[PATCH] coresight: etm4x: Fix etm4_count race using atomic variable

[PATCH] coresight: etm4x: Fix etm4_count race using atomic variable

[Sai Prakash Ranjan]

etm4_count keeps track of number of ETMv4 registered and on some
systems, a race is observed on etm4_count variable which can
lead to multiple calls to cpuhp_setup_state_nocalls_cpuslocked().
This function internally calls cpuhp_store_callbacks() which
prevents multiple registrations of callbacks for a given state
and due to this race, it returns -EBUSY leading to ETM probe
failures like below.

 coresight-etm4x: probe of 7040000.etm failed with error -16

This race can easily be triggered with async probe by setting
probe type as PROBE_PREFER_ASYNCHRONOUS and with ETM power
management property "arm,coresight-loses-context-with-cpu".

Prevent this race by converting etm4_count variable to atomic.

Fixes: 9b6a3f3633a5 ("coresight: etmv4: Fix CPU power management setup in probe() function")
Fixes: 58eb457be028 ("hwtracing/coresight-etm4x: Convert to hotplug state machine")
Suggested-by: Mike Leach <mike.leach@linaro.org>
(Mike: Rootcause and context for commit message)
Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan@codeaurora.org>
---
 drivers/hwtracing/coresight/coresight-etm4x.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/hwtracing/coresight/coresight-etm4x.c b/drivers/hwtracing/coresight/coresight-etm4x.c
index 6d7d2169bfb2..f256ea744c51 100644
--- a/drivers/hwtracing/coresight/coresight-etm4x.c
+++ b/drivers/hwtracing/coresight/coresight-etm4x.c
@@ -49,7 +49,7 @@ MODULE_PARM_DESC(pm_save_enable,
 	"Save/restore state on power down: 1 = never, 2 = self-hosted");
 
 /* The number of ETMv4 currently registered */
-static int etm4_count;
+static atomic_t etm4_count;
 static struct etmv4_drvdata *etmdrvdata[NR_CPUS];
 static void etm4_set_default_config(struct etmv4_config *config);
 static int etm4_set_event_filters(struct etmv4_drvdata *drvdata,
@@ -1403,7 +1403,7 @@ static int etm4_pm_setup_cpuslocked(void)
 {
 	int ret;
 
-	if (etm4_count++)
+	if (atomic_inc_return(&etm4_count))
 		return 0;
 
 	ret = cpu_pm_register_notifier(&etm4_cpu_pm_nb);
@@ -1434,13 +1434,13 @@ static int etm4_pm_setup_cpuslocked(void)
 	cpu_pm_unregister_notifier(&etm4_cpu_pm_nb);
 
 reduce_count:
-	--etm4_count;
+	atomic_dec(&etm4_count);
 	return ret;
 }
 
 static void etm4_pm_clear(void)
 {
-	if (--etm4_count != 0)
+	if (atomic_dec_return(&etm4_count) != 0)
 		return;
 
 	cpu_pm_unregister_notifier(&etm4_cpu_pm_nb);
-- 
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member
of Code Aurora Forum, hosted by The Linux Foundation

Re: [PATCH] coresight: etm4x: Fix etm4_count race using atomic variable

[Sai Prakash Ranjan]

On 2020-07-27 11:37, Sai Prakash Ranjan wrote:
> etm4_count keeps track of number of ETMv4 registered and on some
> systems, a race is observed on etm4_count variable which can
> lead to multiple calls to cpuhp_setup_state_nocalls_cpuslocked().
> This function internally calls cpuhp_store_callbacks() which
> prevents multiple registrations of callbacks for a given state
> and due to this race, it returns -EBUSY leading to ETM probe
> failures like below.
> 
>  coresight-etm4x: probe of 7040000.etm failed with error -16
> 
> This race can easily be triggered with async probe by setting
> probe type as PROBE_PREFER_ASYNCHRONOUS and with ETM power
> management property "arm,coresight-loses-context-with-cpu".
> 
> Prevent this race by converting etm4_count variable to atomic.
> 
> Fixes: 9b6a3f3633a5 ("coresight: etmv4: Fix CPU power management setup
> in probe() function")
> Fixes: 58eb457be028 ("hwtracing/coresight-etm4x: Convert to hotplug
> state machine")
> Suggested-by: Mike Leach <mike.leach@linaro.org>
> (Mike: Rootcause and context for commit message)
> Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan@codeaurora.org>
> ---
>  drivers/hwtracing/coresight/coresight-etm4x.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/hwtracing/coresight/coresight-etm4x.c
> b/drivers/hwtracing/coresight/coresight-etm4x.c
> index 6d7d2169bfb2..f256ea744c51 100644
> --- a/drivers/hwtracing/coresight/coresight-etm4x.c
> +++ b/drivers/hwtracing/coresight/coresight-etm4x.c
> @@ -49,7 +49,7 @@ MODULE_PARM_DESC(pm_save_enable,
>  	"Save/restore state on power down: 1 = never, 2 = self-hosted");
> 
>  /* The number of ETMv4 currently registered */
> -static int etm4_count;
> +static atomic_t etm4_count;
>  static struct etmv4_drvdata *etmdrvdata[NR_CPUS];
>  static void etm4_set_default_config(struct etmv4_config *config);
>  static int etm4_set_event_filters(struct etmv4_drvdata *drvdata,
> @@ -1403,7 +1403,7 @@ static int etm4_pm_setup_cpuslocked(void)
>  {
>  	int ret;
> 
> -	if (etm4_count++)
> +	if (atomic_inc_return(&etm4_count))
>  		return 0;
> 

