Date: Fri, 25 Apr 1997 11:36:32 -0600 From: Steve Passe <smp@csn.net> To: Peter Wemm <peter@spinner.dialix.com> Cc: NISHIO Shuichi <nishio@caleche.kecl.ntt.co.jp>, freebsd-smp@freebsd.org Subject: Re: Tyan ATX1668 problem -- success Message-ID: <199704251736.LAA08547@Ilsa.StevesCafe.com> In-Reply-To: Your message of "Fri, 25 Apr 1997 23:47:06 %2B0800." <199704251547.XAA25776@spinner.DIALix.COM>
next in thread | previous in thread | raw e-mail | index | archive | help
Hi, > Steve Passe wrote: > > As to the improved performance, this is because we are taking a big hit on > > all those TLB invalidations (has to be done), and the old method was competin > g > > for the kernel lock each time. The new method goes right for it without the > > lock. > > No.. The old method wasn't fetching the kernel lock either.. I suspect > that the speedup was because the old method was using the generic > IPI_INTR() handler which was set up for calling C code, while the new one > is a cut down version that doesn't have as much overhead. (Incidently, > the new one doesn't show up on the vmstat interrupt counts) your correct, no lock contention. after looking it over for awhile I'm convinced we want to keep the action in the mainline code, ie NOT call the C routine. Still not sure whether we want to put statistics back in. I'm afraid it will take more work than is justified at this point. The world believes that there are ONLY 32 INT sources, starting @ ICU_LEN, and densely packed thru the folowing 32 positions. I would vote for delaying the restoration of statistics till the re-write of the INT stuff. --- Now that it works I can remove the ipi_hits counter, am I correct in believing that the load of %ds with the kernel selector will still be necessary to address the apic[] for the EOI? -- Steve Passe | powered by smp@csn.net | Symmetric MultiProcessor FreeBSD
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199704251736.LAA08547>