On 23/04/15 13:42, David Vrabel wrote: > On 21/04/15 11:22, Jan Beulich wrote: >>>>> On 21.04.15 at 12:11, <david.vra...@citrix.com> wrote: >>> We have analysed the affect of this series on interrupt latency (by >>> measuring the duration of irq disable/enable regions) and there is no >>> signficant impact. >>> >>> http://xenbits.xen.org/people/dvrabel/bar2_comp.png >> >> Thanks for doing this! >> >>> Interestingly, the biggest offenders for long critical sections are >>> bare irq disable/enable regions, and some of these are really bad (10s >>> of ms)! >>> >>> http://xenbits.xen.org/people/dvrabel/bar_normal_busy.png >> >> Despite both pictures saying micro-seconds at the respective >> axis (and hence the problem not being _as bad_) > > This particular graph doesn't show it very well but there are small > blips around 4, 10, and 31 ms.
This first set of analysis incorrect included some bogus data. 1. The mwait idle driver has interrupts masked for the mwait instruction (with the wake-on-interrupt bit enabled), thus the very long times were measuring idle time. 2. do_IRQ() calls spin_unlock_irq() without a corresponding spin_lock_irq() thus the values for this lock were incorrect. Here's an updated graph (Thanks to Jennifer). http://xenbits.xen.org/people/dvrabel/ticket-locks/busy_byte.png The durations are all more reasonable now. David _______________________________________________ Xen-devel mailing list Xen-devel@lists.xen.org http://lists.xen.org/xen-devel
