Showstoppers for RPI3

[Mark Millard]

On 2020-Feb-25, at 17:57, Kyle Evans <kevans at freebsd.org> wrote:

> On Tue, Feb 25, 2020 at 5:05 PM Mark Millard via freebsd-arm
> <freebsd-arm at freebsd.org> wrote:
>> 
>> On 2020-Feb-25, at 09:54, bob prohaska <fbsd at www.zefox.net> wrote:
>> 
>>> There seem to be a handful of problems for -current on the RPI3 at
>>> the moment. Those I've noticed include not getting multicore operation,
>>> cpu_reset failing and OOMA kills with vm.pfault_oom_attempts="-1" set
>>> in /boot/loader.conf.
>> 
>> If you are seeing vm.pfault_oom_attempts="-1" fail to work on or
>> after head -r357253 (from 2020-Jan-29) that is likely new news.
>> However, for aarch64 RPi*'s, this is after -r356767 where the
>> kernel needs to be told to avoid touching all the armstub8-gic.bin
>> or armstub8.bin RAM. (Previously worked only by accident, i.e.,
>> despite not being told to avoid all that RAM.)
>> 
>> It is messy to make judgments about aarch64 RPi*'s on or after
>> -r356767 (until/unless all the armstub8*.bin is being reported to
>> the kernel as RAM to avoid): too much ends up messed up by the
>> kernel replacing part of armstub8*.bin's RAM content.
>> 
>> The problem -r357253 fixed was for all platforms, not just aarch64,
>> much less being aarch64 RPi* specific. As far as I've seen, the
>> platforms that do not have problems at -r356767 are no longer
>> getting reports of vm.pfault_oom_attempts="-1" problems.
>> 
>> The armstub8-gic.bin/armstub8.bin RAM not being reported to the
>> kernel as RAM to avoid touching is a known problem for aarch64
>> RPi*'s. But, as far as I know, no one has indicated that they are
>> working on getting such RPi*'s to well report armstub8*.bin RAM
>> to the kernel --or that they are planning to do so.
>> 
> 
> I've forwarded your patch on to some relevant U-Boot folk to try and
> get some feedback from them, since they likely understand the process
> better -- I would suspect reserving the second page in U-Boot as well
> is the proper solution.

Cool. Thanks.

I'll note that compared to my replacement of one magic number
by another one that also would not track future armstub8*.bin
page-count changes, Robert Crowston reported the following
about the size of the area to reserve and what the armstub8*.bin
are doing to try  to provide it:

QUOTE
The area to be reserved is already passed in register x1 by armstub to u-boot here:
https://github.com/gonzoua/rpi3-psci-monitor/blob/master/pscimon.S#L178

u-boot can read this register before it does anything else in save_boot_params in lowlevel_init.S. (e.g., https://github.com/RobCrowston/u-boot/blame/7d1d1ce63c1fe50b451ef0c730e1cd870b5bd440/board/raspberrypi/rpi/lowlevel_init.S#L38).
END QUOTE

Being able to build u-boot to make use of such sounds like
it would give armstub8*.bin some self-control without
further u-boot changes also being needed for any armstub8*.bin
page-count change.

(If I remember right, when I looked armstub8-gic.bin was listing
4 pages as the size in x1, despite not needing that much
currently. But that is definitely in the direction of being
sufficient. As I remember, the x1 figure was in Bytes, not pages.)

===
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)