| Message ID | 20190311112216.31391-2-hdegoede@redhat.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | intel-lpss-pci-i2c: Fix oops when there is no ACPI fwnode | expand |
On Mon, Mar 11, 2019 at 12:22:15PM +0100, Hans de Goede wrote: > Before this commit the i2c-designware-platdrv assumes that if the pdev > has an apci-companion it should use a dynamic adapter-nr and otherwise > it will use pdev->id as adapter-nr. > > On some devices e.g. the Apollo Lake using Acer TravelMate Spin B118, > some of the LPSS i2c-adapters are enumerated through PCI and do not have > an ACPI fwnode. These devices are handled as mfd devices so they end up > using the i2c-designware-platdrv driver. > > This results in the i2c-adapter being registered with the mfd generated > pdev->id as adapter-nr, which conflicts with existing adapters, triggering > a WARN(id < 0, "couldn't get idr") in i2c-core-base.c and causing the > adapter registration to fail. > > This commit adds support for setting a "linux,use-dynamic-adapter-nr" > device property on the device to make i2c-designware-platdrv use dynamic > adapter-nrs on devices without an ACPI fwnode, together with changes to > drivers/mfd/intel-lpss-pci.c to set this, this fixes the WARN. > > Before this commit the setting of the adapter.nr was somewhat convoluted, > in the acpi_companion case it was set from dw_i2c_acpi_configure, in the > non acpi_companion case it was set from dw_i2c_set_fifo_size() based on > tx_fifo_depth not being set yet. This commit also cleans this up. Can we split this to two patches, i.e. one is almost the same as this one, except the second one adds a new property check to the conditional? If you agree to do so, you may add mine Rb tag to the first one out of three. > Note the "linux,use-dynamic-adapter-nr" is meant for kernel internal use > only, therefor it is NOT documented under Documents/devicetree/bindings. To the second and third ones, can we rather check if the device has fwnode either ACPI or swnode? AFAIU now we have swnode assigned during MFD device registration and can easily distinguish this w/o any additional properties.
Hi On 3/11/19 1:22 PM, Hans de Goede wrote: > Before this commit the i2c-designware-platdrv assumes that if the pdev > has an apci-companion it should use a dynamic adapter-nr and otherwise > it will use pdev->id as adapter-nr. > > On some devices e.g. the Apollo Lake using Acer TravelMate Spin B118, > some of the LPSS i2c-adapters are enumerated through PCI and do not have > an ACPI fwnode. These devices are handled as mfd devices so they end up > using the i2c-designware-platdrv driver. > > This results in the i2c-adapter being registered with the mfd generated > pdev->id as adapter-nr, which conflicts with existing adapters, triggering > a WARN(id < 0, "couldn't get idr") in i2c-core-base.c and causing the > adapter registration to fail. > I went thinking would we get a regression if we switch the i2c-designware-platdrv to dynamic numbering unconditionally? Only drivers/mfd/intel-lpss.c and drivers/mfd/intel_quark_i2c_gpio.c register platform device "i2c_designware" and otherwise in the driver itself for known ACPI IDs and device tree bindings. Things should be fine for ACPI cases if slave devices are also described in ACPI tables. As far as I've understood with device tree matching adapter number is irrelevant in slave device registration? Andy: could you tell by commit 918fe70cf475 ("mfd: intel_quark_i2c_gpio: support devices behind i2c bus") are those devices described in ACPI or in some i2c_board_infos with referring to fixed adapter number either in or out of kernel tree code? Then drivers/platform/chrome/chromeos_laptop.c is the only code searching for adapter named as "Synopsys DesignWare I2C adapter" without assuming any fixed adapter numbering. What's unclear to me can there be device tree cases where i2c-designware probing comes with pdev->id not starting from zero or in different order? I.e. would it make difference do we use pdev->id or dynamic adapter numbering?
