[Skiboot] [PATCH v10 16/17] core/opal: Use PCI slot and new APIs for slot states
stewart at linux.vnet.ibm.com
Fri May 13 10:48:06 AEST 2016
Alistair Popple <alistair at popple.id.au> writes:
>> My thought that is for P9 (as we can change ABI to kernel there :) we
>> could *always* hotplug - detect nothing before starting linux, just hotplug
>> it all in at runtime.
> We probably wouldn't even have to change the ABI would we? Just get skiboot to
> start up with all the slots "powered down" (even if they aren't physically
> powered down) with no device tree entries for anything plugged into the slots.
> Then once the system is up get the kernel to do a hotplug-in of each slot to
> do the scan and get dt entries...
Well ABI change would be having to have the kernel support hotplug to be
able to boot, so ABI to the skiboot payload kernel at least.
> Is PCI bus enumeration a significant part of skiboot startup time?
[3237993291,5] PCI: Resetting PHBs...
[4912461686,5] PCI: Probing slots...
so 3.2 seconds between those two (on Tuleta) and then:
[5548605580,7] PCI Summary:
1.2 seconds to get to there.
Whole boot time after CHIPTOD sync? 12.8 seconds (Tuleta).
So, on Tuleta, PCI is (one of) the biggest contributor to skiboot boot
time. On OpenPower, it's the damn slow-as-continental-drift flash.
OPAL Architect, IBM.
More information about the Skiboot