next steps with rseq

[mihalicyn]

Thread to cover the next steps with rseq support improvements:

• tests for all architectures (right now we have ZDTM tests only for x86_64)
• move all rseq related code into a separate file (?)
• improvement support of built-in rseq for non-Glibc libraries
• detect built-in rseq support dynamically as Mathieu proposed
• pre-dump tests (?)
• leave-running tests (?)
• crfail test
• fault injection test (?)
• threaded test

https://criu.org/Rseq

[fweimer-rh]

Regarding rseq registration detection, could you fork and obtain the details from the new process using ptrace? That would be library-agnostic. (I don't think self-ptrace works, but it might.) My concern here is not just other libraries, but quite selfishly glibc with a backport of internal rseq functionality (for sched_getcpu) without the public symbols.

[mihalicyn]

Hi Florian,

Regarding rseq registration detection, could you fork and obtain the details from the new process using ptrace?

yes, we can. But I've concerns about performance, because we will need to add extra O(process_tree_size) complexity to the restore.

My concern here is not just other libraries, but quite selfishly glibc with a backport of internal rseq functionality (for sched_getcpu) without the public symbols.

Ugh. I had some idea. We can try to detect that *libc is used rseq by default by forking stub process and check if rseq is registered using ptrace. If yes then go to the expensive option of doing ptrace syscall for each process that we clone in CRIU.

[mihalicyn]

Our friends from https://github.com/containers/podman met the problem containers/podman#13958

It looks related to this ABI change:
https://sourceware.org/pipermail/libc-alpha/2022-February/136024.html

Mathieu already suggested to use dlsym to get rseq __rseq_offset dynamically:
https://github.com/compudj/librseq/blob/170f840b498e1aff068b90188727a656111bfc2f/src/rseq.c#L109

But at the moment I don't know how to get "symbol size" using dlsym. I will read dl* implementation for getting more information.

Upd: okay it looks like dladdr1(..., RTLD_DL_SYMENT) is what we need.

[mihalicyn]

#1937

[mihalicyn]

Threaded test is implemented in a scope of
#2147
by @mclapinski