Using GDB to debug a user space process in UML

[Johannes Berg]

On Fri, 2021-10-08 at 12:01 -0500, Glenn Washburn wrote:
> 
> Do you have an idea of what is not working? The only thing I see
> referenced in the cover letter[1] seems to be more of a security issue
> and a functionality issue. So don't run the patch on untrusted code.
> Anything else I should be aware of?

Last I checked, it didn't even compile entirely well, and when I looked
a bit, it seemed the reason it worked for him was that he had some half-
compiled version before/after the changes, likely due to some missing
dependencies in the build system (possibly [2]).

I'm sure it's fixable, and I have an interest in this as well (due to
some payloads I run in UML for my testing), but so far neither of us had
a chance to actually sit down and do it. It's probably not even that
hard.

> According to the cover letter, this could be an answer to the "is
> exec still slow" question I posted earlier. Very cool!

To some extent. It's still slowish I believe - still has to entirely
tear down all the mappings and recreate etc., but not quite as bad any
more because it happens directly in the userspace process, not with a
ton of overhead through executing individual syscalls there. But yeah,
it's much better on fork/exec heavy loads for sure.

johannes

> [1] http://lists.infradead.org/pipermail/linux-um/2021-March/001167.html
[2] https://patchwork.ozlabs.org/project/linux-um/patch/20210928115146.277bec2328bb.If70926505d16d6c6a229002eb6d3f6838b546175@changeid/