On Fri, 14 Oct 2005, Esben Nielsen wrote:I set up rtc_wakeup and got a jitter up 1.6ms!
It came when I cd'en into a nfs-mount and typed ls.
ls-11239 0Dn.. 4us : profile_hit (__schedule)
ls-11239 0Dn.1 4us : sched_clock (__schedule)
ls-11239 0Dn.1 5us : check_tsc_unstable (sched_clock)
ls-11239 0Dn.1 5us : tsc_read_c3_time (sched_clock)
IRQ 8-775 0D..2 6us : __switch_to (__schedule)
IRQ 8-775 0D..2 7us!: __schedule <ls-11239> (75 0)
IRQ 8-775 0...1 1594us : trace_stop_sched_switched (__schedule)
IRQ 8-775 0D..2 1594us : trace_stop_sched_switched <IRQ 8-775> (0 0)
IRQ 8-775 0D..2 1595us : trace_stop_sched_switched (__schedule)
ouch! This very much looks like a hardware induced latency, because the codepath from those two __schedule points is extremely short and there is no loop there. Have you tested this particular box before too? If not, can you reproduce this latency with older versions of -rt too on the same box, or is this completely new?
I have been testing on this box before but never with such long latencies
before.
I could not reproduce it on linux-2.6.14-rc3-rt10. The maximum meassured
rtc_wakeup is 146us with the same config options. Running linux-2.6.14-rc4-rt1 with ethernet disabled gave a maximum latency
of 1468us while building the kernel: