From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1764928AbYDVTXc (ORCPT ); Tue, 22 Apr 2008 15:23:32 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1755860AbYDVTXT (ORCPT ); Tue, 22 Apr 2008 15:23:19 -0400 Received: from smtp1.linux-foundation.org ([140.211.169.13]:45407 "EHLO smtp1.linux-foundation.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1758246AbYDVTXR (ORCPT ); Tue, 22 Apr 2008 15:23:17 -0400 Date: Tue, 22 Apr 2008 12:22:32 -0700 (PDT) From: Linus Torvalds To: Ingo Molnar cc: Jens Axboe , linux-arch@vger.kernel.org, Linux Kernel Mailing List , npiggin@suse.de, peterz@infradead.org, sam@ravnborg.org Subject: Re: [PATCH 2/11] x86: convert to generic helpers for IPI function calls In-Reply-To: <20080422191213.GA6370@elte.hu> Message-ID: References: <1208890227-24808-1-git-send-email-jens.axboe@oracle.com> <1208890227-24808-3-git-send-email-jens.axboe@oracle.com> <20080422191213.GA6370@elte.hu> User-Agent: Alpine 1.10 (LFD 962 2008-03-14) MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, 22 Apr 2008, Ingo Molnar wrote: > > > Or is it just a performance optimization? [...] > > yes and i gave in - Nick and Jens wants to do some crazy stuff and they > had the numbers. Here's the previous discussion: > > http://lkml.org/lkml/2008/2/27/125 No, the previous discussion was about single *queues* vs single *vectors*. I agree unconditionally with the decision to use a separate per-cpu queue from the shared queue (in fact, I would argue that the "mask" code might want to notice when the mask is just a single CPU, and turn a mask request into a targeted request). But I wonder why we want to then have two IPI target vectors, when it would appear to be perfectly fine and cheap to have just a single vector that can handle both the per-cpu case and the shared queue case (since the thing would tend to be one or the other, not both). A single vector is still pefectly fine, if 99% of all usage cases are the targeted-to-a-single-cpu thing, because the shared queue will basically be empty (and you can test that without even taking any locks). Linus