Re: [PATCH v2 0/4] futex: Drop ROBUST_LIST_LIMIT

[Florian Weimer <fweimer@redhat.com>]

* André Almeida:

> Hi Florian,
>
> Em 28/01/2025 04:50, Florian Weimer escreveu:
>> * André Almeida:
>> 
>>> As requested by Peter at [1], this patchset drops the
>>> ROBUST_LIST_LIMIT. This is achieve by simply rewriting the processed
>>> list element ->next to point to the head->list address, destroying the
>>> linked list to avoid any circular list.
>> Doesn't this turn a robust mutex overwrite or a TCB overwrite into a
>> write-anything-anywhere primitive?  Furthermore, I'm not entirely sure
>> if this is entirely backwards-compatible.
>> 
>
> The robust list is meant to be a private resource, per-process, and
> this patch only rewrites it after the process exits, so I believe that
> any changes done in this memory should be safe given that the process
> will soon disappear anyway, right?

At least in the glibc implementation, we let the kernel handle robust
mutex notification on thread exit, and that's observable.

Beyond that, process-shared robust mutexes exist, too, and those updates
will be observable, too.

> Do you think you can point out a scenario that wouldn't be
> backwards-compatible? I would like to try to test it.

I think it should be okay for the glibc implementation.  The robust list
is libc-owned (at least in glibc implementation), so it should not
matter, but the are other libs out there.

>> Could you use the tortoise/hare approach instead?

> I believe that you want the approach to be "slow and steady" but I'm
> not sure what you have in mind, if you could you please elaborate :)

I meant cycle detection using Floyd's algorithm.

Thanks,
Florian