Re: [PATCH] cpufreq: intel_pstate: Fix for HWP interrupt before driver is ready

On 9/6/21 10:17 AM, Rafael J. Wysocki wrote:

On Sat, Sep 4, 2021 at 7:37 AM Srinivas Pandruvada
[off-list ref] wrote:

quoted

In Lenovo X1 gen9 laptop, HWP interrupts arrive before driver is ready
to handle on that CPU. Basically didn't even allocated memory for per
cpu data structure and not even started interrupt enable process on that
CPU. So interrupt handler observes a NULL pointer to schedule work.

This interrupt was probably for SMM, but since it is redirected to
OS by OSC call, OS receives it, but not ready to handle. That redirection
of interrupt to OS was also done to solve one SMM crash on Yoga 260 for
HWP interrupt a while back.

To solve this the HWP interrupt handler should ignore such request if the
driver is not ready. This will require some flag to wait till the driver
setup a workqueue to handle on a CPU. We can't simply assume cpudata to
be NULL and avoid processing as it may not be NULL but data structure is
not in consistent state.

So created a cpumask which sets the CPU on which interrupt was setup. If
not setup, simply clear the interrupt status and return. Since the
similar issue can happen during S3 resume, clear the bit during offline.

Since interrupt timing may be before HWP is enabled, use safe MSR read
writes as before the change for HWP interrupt.

Fixes: d0e936adbd22 ("cpufreq: intel_pstate: Process HWP Guaranteed change notification")
Reported-and-tested-by: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>
---
 drivers/cpufreq/intel_pstate.c | 23 ++++++++++++++++++++++-
 1 file changed, 22 insertions(+), 1 deletion(-)

diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index b4ffe6c8a0d0..5ac86bfa1080 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c

@@ -298,6 +298,8 @@ static bool hwp_boost __read_mostly;

 static struct cpufreq_driver *intel_pstate_driver __read_mostly;

+static cpumask_t hwp_intr_enable_mask;
+
 #ifdef CONFIG_ACPI
 static bool acpi_ppc;
 #endif

@@ -1067,11 +1069,15 @@ static void intel_pstate_hwp_set(unsigned int cpu)
        wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 }

+static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
+
 static void intel_pstate_hwp_offline(struct cpudata *cpu)
 {
        u64 value = READ_ONCE(cpu->hwp_req_cached);
        int min_perf;

+       intel_pstate_disable_hwp_interrupt(cpu);
+
        if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
                /*
                 * In case the EPP has been set to "performance" by the

@@ -1645,20 +1651,35 @@ void notify_hwp_interrupt(void)
        if (!hwp_active || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
                return;

-       rdmsrl(MSR_HWP_STATUS, value);
+       rdmsrl_safe(MSR_HWP_STATUS, &value);
        if (!(value & 0x01))
                return;

+       if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask)) {
+               wrmsrl_safe(MSR_HWP_STATUS, 0);
+               return;
+       }

Without additional locking, there is a race between this and
intel_pstate_disable_hwp_interrupt().

1. notify_hwp_interrupt() checks hwp_intr_enable_mask() and the target
CPU is in there, so it will go for scheduling the delayed work.
2. intel_pstate_disable_hwp_interrupt() runs between the check and the
cpudata load below.
3. hwp_notify_work is scheduled on the CPU that isn't there in the
mask any more.

I noticed that too, not clear to me how much of an issue that is in
practice. But there's definitely a race there.

Can we just revert d0e936adbd22 for now so this can get solved in due
time? I doubt I'm the only one that got bit by this, but I might be the
only one that bothered to actually figured out what caused my laptop to
crash with a black screen on boot.

-- 
Jens Axboe