Re: [PATCH] Parallelize mtrr init between cpus

["zhenzhong.duan" <zhenzhong.duan@oracle.com>]

On 2012-08-08 11:53, H. Peter Anvin wrote:
> On 08/07/2012 07:08 PM, zhenzhong.duan wrote:
>>
>>
>>   2012-08-08 00:32, H. Peter Anvin wrote:
>>> On 08/07/2012 12:29 AM, zhenzhong.duan wrote:
>>>> Current code serialize mtrr init with set_atomicity_lock.
>>>> Mtrr init is quite slow when we bootup on a hvm with large mem, vcpus
>>>> and pci passthroughed devices(eg. 24 vcpus + 90G mem).
>>>> It took about ~30 mins to bootup, after patch, it took ~2 min.
>>> When you're saying "HVM" do you mean "Xen HVM"?  How does it behave on
>>> native hardware?
>>>
>>>     -hpa
>> Yes, I mean Xen HVM. Bootup at same speed as before on baremetal.
>
> Then I would like to know why Xen HVM takes so infernally long.  It 
> isn't a good idea to make gratuitous changes in the main kernel to 
> work around defects in Xen.
>
>     -hpa
Hi,
Sorrry for late response, I have an opportunity to test the patch on an 
old G5 with 128 physical CPUS and 1 TB of memory. Arround 0.82 second 
time saved.
Depends on cpu count on the system.
I'm curious what's the usage of set_atomicity_lock here? There may be 
something I missed. thanks

Before patch:
[   13.176038] NMI watchdog enabled, takes one hw-pmu counter.
[   13.176444] Brought up 128 CPUs
[   13.176688] Total of 128 processors activated (578760.49 BogoMIPS).
[   14.035223] devtmpfs: initialized
[   17.277652] PM: Registering ACPI NVS region at 7f76e000 (401408 bytes)
[   17.279541] print_constraints: dummy:
[   17.280576] NET: Registered protocol family 16
After patch:
[   13.094227] smpboot cpu 127: start_ip = 90000
[   13.185142] NMI watchdog enabled, takes one hw-pmu counter.
[   13.185599] Brought up 128 CPUs
[   13.185851] Total of 128 processors activated (578871.80 BogoMIPS).
[   13.224783] devtmpfs: initialized
[   16.468002] PM: Registering ACPI NVS region at 7f76e000 (401408 bytes)
[   16.469706] print_constraints: dummy:
[   16.470745] NET: Registered protocol family 16