Sorry, I messed up here, will send a next version fixing this.

Thanks,
Sai

-- 
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a 
member
of Code Aurora Forum, hosted by The Linux Foundation

Re: [PATCH] coresight: etm4x: Fix etm4_count race using atomic variable

[Suzuki K Poulose]

On 07/27/2020 07:07 AM, Sai Prakash Ranjan wrote:
> etm4_count keeps track of number of ETMv4 registered and on some
> systems, a race is observed on etm4_count variable which can
> lead to multiple calls to cpuhp_setup_state_nocalls_cpuslocked().
> This function internally calls cpuhp_store_callbacks() which
> prevents multiple registrations of callbacks for a given state
> and due to this race, it returns -EBUSY leading to ETM probe
> failures like below.
> 
>   coresight-etm4x: probe of 7040000.etm failed with error -16
> 
> This race can easily be triggered with async probe by setting
> probe type as PROBE_PREFER_ASYNCHRONOUS and with ETM power
> management property "arm,coresight-loses-context-with-cpu".
> 
> Prevent this race by converting etm4_count variable to atomic.
> 
> Fixes: 9b6a3f3633a5 ("coresight: etmv4: Fix CPU power management setup in probe() function")
> Fixes: 58eb457be028 ("hwtracing/coresight-etm4x: Convert to hotplug state machine")
> Suggested-by: Mike Leach <mike.leach@linaro.org>
> (Mike: Rootcause and context for commit message)
> Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan@codeaurora.org>

Please could we leave the hotplug notifier installed with the driver
init and don't worry about this at all ? We bail out early in the
notifier anyways, if the CPU is not registered with its ETM.

Cheers
Suzuki

Re: [PATCH] coresight: etm4x: Fix etm4_count race using atomic variable

[Sai Prakash Ranjan]

On 2020-07-27 15:09, Suzuki K Poulose wrote:
> On 07/27/2020 07:07 AM, Sai Prakash Ranjan wrote:
>> etm4_count keeps track of number of ETMv4 registered and on some
>> systems, a race is observed on etm4_count variable which can
>> lead to multiple calls to cpuhp_setup_state_nocalls_cpuslocked().
>> This function internally calls cpuhp_store_callbacks() which
>> prevents multiple registrations of callbacks for a given state
>> and due to this race, it returns -EBUSY leading to ETM probe
>> failures like below.
>> 
>>   coresight-etm4x: probe of 7040000.etm failed with error -16
>> 
>> This race can easily be triggered with async probe by setting
>> probe type as PROBE_PREFER_ASYNCHRONOUS and with ETM power
>> management property "arm,coresight-loses-context-with-cpu".
>> 
>> Prevent this race by converting etm4_count variable to atomic.
>> 
>> Fixes: 9b6a3f3633a5 ("coresight: etmv4: Fix CPU power management setup 
>> in probe() function")
>> Fixes: 58eb457be028 ("hwtracing/coresight-etm4x: Convert to hotplug 
>> state machine")
>> Suggested-by: Mike Leach <mike.leach@linaro.org>
>> (Mike: Rootcause and context for commit message)
>> Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan@codeaurora.org>
> 
> Please could we leave the hotplug notifier installed with the driver
> init and don't worry about this at all ? We bail out early in the
> notifier anyways, if the CPU is not registered with its ETM.
> 

Sure thing, sorry for not taking the cue from earlier discussion.

Thanks,
Sai

-- 
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a 
member
of Code Aurora Forum, hosted by The Linux Foundation