Hi, On 11-03-19 13:52, Andy Shevchenko wrote: > On Mon, Mar 11, 2019 at 12:22:15PM +0100, Hans de Goede wrote: >> Before this commit the i2c-designware-platdrv assumes that if the pdev >> has an apci-companion it should use a dynamic adapter-nr and otherwise >> it will use pdev->id as adapter-nr. >> >> On some devices e.g. the Apollo Lake using Acer TravelMate Spin B118, >> some of the LPSS i2c-adapters are enumerated through PCI and do not have >> an ACPI fwnode. These devices are handled as mfd devices so they end up >> using the i2c-designware-platdrv driver. >> >> This results in the i2c-adapter being registered with the mfd generated >> pdev->id as adapter-nr, which conflicts with existing adapters, triggering >> a WARN(id < 0, "couldn't get idr") in i2c-core-base.c and causing the >> adapter registration to fail. >> >> This commit adds support for setting a "linux,use-dynamic-adapter-nr" >> device property on the device to make i2c-designware-platdrv use dynamic >> adapter-nrs on devices without an ACPI fwnode, together with changes to >> drivers/mfd/intel-lpss-pci.c to set this, this fixes the WARN. >> > >> Before this commit the setting of the adapter.nr was somewhat convoluted, >> in the acpi_companion case it was set from dw_i2c_acpi_configure, in the >> non acpi_companion case it was set from dw_i2c_set_fifo_size() based on >> tx_fifo_depth not being set yet. This commit also cleans this up. > > Can we split this to two patches, i.e. one is almost the same as this one, > except the second one adds a new property check to the conditional? Ok, I've split the patch for v2 of the series. > If you agree to do so, you may add mine Rb tag to the first one out of three. > >> Note the "linux,use-dynamic-adapter-nr" is meant for kernel internal use >> only, therefor it is NOT documented under Documents/devicetree/bindings. > > To the second and third ones, can we rather check if the device has fwnode > either ACPI or swnode? AFAIU now we have swnode assigned during MFD device > registration and can easily distinguish this w/o any additional properties. Detecting a swnode is non trivial, it may be either the primary or secondary fwnode depending if the device already had a node before calling device_add_properties. More importantly deciding this based on the presence of a swnode feels like its "wrong". But I've come up with a better solution. I've analyzed all ways a designware_i2c compatible platform-device can be instantiated and all cases except for: drivers/mfd/intel_quark_i2c_gpio.c drivers/mfd/intel-lpss-pci.c Are always using dynamic adapter numbers already. The Quark X1000 has only 1 i2c controller, so there using dynamic adapter numbers will not make a difference. And the intel-lpss-pci.c case is the one which we actually want to fix, devices using this code have many i2c busses so using a fixed bus number leads to the bus-number conflicts this patch is trying to fix. So we can simply always use dynamic adapter numbers, see the commit message of the second patch in v2 of this series for the full analysis. Regards, Hans
Hi, On 12-03-19 15:47, Jarkko Nikula wrote: > Hi > > On 3/11/19 1:22 PM, Hans de Goede wrote: >> Before this commit the i2c-designware-platdrv assumes that if the pdev >> has an apci-companion it should use a dynamic adapter-nr and otherwise >> it will use pdev->id as adapter-nr. >> >> On some devices e.g. the Apollo Lake using Acer TravelMate Spin B118, >> some of the LPSS i2c-adapters are enumerated through PCI and do not have >> an ACPI fwnode. These devices are handled as mfd devices so they end up >> using the i2c-designware-platdrv driver. >> >> This results in the i2c-adapter being registered with the mfd generated >> pdev->id as adapter-nr, which conflicts with existing adapters, triggering >> a WARN(id < 0, "couldn't get idr") in i2c-core-base.c and causing the >> adapter registration to fail. >> > I went thinking would we get a regression if we switch the i2c-designware-platdrv to dynamic numbering unconditionally? > > Only drivers/mfd/intel-lpss.c and drivers/mfd/intel_quark_i2c_gpio.c register platform device "i2c_designware" and otherwise in the driver itself for known ACPI IDs and device tree bindings. > > Things should be fine for ACPI cases if slave devices are also described in ACPI tables. As far as I've understood with device tree matching adapter number is irrelevant in slave device registration? > > Andy: could you tell by commit 918fe70cf475 ("mfd: intel_quark_i2c_gpio: support devices behind i2c bus") are those devices described in ACPI or in some i2c_board_infos with referring to fixed adapter number either in or out of kernel tree code? > > Then drivers/platform/chrome/chromeos_laptop.c is the only code searching for adapter named as "Synopsys DesignWare I2C adapter" without assuming any fixed adapter numbering. > > What's unclear to me can there be device tree cases where i2c-designware probing comes with pdev->id not starting from zero or in different order? I.e. would it make difference do we use pdev->id or dynamic adapter numbering? I've just completed analyzing all ways a designware_i2c (or compatible) platform device can be instantiated and I've come to the conclusion that always using dynamic adapter-nrs is fine and is a proper, clean fix for this. Also see my reply to Andy which I send about the same time you send your reply :) Regards, Hans
On Tue, Mar 12, 2019 at 04:47:48PM +0200, Jarkko Nikula wrote: > On 3/11/19 1:22 PM, Hans de Goede wrote: > > Before this commit the i2c-designware-platdrv assumes that if the pdev > > has an apci-companion it should use a dynamic adapter-nr and otherwise > > it will use pdev->id as adapter-nr. > > > > On some devices e.g. the Apollo Lake using Acer TravelMate Spin B118, > > some of the LPSS i2c-adapters are enumerated through PCI and do not have > > an ACPI fwnode. These devices are handled as mfd devices so they end up > > using the i2c-designware-platdrv driver. > > > > This results in the i2c-adapter being registered with the mfd generated > > pdev->id as adapter-nr, which conflicts with existing adapters, triggering > > a WARN(id < 0, "couldn't get idr") in i2c-core-base.c and causing the > > adapter registration to fail. > > > I went thinking would we get a regression if we switch the > i2c-designware-platdrv to dynamic numbering unconditionally? > > Only drivers/mfd/intel-lpss.c and drivers/mfd/intel_quark_i2c_gpio.c > register platform device "i2c_designware" and otherwise in the driver itself > for known ACPI IDs and device tree bindings. > > Things should be fine for ACPI cases if slave devices are also described in > ACPI tables. As far as I've understood with device tree matching adapter > number is irrelevant in slave device registration? Seems like Hans came to the same conclusion. > Andy: could you tell by commit 918fe70cf475 ("mfd: intel_quark_i2c_gpio: > support devices behind i2c bus") are those devices described in ACPI or in > some i2c_board_infos with referring to fixed adapter number either in or out > of kernel tree code? As far as I remember they are coming from ACPI, but you may easily check on real hardware we have in our lab. > Then drivers/platform/chrome/chromeos_laptop.c is the only code searching > for adapter named as "Synopsys DesignWare I2C adapter" without assuming any > fixed adapter numbering. > What's unclear to me can there be device tree cases where i2c-designware > probing comes with pdev->id not starting from zero or in different order? > I.e. would it make difference do we use pdev->id or dynamic adapter > numbering?
