From: Peter Zijlstra Subject: Re: [patch 1/2] genirq: Provide synchronize_hardirq() Date: Sat, 15 Feb 2014 10:50:08 +0100 Message-ID: <20140215095008.GG14089@laptop.programming.kicks-ass.net> References: <20140215002742.507053194@linutronix.de> <20140215003823.653236081@linutronix.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from merlin.infradead.org ([205.233.59.134]:33216 "EHLO merlin.infradead.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751754AbaBOJuX (ORCPT ); Sat, 15 Feb 2014 04:50:23 -0500 Content-Disposition: inline In-Reply-To: <20140215003823.653236081@linutronix.de> Sender: linux-mmc-owner@vger.kernel.org List-Id: linux-mmc@vger.kernel.org To: Thomas Gleixner Cc: LKML , Russell King , Chris Ball , linux-mmc@vger.kernel.org, Ingo Molnar On Sat, Feb 15, 2014 at 12:55:18AM -0000, Thomas Gleixner wrote: > synchronize_irq() waits for hard irq and threaded handlers to complete > before returning. For some special cases we only need to make sure > that the hard interrupt part of the irq line is not in progress when > we disabled the - possibly shared - interrupt at the device level. > > A proper use case for this was provided by Russell. The sdhci driver > requires some irq triggered functions to be run in thread context. The > current implementation of the thread context is a sdio private kthread > construct, which has quite some shortcomings. These can be avoided > when the thread is directly associated to the device interrupt via the > generic threaded irq infrastructure. > > Though there is a corner case related to run time power management > where one side disables the device interrupts at the device level and > needs to make sure, that an already running hard interrupt handler has > completed before proceeding further. Though that hard interrupt > handler might wake the associated thread, which in turn can request > the runtime PM to reenable the device. Using synchronize_irq() leads > to an immediate deadlock of the irq thread waiting for the PM lock and > the synchronize_irq() waiting for the irq thread to complete. > > Due to the fact that it is sufficient for this case to ensure that no > hard irq handler is executing a new function which avoids the check > for the thread is required. > > Add a function, which just monitors the hard irq parts and ignores the > threaded handlers. > > Signed-off-by: Thomas Gleixner Acked-by: Peter Zijlstra