Re: earm(v5) not working on RPi

[Robert Swindells <rjs%fdy2.co.uk@localhost>]

Nick Hudson <nick.hudson%gmx.co.uk@localhost> wrote:
>On 15/07/2019 14:15, Robert Swindells wrote:
>>
>> Using ktrace(1), I can see that the last thing done before the crash for
>> 8.1 applications is a call to _lwp_ctl(2). This looks to be in
>> pthread__init().
>>
>> Doing ktrace on hpcarm 7.1, _lwp_ctl() never gets called.
>>
>> Do we have some optimization for reading pthread private data that is
>> different for armv7 and armv4 ?
>
>https://nxr.netbsd.org/xref/src/lib/libc/arch/arm/sys/__aeabi_read_tp.S

Thanks, I was probably looking for the wrong thing though.

A 7.1 tree works better than an 8.1 one because fewer applications were
linked with libpthread back then. Running a 7.1 binary that was linked
against libpthread dumps core in the same way as the 8.1 binaries.

The arm MD code raises SIGILL for several things that are not really
instruction related, will try adding some kernel debug to see which
one is being hit. Several seem related to the stack having been
corrupted.

Trying to chroot from -current earmv7hf to a -current earmv4 I get:

 # chroot /u8/build/cats
<jemalloc>: Unsupported system page size
Abort trap (core dumped)

Obviously this is down to the difference in page size but I'm not sure
that adding page size to libkver as a changeable option will work.

Maybe we need to document/decide what combinations of things are
expected to work together, including for arm64 host running arm32
binaries of different kinds.