Messages in this thread | | | Date | Tue, 14 Feb 2012 02:25:33 +0530 | From | "Srivatsa S. Bhat" <> | Subject | Re: [PATCH] Avoid mask based num_possible_cpus and num_online_cpus -v5 |
| |
On 02/14/2012 02:13 AM, Venki Pallipadi wrote:
> On Mon, Feb 13, 2012 at 12:25 PM, Srivatsa S. Bhat > <srivatsa.bhat@linux.vnet.ibm.com> wrote: >> On 02/14/2012 01:24 AM, Tony Luck wrote: >> >>> On Thu, Feb 2, 2012 at 12:03 PM, Rusty Russell <rusty@rustcorp.com.au> wrote: >>>> IIRC playing with 3 archs boot code seemed like a recipe for disaster. >>>> Feel free to try to fix this in -next though, and see what breaks... >>> >>> ia64 is what breaks ... well not actually broken ... but some very >>> weird delays that >>> show up in different places depending on whether this patch is present. >>> >>> First linux-next kernel to be blessed with this patch was >>> next-20120210. Booting it >>> I see: >>> [ 7.164233] Switching to clocksource itc >>> [ 146.077315] pnp: PnP ACPI init >>> >>> An ugly 138.913 second delay. Digging in the code showed that the bad bits >>> happened inside stop_machine() >>> >>> Reverting just this patch makes this big delay disappear: >>> >>> [ 32.780232] Switching to clocksource itc >>> [ 32.832100] pnp: PnP ACPI init >>> >>> but notice that it takes 25 extra seconds to get to this point in the >>> boot (and while >>> we expect to save some time by not re-computing num_online_cpus each time we >>> need it ... this looks to be a lot more than I'd expect!) >>> >> >> >> Oh no!! ia64 directly uses cpu_set() and cpu_clear() on cpu_online_map!! >> Grr.. It means num_online_cpus can be different from the actual number of >> online cpus because it doesn't go through the set_cpu_online() path.. I haven't >> yet pin-pointed the exact problem, but this definitely doesn't look good... >> > > This feels like a minefield in general. ia64, mips and um seems to > have cpu_set and cpu_clear of cpu_online_map and/or cpu_possible_map > in there. >
Since I had almost never seen code using "cpu_online_map" instead of "cpu_online_mask", I hadn't checked it while reviewing your patch... :-( Honestly, it is only now that I realized that there is this other variant too!
Regards, Srivatsa S. Bhat
| |