lkml.org 
[lkml]   [2003]   [Jun]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: 2.5.73-mm2 - odd audio problem, bad intel8x0/ac97 clocking.
From
Date
On Sat, 28 Jun 2003 17:10:36 PDT, Andrew Morton said:

> > The 'clocking to 51084' is *VERY* suspicious
>
> It could be that do_gettimeofday() has gone silly. Could you
> add this patch and see what it says?

Woo woo. Good catch, Andrew. It says:

intel8x0_measure_ac97_clock: measured 39909 usecs

Hmm.. wonder why it's 40K rather than the expected 50K...

Turns out we're running at 1.2G rather than 1.6G.. Or at least /proc/cpuinfo
says we are. However, things are not always what they seem....

Some testing indicates it's probably b0rkage in Dave Jone's recent work
splitting/reorging the speedstep driver....

cset-davej@codemonkey.org.uk|ChangeSet|20030626004701|05850.txt

(although the comments in the cset say Dominik Brodowski did the cleanups?)

after enabling debugging in arch/i386/kernel/cpu/cpufreq/cpufreq-ich.c
(which required the attached patch), we have this (annotated) dmesg output:

Machine check exception polling timer started.
OK, so we know approx where in boot we are now..
speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0x10300410 0x0
from speedstep_get_freqs()

speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0x10300410 0x0
cpufreq: read at pmbase 0x800 + 0x50 returned 0x0
cpufreq: writing 0x1 to pmbase 0x800 + 0x50
cpufreq: read at pmbase 0x800 + 0x50 returned 0x1
cpufreq: change to 0 MHz succeeded
all this from first call from get_freqs() to set_state(SPEEDSTEP_LOW,0);
Does that 'to 0mz' give you warm-n-fuzzies? Not me.... freqs.new isn't initialized yet...

speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0xc300410 0x0
speedstep_get_freqs() checking that we actually went low

speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0xc300410 0x0
cpufreq: read at pmbase 0x800 + 0x50 returned 0x1
cpufreq: writing 0x0 to pmbase 0x800 + 0x50
cpufreq: read at pmbase 0x800 + 0x50 returned 0x0
cpufreq: change to 1200 MHz succeeded
set_state(SPEEDSTEP_HIGH,0); and again we got the wrong value in
the message. We went *from* 1200 *to* 1600.

speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0x10300410 0x0
speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0x10300410 0x0
cpufreq: currently at high speed setting - 1600 MHz
speedstep_cpu_init();
cpufreq: speed=1600000 low=1200000 high=1600000
my debugging to make sure low/high were set right. All is good at this point,
and we return from speedstep_cpu_init().

speedstep-lib: P4 - MSR_EBC_FREQUENCY_ID: 0x10300410 0x0
cpufreq: read at pmbase 0x800 + 0x50 returned 0x0
cpufreq: writing 0x0 to pmbase 0x800 + 0x50
(remember - LOW is 1 and HIGH is 0 here)
cpufreq: read at pmbase 0x800 + 0x50 returned 0x0
cpufreq: change to 1200 MHz succeeded
Umm.. No. We went from 1.6G to 1.6G.

OK.. Who's the wise guy? ;)

Looks to me like cpufreq_add_device() calls cpufreq_set_policy(), which ends up
calling speedstep_set_state() with notify=1. Rememer the missing warm-n-fuzzies?

freqs.old = speedstep_get_processor_frequency(speedstep_processor);
freqs.new = speedstep_freqs[SPEEDSTEP_LOW].frequency;
freqs.cpu = 0; /* speedstep.c is UP only driver */

if (notify)
cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);

which ends up calling notify_transition() ends up calling adjust_jiffies(),
and things get pear-shaped ;)

If we're going from 1200 to 1600, this Does The Wrong Thing because
adjust_jiffies will be passed old == new == 1200 and fail to actually change
the loops_per_jiffy because the tests are > and <.

More subtly evil, when cpu_set_policy() runs, we go from 1600 to 1600, and
notify_transition is called with bogus values. So as a result,
adjust_jiffies() gets called with old=1600 new=1200, and we reset the jiffies
inappropriately. Meanwhile, similar evil happens in time_cpufreq_notifier(),
where cpu_khz will be incorrectly set to 1200. So cat /proc/cpuinfo
says we're at 1.2G when we're really still at 1.6G.

End result? Processor is running at 1.6G, /proc/cpuinfo *says* 1.2G, and
jiffies_per_loop is set for 1.2G. We call mdelay(50) at 1.6G and enough loops
for 1.2G, and end up at 80%. So the 50ms is really 40ms, and i810 gets the wrong
clocking.

And there's this remaining nit in speedstep_set_state():
if (state == (value & 0x1)) {
dprintk (KERN_INFO "cpufreq: change to %u MHz succeeded\n", (freqs.new / 1000));
} else {

Not initialized the first call, and even after will print the SPEEDSTEP_LOW number no matter what.

Oh, and the debugging patch:

--- arch/i386/kernel/cpu/cpufreq/speedstep-ich.c.dist 2003-06-29 10:34:17.949264614 -0400
+++ arch/i386/kernel/cpu/cpufreq/speedstep-ich.c 2003-06-29 10:36:34.833000667 -0400
@@ -295,7 +295,7 @@
return -EIO;

dprintk(KERN_INFO "cpufreq: currently at %s speed setting - %i MHz\n",
- (speed == speedstep_low_freq) ? "low" : "high",
+ (speed == speedstep_freqs[SPEEDSTEP_LOW].frequency) ? "low" : "high",
(speed / 1000));

/* cpuinfo and default policy values */

[unhandled content-type:application/pgp-signature]
\
 
 \ /
  Last update: 2005-03-22 13:43    [W:0.129 / U:0.148 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site