lkml.org 
[lkml]   [2008]   [Jan]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectPIT clocksource makes invalid assumptions
Looking at pit_read() in arch/x86/kernel/i8253.c, it seems that the PIT 
clocksource code assumes that the PIT CH0 is in periodic mode. With
clockevents, this assumption is no longer valid. There are at least two
places that make this assumption:

1) The calculation at the end of pit_read() assumes that the PIT is in
periodic mode. This isn't true unless the PIT is the current clockevent
and nohz is inactive. (Though #2 can end up forcing the PIT to be
reprogrammed).

2) The PIT clockevent is shutdown by using PIT mode 0 (interrupt on
terminal count) -- doesn't the PIT counter continue to count (even
though it won't be raising an interrupt)? If so, the test in pit_read()
under the VIA686a comment can succeed after the PIT clockevent has been
shutdown, and the PIT hardware may be reprogrammed to start firing
interrupts again. This doesn't seem intentional, and can defeat nohz
since now the PIT is firing periodically.

Seems these problems can happen when the PIT is used as the clocksource
or even just the clocksource watchdog. It looks like there is some code
in clocksource.c that checks for CLOCK_SOURCE_IS_CONTINUOUS, which is
not set for the PIT clocksource, but it doesn't seem to be strong enough
to prevent these problematic scenarios (and it's not clear if that is
the intent of IS_CONTINUOUS anyway).

To verify this really can happen, when I boot a kernel, I can see this
sequence:

init_pit_timer (with mode==CLOCK_EVT_MODE_PERIODIC)
init_pit_timer (with mode==CLOCK_EVT_MODE_UNUSED)
init_pit_timer (with mode==CLOCK_EVT_MODE_SHUTDOWN)
pit_read() and count > LATCH (I believe the PIT is the watchdog at
this point), which causes the PIT to raise periodic interrupts.

(Shortly after, the acpi pm clocksource is registered and replaces the
PIT as the watchdog. Later, the PIT clockevent is used as the broadcast
clockevent and reprogrammed into one-shot mode, stopping the PIT
interrupts.)

Also, the user could force the PIT clocksource to be current_clocksource
even though the PIT is in one-shot mode (and therefore the calculation
in pit_read is bogus).

Of course, all this can only happen for 32-bit UP. I'm not sure what
the preferred fix for this is...

Dan



\
 
 \ /
  Last update: 2008-01-04 00:55    [W:0.040 / U:0.620 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site