On Fri, 19 Oct 2001, Takashi Iwai wrote: [preemptive kernel patch] > However even PE patched kernel can cause unacceptably high latency. > PE patch checks context switching with help of spinlock/unlock hacks. > Thus it cannot overcome the long spinlock itself, which can be seen in > VM and fs. Consequently PE patch requires also the LL'ish > preemption/scheduling checks inside the long locked loops to reduce > the high peaks.
And given that a) PE-patch "just" uses the existing SMP-lock points, and b) I just reported +3000msec breaks running a SMP-kernel. But of course, it's clear that if a piece of kernel code halts an SMP-machine for that long, it's not just a low-latency issue anymore - the code-path in question needs tuning. -- http://www.eca.cx Audio software for Linux!
