Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Mon, 28 Feb 2022 16:07:02 +0000
From: Lee Shallis <>
Subject: Re: Suggestion for thread safety

On Mon, 28 Feb 2022 at 15:51, Joakim Sindholt <> wrote:
> On Mon, 28 Feb 2022 14:43:36 +0000, Lee Shallis <> wrote:
> > Seems the wait just wasn't long enough, at about 4 yields onwards the
> > results become consistent success, I've attached the file I did the
> > experiments in, I even tried it under -O3 and no exits were
> > encountered, so yes my method works, just needs a bit more wait time
> > for extreme cases
> Between the lines
> > if ( !(shared->tid) )
> and
> > shared->tid = tid;
> the kernel might suspend the running thread and allow the other to run,
> or you might simply get unlucky and have the two threads do the checks
> close enough to simultaneously that the memory hasn't been synchronized
> yet. Either way you end up with both threads seeing that shared->tid is
> zero and both of them writing their tids to it, and thus both enter the
That's the point of the loop, to check it's the same as what they
wrote, if it's not then it's either locked to another thread or empty,
the point in doing the yield after the write is to allow that failure
to occur, basically I'm using the race condition itself as the point
of success, rather than expect the CPU to perform an atomic lock that
could be just as broken as timing based locks, I already have my ideas
on how to fix the need for many yields to need only 2, I'm about to
try it now
> critical section at the same time. And so the lock fails at the very
> first hurdle: mutual exclusion. No amount of sleeping will make the bug
> go away, only slightly more difficult to trigger.

No it doesn't, think through the loop properly and you'll see that the
concept is the best one to go with, implementation just needs a little

> The point of the clock_nanosleep call was to force a reschedule while
> holding the lock. This also increases the runtime inside the lock which
> in this case increases the likelihood that the thread trying to take the
> lock will be waiting for it and end up racing with the thread that
> currently has it when it unlocks and tries to relock it.

How so? It still takes time for the jump condition to be evaluated and
the call to LockSiData to start, the other thread will already be in
the call loop ready to lock it, I designed this function specifically
around the idea that multiple threads could see an empty tid at the
same time, that's the reason for the yield call, so that all those
writes get in before the execution resumes.

> Now that you've inserted lots of sched_yield()s your lock is not only
> still broken (in more ways than the one we've been trying to get you to
> understand) but also extremely slow.
> As a hint for your future education: the first (and far from only) thing
> you'll need is compare-and-swap, aka. CAS.
> You can read up on this class of bugs if you'd like. It's called "Time
> Of Check to Time Of Use" or "TOCTOU" for short.
> I didn't even need to poke at the code this time as the code you sent
> breaks just the same on my machine.
> I hope you'll learn from this.

I hope you'll learn to think through the code before you speak out of
your ass, the concept is perfect, it's only that my implementation of
that concept isn't

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.