Fix Arm system PMU hotplug issues
v2 cover Fix Arm system PMU hotplug issues
Hi all,
After a converstaion with Thomas a while back, it felt like the best way forward here is just to resolve the pressing preemption violation, and not make things more complicated by trying to manage theoretical races within perf core from a distance.
Thus v2 simply nukes the offending logic and brings those drivers in line with the status quo. Due to the abrupt change in tactic I've not kept the review tags given for v1.
Robin.
Robin Murphy (2):
perf/arm-cci: Remove broken race mitigation
perf/arm-ccn: Remove broken race mitigation
drivers/perf/arm-cci.c | 16 +++++++--------- drivers/perf/arm-ccn.c | 23 +++++++++++------------ 2 files changed, 18 insertions(+), 21 deletions(-)
Signature · 1 line
2.21.0.dirty
PATCH v2 1/2 perf/arm-cci: Remove broken race mitigation
Uncore PMU drivers face an awkward cyclic dependency wherein:
- They have to pick a valid online CPU to associate with before registering the PMU device, since it will get exposed to userspace immediately.
- The PMU registration has to be be at least partly complete before hotplug events can be handled, since trying to migrate an uninitialised context would be bad.
- The hotplug handler has to be ready as soon as a CPU is chosen, lest it go offline without the user-visible cpumask value getting updated.
The arm-cci driver has tried to solve this by using get_cpu() to pick the current CPU and prevent it from disappearing while both registrations are performed, but that results in taking mutexes with preemption disabled, which makes certain configurations very unhappy:
[ 1.983337] BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:2004 [ 1.983340] in_atomic(): 1, irqs_disabled(): 0, pid: 1, name: swapper/0 [ 1.983342] Preemption disabled at: [ 1.983353] [<ffffff80089801f4>] cci_pmu_probe+0x1dc/0x488 [ 1.983360] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.18.20-rt8-yocto-preempt-rt #1 [ 1.983362] Hardware name: ZynqMP ZCU102 Rev1.0 (DT) [ 1.983364] Call trace: [ 1.983369] dump_backtrace+0x0/0x158 [ 1.983372] show_stack+0x24/0x30 [ 1.983378] dump_stack+0x80/0xa4 [ 1.983383] ___might_sleep+0x138/0x160 [ 1.983386] __might_sleep+0x58/0x90 [ 1.983391] __rt_mutex_lock_state+0x30/0xc0 [ 1.983395] _mutex_lock+0x24/0x30 [ 1.983400] perf_pmu_register+0x2c/0x388 [ 1.983404] cci_pmu_probe+0x2bc/0x488 [ 1.983409] platform_drv_probe+0x58/0xa8
It is not feasible to resolve all the possible races outside of the perf core itself, so address the immediate bug by following the example of nearly every other PMU driver and not even trying to do so. Registering the hotplug notifier first should minimise the window in which things can go wrong, so that's about as much as we can reasonably do here. This also revealed an additional race in assigning the global pointer too late relative to the hotplug notifier, which gets fixed in the process.
--- drivers/perf/arm-cci.c | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-)
drivers/perf/arm-cci.c +7 -9
@@ -1684,19 +1684,17 @@ static int cci_pmu_probe(struct platform_device *pdev)raw_spin_lock_init (&cci_pmu->hw_events.pmu_lock);mutex_init (&cci_pmu->reserve_mutex);atomic_set (&cci_pmu->active_events,0 ); - cci_pmu->cpu =get_cpu (); - - ret =cci_pmu_init (cci_pmu, pdev); -if (ret) { -put_cpu (); -return ret; - } + cci_pmu->cpu =raw_smp_processor_id (); + g_cci_pmu = cci_pmu;cpuhp_setup_state_nocalls (CPUHP_AP_PERF_ARM_CCI_ONLINE ,"perf/arm/cci:online" ,NULL , cci_pmu_offline_cpu); -put_cpu (); - g_cci_pmu = cci_pmu; + + ret =cci_pmu_init (cci_pmu, pdev); +if (ret) +return ret; +pr_info ("ARM %s PMU driver probed" , cci_pmu->model->name);return 0 ; }
Signature · 1 line
2.21.0.dirty
Re: [PATCH v2 1/2] perf/arm-cci: Remove broken race mitigation
Quoted 42 lines · Robin Murphy
Uncore PMU drivers face an awkward cyclic dependency wherein:
- They have to pick a valid online CPU to associate with before registering the PMU device, since it will get exposed to userspace immediately.
- The PMU registration has to be be at least partly complete before hotplug events can be handled, since trying to migrate an uninitialised context would be bad.
