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

[Mathieu Desnoyers]

----- On Oct 16, 2017, at 12:46 PM, Andi Kleen andi@xxxxxxxxxxxxxx wrote:

>> How you collect, summarize, and analyze that overwhelming evidence
>> is up to you, specific to each change, and difficult to do accurately
>> and with any large measure of statistical confidence. The reviewer
>> has to basically trust you to some degree :-)
> 
> I think Linus' just asked for some working "real world, not micro" code that
> demonstrates use.
> 
> A prototype type implementation of the glibc malloc cache using this may
> be good enough.
> 
> Even if the API still changes slightly later in review I would assume
> the basic concepts will stay the same, so it would be likely not
> too difficult to convert that prototype to the later final API.

In that respect, I have working prototypes of two non-trivial library
projects using rseq within the same process.

Those can be considered as being "early adopters" of rseq, before it
becomes available in glibc.

- liburcu per-cpu flavor prototype [1]
  Interesting bits at
  https://github.com/compudj/userspace-rcu-dev/blob/urcu-percpu/include/urcu/static/urcu-percpu.h
  https://github.com/compudj/userspace-rcu-dev/blob/urcu-percpu/src/urcu-percpu.c
  (it also has its own copy of rseq and cpu-opv helper libraries)

- lttng-ust tracer rseq prototype [2, 3]
  Interesting bits at
  https://github.com/compudj/lttng-ust-dev/blob/rseq-integration-oct-2017/libringbuffer/getcpu.h#L85
  https://github.com/compudj/lttng-ust-dev/blob/rseq-integration-oct-2017/libringbuffer/vatomic.h#L60
  (it also has its own copy of rseq and cpu-opv helper libraries)

They use a slightly updated version of the rseq patchset, which I
plan to push into a new "rseq" tree on kernel.org soon. It takes care
of the comments I received in the past few days.

They end up sharing the "__rseq_abi" TLS weak symbol (initial state of
cpu_id = -1). They lazy-detect whether rseq needs to be registered for
the current thread by checking if the cpu_id read from the rseq TLS
is < 0. If rseq registration fails, they set its value to -2 and won't
try to register again (will use their fallback). When they successfully
register, they setup a pthread_key so rseq is unregistered when the
thread exits.

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.

If those are considered acceptable limitations, then we can stick to
the "single rseq TLS per thread" rule, and we don't have to implement
a linked-list of rseq TLS per thread.

When glibc eventually adds support for rseq, I expect it to deal with
rseq TLS registration and unregistration at thread creation/exit.
Therefore, the checks for negative cpu_id performed by lttng-ust and
liburcu will figure out that rseq is already registered, and skip
registration altogether when it's already performed by glibc.

Thoughts ?

Thanks,

Mathieu

[1] https://github.com/compudj/userspace-rcu-dev/tree/urcu-percpu
[2] https://github.com/compudj/lttng-ust-dev/tree/rseq-integration-oct-2017
[3] https://github.com/compudj/lttng-tools-dev/tree/urcu-percpu


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