Re: [RFC PATCH glibc 1/3] glibc: Perform rseq(2) registration at C startup and thread creation (v18)

[Florian Weimer <fweimer@redhat.com>]

* Mathieu Desnoyers:

> @deftypevar {struct rseq} __rseq_abi
> @standards{Linux, sys/rseq.h}
> @Theglibc{} implements a @code{__rseq_abi} TLS symbol to interact with the
> Restartable Sequences system call (Linux-specific).  The layout of this
> structure is defined by the @file{sys/rseq.h} header.  Registration of each
> thread's @code{__rseq_abi} is performed by @theglibc{} at libc library
> initialization and thread creation.

s/libc library/library/

> The main executable and shared libraries may either have an undefined
> @code{__rseq_abi} TLS symbol, or define their own, with the same
> declaration as the one present in @file{sys/rseq.h}.  The dynamic linker
> will ensure that only one of those available symbols will be used at
> runtime across the process.
>
> If the main executable or shared libraries observe an uninitialized
> @code{__rseq_abi.cpu_id} field (value @code{RSEQ_CPU_ID_UNINITIALIZED}), they
> may perform rseq registration to the kernel: this means either glibc was
> prevented from doing the registration, or an older glibc version, which does
> not include rseq support, is in use.  When the main executable or a library
> thus takes ownership of the registration, the memory used to hold the
> @code{__rseq_abi} TLS variable must stay allocated, and is not re-used, until
> the very end of the thread lifetime or until an explicit rseq unregistration
> for that thread is performed.  It is not recommended to dlclose() libraries
> owning the @code{__rseq_abi} TLS variable.

s/dlclose()/@code{dlclose}/ (no parentheses)

Rest looks okay.

>>> +  if (__rseq_abi.cpu_id == RSEQ_CPU_ID_REGISTRATION_FAILED)
>>> +    return;
>>> +  ret = INTERNAL_SYSCALL_CALL (rseq, &__rseq_abi, sizeof (struct rseq),
>>> +                              0, RSEQ_SIG);
>>> +  if (INTERNAL_SYSCALL_ERROR_P (ret) &&
>>> +      INTERNAL_SYSCALL_ERRNO (ret) != EBUSY)
>>> +    __rseq_abi.cpu_id = RSEQ_CPU_ID_REGISTRATION_FAILED;
>> 
>> Sorry, I forgot: Please add a comment that the EBUSY error is ignored
>> because registration may have already happened in a legacy library.
>
> Considering that we now disable signals across thread creation, and that
> glibc's initialization happens before other libraries' constructors
> (as far as I remember even before LD_PRELOADed library constructors),
> in which scenario can we expect to have EBUSY here ?

That's a good point.

> Not setting __rseq_abi.cpu_id to RSEQ_CPU_ID_REGISTRATION_FAILED in case
> of EBUSY is more a way to handle "unforeseen" scenarios where somehow the
> registration would already be done. But I cannot find an "expected"
> scenario which would lead to this now.
>
> So if EBUSY really is unexpected, how should we treat that ? I don't think
> setting REGISTRATION_FAILED would be appropriate, because then it would
> break assumption of the prior successful registration that have already
> been done by this thread.

You could call __libc_fatal with an error message.  ENOSYS is definitely
an expected error code here, and EPERM (and perhaps EACCES) can happen
with seccomp filters.

Thanks,
Florian