lkml.org 
[lkml]   [2009]   [Jan]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH -v9][RFC] mutex: implement adaptive spinning

    * Linus Torvalds <torvalds@linux-foundation.org> wrote:

    > On Tue, 13 Jan 2009, Ingo Molnar wrote:
    > >
    > > And v8 is rock solid in all my testing - and i'll give v10 a similar
    > > workout as well.
    >
    > The differences between v8 and v10 are very fundamental, since v8 does
    > the spinning inside the spinlock'ed loop (the spinning itself is not
    > inside the spinlock, but all the "real action" is). So v8 not showing
    > problems doesn't really say much about v10 - totally different
    > algorithms that share only some of the support code.
    >
    > So even if many lines look the same, those code-lines aren't the really
    > interesting ones. The only really interesting once is really the
    > atomic_cmpxchg (outside spinlock) vs atomic_xchg (inside spinlock), and
    > those are almost diametrically opposite.

    yeah. What i thought they would be useful for are testing and experiments
    like this:

    " what if you switch the spinning to more fair by typing this in your
    current tree:

    git revert c10b491
    "

    ... but that's a pretty narrow purpose.

    > > Would you prefer a single commit or is this kind of delta development
    > > history useful, with all the variants (at least the later, more
    > > promising ones) included?
    >
    > I'm not sure it makes sense to show the history here, especially as
    > there really were two different approaches, and while they share many
    > issues, they sure aren't equivalent nor are we really talking about any
    > evolution of the patch except in the sense of one being the kick-starter
    > for the alternative approach.
    >
    > What _can_ make sense is to commit some of the infrastructure helper
    > code separately, ie the lock ownership and preemption changes, since
    > those really are independent of the spinning code, and at least the
    > preemption thing is interesting and relevant even without it.

    ok, we'll improve the splitup.

    Ingo


    \
     
     \ /
      Last update: 2009-01-13 19:27    [W:0.028 / U:0.732 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site