Temporarily pinning a thread to a capability

Ryan Yates fryguybob at gmail.com
Wed Oct 28 20:28:10 UTC 2015


A thread with TSO_LOCKED can be migrated if the number of capabilities
decreases.

Ryan

On Tue, Oct 27, 2015 at 11:35 PM, Edward Kmett <ekmett at gmail.com> wrote:

> Would anything go wrong with a thread id if I pinned it to a capability
> after the fact?
>
> I could in theory do so just by setting
>
> tso->flags |= TSO_LOCKED
>
> and then disabling this later by restoring the TSO flags.
>
> I can't think of anything but I figured folks here might be able to think
> of invariants I don't know about.
>
> Usage scenario:
>
> I have a number of things where I can't afford a map from a ThreadId# or
> even its internal id to a per-thread value for bounded wait-free
> structures.
>
> On the other hand, I can afford one entry per capability and to make a
> handful of primitives that can't be preempted, letting me use normal
> writes, not even a CAS, to update the capability-local variable in a
> primitive (indexing into an array of size based on the number of
> capabilities). This lets me bound the amount of "helpers" to consider by
> the capability count rather than the potentially much larger and much more
> variable number of live threads.
>
> However, I may need to access this stuff in "pure" code that wasn't
> written with my needs in mind, so I need to at least temporarily pin the
> current thread to a fixed capability for the duration when that happens.
>
> This isn't perfect, it won't react to a growing number of capabilities
> nicely in the future, but it does handle a lot of things I can't do now at
> all without downgrading to lock-free and starving a lot of computations, so
> I'm hoping the answer is "it all works". =)
>
> -Edward
>
> _______________________________________________
> ghc-devs mailing list
> ghc-devs at haskell.org
> http://mail.haskell.org/cgi-bin/mailman/listinfo/ghc-devs
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20151028/e3811219/attachment-0001.html>


More information about the ghc-devs mailing list