Date: Fri, 11 Apr 2003 12:10:05 -0700 From: Peter Wemm <peter@wemm.org> To: "David Xu" <davidxu@freebsd.org> Cc: freebsd-threads@freebsd.org Subject: Re: patch for %gs saving Message-ID: <20030411191005.68C322A7EA@canning.wemm.org> In-Reply-To: <012401c2ffef$e50657e0$f001a8c0@davidw2k>
next in thread | previous in thread | raw e-mail | index | archive | help
"David Xu" wrote: > > ----- Original Message -----=20 > From: "Peter Wemm" <peter@wemm.org> > To: "David Xu" <davidxu@freebsd.org> > Cc: <freebsd-threads@freebsd.org> > Sent: Friday, April 11, 2003 1:37 PM > Subject: Re: patch for %gs saving=20 > > > > "David Xu" wrote: > > > Here is the patch for kernel to save %gs, > > > it works well on my machine. > > > http://people.freebsd.org/~davidxu/i386_gs.diff > > > Daniel, is this the reason in your libpthread > > > patch that doesn't use getcontext syscall ? > >=20 > > To put some background on the issue, there is a reason why we did not > > do this. %gs is not used by the kernel, so it does not normally need = > to > > be saved and restored on every trap into the kernel. Setting a = > segment > > register is Really Slow - measured in hundreds of clock cycles. > >=20 > > So, we normally only touch %gs when we context switch to a different = > process > > that may have a different %gs. Or when one of the context syscalls = > wants > > it changed. We cannot avoid touching %fs because we use it for kernel > > private data. But if it wasn't for that, we wouldn't be touching > > %fs for regular traps/syscalls/etc either. > >=20 > > Bruce Evans understands this better than I do, I would suggest not = > making > > this change without talking about it with him first. > >=20 > > Yes, I know loading a descriptor is slow, but in real world, such = > optimization > will be lost in real work noise. And setcontext syscall is broken by = > this > optimization, userland will fail to set his context in atomic operation, > set pcb->gs =3D userland_gs does not work, so the cost is obviously, I = > can optimize > my patch to only save and restore %gs at trap and interrupt time, when = > entering > kernel, I can always zero %gs because kernel does not use it, I think = > this can > reduce clock cycles at context switch, this might be better than current = > code which > loading %gs at every context switch. Currently we only change %gs when switching one entire process to another seperate process. Your patch changes it so that it changes %gs on every single fault/trap/syscall/etc. I regularly see process traces where the ratio of syscalls to context switches is over 1000 to 1. ie: your patch would increase the cost of the %gs activity over a thousand times. Cheers, -Peter -- Peter Wemm - peter@wemm.org; peter@FreeBSD.org; peter@yahoo-inc.com "All of this is for nothing if we don't go to the stars" - JMS/B5
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030411191005.68C322A7EA>