Date: Thu, 29 Nov 2007 11:08:41 -0800 (PST)
From: Linus Torvalds <torvalds@linux-foundation.org>
To: "H. Peter Anvin" <hpa@zytor.com>
cc: Chuck Ebbert <cebbert@redhat.com>, Roland McGrath <roland@redhat.com>,
       Andrew Morton <akpm@linux-foundation.org>, linux-kernel@vger.kernel.org,
       Thomas Gleixner <tglx@linutronix.de>, Ingo Molnar <mingo@elte.hu>
Subject: Re: [PATCH x86/mm 6/6] x86-64 ia32 ptrace get/putreg32 current
 task
In-Reply-To: <474F08E1.2090806@zytor.com>
Message-ID: <alpine.LFD.0.9999.0711291057320.8458@woody.linux-foundation.org>
References: <20071129003849.428E026F8E7@magilla.localdomain> <20071129004222.E49AD26F8E7@magilla.localdomain> <474EF824.3020806@redhat.com> <alpine.LFD.0.9999.0711290958110.8458@woody.linux-foundation.org> <474F01F6.2030509@zytor.com>
 <alpine.LFD.0.9999.0711291027070.8458@woody.linux-foundation.org> <474F08E1.2090806@zytor.com>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1706
Lines: 38


On Thu, 29 Nov 2007, H. Peter Anvin wrote:
> Linus Torvalds wrote:
> > 
> > > It is advantageous for user space to use the register the kernel typically
> > > won't, in order to speed up system call entry/exit.
> > 
> > but I'm not seeing the reason for that one. Care to comment more? (Yes,
> > there is often a latency from segment reload to use, but the reload latency
> > for system call exit *should* be entirely covered by the cost of doing the
> > system call return itself, no?)
> 
> I do seem to recall that some processor implementations can load a NULL
> segment faster than a non-NULL segment.  This was significant enough that we
> wanted to use %fs in x86-64 userspace, as opposed to the original ABI which
> used %gs both in userspace and in the kernel.

Ahh, I think you may be right for some CPUs. The zero selector is indeed 
potentially faster to load, since it doesn't have to even bother looking 
at the GDT/LDT.

That said, I doubt it's very noticeable. I just ran tests on both an old 
P4 and on a more modern Core 2 machine, and for both of those the 
performance was identical between loading a NUL selector and loading it 
with a non-zero one.

But I could well imagine that it matters a few cycles on other CPU's. But 
from my testing, it definitely isn't noticeable, and I think the 
maintenance advantage of using the same segment setup would more than make 
up for the fact that maybe some odd CPU can see a difference.

			Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/