Messages in this thread | | | Subject | Re: [RFC PATCH] perf/x86/intel/rapl: avoid access unallocate memory | From | "Charles (Chas) Williams" <> | Date | Mon, 7 Nov 2016 11:59:52 -0500 |
| |
On 11/07/2016 11:19 AM, Thomas Gleixner wrote: > On Wed, 2 Nov 2016, Charles (Chas) Williams wrote: > >> On 11/02/2016 08:25 AM, Sebastian Andrzej Siewior wrote: >>> I am not sure if this a race with the new hotplug code or something that was >>> always there. Both (M. Vefa Bicakc and Charles) say that the box boots >>> sometimes fine (without the patch). smp_store_boot_cpu_info() should have >>> run >>> before the notofoert and thus should have set the info properly. However I >>> got >>> the following bootlog from Charles with this patch: >> >> I don't this this is a race. Here is some debugging from the two CPU VM >> (2 sockets, 1 core per socket). In identify_cpu() we have: >> >> /* The boot/hotplug time assigment got cleared, restore it */ >> c->logical_proc_id = topology_phys_to_logical_pkg(c->phys_proc_id); >> >> The values just after this: >> >> [ 0.228306] identify_cpu: c ffff88023fd0a040 logical_proc_id 65535 >> c->phys_proc_id 2 >> >> So what's interesting here, is the phys_proc_id of 2 for CPU1: >> >> int topology_phys_to_logical_pkg(unsigned int phys_pkg) >> { >> if (phys_pkg >= max_physical_pkg_id) >> return -1; >> return physical_to_logical_pkg[phys_pkg]; >> } >> >> And we happen to know the max_physical_pkg_id is 2 in this case. >> So apparently, topology_phys_to_logical_pkg() returns -1 and it gets >> assigned to the logical_proc_id. >> >> I don't know why the CPU's phys_proc_id is 2. > > max_physical_pkg_id gets initialized via: > > cpus = boot_cpu_data.x86_max_cores; > max_physical_pkg_id = DIV_ROUND_UP(MAX_LOCAL_APIC, ncpus); > > What's the value of boot_cpu_data.x86_max_cores and MAX_LOCAL_APIC?
I have discovered that that is not the problem. smp_init_package_map() is calculating the physical core id using the following:
for_each_present_cpu(cpu) { unsigned int apicid = apic->cpu_present_to_apicid(cpu);
... if (!topology_update_package_map(apicid, cpu)) continue;
... int topology_update_package_map(unsigned int apicid, unsigned int cpu) { unsigned int new, pkg = apicid >> boot_cpu_data.x86_coreid_bits;
But later when the secondary CPU's are identified they use a different calculation using the local APIC ID from the CPU's registers:
static void generic_identify(struct cpuinfo_x86 *c) ... if (c->cpuid_level >= 0x00000001) { c->initial_apicid = (cpuid_ebx(1) >> 24) & 0xFF; ... c->phys_proc_id = c->initial_apicid;
So at the end of identify_cpu() when the boot/hotplug assignment is put back:
c->logical_proc_id = topology_phys_to_logical_pkg(c->phys_proc_id);
topology_phys_to_logical_pkg() is returning an invalid logical processor since one isn't configured.
It's not clear to me what the right thing to do is or which is right.
| |