
 
+       rcu_idle_enter();
+
        acpi_idle_do_entry(cx);
 
+       rcu_idle_exit();
+
        /* Re-enable bus master arbitration */
-       if (pr->flags.bm_control) {
+       if (dis_bm) {
                raw_spin_lock(&c3_lock);
                acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
                c3_cpu_count--;
                raw_spin_unlock(&c3_lock);
        }
+
+       return index;
 }
 
 static int acpi_idle_enter(struct cpuidle_device *dev,
@@ -604,20 +631,13 @@ static int acpi_idle_enter(struct cpuidle_device *dev,
                return -EINVAL;
 
        if (cx->type != ACPI_STATE_C1) {
+               if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check)
+                       return acpi_idle_enter_bm(drv, pr, cx, index);
+
+               /* C2 to C1 demotion. */
                if (acpi_idle_fallback_to_c1(pr) && num_online_cpus() > 1) {
                        index = ACPI_IDLE_STATE_START;
                        cx = per_cpu(acpi_cstate[index], dev->cpu);
-               } else if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check) {
-                       if (cx->bm_sts_skip || !acpi_idle_bm_check()) {
-                               acpi_idle_enter_bm(pr, cx);
-                               return index;
-                       } else if (drv->safe_state_index >= 0) {
-                               index = drv->safe_state_index;
-                               cx = per_cpu(acpi_cstate[index], dev->cpu);
-                       } else {
-                               acpi_safe_halt();
-                               return -EBUSY;
-                       }
                }
        }
 
@@ -641,7 +661,13 @@ static int acpi_idle_enter_s2idle(struct cpuidle_device *dev,
                        return 0;
 
                if (pr->flags.bm_check) {
-                       acpi_idle_enter_bm(pr, cx);
+                       u8 bm_sts_skip = cx->bm_sts_skip;
+
+                       /* Don't check BM_STS, do an unconditional ARB_DIS for S2IDLE */
+                       cx->bm_sts_skip = 1;
+                       acpi_idle_enter_bm(drv, pr, cx, index);
+                       cx->bm_sts_skip = bm_sts_skip;
+
                        return 0;
                } else {
                        ACPI_FLUSH_CPU_CACHE();
@@ -674,8 +700,11 @@ static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
                if (lapic_timer_needs_broadcast(pr, cx))
                        state->flags |= CPUIDLE_FLAG_TIMER_STOP;
 
-               if (cx->type == ACPI_STATE_C3)
+               if (cx->type == ACPI_STATE_C3) {
                        state->flags |= CPUIDLE_FLAG_TLB_FLUSHED;
+                       if (pr->flags.bm_check)
+                               state->flags |= CPUIDLE_FLAG_RCU_IDLE;
+               }
 
                count++;
                if (count == CPUIDLE_STATE_MAX)