Re: [PATCH v2] powerpc/perf: Fix memory allocation for core-imc based on num_possible_cpus()
On Wednesday 16 May 2018 12:18 PM, ppaidipe wrote: On 2018-05-16 12:05, Anju T Sudhakar wrote: Currently memory is allocated for core-imc based on cpu_present_mask, which has bit 'cpu' set iff cpu is populated. We use (cpu number / threads per core) as the array index to access the memory. Under some circumstances firmware marks a CPU as GUARDed CPU and boot the system, until cleared of errors, these CPU's are unavailable for all subsequent boots. GUARDed CPUs are possible but not present from linux view, so it blows a hole when we assume the max length of our allocation is driven by our max present cpus, where as one of the cpus might be online and be beyond the max present cpus, due to the hole. So (cpu number / threads per core) value bounds the array index and leads to memory overflow. Call trace observed during a guard test: Faulting instruction address: 0xc0149f1c cpu 0x69: Vector: 380 (Data Access Out of Range) at [c03fea303420] pc:c0149f1c: prefetch_freepointer+0x14/0x30 lr:c014e0f8: __kmalloc+0x1a8/0x1ac sp:c03fea3036a0 msr:90009033 dar:c9c54b2c91dbf6b7 current = 0xc03fea2c paca = 0xcfddd880 softe: 3 irq_happened: 0x01 pid = 1, comm = swapper/104 Linux version 4.16.7-openpower1 (smc@smc-desktop) (gcc version 6.4.0 (Buildroot 2018.02.1-6-ga8d1126)) #2 SMP Fri May 4 16:44:54 PDT 2018 enter ? for help call trace: __kmalloc+0x1a8/0x1ac (unreliable) init_imc_pmu+0x7f4/0xbf0 opal_imc_counters_probe+0x3fc/0x43c platform_drv_probe+0x48/0x80 driver_probe_device+0x22c/0x308 __driver_attach+0xa0/0xd8 bus_for_each_dev+0x88/0xb4 driver_attach+0x2c/0x40 bus_add_driver+0x1e8/0x228 driver_register+0xd0/0x114 __platform_driver_register+0x50/0x64 opal_imc_driver_init+0x24/0x38 do_one_initcall+0x150/0x15c kernel_init_freeable+0x250/0x254 kernel_init+0x1c/0x150 ret_from_kernel_thread+0x5c/0xc8 Allocating memory for core-imc based on cpu_possible_mask, which has bit 'cpu' set iff cpu is populatable, will fix this issue. Reported-by: Pridhiviraj PaidipeddiSigned-off-by: Anju T Sudhakar Reviewed-by: Balbir Singh I have verified this fix with both normal and kexec boot multiple times when system is having GARDed cores. Not seen any crash/memory corruption issues with this. Tested-by: Pridhiviraj Paidipeddi Thanks. -Anju --- arch/powerpc/perf/imc-pmu.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c index d7532e7..75fb23c 100644 --- a/arch/powerpc/perf/imc-pmu.c +++ b/arch/powerpc/perf/imc-pmu.c @@ -1146,7 +1146,7 @@ static int init_nest_pmu_ref(void) static void cleanup_all_core_imc_memory(void) { - int i, nr_cores = DIV_ROUND_UP(num_present_cpus(), threads_per_core); + int i, nr_cores = DIV_ROUND_UP(num_possible_cpus(), threads_per_core); struct imc_mem_info *ptr = core_imc_pmu->mem_info; int size = core_imc_pmu->counter_mem_size; @@ -1264,7 +1264,7 @@ static int imc_mem_init(struct imc_pmu *pmu_ptr, struct device_node *parent, if (!pmu_ptr->pmu.name) return -ENOMEM; - nr_cores = DIV_ROUND_UP(num_present_cpus(), threads_per_core); + nr_cores = DIV_ROUND_UP(num_possible_cpus(), threads_per_core); pmu_ptr->mem_info = kcalloc(nr_cores, sizeof(struct imc_mem_info), GFP_KERNEL);
