Le 27/08/2020 à 10:28, Giuseppe Sacco a écrit :
Il giorno gio, 27/08/2020 alle 09.46 +0200, Giuseppe Sacco ha scritto:
Il giorno gio, 27/08/2020 alle 00.28 +0200, Giuseppe Sacco ha scritto:
Hello Christophe,

Il giorno mer, 26/08/2020 alle 15.53 +0200, Christophe Leroy ha
If there is no warning, then the issue is something else, bad luck.

Could you increase the loglevel and try again both with and without
VMAP_STACK ? Maybe we'll get more information on where it stops.

The problem is related to the CPU frequency changes. This is where the
system stop: cpufreq get the CPU frequency limits and then start the
default governor (performance) and then calls function
cpufreq_gov_performance_limits() that never returns.

Rebuilding after enabling pr_debug for cpufreq.c, I've got some more
lines of output:

cpufreq: setting new policy for CPU 0: 667000 - 867000 kHz
cpufreq: new min and max freqs are 667000 - 867000 kHz
cpufreq: governor switch
cpufreq: cpufreq_init_governor: for CPU 0
cpufreq: cpufreq_start_governor: for CPU 0
cpufreq: target for CPU 0: 867000 kHz, relation 1, requested 867000 kHz
cpufreq: __target_index: cpu: 0, oldfreq: 667000, new freq: 867000
cpufreq: notification 0 of frequency transition to 867000 kHz
cpufreq: saving 133328 as reference value for loops_per_jiffy; freq is 667000 

no more lines are printed. I think this output only refers to the
notification sent prior to the frequency change.

I was thinking that selecting a governor that would run at 667mHz would
probably skip the problem. I added cpufreq.default_governor=powersave
to the command line parameters but it did not work: the selected
governor was still performance and the system crashed.

I kept following the thread and found that CPU frequency is changed in
function pmu_set_cpu_speed() in file drivers/cpufreq/pmac32-cpufreq.c.
As first thing, the function calls the macro preempt_disable() and this
is where it stops.

Sorry, I made a mistake. The real problem is down, on the same
function, when it calls low_sleep_handler(). This is where the problem
probably is.

Great, you spotted the problem.

I see what it is, it is in low_sleep_handler() in arch/powerpc/platforms/powermac/sleep.S

All critical registers are saved on the stack. At restore, they are restore BEFORE re-enabling MMU (because they are needed for that). But when we have VMAP_STACK, the stack can hardly be accessed without the MMU enabled. tophys() doesn't work for virtual stack addresses.

Therefore, the low_sleep_handler() has to be reworked for using an area in the linear mem instead of the stack.


Reply via email to