CALL TO COMMITTERS! HELPING WITH SPL REMOVAL!
dillon at apollo.backplane.com
Thu Jun 2 10:35:07 PDT 2005
:Matt, I don't think this is a good to rush. I think it can seriously degrade
:performance, esp. since some of the spl() protections are hold over longer
:areas of time.
:I don't have a problem with replacing critical short usage areas with
:critical sections, but flatting it directly is not good.
Joerg, we've been juggling the SPL issue since day 1. The time has come
to remove them.
And, again, I completely disagree with your performance concerns. I
challenge you to find a case where performance would be degraded.
The most widely used long-held SPLs in the system are SPLBIO and SPLVM,
and guess what? They pretty much mask ALL MAJOR DEVICES anyway. So
what's the point of having a variable mask in the first place? No
point at all...
Even more importantly, many of these spl/critical-sections are going to
eventually be replaced by serialization in order to become MP safe and
remove Giant. So even if there were some performance degredation (and
there won't be), it's a moot point. We need to remove the SPLs to
be able to progress.
<dillon at xxxxxxxxxxxxx>
More information about the Kernel