From mboxrd@z Thu Jan 1 00:00:00 1970 From: Arnd Bergmann Subject: Re: [PATCH] asm-generic/io.h: Fix io{read,write}{16,32}be for big endian systems Date: Tue, 18 Jan 2011 20:56:35 +0100 Message-ID: <201101182056.35673.arnd@arndb.de> References: <1295374261-19609-1-git-send-email-lars@metafoo.de> <201101181944.07146.arnd@arndb.de> <4D35E378.70300@metafoo.de> Mime-Version: 1.0 Content-Type: Text/Plain; charset="iso-8859-15" Content-Transfer-Encoding: 7bit Return-path: Received: from moutng.kundenserver.de ([212.227.17.9]:64094 "EHLO moutng.kundenserver.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751745Ab1ART4l (ORCPT ); Tue, 18 Jan 2011 14:56:41 -0500 In-Reply-To: <4D35E378.70300@metafoo.de> Sender: linux-arch-owner@vger.kernel.org List-ID: To: Lars-Peter Clausen Cc: linux-arch@vger.kernel.org, linux-kernel@vger.kernel.org On Tuesday 18 January 2011 20:01:12 Lars-Peter Clausen wrote: > Well, i've though about that as well, but in the current asm-generic/io.h readl is > unconditionally defined as cpu_to_le32(__raw_readl(addr)) and ioread32 is defined as > readl. > > So unless an arch io.h undefines those macros and redefines them (which none of the > current archs does, as far as i can see), we are o > > If an arch chooses to redefine ioread or readl, it should probably also redefine > ioread{16,32}be. Right, but the header file also serves as a template for new architectures that cannot directly use it. I would prefer not to give a possibly bad example here, especially when it's in a rarely used function. > > The right solution is probably to use swab16/swab32 for the > > big-endian functions. This also corrects the iowrite functions > > which really should be using cpu_to_be32 instead of be32_to_cpu > > (although they are always defined to be the same afaict. > > This would first cause a conversion to little-endian, which is a swap() in the > generic case and then you would call swap() again on the result. Which is basically a > noop, but I'm not sure if compilers will detect this. The overhead of the swab() is certainly dwarfed by the long time spent in readl(). I would prefer to swap twice in this case and let the compiler work it out if possible. The next best alternative would probably be to define both ioread and ioread_be using __raw_readl in combination with a le32_to_cpu or be32_to_cpu. Arnd