From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756882AbaIRJ6V (ORCPT ); Thu, 18 Sep 2014 05:58:21 -0400 Received: from eddie.linux-mips.org ([148.251.95.138]:38880 "EHLO cvs.linux-mips.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756092AbaIRJ6U (ORCPT ); Thu, 18 Sep 2014 05:58:20 -0400 Date: Thu, 18 Sep 2014 11:58:13 +0200 From: Ralf Baechle To: minyard@acm.org Cc: linux-mips@linux-mips.org, linux-kernel@vger.kernel.org, Corey Minyard Subject: Re: [PATCH] mips: Save all registers when saving the frame Message-ID: <20140918095813.GA9804@linux-mips.org> References: <1410903925-10744-1-git-send-email-minyard@acm.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1410903925-10744-1-git-send-email-minyard@acm.org> User-Agent: Mutt/1.5.23 (2014-03-12) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, Sep 16, 2014 at 04:45:25PM -0500, minyard@acm.org wrote: > From: Corey Minyard > > The MIPS frame save code was just saving a few registers, enough to > do a backtrace if every function set up a frame. However, this is > not working if you are using DWARF unwinding, because most of the > registers are wrong. This was causing kdump backtraces to be short > or bogus. > > So save all the registers. The stratey of partial and full stack frames was developed in '97 to bring down the syscall overhead. It certaily was very effective - it brought down the syscall latency to the level of Alphas running at much higher clock. That certainly worked well back then for kernel 2.0 / 2.2. But the syscall code has become much more complex. Since then support for 64 bit kernels, two 32 bit ABIs running on a 64 bit kernels and numerous features that changed the once simple syscall path have been implemented. My gut feeling is it might be worth to yank out the whole optimization to see how much code complexity we get rid of in exchange for how much extra syscall latency. Ralf