- The hotplug handler has to be ready as soon as a CPU is chosen, lest it go offline without the user-visible cpumask value getting updated.The arm-cci driver has tried to solve this by using
get_cpu()to pick the current CPU and prevent it from disappearing while both registrations are performed, but that results in taking mutexes with preemption disabled, which makes certain configurations very unhappy:[ 1.983337] BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:2004 [ 1.983340] in_atomic(): 1, irqs_disabled(): 0, pid: 1, name: swapper/0 [ 1.983342] Preemption disabled at: [ 1.983353] [<ffffff80089801f4>] cci_pmu_probe+0x1dc/0x488 [ 1.983360] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.18.20-rt8-yocto-preempt-rt #1 [ 1.983362] Hardware name: ZynqMP ZCU102 Rev1.0 (DT) [ 1.983364] Call trace: [ 1.983369] dump_backtrace+0x0/0x158 [ 1.983372] show_stack+0x24/0x30 [ 1.983378] dump_stack+0x80/0xa4 [ 1.983383] ___might_sleep+0x138/0x160 [ 1.983386] __might_sleep+0x58/0x90 [ 1.983391] __rt_mutex_lock_state+0x30/0xc0 [ 1.983395] _mutex_lock+0x24/0x30 [ 1.983400] perf_pmu_register+0x2c/0x388 [ 1.983404] cci_pmu_probe+0x2bc/0x488 [ 1.983409] platform_drv_probe+0x58/0xa8It is not feasible to resolve all the possible races outside of the perf core itself, so address the immediate bug by following the example of nearly every other PMU driver and not even trying to do so. Registering the hotplug notifier first should minimise the window in which things can go wrong, so that's about as much as we can reasonably do here. This also revealed an additional race in assigning the global pointer too late relative to the hotplug notifier, which gets fixed in the process.
Quoted 3 lines
Reported-by "Li, Meng" <[email protected]>Reported-by Corentin Labbe <[email protected]>Signed-off-by Robin Murphy <[email protected]>
Thanks for fixing this.
PATCH v2 2/2 perf/arm-ccn: Remove broken race mitigation
In reply to Robin MurphyLike arm-cci, arm-ccn has the same issue of disabling preemption around operations which can take mutexes. Again, remove the definite bug by simply not trying to fight the theoretical races. And since we are touching the hotplug handling code, take the opportunity to streamline it, as there's really no need to store a full-sized cpumask to keep track of a single CPU ID.
--- drivers/perf/arm-ccn.c | 23 +++++++++++------------ 1 file changed, 11 insertions(+), 12 deletions(-)
drivers/perf/arm-ccn.c +11 -12
@@ -167,7 +167,7 @@ struct arm_ccn_dt {struct hrtimer hrtimer; -cpumask_t cpu; +unsigned int cpu;struct hlist_node node;struct pmu pmu;
@@ -559,7 +559,7 @@ static ssize_t arm_ccn_pmu_cpumask_show(struct device *dev, {struct arm_ccn *ccn =pmu_to_arm_ccn (dev_get_drvdata (dev)); -return cpumap_print_to_pagebuf (true, buf, &ccn->dt.cpu); +return cpumap_print_to_pagebuf (true, buf,cpumask_of (ccn->dt.cpu)); } static struct device_attribute arm_ccn_pmu_cpumask_attr =
@@ -759,7 +759,7 @@ static int arm_ccn_pmu_event_init(struct perf_event *event)* mitigate this, we enforce CPU assignment to one, selected * processor (the one described in the "cpumask" attribute). */ - event->cpu =cpumask_first (&ccn->dt.cpu); + event->cpu = ccn->dt.cpu; node_xp =CCN_CONFIG_NODE (event->attr.config); type =CCN_CONFIG_TYPE (event->attr.config);
@@ -1215,15 +1215,15 @@ static int arm_ccn_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node) struct arm_ccn *ccn = container_of(dt, struct arm_ccn, dt);unsigned int target; -if (!cpumask_test_and_clear_cpu (cpu, &dt->cpu)) +if (cpu != dt->cpu)return 0 ; target =cpumask_any_but (cpu_online_mask, cpu);if (target >= nr_cpu_ids)return 0 ;perf_pmu_migrate_context (&dt->pmu, cpu, target); -cpumask_set_cpu (target, &dt->cpu); + dt->cpu = target;if (ccn->irq) -WARN_ON (irq_set_affinity_hint (ccn->irq, &dt->cpu) !=0 ); +WARN_ON (irq_set_affinity_hint (ccn->irq,cpumask_of (dt->cpu)));return 0 ; }
