From mboxrd@z Thu Jan 1 00:00:00 1970 From: Guenter Roeck Subject: Re: [PATCH -next] smsc911x: Fix crash seen if neither ACPI nor OF is configured or used Date: Wed, 26 Aug 2015 11:34:04 -0700 Message-ID: <55DE069C.7000402@roeck-us.net> References: <1439844336-21596-1-git-send-email-linux@roeck-us.net> <20150826170410.GP4215@atomide.com> Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: 7bit Cc: Steve Glendinning , netdev@vger.kernel.org, linux-kernel@vger.kernel.org, Jeremy Linton , linux-omap@vger.kernel.org To: Tony Lindgren Return-path: Received: from bh-25.webhostbox.net ([208.91.199.152]:36145 "EHLO bh-25.webhostbox.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757061AbbHZSeH (ORCPT ); Wed, 26 Aug 2015 14:34:07 -0400 In-Reply-To: <20150826170410.GP4215@atomide.com> Sender: netdev-owner@vger.kernel.org List-ID: On 08/26/2015 10:04 AM, Tony Lindgren wrote: > Hi, > > * Guenter Roeck [150817 13:48]: >> Commit 0b50dc4fc971 ("Convert smsc911x to use ACPI as well as DT") makes >> the call to smsc911x_probe_config() unconditional, and no longer fails if >> there is no device node. device_get_phy_mode() is called unconditionally, >> and if there is no phy node configured returns an error code. This error >> code is assigned to phy_interface, and interpreted elsewhere in the code >> as valid phy mode. This in turn causes qemu to crash when running a >> variant of realview_pb_defconfig. >> >> qemu: hardware error: lan9118_read: Bad reg 0x86 >> >> Fixes: 0b50dc4fc971 ("Convert smsc911x to use ACPI as well as DT") >> Cc: Jeremy Linton >> Cc Graeme Gregory >> Signed-off-by: Guenter Roeck >> --- >> drivers/net/ethernet/smsc/smsc911x.c | 7 ++++++- >> 1 file changed, 6 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/net/ethernet/smsc/smsc911x.c b/drivers/net/ethernet/smsc/smsc911x.c >> index 0f21aa3bb537..34f97684506b 100644 >> --- a/drivers/net/ethernet/smsc/smsc911x.c >> +++ b/drivers/net/ethernet/smsc/smsc911x.c >> @@ -2367,12 +2367,17 @@ static const struct smsc911x_ops shifted_smsc911x_ops = { >> static int smsc911x_probe_config(struct smsc911x_platform_config *config, >> struct device *dev) >> { >> + int phy_interface; >> u32 width = 0; >> >> if (!dev) >> return -ENODEV; >> >> - config->phy_interface = device_get_phy_mode(dev); >> + phy_interface = device_get_phy_mode(dev); >> + if (phy_interface < 0) >> + return phy_interface; >> + >> + config->phy_interface = phy_interface; >> >> device_get_mac_address(dev, config->mac, ETH_ALEN); > > Looks like this change makes at least omap boards using smsc911x > fail with -22 for me in Linux next. > > Do any of the the device tree configured smsc911x devices actually > have a phy configured? > Ok, this is more subtle than I thought. Previously, the code would not attempt any devicetree configuration if devicetree was not configured. Now it does. The error return from device_get_phy_mode() isn't the actual problem. Apparently it doesn't really matter if a nonsensical value is assigned to phy_interface. The problem is that the reg-io-width property is obviously not present in the non-dt and non-acpi case. This overwrites the existing platform data configuration and selects 16 bit mode, to which the (simulated) hardware obviously reacts less than enthusiastic. Fixing this properly won't be easy. If the "reg-io-width" property is not present or wrong, the default register width is 16 bit. Obviously, if neither DT nor ACPI is available, it won't be present. This causes the crash I had observed. Bad part is that there does not seem to be a reliable means to detect that platform data should be used in that situation. Other device_get_XXX functions return -ENXIO if that happens, but not device_property_read_u32(). It is _supposed_ to return it per its API, but it doesn't (it returns -ENODATA). We may need two separate patches, one to fix up device_property_read_u32() to return -ENXIO, and one to fix smsc911x_probe_config() to ignore the error from device_get_phy_mode(), and to bail out if device_property_read_u32() returns -ENXIO. The simpler alternative would be to check the return value from device_property_read_u32() for both -ENXIO and -ENODATA. This would make the code independent of the necessary core changes (which may take a while). I tested this variant, and it works, at least for the non-DT case. Does this make sense ? Thanks, Guenter