Messages in this thread | | | Subject | Re: [PATCH v5 2/3] x86/topology: Avoid wasting 128k for package id array | From | Prarit Bhargava <> | Date | Sun, 12 Nov 2017 11:55:25 -0500 |
| |
On 11/12/2017 08:36 AM, Thomas Gleixner wrote: > On Fri, 10 Nov 2017, Andi Kleen wrote: >>>> All of that works. There is no way to make sure that a lookup is fully >>>> serialized against a concurrent update. Even if the lookup holds >>>> cpu_read_lock() the new package might arrive right after the unlock. >>>> >>> >>> Thanks Thomas. >>> >>> Andi, do you want to take a look at this? >> >> I was originally worried about races, that is why i tried to put >> everything into cpu_data. But that didn't work out because something >> clears it. Perhaps the right solution would be some extra per_cpu >> data variables, and search for the first match. I suspect that would >> be simpler. But if that doesn't work I guess something like Thomas' >> example will work. > > Sure, we can use a separate per cpu variable. The race for looking up phys > -> logical will always be there if that handles stuff like the uncore PCI > physid one. There is not much which can prevent that. > > The other option is to figure out what clears cpu_data on online and just > preserve the logical/physcial translation across that clear. > > One thing you need to be careful about (in both cases) is the value. The > data is zeroed on boot, so we either need to fill that with UINT_MAX at > boot time in one of the functions which does a for_each_possible_cpu() loop > anyway or just leave 0 as the 'not initialized' value and make the first > logical package be '1'. The readout functions (percpu, translation, > etc.) just can subtract 1.
I'll look into this (code & test) and get back with a v6.
Thanks Thomas,
P.
> > Thanks, > > tglx >
| |