Quoting Heikki Lindholm <[EMAIL PROTECTED]>:
Hello,
While processing __adeos_handle_event enables interrupts for the
ectual event handler periods. Interrupts might happen then and only
logged for domains lower than the current domain in the
__adeos_handle_event loop. The logged irqs would normally get
replayed at the next real interrupt if the domain that caused the
event was lower than the ones the interrupts were logged in or if it
wasn't lower, at the next suspend_domain. On powersave-enabled ppc
machines this might cause a chain where the cpu goes napping and
wakes at the next timer tick which, having possibly been missed at
*handle_event, happens after maximum decrementer period (~minutes).
The following patch seems to help in the described case, but I've
also observed another case where interrupt seems to get dropped
altogether. Performance didn't seem to suffer much from the patch. If
anything, the latency/cruncher results got better. This is for the
non-threaded domains case.
Good spot, thanks. I'd also suggest a different implementation to solve this,
still in adeos_handle_event() though:
- you only need to sync the next domain when it does not handle the current
event; if it does, then it would suspend calling adeos_suspend_domain() when
the event is processed, hence switching to the next domain down the pipeline
that has IRQs to process anyway. Additionally, I don't see how notfirst goes
non-zero in this patch, and why we'd need to make some exception case for the
root domain (i.e. you may have domains below it down the pipeline).
- in the !threaded case, calling __adeos_switch_to() instead of changing the
domain descriptor by hand then syncing the stage would be better, since it
would perform all the additional housekeeping chores, like calling the switch
hook and resetting the current descriptor pointer.
-- Heikki Lindholm
