From mboxrd@z Thu Jan 1 00:00:00 1970 From: Robin Holt Date: Tue, 18 Nov 2008 11:58:12 +0000 Subject: Re: [PATCH] Ensure PSR.ac is cleared for early userspace Message-Id: <20081118115812.GA8778@sgi.com> List-Id: References: <200811120135.mAC1ZoSd017352@agluck-lia64.sc.intel.com> In-Reply-To: <200811120135.mAC1ZoSd017352@agluck-lia64.sc.intel.com> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: linux-ia64@vger.kernel.org On Tue, Nov 18, 2008 at 12:12:55AM -0700, Matthew Wilcox wrote: > On Tue, Nov 18, 2008 at 08:06:53AM +0100, Petr Tesarik wrote: > > Well, then let's make it a boot option (or a sysctl). The default must > > be psr.ac = 1, but if the admin sees "unaligned access" messages and > > knows that there is no way of getting a (timely) fix for that particular > > production system, s/he should be able to turn them off. > > Or how about ia64 does what every other Linux port does and turn off > alignment warnings altogether? Are they *particularly* expensive > on Itanium? I bet they aren't. They are expensive enough that our customers complain. Many of SGI's customers are concerned about that last 1% of performance. Having a way to quickly track down the customer's issue without having to recreate their exact setup is helpful. The last three kernel unaligned access issues I can recall being reported by our customers were in the IDE(1) and IPv4(2) areas. Neither of these areas is SGI specific and the three reports were customer workload specific. I doubt we would have gotten these isolated without frustrating the customer with cycles of try this, try this, oh and try this, now try this kernel with different options and you need to rebuild all those other things now too. As it was, we merely took their console log and /proc/kallsyms and lined up the source of the problem, scratched our heads for a bit, and isolated the issue. On a kernel modules note, we are seeing more and more of our customers running stuff like panasys, lustre, etc which have external kernel modules built either by the customer's support staff or other vendors. Putting custom kernels on a site is becoming much more difficult. If we have a choice, please consider failing on the side of easily isolating problems using a standard config. I like Petr's suggestion of making it a sysctl/boot option with the default being on. Thanks, Robin