Re: [PATCH] powerpc/perf: Fix core-imc hotplug callback failure during imc initialization

From: Madhavan Srinivasan
Date: Wed Nov 01 2017 - 05:27:44 EST



On Wednesday 01 November 2017 06:22 AM, Michael Ellerman wrote:
Anju T Sudhakar <anju@xxxxxxxxxxxxxxxxxx> writes:

Call trace observed during boot:
What's the actual oops?

I could recreate this in mambo with CPUS=2 and THREAD=2
Here is the complete stack trace.

[ÂÂÂ 0.045367] core_imc memory allocation for cpu 2 failed
[ÂÂÂ 0.045408] Unable to handle kernel paging request for data at address 0x7d20e2a6f92d03b8
[ÂÂÂ 0.045443] Faulting instruction address: 0xc0000000000dde18
cpu 0x0: Vector: 380 (Data Access Out of Range) at [c0000000fd1cb890]
ÂÂÂ pc: c0000000000dde18: event_function_call+0x28/0x14c
ÂÂÂ lr: c0000000000dde00: event_function_call+0x10/0x14c
ÂÂÂ sp: c0000000fd1cbb10
ÂÂ msr: 9000000000009033
ÂÂ dar: 7d20e2a6f92d03b8
 current = 0xc0000000fd15da00
 paca = 0xc00000000fff0000 softe: 0 irq_happened: 0x01
ÂÂÂ pidÂÂ = 11, comm = cpuhp/0
Linux version 4.14.0-rc7-00014-g0a08377b127b (maddy@SrihariSrinidhi) (gcc version 5.4.0 20160609 (Ubuntu/IBM 5.4.0-6ubuntu1~16.04.1)) #5 SMP Wed Nov 1 14:12:27 IST 2017
enter ? for help
[c0000000fd1cbb10] 0000000000000000 (unreliable)
[c0000000fd1cbba0] c0000000000de180 perf_remove_from_context+0x30/0x9c
[c0000000fd1cbbe0] c0000000000e9108 perf_pmu_migrate_context+0x9c/0x224
[c0000000fd1cbc60] c0000000000682e0 ppc_core_imc_cpu_offline+0xdc/0x144
[c0000000fd1cbcb0] c000000000070568 cpuhp_invoke_callback+0xe4/0x244
[c0000000fd1cbd10] c000000000070824 cpuhp_thread_fun+0x15c/0x1b0
[c0000000fd1cbd60] c00000000008e8cc smpboot_thread_fn+0x1e0/0x200
[c0000000fd1cbdc0] c00000000008ae58 kthread+0x150/0x158
[c0000000fd1cbe30] c00000000000b464 ret_from_kernel_thread+0x5c/0x78



[c000000ff38ffb80] c0000000002ddfac perf_pmu_migrate_context+0xac/0x470
[c000000ff38ffc40] c00000000011385c ppc_core_imc_cpu_offline+0x1ac/0x1e0
[c000000ff38ffc90] c000000000125758 cpuhp_invoke_callback+0x198/0x5d0
[c000000ff38ffd00] c00000000012782c cpuhp_thread_fun+0x8c/0x3d0
[c000000ff38ffd60] c0000000001678d0 smpboot_thread_fn+0x290/0x2a0
[c000000ff38ffdc0] c00000000015ee78 kthread+0x168/0x1b0
[c000000ff38ffe30] c00000000000b368 ret_from_kernel_thread+0x5c/0x74

While registering the cpuhoplug callbacks for core-imc, if we fails
in the cpuhotplug online path for any random core (either because opal call to
initialize the core-imc counters fails or because memory allocation fails for
that core), ppc_core_imc_cpu_offline() will get invoked for other cpus who
successfully returned from cpuhotplug online path.

But in the ppc_core_imc_cpu_offline() path we are trying to migrate the event
context, when core-imc counters are not even initialized. Thus creating the
above stack dump.

Add a check to see if core-imc counters are enabled or not in the cpuhotplug
offline path before migrating the context to handle this failing scenario.
Why do we need a bool to track this? Can't we just check the data
structure we're deinitialising has been initialised?

My bad. yes we could do that. Something like this will work?

@@ -606,6 +608,20 @@ static int ppc_core_imc_cpu_offline(unsigned int cpu)
ÂÂÂÂÂÂÂ if (!cpumask_test_and_clear_cpu(cpu, &core_imc_cpumask))
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return 0;

+ÂÂÂÂÂÂ /*
+ÂÂÂÂÂÂÂ * Check whether core_imc is registered. We could end up here
+ÂÂÂÂÂÂÂ * if the cpuhotplug callback registration fails. i.e, callback
+ÂÂÂÂÂÂÂ * invokes the offline path for all sucessfully registered cpus.
+ÂÂÂÂÂÂÂ * At this stage, core_imc pmu will not be registered and we
+ÂÂÂÂÂÂÂ * should return here.
+ÂÂÂÂÂÂÂ *
+ÂÂÂÂÂÂÂ * We return with a zero since this is not a offline failure.
+ÂÂÂÂÂÂÂ * And cpuhp_setup_state() returns the actual failure reason
+ÂÂÂÂÂÂÂ * to the caller, which inturn will call the cleanup routine.
+ÂÂÂÂÂÂÂ */
+ÂÂÂÂÂÂ if (!core_imc_pmu->pmu.event_init)
+ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return 0;
+
ÂÂÂÂÂÂÂ /* Find any online cpu in that core except the current "cpu" */
ÂÂÂÂÂÂÂ ncpu = cpumask_any_but(cpu_sibling_mask(cpu), cpu);



Doesn't this also mean we won't cleanup the initialisation for any CPUs
that have been initialised?

No, we do. cpuhp_setup_state() returns the actual failure reason
which in-turn calls the cleanup routine.

Thanks for review comments
Maddy


cheers

diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c
index 8812624..08139f9 100644
--- a/arch/powerpc/perf/imc-pmu.c
+++ b/arch/powerpc/perf/imc-pmu.c
@@ -30,6 +30,7 @@ static struct imc_pmu *per_nest_pmu_arr[IMC_MAX_PMUS];
static cpumask_t nest_imc_cpumask;
struct imc_pmu_ref *nest_imc_refc;
static int nest_pmus;
+static bool core_imc_enabled;
/* Core IMC data structures and variables */
@@ -607,6 +608,19 @@ static int ppc_core_imc_cpu_offline(unsigned int cpu)
if (!cpumask_test_and_clear_cpu(cpu, &core_imc_cpumask))
return 0;
+ /*
+ * See if core imc counters are enabled or not.
+ *
+ * Suppose we reach here from core_imc_cpumask_init(),
+ * since we failed at the cpuhotplug online path for any random
+ * core (either because opal call to initialize the core-imc counters
+ * failed or because memory allocation failed).
+ * We need to check whether core imc counters are enabled or not before
+ * migrating the event context from cpus in the other cores.
+ */
+ if (!core_imc_enabled)
+ return 0;
+
/* Find any online cpu in that core except the current "cpu" */
ncpu = cpumask_any_but(cpu_sibling_mask(cpu), cpu);
@@ -1299,6 +1313,7 @@ int init_imc_pmu(struct device_node *parent, struct imc_pmu *pmu_ptr, int pmu_id
return ret;
}
+ core_imc_enabled = true;
break;
case IMC_DOMAIN_THREAD:
ret = thread_imc_cpu_init();
--
2.7.4