diff --git a/drivers/i2c/busses/i2c-designware-platdrv.c b/drivers/i2c/busses/i2c-designware-platdrv.c index ead5e7de3e4d..9f28159efdb2 100644 --- a/drivers/i2c/busses/i2c-designware-platdrv.c +++ b/drivers/i2c/busses/i2c-designware-platdrv.c @@ -86,7 +86,6 @@ static int dw_i2c_acpi_configure(struct platform_device *pdev) struct i2c_timings *t = &dev->timings; u32 ss_ht = 0, fp_ht = 0, hs_ht = 0, fs_ht = 0; - dev->adapter.nr = -1; dev->tx_fifo_depth = 32; dev->rx_fifo_depth = 32; @@ -219,7 +218,7 @@ static void i2c_dw_configure_slave(struct dw_i2c_dev *dev) dev->mode = DW_IC_SLAVE; } -static void dw_i2c_set_fifo_size(struct dw_i2c_dev *dev, int id) +static void dw_i2c_set_fifo_size(struct dw_i2c_dev *dev) { u32 param, tx_fifo_depth, rx_fifo_depth; @@ -233,7 +232,6 @@ static void dw_i2c_set_fifo_size(struct dw_i2c_dev *dev, int id) if (!dev->tx_fifo_depth) { dev->tx_fifo_depth = tx_fifo_depth; dev->rx_fifo_depth = rx_fifo_depth; - dev->adapter.nr = id; } else if (tx_fifo_depth >= 2) { dev->tx_fifo_depth = min_t(u32, dev->tx_fifo_depth, tx_fifo_depth); @@ -324,6 +322,13 @@ static int dw_i2c_plat_probe(struct platform_device *pdev) if (has_acpi_companion(&pdev->dev)) dw_i2c_acpi_configure(pdev); + if (has_acpi_companion(&pdev->dev) || + device_property_read_bool(&pdev->dev, + "linux,use-dynamic-adapter-nr")) + dev->adapter.nr = -1; + else + dev->adapter.nr = pdev->id; + /* * Only standard mode at 100kHz, fast mode at 400kHz, * fast mode plus at 1MHz and high speed mode at 3.4MHz are supported. @@ -358,7 +363,7 @@ static int dw_i2c_plat_probe(struct platform_device *pdev) div_u64(clk_khz * t->sda_hold_ns + 500000, 1000000); } - dw_i2c_set_fifo_size(dev, pdev->id); + dw_i2c_set_fifo_size(dev); adap = &dev->adapter; adap->owner = THIS_MODULE;
Before this commit the i2c-designware-platdrv assumes that if the pdev has an apci-companion it should use a dynamic adapter-nr and otherwise it will use pdev->id as adapter-nr. On some devices e.g. the Apollo Lake using Acer TravelMate Spin B118, some of the LPSS i2c-adapters are enumerated through PCI and do not have an ACPI fwnode. These devices are handled as mfd devices so they end up using the i2c-designware-platdrv driver. This results in the i2c-adapter being registered with the mfd generated pdev->id as adapter-nr, which conflicts with existing adapters, triggering a WARN(id < 0, "couldn't get idr") in i2c-core-base.c and causing the adapter registration to fail. This commit adds support for setting a "linux,use-dynamic-adapter-nr" device property on the device to make i2c-designware-platdrv use dynamic adapter-nrs on devices without an ACPI fwnode, together with changes to drivers/mfd/intel-lpss-pci.c to set this, this fixes the WARN. Before this commit the setting of the adapter.nr was somewhat convoluted, in the acpi_companion case it was set from dw_i2c_acpi_configure, in the non acpi_companion case it was set from dw_i2c_set_fifo_size() based on tx_fifo_depth not being set yet. This commit also cleans this up. Note the "linux,use-dynamic-adapter-nr" is meant for kernel internal use only, therefor it is NOT documented under Documents/devicetree/bindings. BugLink: https://bugzilla.redhat.com/show_bug.cgi?id=1687065 Signed-off-by: Hans de Goede <hdegoede@redhat.com> --- drivers/i2c/busses/i2c-designware-platdrv.c | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-)