Re: [RFC] X86 SMP cpu enumeration changes

Re: [RFC] X86 SMP cpu enumeration changes

[Mark Rustad]

As a followup to the message I posted earlier, I found that the  
following patch restores the old behavior of allowing maxcpus to  
effectively turn off hyperthreading. In case you missed my earlier  
message, in 2.6.16 through at least 2.6.19 using the maxcpus trick  
will fail to use all of your physical cpus.

Although this patch works for me, I am quite sure that this patch is  
unacceptable to folks using CPU hotplug, however. One alternative  
would be to make the two lines of code in each location conditional  
on CONFIG_HOTPLUG_CPU and !CONFIG_HOTPLUG_CPU respectively, but that  
seems a little ugly.

Does anyone have a better fix?

I know my mailer smashes the patch, but this is intended for  
discussion rather than application.

--- a/arch/i386/kernel/mpparse.c        2006-11-29 15:57:37.000000000  
-0600
+++ b/arch/i386/kernel/mpparse.c	2007-06-06 16:41:06.635814335 -0500
@@ -180,9 +180,6 @@ static void __devinit MP_processor_info
         }
         apic_version[m->mpc_apicid] = ver;
-       phys_cpu = apicid_to_cpu_present(apicid);
-       physids_or(phys_cpu_present_map, phys_cpu_present_map,  
phys_cpu);
-
         if (num_processors >= NR_CPUS) {
                 printk(KERN_WARNING "WARNING: NR_CPUS limit of %i  
reached."
                         "  Processor ignored.\n", NR_CPUS);
@@ -195,6 +192,9 @@ static void __devinit MP_processor_info
                 return;
         }
+       phys_cpu = apicid_to_cpu_present(apicid);
+       physids_or(phys_cpu_present_map, phys_cpu_present_map,  
phys_cpu);
+
         cpu_set(num_processors, cpu_possible_map);
         num_processors++;


-- 
Mark Rustad, MRustad@gmail.com

[PATCH] [RFC] wrong cpus started

[Mark Rustad]

I have noticed that recent kernels, 2.6.16 through at least 2.6.19, hyperthreading
can no longer be effectively disabled by using maxcpus. This means that anyone that
was using maxcpus to disable hyperthreading when running Linux will no longer be
using all of their physical cpus.

The following patch restores the earlier behavior, but I am fairly sure that some
CPU hotplug environments would not want this change. I can imagine making the two
lines moved in the patch below to instead be conditional on CONFIG_HOTPLUG_CPU and
!CONFIG_HOTPLUG_CPU respectively. Can someone more familiar with hotplug CPU come
up with a better fix?

--- a/arch/i386/kernel/mpparse.c	2006-11-29 15:57:37.000000000 -0600
+++ b/arch/i386/kernel/mpparse.c	2007-06-06 16:41:06.635814335 -0500
@@ -180,9 +180,6 @@ static void __devinit MP_processor_info 
 	}
 	apic_version[m->mpc_apicid] = ver;
 
-	phys_cpu = apicid_to_cpu_present(apicid);
-	physids_or(phys_cpu_present_map, phys_cpu_present_map, phys_cpu);
-
 	if (num_processors >= NR_CPUS) {
 		printk(KERN_WARNING "WARNING: NR_CPUS limit of %i reached."
 			"  Processor ignored.\n", NR_CPUS);
@@ -195,6 +192,9 @@ static void __devinit MP_processor_info 
 		return;
 	}
 
+	phys_cpu = apicid_to_cpu_present(apicid);
+	physids_or(phys_cpu_present_map, phys_cpu_present_map, phys_cpu);
+
 	cpu_set(num_processors, cpu_possible_map);
 	num_processors++;
 
-- 
Mark Rustad, mrustad@gmail.com