Re: [PATCH v2] powerpc/perf: Fix memory allocation for core-imc based on num_possible_cpus()
On 2018-05-16 12:05, Anju T Sudhakar wrote: Currently memory is allocated for core-imc based on cpu_present_mask, which has bit 'cpu' set iff cpu is populated. We use (cpu number / threads per core) as the array index to access the memory. Under some circumstances firmware marks a CPU as GUARDed CPU and boot the system, until cleared of errors, these CPU's are unavailable for all subsequent boots. GUARDed CPUs are possible but not present from linux view, so it blows a hole when we assume the max length of our allocation is driven by our max present cpus, where as one of the cpus might be online and be beyond the max present cpus, due to the hole. So (cpu number / threads per core) value bounds the array index and leads to memory overflow. Call trace observed during a guard test: Faulting instruction address: 0xc0149f1c cpu 0x69: Vector: 380 (Data Access Out of Range) at [c03fea303420] pc:c0149f1c: prefetch_freepointer+0x14/0x30 lr:c014e0f8: __kmalloc+0x1a8/0x1ac sp:c03fea3036a0 msr:90009033 dar:c9c54b2c91dbf6b7 current = 0xc03fea2c paca= 0xcfddd880 softe: 3irq_happened: 0x01 pid = 1, comm = swapper/104 Linux version 4.16.7-openpower1 (smc@smc-desktop) (gcc version 6.4.0 (Buildroot 2018.02.1-6-ga8d1126)) #2 SMP Fri May 4 16:44:54 PDT 2018 enter ? for help call trace: __kmalloc+0x1a8/0x1ac (unreliable) init_imc_pmu+0x7f4/0xbf0 opal_imc_counters_probe+0x3fc/0x43c platform_drv_probe+0x48/0x80 driver_probe_device+0x22c/0x308 __driver_attach+0xa0/0xd8 bus_for_each_dev+0x88/0xb4 driver_attach+0x2c/0x40 bus_add_driver+0x1e8/0x228 driver_register+0xd0/0x114 __platform_driver_register+0x50/0x64 opal_imc_driver_init+0x24/0x38 do_one_initcall+0x150/0x15c kernel_init_freeable+0x250/0x254 kernel_init+0x1c/0x150 ret_from_kernel_thread+0x5c/0xc8 Allocating memory for core-imc based on cpu_possible_mask, which has bit 'cpu' set iff cpu is populatable, will fix this issue. Reported-by: Pridhiviraj PaidipeddiSigned-off-by: Anju T Sudhakar Reviewed-by: Balbir Singh I have verified this fix with both normal and kexec boot multiple times when system is having GARDed cores. Not seen any crash/memory corruption issues with this. Tested-by: Pridhiviraj Paidipeddi --- arch/powerpc/perf/imc-pmu.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c index d7532e7..75fb23c 100644 --- a/arch/powerpc/perf/imc-pmu.c +++ b/arch/powerpc/perf/imc-pmu.c @@ -1146,7 +1146,7 @@ static int init_nest_pmu_ref(void) static void cleanup_all_core_imc_memory(void) { - int i, nr_cores = DIV_ROUND_UP(num_present_cpus(), threads_per_core); + int i, nr_cores = DIV_ROUND_UP(num_possible_cpus(), threads_per_core); struct imc_mem_info *ptr = core_imc_pmu->mem_info; int size = core_imc_pmu->counter_mem_size; @@ -1264,7 +1264,7 @@ static int imc_mem_init(struct imc_pmu *pmu_ptr, struct device_node *parent, if (!pmu_ptr->pmu.name) return -ENOMEM; - nr_cores = DIV_ROUND_UP(num_present_cpus(), threads_per_core); + nr_cores = DIV_ROUND_UP(num_possible_cpus(), threads_per_core); pmu_ptr->mem_info = kcalloc(nr_cores, sizeof(struct imc_mem_info), GFP_KERNEL);
