PM: s2idle: ACPI: Fix wakeup interrupts handling
authorRafael J. Wysocki <rafael.j.wysocki@intel.com>
Fri, 4 Feb 2022 17:35:22 +0000 (18:35 +0100)
committerRafael J. Wysocki <rafael.j.wysocki@intel.com>
Mon, 7 Feb 2022 20:02:31 +0000 (21:02 +0100)
After commit e3728b50cd9b ("ACPI: PM: s2idle: Avoid possible race
related to the EC GPE") wakeup interrupts occurring immediately after
the one discarded by acpi_s2idle_wake() may be missed.  Moreover, if
the SCI triggers again immediately after the rearming in
acpi_s2idle_wake(), that wakeup may be missed too.

The problem is that pm_system_irq_wakeup() only calls pm_system_wakeup()
when pm_wakeup_irq is 0, but that's not the case any more after the
interrupt causing acpi_s2idle_wake() to run until pm_wakeup_irq is
cleared by the pm_wakeup_clear() call in s2idle_loop().  However,
there may be wakeup interrupts occurring in that time frame and if
that happens, they will be missed.

To address that issue first move the clearing of pm_wakeup_irq to
the point at which it is known that the interrupt causing
acpi_s2idle_wake() to tun will be discarded, before rearming the SCI
for wakeup.  Moreover, because that only reduces the size of the
time window in which the issue may manifest itself, allow
pm_system_irq_wakeup() to register two second wakeup interrupts in
a row and, when discarding the first one, replace it with the second
one.  [Of course, this assumes that only one wakeup interrupt can be
discarded in one go, but currently that is the case and I am not
aware of any plans to change that.]

Fixes: e3728b50cd9b ("ACPI: PM: s2idle: Avoid possible race related to the EC GPE")
Cc: 5.4+ <stable@vger.kernel.org> # 5.4+
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
drivers/acpi/sleep.c
drivers/base/power/wakeup.c
include/linux/suspend.h
kernel/power/main.c
kernel/power/process.c
kernel/power/suspend.c

index fac7c9d..d4fbea9 100644 (file)
@@ -758,6 +758,7 @@ bool acpi_s2idle_wake(void)
                        return true;
                }
 
+               pm_wakeup_clear(acpi_sci_irq);
                rearm_wake_irq(acpi_sci_irq);
        }
 
index 99bda0d..8666590 100644 (file)
@@ -34,7 +34,8 @@ suspend_state_t pm_suspend_target_state;
 bool events_check_enabled __read_mostly;
 
 /* First wakeup IRQ seen by the kernel in the last cycle. */
-unsigned int pm_wakeup_irq __read_mostly;
+static unsigned int wakeup_irq[2] __read_mostly;
+static DEFINE_RAW_SPINLOCK(wakeup_irq_lock);
 
 /* If greater than 0 and the system is suspending, terminate the suspend. */
 static atomic_t pm_abort_suspend __read_mostly;
@@ -942,19 +943,45 @@ void pm_system_cancel_wakeup(void)
        atomic_dec_if_positive(&pm_abort_suspend);
 }
 
-void pm_wakeup_clear(bool reset)
+void pm_wakeup_clear(unsigned int irq_number)
 {
-       pm_wakeup_irq = 0;
-       if (reset)
+       raw_spin_lock_irq(&wakeup_irq_lock);
+
+       if (irq_number && wakeup_irq[0] == irq_number)
+               wakeup_irq[0] = wakeup_irq[1];
+       else
+               wakeup_irq[0] = 0;
+
+       wakeup_irq[1] = 0;
+
+       raw_spin_unlock_irq(&wakeup_irq_lock);
+
+       if (!irq_number)
                atomic_set(&pm_abort_suspend, 0);
 }
 
 void pm_system_irq_wakeup(unsigned int irq_number)
 {
-       if (pm_wakeup_irq == 0) {
-               pm_wakeup_irq = irq_number;
+       unsigned long flags;
+
+       raw_spin_lock_irqsave(&wakeup_irq_lock, flags);
+
+       if (wakeup_irq[0] == 0)
+               wakeup_irq[0] = irq_number;
+       else if (wakeup_irq[1] == 0)
+               wakeup_irq[1] = irq_number;
+       else
+               irq_number = 0;
+
+       raw_spin_unlock_irqrestore(&wakeup_irq_lock, flags);
+
+       if (irq_number)
                pm_system_wakeup();
-       }
+}
+
+unsigned int pm_wakeup_irq(void)
+{
+       return wakeup_irq[0];
 }
 
 /**
index 3e8ecde..300273f 100644 (file)
@@ -497,14 +497,14 @@ extern void ksys_sync_helper(void);
 
 /* drivers/base/power/wakeup.c */
 extern bool events_check_enabled;
-extern unsigned int pm_wakeup_irq;
 extern suspend_state_t pm_suspend_target_state;
 
 extern bool pm_wakeup_pending(void);
 extern void pm_system_wakeup(void);
 extern void pm_system_cancel_wakeup(void);
-extern void pm_wakeup_clear(bool reset);
+extern void pm_wakeup_clear(unsigned int irq_number);
 extern void pm_system_irq_wakeup(unsigned int irq_number);
+extern unsigned int pm_wakeup_irq(void);
 extern bool pm_get_wakeup_count(unsigned int *count, bool block);
 extern bool pm_save_wakeup_count(unsigned int count);
 extern void pm_wakep_autosleep_enabled(bool set);
index 44169f3..7e64607 100644 (file)
@@ -504,7 +504,10 @@ static ssize_t pm_wakeup_irq_show(struct kobject *kobj,
                                        struct kobj_attribute *attr,
                                        char *buf)
 {
-       return pm_wakeup_irq ? sprintf(buf, "%u\n", pm_wakeup_irq) : -ENODATA;
+       if (!pm_wakeup_irq())
+               return -ENODATA;
+
+       return sprintf(buf, "%u\n", pm_wakeup_irq());
 }
 
 power_attr_ro(pm_wakeup_irq);
index b7e7798..11b570f 100644 (file)
@@ -134,7 +134,7 @@ int freeze_processes(void)
        if (!pm_freezing)
                atomic_inc(&system_freezing_cnt);
 
-       pm_wakeup_clear(true);
+       pm_wakeup_clear(0);
        pr_info("Freezing user space processes ... ");
        pm_freezing = true;
        error = try_to_freeze_tasks(true);
index 80cc1f0..6fcdee7 100644 (file)
@@ -136,8 +136,6 @@ static void s2idle_loop(void)
                        break;
                }
 
-               pm_wakeup_clear(false);
-
                s2idle_enter();
        }