On Mon, May 13, 2024 at 12:14:32AM -0300, Leonardo Bras wrote: > On Sun, May 12, 2024 at 06:44:23PM -0300, Marcelo Tosatti wrote: > > On Fri, May 10, 2024 at 11:05:56PM -0300, Leonardo Bras wrote: > > > As of today, KVM notes a quiescent state only in guest entry, which is good > > > as it avoids the guest being interrupted for current RCU operations. > > > > > > While the guest vcpu runs, it can be interrupted by a timer IRQ that will > > > check for any RCU operations waiting for this CPU. In case there are any of > > > such, it invokes rcu_core() in order to sched-out the current thread and > > > note a quiescent state. > > > > > > This occasional schedule work will introduce tens of microsseconds of > > > latency, which is really bad for vcpus running latency-sensitive > > > applications, such as real-time workloads. > > > > > > So, note a quiescent state in guest exit, so the interrupted guests is able > > > to deal with any pending RCU operations before being required to invoke > > > rcu_core(), and thus avoid the overhead of related scheduler work. > > > > This does not properly fix the current problem, as RCU work might be > > scheduled after the VM exit, followed by a timer interrupt. > > > > Correct? > > Correct, for this case, check the note below: > > > > > > > > > Signed-off-by: Leonardo Bras <leobras@xxxxxxxxxx> > > > --- > > > > > > ps: A patch fixing this same issue was discussed in this thread: > > > https://lore.kernel.org/all/20240328171949.743211-1-leobras@xxxxxxxxxx/ > > > > > > Also, this can be paired with a new RCU option (rcutree.nocb_patience_delay) > > > to avoid having invoke_rcu() being called on grace-periods starting between > > > guest exit and the timer IRQ. This RCU option is being discussed in a > > > sub-thread of this message: > > > https://lore.kernel.org/all/5fd66909-1250-4a91-aa71-93cb36ed4ad5@paulmck-laptop/ > > ^ This one above. > The idea is to use this rcutree.nocb_patience_delay=N : > a new option we added on RCU that allow us to avoid invoking rcu_core() if > the grace_period < N miliseconds. This only works on nohz_full cpus. > > So with both the current patch and the one in above link, we have the same > effect as we previously had with last_guest_exit, with a cherry on top: we > can avoid rcu_core() getting called in situations where a grace period just > started after going into kernel code, and a timer interrupt happened before > it can report quiescent state again. > > For our nohz_full vcpu thread scenario, we have: > > - guest_exit note a quiescent state > - let's say we start a grace period in the next cycle > - If timer interrupts, it requires the grace period to be older than N > miliseconds > - If we configure a proper value for patience, it will never reach the > end of patience before going guest_entry, and thus noting a quiescent > state > > What do you think? I don't fully understand all of the RCU details, but since RCU quiescent state marking happens in IRQ disabled section, there is no chance for a timer interrupt to conflict with the marking of quiescent state. So seem to make sense to me.