[PATCH v2] powerpc/perf: Fix memory allocation for core-imc based on num_possible_cpus()
Currently memory is allocated for core-imc based on cpu_present_mask, which has bit 'cpu' set iff cpu is populated. We use (cpu number / threads per core) as the array index to access the memory. Under some circumstances firmware marks a CPU as GUARDed CPU and boot the system, until cleared of errors, these CPU's are unavailable for all subsequent boots. GUARDed CPUs are possible but not present from linux view, so it blows a hole when we assume the max length of our allocation is driven by our max present cpus, where as one of the cpus might be online and be beyond the max present cpus, due to the hole. So (cpu number / threads per core) value bounds the array index and leads to memory overflow. Call trace observed during a guard test: Faulting instruction address: 0xc0149f1c cpu 0x69: Vector: 380 (Data Access Out of Range) at [c03fea303420] pc:c0149f1c: prefetch_freepointer+0x14/0x30 lr:c014e0f8: __kmalloc+0x1a8/0x1ac sp:c03fea3036a0 msr:90009033 dar:c9c54b2c91dbf6b7 current = 0xc03fea2c paca= 0xcfddd880 softe: 3irq_happened: 0x01 pid = 1, comm = swapper/104 Linux version 4.16.7-openpower1 (smc@smc-desktop) (gcc version 6.4.0 (Buildroot 2018.02.1-6-ga8d1126)) #2 SMP Fri May 4 16:44:54 PDT 2018 enter ? for help call trace: __kmalloc+0x1a8/0x1ac (unreliable) init_imc_pmu+0x7f4/0xbf0 opal_imc_counters_probe+0x3fc/0x43c platform_drv_probe+0x48/0x80 driver_probe_device+0x22c/0x308 __driver_attach+0xa0/0xd8 bus_for_each_dev+0x88/0xb4 driver_attach+0x2c/0x40 bus_add_driver+0x1e8/0x228 driver_register+0xd0/0x114 __platform_driver_register+0x50/0x64 opal_imc_driver_init+0x24/0x38 do_one_initcall+0x150/0x15c kernel_init_freeable+0x250/0x254 kernel_init+0x1c/0x150 ret_from_kernel_thread+0x5c/0xc8 Allocating memory for core-imc based on cpu_possible_mask, which has bit 'cpu' set iff cpu is populatable, will fix this issue. Reported-by: Pridhiviraj PaidipeddiSigned-off-by: Anju T Sudhakar Reviewed-by: Balbir Singh --- arch/powerpc/perf/imc-pmu.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/arch/powerpc/perf/imc-pmu.c b/arch/powerpc/perf/imc-pmu.c index d7532e7..75fb23c 100644 --- a/arch/powerpc/perf/imc-pmu.c +++ b/arch/powerpc/perf/imc-pmu.c @@ -1146,7 +1146,7 @@ static int init_nest_pmu_ref(void) static void cleanup_all_core_imc_memory(void) { - int i, nr_cores = DIV_ROUND_UP(num_present_cpus(), threads_per_core); + int i, nr_cores = DIV_ROUND_UP(num_possible_cpus(), threads_per_core); struct imc_mem_info *ptr = core_imc_pmu->mem_info; int size = core_imc_pmu->counter_mem_size; @@ -1264,7 +1264,7 @@ static int imc_mem_init(struct imc_pmu *pmu_ptr, struct device_node *parent, if (!pmu_ptr->pmu.name) return -ENOMEM; - nr_cores = DIV_ROUND_UP(num_present_cpus(), threads_per_core); + nr_cores = DIV_ROUND_UP(num_possible_cpus(), threads_per_core); pmu_ptr->mem_info = kcalloc(nr_cores, sizeof(struct imc_mem_info), GFP_KERNEL); -- 2.7.4