Re: Bringing rseq back into glibc

----- On Nov 18, 2021, at 5:17 AM, Florian Weimer fweimer@redhat.com wrote:

I would like to bring back rseq for glibc 2.35.

That's excellent news ! Thanks for looking into this.

I propose the following steps:

1. Enable rseq registration in glibc, for internal use only.  This time,
  put the rseq area into struct pthread, not into a initial-exec TLS
  symbol.  (This helps to avoid with initial-exec TLS bloat with dlopen
  and simplifies initialization somewhat.)

That works for me.

2. Add a tunable to disable rseq registration in glibc.  This way, if
  there is already an rseq user, it can be made to work again by
  setting the tunable.

Out of curiosity, how is the glibc tunable exposed ? Can it be called
from the application, or is it an environment variable which needs to
be set before running the application ?

3. Implement sched_getcpu on top of rseq.

4. Add public symbols __rseq_abi_offset, __rseq_abi_size (currently 32
  or 0), __rseq_abi_flags (currently 0).  __rseq_abi_offset is the
  offset to add to the thread pointer (see __builtin_thread_pointer) to
  get to the rseq area.  They will be public ABI symbols.  These
  variables are initialized before user code runs, and changing the
  results in undefined behavior.

Works for me. So if the Linux kernel eventually implements something along
the lines of an extensible kTLS, we can could use that underneath.

Small bike-shedding comment: I wonder if we want those public glibc
symbols to be called "__rseq_abi_{offset,size,flags}", or if a name like
"__ktls_{offset,size,flags}" might be more appropriate and future-proof
from a glibc ABI standpoint ?

Under this model, the rseq area offset is clearly constant across all
threads.  (This was previously implied by using initial-exec TLS
memory.)  rseq registration failure is indicated by __rseq_abi_size ==
0.  If the size is non-zero, rseq will be registered on all threads
created by glibc, and all the time as far as user code is concernes.
(This assumes that if rseq registration succeeds on the main thread, it
will succeed on all other threads.  We will terminate the process if
not.)  For example, if a JIT compiler sees __rseq_abi_size >= 32, in
generated code, it can inline a version of sched_getcpu that
materializes the thread pointer and loads the cpu_id field from the rseq
area, without further checks.  Under the old TLS-based model, it was
less clear that this was a valid optimization.

Sounds good.

Note that multiple applications wishing to use rseq on a shared memory
area may find themselves in a situation where some applications support
rseq, and others don't. So it would be up to the application to negotiate
whether they can use rseq in a shared memory area or not.

Furthermore, I believe this approach will be more compatible with
potential future kernel changes in this area.  If the kernel tells us
some day through the auxiliary vector that we should register a 128-byte
rseq area with 64-byte alignment, we can make that happen and change
__rseq_abi_offset and __rseq_abi_size accordingly.

Yes, hence my question about __ktls_* naming for the glibc symbols.

Steps 1 to 3 are backportable to previous glibc version, especially to
2.34 with its integrated libpthread.

So if we have an application or library already using rseq directly through
the system call, upgrading glibc may cause it to fail. Arguably, no new
symbol are exposed, so I guess it's OK with the backport guide-lines.
My question here is: is it OK for a backported patch to break an
application which uses the Linux kernel system calls directly ?

Comments?  As I said, I'd like to bring these changes into glibc 2.35,
hopefully in early December.

I won't have time to do the implementation effort myself this time due to
other commitments, but I will try to free up some time for review. Feel
free to grab whatever code you feel is useful from my earlier rseq
integration patches (if any).

Thanks,

Mathieu

Thanks,
Florian

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com