Re: [Qemu-devel] [PATCH v2] rtc: placing RTC memory region outside BQL

[Gonglei (Arei)]

> -----Original Message-----
> From: Peter Maydell [mailto:address@hidden
> Sent: Tuesday, February 06, 2018 10:36 PM
> To: Gonglei (Arei)
> Cc: QEMU Developers; Paolo Bonzini; Huangweidong (C)
> Subject: Re: [PATCH v2] rtc: placing RTC memory region outside BQL
> 
> On 6 February 2018 at 14:07, Gonglei <address@hidden> wrote:
> > As windows guest use rtc as the clock source device,
> > and access rtc frequently. Let's move the rtc memory
> > region outside BQL to decrease overhead for windows guests.
> > Meanwhile, adding a new lock to avoid different vCPUs
> > access the RTC together.
> >
> > $ cat strace_c.sh
> > strace -tt -p $1 -c -o result_$1.log &
> > sleep $2
> > pid=$(pidof strace)
> > kill $pid
> > cat result_$1.log
> >
> > Before appling this change:
> > $ ./strace_c.sh 10528 30
> > % time     seconds  usecs/call     calls    errors syscall
> > ------ ----------- ----------- --------- --------- ----------------
> >  93.87    0.119070          30      4000           ppoll
> >   3.27    0.004148           2      2038           ioctl
> >   2.66    0.003370           2      2014           futex
> >   0.09    0.000113           1       106           read
> >   0.09    0.000109           1       104           io_getevents
> >   0.02    0.000029           1        30           poll
> >   0.00    0.000000           0         1           write
> > ------ ----------- ----------- --------- --------- ----------------
> > 100.00    0.126839                  8293           total
> >
> > After appling the change:
> > $ ./strace_c.sh 23829 30
> > % time     seconds  usecs/call     calls    errors syscall
> > ------ ----------- ----------- --------- --------- ----------------
> >  92.86    0.067441          16      4094           ppoll
> >   4.85    0.003522           2      2136           ioctl
> >   1.17    0.000850           4       189           futex
> >   0.54    0.000395           2       202           read
> >   0.52    0.000379           2       202           io_getevents
> >   0.05    0.000037           1        30           poll
> > ------ ----------- ----------- --------- --------- ----------------
> > 100.00    0.072624                  6853           total
> >
> > The futex call number decreases ~90.6% on an idle windows 7 guest.
> 
> These are the same figures as from v1 -- it would be interesting
> to check whether the additional locking that v2 adds has affected
> the results.
> 
Oh, yes. the futex number of v2 don't decline compared too much to v1 because it
takes the BQL before raising the outbound IRQ line now.

Before applying v2:
# ./strace_c.sh 8776 30
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 78.01    0.164188          26      6436           ppoll
  8.39    0.017650           5      3700        39 futex
  7.68    0.016157           6      2758           ioctl
  5.48    0.011530           3      4586      1113 read
  0.30    0.000640          20        32           io_submit
  0.15    0.000317           4        89           write
------ ----------- ----------- --------- --------- ----------------
100.00    0.210482                 17601      1152 total

After applying v2:
# ./strace_c.sh 15968 30
% time     seconds  usecs/call     calls    errors syscall
------ ----------- ----------- --------- --------- ----------------
 78.28    0.171117          27      6272           ppoll
  8.50    0.018571           5      3663        21 futex
  7.76    0.016973           6      2732           ioctl
  4.85    0.010597           3      4115       853 read
  0.31    0.000672          11        63           io_submit
  0.30    0.000659           4       180           write
------ ----------- ----------- --------- --------- ----------------
100.00    0.218589                 17025       874 total

> Does the patch improve performance in a more interesting use
> case than "the guest is just idle" ?
> 
I think so, after all, the scope of the locking is reduced . 
Besides this, can we optimize the rtc timer to avoid to hold BQL 
by separate threads?

> > +static void rtc_rasie_irq(RTCState *s)
> 
> Typo: should be "raise".
> 
Good catch. :)

Thanks,
-Gonglei