Re: Practical libc-free threading on Linux

Konstantin Andreev <grapvar@gmail.com>
Message ID
DKIM signature
Download raw message
Hi. I have a couple of remarks on your demo.

1) in threadentry(), the __atomic_store_n( join_futex, … )
      is redundant for the goals of this demo.

      this is because futex( join_futex, FUTEX_WAKE )
      does not care about join_futex contents,
      it wakes up waiters unconditionally

2) when you `ret' from `newthread' into `threadentry',
      the `struct stack_head *stack'
      starts to point to the free space ABOVE the stack

      Probably, this does not harm, because
      `entry' member of `struct stack_head'
      is not needed anymore, but ...
      this unpleasantly contrasts to common practice

Re: Practical libc-free threading on Linux

Message ID
<9ccd39f3-ac09-9ca4-4635-87b28387db37@gmail.com> (view parent)
DKIM signature
Download raw message
> is redundant for the goals of this demo

That atomic store prevents a race condition and is essential. The futex 
syscall atomically loads the futex and compares it to the expected value, 
in this case zero. If the waiter arrives late, it misses the futex wakeup, 
but it won't wait because the value has been changed away from zero. What 
if it's non-zero in the first place? Then if it arrives early it won't 
wait at all, which is bad for a different reason.

So it must be an atomic update followed by a wake-up in order to cover 
both possibilities. If the futex was used frequently, perhaps an important 
and mutex on a hot path, I'd want to avoid the syscall costs when there is 
no contention. I would signal the presence of a waiter through the futex, 
then conditionally wake based on the value. Similarly, on the other side 
I'd check the futex in userspace before transitioning to the kernel. The 
futex(7) man page describes (quite poorly) such an algorithm. Here's my 
implementation of that concept:


Spawning and joining a thread is rare (or should be), so I'm unconcerned 
about the potentially unnecessary syscall on each side, opting instead for 

> starts to point to the free space ABOVE the stack

Yup, true, though I'd describe it as the structure partially overlapping 
the stack. The appropriate warning would be "do not touch the entry field 
after newthread." In practice it contains the address of the last function 
called directly by the entry function, but due to aliasing — the field is 
modified indirectly in ways the compiler doesn't understand — it's 
uncertain what might be read even by the thread running on that stack.
Reply to thread Export thread (mbox)