Re: [RFC PATCH v9 for 4.15 01/14] Restartable sequences system call

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Oct 17, 2017, at 12:19 PM, Ben Maurer bmaurer@fb.com wrote:

> Hey,
> 
>> So far the restrictions I see for libraries using this symbol are:
>> - They should never be unloaded,
>> - They should never be loaded with dlopen RTLD_LOCAL flag.
> 
> We talked a bit about this off-list but I wanted to state publicly that I think
> this model works well for our use case. Specifically,
> 
> (1) It reduces complexity by focusing on the common case -- long term we expect
> glibc to manage the process of using this feature and registering/deregistering
> threads for rseq. Unloading isn't a challenge in these situations, so why add
> the complexity for it?
> 
> (2) This still allows for early adopters to use rseq before there is glibc
> support. I believe the vast majority of real world applications meet these two
> criteria you've listed. If not, they can create a thin shared library that has
> the sole purpose of providing the weak symbol and that never gets unloaded
> 
> (3) This allows for applications to provide the __rseq_abi so that they can
> ensure it uses the initial_exec tls model and optimize in-application assembly
> code for it. This is a good optimization for server applications that tend to
> statically link.

Agreed with all the above,

> 
> If others agree with this, would it make sense to remove the concept of
> reference counting in the system call that defines and redefines the per-thread
> area? Seems like it would remove complexity.

I have a use-case for keeping the reference counting in place though. It's
use of rseq in signal handlers.

If we have two early-adopter libraries trying to lazy-register rseq, and
one of those libraries can be called within a signal handler (e.g. lttng-ust),
we run into a situation where signal handler could nest on top of the
first library lazy-register (test, branch, register), and race against it.
So having reference counting in place allows the kernel to deal with
those multi-lib use-cases atomically wrt signal handlers from a thread
perspective.

And I don't want to require every early-adopter library to disable signals
just in case some _other_ library would be invoked in a signal handler.

Thoughts ?

Thanks,

Mathieu

> 
> -b

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