Return-Path: Date: Wed, 26 Nov 2014 11:53:46 +0100 Message-ID: From: Takashi Iwai To: Oliver Neukum Cc: Marcel Holtmann , Dave Jones , Linux Kernel , pgynther@google.com, linux-bluetooth@vger.kernel.org Subject: Re: bluetooth related firmware loader spew on resume. In-Reply-To: <1416998616.3171.13.camel@linux-0dmf.site> References: <20141111181228.GA27815@redhat.com> <1416996623.3171.7.camel@linux-0dmf.site> <1416998616.3171.13.camel@linux-0dmf.site> MIME-Version: 1.0 (generated by SEMI 1.14.6 - "Maruoka") Content-Type: text/plain; charset=US-ASCII List-ID: At Wed, 26 Nov 2014 11:43:36 +0100, Oliver Neukum wrote: > > On Wed, 2014-11-26 at 11:31 +0100, Takashi Iwai wrote: > > At Wed, 26 Nov 2014 11:10:23 +0100, > > Oliver Neukum wrote: > > > > > > On Wed, 2014-11-26 at 09:52 +0100, Takashi Iwai wrote: > > > > At Wed, 26 Nov 2014 14:15:27 +0900, > > > > > > > In order to paper over this, we may also remember the failing firmware > > > > and avoid loading it. This might be an easer way than the endless > > > > fight against UMH race... > > > > > > Hi, > > > > > > the full fix would be to implement reset_resume() for btusb. > > > It seems to me that setup() should be split in two methods, > > > one to request the firmware from user space and the second > > > to transfer it to the device. reset_resume() would just need > > > to repeat the second operation. > > > > I'm not against it, but one slight drawback is that you'll have to > > remember the firmware content to transfer by the driver itself in this > > scenario. In the firmware loader framework, the content is re-read > > at resume so that the largish content isn't kept unnecessarily during > > the whole operation. > > That isn't a drawback but an advantage. Firmware for devices that > do power management needs to be in RAM. The right time to free it > is in disconnect(). But why does that mean that the driver has to > manage the firmware? Can't the firmware layer do it? The f/w loader remembers the f/w names of the successful loads, and retries to load them automatically at the suspend time. But it doesn't remember/cache the failed loads. So, when the driver retires request_firmware() for a non-existing file in the resume path, it actually reaches to the f/w loading part again unexpectedly. Takashi