@@ -1299,29 +1299,28 @@ static int arm_ccn_pmu_init(struct arm_ccn *ccn) }/* Pick one CPU which we will use to collect data from CCN... */ -cpumask_set_cpu (get_cpu (), &ccn->dt.cpu); + ccn->dt.cpu =raw_smp_processor_id ();/* Also make sure that the overflow interrupt is handled by this CPU */ if (ccn->irq) { - err =irq_set_affinity_hint (ccn->irq, &ccn->dt.cpu); + err =irq_set_affinity_hint (ccn->irq,cpumask_of (ccn->dt.cpu));if (err) {dev_err (ccn->dev,"Failed to set interrupt affinity!\n" ); goto error_set_affinity; } } +cpuhp_state_add_instance_nocalls (CPUHP_AP_PERF_ARM_CCN_ONLINE , + &ccn->dt.node); + err =perf_pmu_register (&ccn->dt.pmu, name,-1 );if (err) goto error_pmu_register; -cpuhp_state_add_instance_nocalls (CPUHP_AP_PERF_ARM_CCN_ONLINE , - &ccn->dt.node); -put_cpu ();return 0 ; error_pmu_register: error_set_affinity: -put_cpu (); error_choose_name:ida_simple_remove (&arm_ccn_pmu_ida, ccn->dt.id);for (i =0 ; i < ccn->num_xps; i++)
Signature · 1 line
2.21.0.dirty
Re: [PATCH v2 2/2] perf/arm-ccn: Remove broken race mitigation
Hi Robin,
Quoted 27 lines · Robin Murphy
Like arm-cci, arm-ccn has the same issue of disabling preemption around operations which can take mutexes. Again, remove the definite bug by simply not trying to fight the theoretical races. And since we are touching the hotplug handling code, take the opportunity to streamline it, as there's really no need to store a full-sized cpumask to keep track of a single CPU ID.
Signed-off-by Robin Murphy <[email protected]>--- drivers/perf/arm-ccn.c | 23 +++++++++++------------ 1 file changed, 11 insertions(+), 12 deletions(-)drivers/perf/arm-ccn.c +2
-2/* Pick one CPU which we will use to collect data from CCN... */ -cpumask_set_cpu (get_cpu (), &ccn->dt.cpu); + ccn->dt.cpu =raw_smp_processor_id ();/* Also make sure that the overflow interrupt is handled by this CPU */ if (ccn->irq) { - err =irq_set_affinity_hint (ccn->irq, &ccn->dt.cpu); + err =irq_set_affinity_hint (ccn->irq,cpumask_of (ccn->dt.cpu));if (err) {dev_err (ccn->dev,"Failed to set interrupt affinity!\n" );
drivers/perf/arm-ccn.c static int arm_ccn_pmu_init(struct arm_ccn *ccn) · line 1318
Quoted 10 lines
drivers/perf/arm-ccn.c +3
-0goto error_set_affinity; } } +cpuhp_state_add_instance_nocalls (CPUHP_AP_PERF_ARM_CCN_ONLINE , + &ccn->dt.node); + err =perf_pmu_register (&ccn->dt.pmu, name,-1 );if (err) goto error_pmu_register;
Should we not remove the above instance, in case we fail to register the PMU ? Similarly for the CCI driver, we may have to reset the g_cci_pmu if we fail.
Cheers
Suzuki
Quoted 13 lines
drivers/perf/arm-ccn.c +0
-4-cpuhp_state_add_instance_nocalls (CPUHP_AP_PERF_ARM_CCN_ONLINE , - &ccn->dt.node); -put_cpu ();return 0 ; error_pmu_register: error_set_affinity: -put_cpu (); error_choose_name:ida_simple_remove (&arm_ccn_pmu_ida, ccn->dt.id);for (i =0 ; i < ccn->num_xps; i++)
Quoted 14 lines · Robin Murphy
Hi all,
After a converstaion with Thomas a while back, it felt like the best way forward here is just to resolve the pressing preemption violation, and not make things more complicated by trying to manage theoretical races within perf core from a distance.
Thus v2 simply nukes the offending logic and brings those drivers in line with the status quo. Due to the abrupt change in tactic I've not kept the review tags given for v1.
Robin.
Quoted 11 lines
Robin Murphy (2):
perf/arm-cci: Remove broken race mitigation
perf/arm-ccn: Remove broken race mitigationdrivers/perf/arm-cci.c | 16 +++++++--------- drivers/perf/arm-ccn.c | 23 +++++++++++------------ 2 files changed, 18 insertions(+), 21 deletions(-)Signature · 1 line
2.21.0.dirty
Hello
Tested-on: sun8i-a83t-bananapi-m3
Thanks for having fixed this problem.
Regards