Paul E. McKenney wrote: > On Tue, Jan 12, 2010 at 01:23:15AM +0100, Jan Kiszka wrote: >> Jan Kiszka wrote: >>> Avi Kivity wrote: >>>> On 01/11/2010 12:13 PM, Jan Kiszka wrote: >>>>> BTW, does anybody know how to back-port synchronize_srcu_expedited best? >>>>> It looked like a simple mapping to synchronize_srcu was not sufficient >>>>> to achieve the same performance as with the pre-srcu locking (e.g. >>>>> guest&host stalled during guest's framebuffer setup). >>>>> >>>> Isn't it sufficient to backport kernel/srcu.c? I thought no sched.c >>>> changes were necessary. >>> Haven't looked yet, but if that's the case, it would indeed be >>> straightforward. >> It's far away from being straightforward: synchronize_rcu_expedited is >> based on synchronize_sched_expedited, introduced to 2.6.32. But that >> services is hooked deep into the scheduler, fiddling directly with >> runqueues (which are completely private to sched.c). This path looks >> like a dead end, specifically when its about supporting ~8 major Linux >> releases backwards. >> >> Paul, we have a problem here on the KVM-for-older-kernels front: We need >> synchronize_rcu_expedited for acceptable write-side performance (there >> are certain phases with lots of changes, plain synchronize_rcu just >> stalls both guest and host for several seconds). Our target kernels >> (down to 2.6.27, unofficially even 2.6.24) do not have the expedited >> service. Can you think of a poor man's solution for those kernels? >> >> Unfortunately, I don't think there is mechanical patching possible to >> role-back our srcu use to a rw-sem. But I will check this once again >> tomorrow. > > Would it help if I backported the RCU expedited stuff? Yes, it does > involve kernel/sched.c, but it is reasonably hands-off. > > If this would help, please let me know to which kernel version you need > the backport. Thanks a lot. Mmm, it's probably already sufficient when you tell me if I'm now on the right tracks: The idea behind synchronize_sched_expedited is, instead of waiting for random task switches on all CPUs, to enforce them. It now just reuses the migration_thread, locks and request queues from the rqs for this purpose. So it should have the same effect if we create our own high-prio srcu_expedited kthreads that include the srcu logic which just complete() the queued requests of some compat kvm_synchronize_sched_expedited. Did I get it? If so, I will try to write something like this the next days. Will surely appreciate your review afterwards! Thanks, Jan
Attachment:
signature.asc
Description: OpenPGP digital signature
