Re: [PATCH] gdb: relax requirement for the map_failed stap probe to be present

[Andrew Burgess <aburgess@redhat.com>]

Luis Machado <luis.machado@arm.com> writes:

> On 12/5/22 10:09, Andrew Burgess wrote:
>> Luis Machado <luis.machado@arm.com> writes:
>> 
>>> On 11/29/22 08:27, Luis Machado wrote:
>>>> Hi Andrew,
>>>>
>>>> This seems to have broken armhf on Ubuntu 22.04.
>>>>
>>>> https://builder.sourceware.org/buildbot/#/builders/169/builds/1163
>>>>
>>>
>>> I haven't investigated this. I only spotted it in the sourceware buildbot page. But I'd guess it is something to do
>>> with thumb mode detection early in the startup.
>>>
>>> Ubuntu has moved to not stripping ld.so for armhf because the probe mechanism is not capable of conveying the thumb mode
>>> information, so gdb has to rely on symbols instead. If the changes have touched this fragile area, it may have broken the
>>> delicate balance.
>> 
>> I somehow missed this last week.  I'll take a look today.
>> 
>> Sorry for the breakage.
>
> No problem. For the record, I haven't managed to reproduce this on my Ubuntu 20.04/22.04 setup. I'll have to give it a try on the
> buildbot machine.
>
> Just a heads-up in case this doesn't reproduce for you either.

I don't have immediate access to a suitable ARM machine.  I'm trying to
set something up that might work, but given what you report, I'm not too
hopeful.

My guess for what is happening though is this:

  - Previously, when Ubuntu was using glibc 2.35 / 2.36, one of the
    shared library probes was missing.  As a result GDB would fall back
    to use the symbol/breakpoint approach.  This worked out just fine
    for the ARM case as is appears we shouldn't be using the probes
    anyway.

  - After my patch GDB no longer cares about the missing probe, and so
    started trying to use the probes based interface, as this is the
    preferred strategy.  Unfortunately for ARM this doesn't work :/

I guess the ideal solution would be that glibc didn't include probe
information if those probes can't then be used for some reason.  I guess
that will need someone who understands the problem to either raise a
glibc bug, or propose a glibc fix.

Until then we probably need to consider GDB based solutions.

Currently we don't have anyway to avoid particular probes on a
per-architecture basis, this will probably need a new gdbarch hook
adding.  I'll see if I can write something like this and test it on the
buildbot.


Thanks,
Andrew