[lttng-dev] URCU background threads vs signalfd

[Mathieu Desnoyers]

On 2022-09-26 15:58, Eric Wong wrote:
> Mathieu Desnoyers <mathieu.desnoyers at efficios.com> wrote:
>> Do you mind if I fold our 2 patches together with a Co-developed-by tag and
>> use your Signed-off-by ?
> 
> Looks good to me, thanks.
> 
>> Then the question that will arise is whether this change is sufficiently
>> self-contained that we can push it to stable branches, or if it's master
>> branch material only.
> 
> It seems self-contained enough and fixes a real problem
> encountered in real-world usage, so I think it's suitable for stable.

That's tricky. I try very hard not to introduce any user-observable 
regressions in stable branches. I can very well imagine an application 
which, for its own peculiar reasons, would expect signals to be 
unblocked in call_rcu worker thread callbacks.

On the other hand, those worker threads have always had the behavior of 
having signals unblocked so far, therefore an application that wants to 
use signalfd() today with those liburcu worker threads already needs to 
take some care about how those threads get created.

Therefore, it's not a small self-contained change: it has effects on the 
blocked signal masks that are observable from the application call_rcu 
callbacks. So far, applications which expect to use signalfd() in a 
non-broken way need to ensure they either have the work-around in place 
around thread creation, or will need to add a dependency on an upcoming 
liburcu version (e.g. 0.14).

For all those reasons, I think it would be a good time to do a liburcu 
0.14 release soon, and I don't think backporting this patch makes sense 
for stable branches of liburcu.

Thanks,

Mathieu

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