[Xen-devel] Tracking down a boot speed issue

[Roger Cruz <roger.cruz@xxxxxxxxxx>]

With a Linux 3.2.23 PVOPs and Xen 4.0.4, I am observing that boot speed takes a 
hit when the following per_cpu access is made  in 
arch/x86/xen/enlighthen.c:xen_vcpu_setup().  The slow down is caused by the 
term on the left of the equal sign, ie, access to the per cpu area.  As you see 
from the timed samples below, the access can take anywhere from .6 to 1.2 secs 
(though I have seen it take 1.6 secs every single time with  Linux 3.2.16 and 
Xen 4.0.3).  Subsequent iterations from the loop do not incur the expense.  It 
also doesn't happen on older CPUs from what I can tell, only in the newer 
Sandybridge Intel i-series processors.  My initial guess is that there is some 
sort of page fault taking place but I haven't been able to confirm that.  Any 
ideas on how to track this farther down into Xen to isolate and understand what 
is going on?

Thank you.  
Roger R. Cruz

            printk(KERN_ALERT "in cpu loop %d for shared info", cpu);
===>    per_cpu(xen_vcpu,cpu) = &HYPERVISOR_shared_info->vcpu_info[cpu];
            printk(KERN_ALERT "in cpu loop %d computed shared info", cpu);

Oct 25 20:24:11 23445LU-PBBFMWY kernel: [    0.000000] in cpu loop 0 for shared 
info
Oct 25 20:24:11 23445LU-PBBFMWY kernel: [    1.238877] in cpu loop 0 computed 
shared info

Oct 25 20:52:53 23445LU-PBBFMWY kernel: [    0.000000] in cpu loop 0 for shared 
info
Oct 25 20:52:53 23445LU-PBBFMWY kernel: [    0.718365] in cpu loop 0 computed 
shared info

Oct 25 20:41:27 23445LU-PBBFMWY kernel: [    0.000000] in cpu loop 0 for shared 
info
Oct 25 20:41:27 23445LU-PBBFMWY kernel: [    0.671690] in cpu loop 0 computed 
shared info



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel