Message ID | 20230820124910.71526-1-biju.das.jz@bp.renesas.com (mailing list archive) |
---|---|
Headers | show |
Series | Convert enum->pointer for data in the tmp51x match tables | expand |
On Sun, Aug 20, 2023 at 01:49:08PM +0100, Biju Das wrote: > Convert enum->pointer for data in the match tables, so that > device_get_match_data() can do match against OF/ACPI/I2C tables, once i2c > bus type match support added to it. > I don't see why this would be necessary. You don't explain why the current implementation would no longer work. Various other drivers implement the same mechanism as this driver, i.e., type cast the return value of device_get_match_data() to a non-pointer. I'd argue that changing the functionality of device_get_match_data() such that this is no longer possible would be inherently flawed and would introduce unnecessary complexity to drivers using that mechanism today. If device_get_match_data() is enhanced to include the functionality of i2c_match_id(), it should be done in a way that doesn't mandate such an API change. > This patch series is only compile tested. I assume that means you don't have access to the chip. Is this correct ? Just asking, because it would be great to have a register dump which would enable me to write unit test code. Thanks, Guenter > > Biju Das (2): > hwmon: tmp513: Convert enum->pointer for data in the match tables > hwmon: tmp513: Add temp_config to struct tmp51x_info > > drivers/hwmon/tmp513.c | 51 ++++++++++++++++++++++-------------------- > 1 file changed, 27 insertions(+), 24 deletions(-) > > -- > 2.25.1 >
Hi Guenter Roeck, Thanks for the feedback. > Subject: Re: [PATCH 0/2] Convert enum->pointer for data in the tmp51x match > tables > > On Sun, Aug 20, 2023 at 01:49:08PM +0100, Biju Das wrote: > > Convert enum->pointer for data in the match tables, so that > > device_get_match_data() can do match against OF/ACPI/I2C tables, once > > i2c bus type match support added to it. > > > > I don't see why this would be necessary. You don't explain why the current > implementation would no longer work. Various other drivers implement the > same mechanism as this driver, i.e., type cast the return value of > device_get_match_data() to a non-pointer. I'd argue that changing the > functionality of device_get_match_data() such that this is no longer > possible would be inherently flawed and would introduce unnecessary > complexity to drivers using that mechanism today. If > device_get_match_data() is enhanced to include the functionality of > i2c_match_id(), it should be done in a way that doesn't mandate such an API > change. Currently nothing is broken. There is a helper function to do OF/ACPI/ID match(i2c_get_match_data). But Dmitry proposed to extend device_get_match_data() to buses so that we can get rid of i2c_get_match_data and future it can be extended to SPI aswell. see [1]. While doing this Jonathan found an issue where it won't work if OF table is using pointer and ID is using enum in the match data. Moreover,the proposed API returns NULL for non-match. So Andy proposed to convert the valid enums to non-zero or use a pointer. In this case, pointer makes sense as the hardware differences between the chips are compared against type rather than using feature and driver data. In the second patch, I just used a driver data to avoid one such case. [1] https://lore.kernel.org/all/20230804161728.394920-1-biju.das.jz@bp.renesas.com/ > > > This patch series is only compile tested. > > I assume that means you don't have access to the chip. Is this correct ? > Just asking, because it would be great to have a register dump which would > enable me to write unit test code. That is correct, I don't have access to the chip. Cheers, Biju > > > > > Biju Das (2): > > hwmon: tmp513: Convert enum->pointer for data in the match tables > > hwmon: tmp513: Add temp_config to struct tmp51x_info > > > > drivers/hwmon/tmp513.c | 51 > > ++++++++++++++++++++++-------------------- > > 1 file changed, 27 insertions(+), 24 deletions(-) > > > > -- > > 2.25.1 > >
On Sun, Aug 20, 2023 at 02:55:08PM +0000, Biju Das wrote: > Hi Guenter Roeck, > > Thanks for the feedback. > > > Subject: Re: [PATCH 0/2] Convert enum->pointer for data in the tmp51x match > > tables > > > > On Sun, Aug 20, 2023 at 01:49:08PM +0100, Biju Das wrote: > > > Convert enum->pointer for data in the match tables, so that > > > device_get_match_data() can do match against OF/ACPI/I2C tables, once > > > i2c bus type match support added to it. > > > > > > > I don't see why this would be necessary. You don't explain why the current > > implementation would no longer work. Various other drivers implement the > > same mechanism as this driver, i.e., type cast the return value of > > device_get_match_data() to a non-pointer. I'd argue that changing the > > functionality of device_get_match_data() such that this is no longer > > possible would be inherently flawed and would introduce unnecessary > > complexity to drivers using that mechanism today. If > > device_get_match_data() is enhanced to include the functionality of > > i2c_match_id(), it should be done in a way that doesn't mandate such an API > > change. > > Currently nothing is broken. There is a helper function > to do OF/ACPI/ID match(i2c_get_match_data). But Dmitry proposed > to extend device_get_match_data() to buses so that we can get > rid of i2c_get_match_data and future it can be extended to SPI aswell. see [1]. > > While doing this Jonathan found an issue where it won't work if > OF table is using pointer and ID is using enum in the match data. Moreover,the proposed API returns NULL for non-match. I think that is may be problem because many drivers don't have a value in the driver_data field. Maybe that doesn't matter because such drivers would normally not call device_get_match_data() or i2c_match_id(), but it would create some ambiguity because those functions would no longer work for all cases. > > So Andy proposed to convert the valid enums to non-zero or use a pointer. > There are _lots_ of drivers where the chip ID is in driver_data and starts with 0, or with the field not used. It is not my call to make, but I really think that demanding that this field is always != 0 is just wrong. > In this case, pointer makes sense as the hardware differences between > the chips are compared against type rather than using feature and > driver data. In the second patch, I just used a driver data to > avoid one such case. > > [1] https://lore.kernel.org/all/20230804161728.394920-1-biju.das.jz@bp.renesas.com/ > The suggested changes do make some sense independently of the API change, but not as suggested. I'd be inclined to accept the changes if the number of channels is kept in the configuration data. With that, the chip ID would no longer be needed. TMP513_TEMP_CONFIG_DEFAULT and TMP512_TEMP_CONFIG_DEFAULT are not really needed in the configuration data since the value can be derived from the number of channels instead of the chip type. According to what you said earlier, that suggests that the only necessary change would be to report the number of channels in driver_data / data because that would never be 0. Either case, I would want to keep temp_config and the number of channels in struct tmp51x_data to avoid repeated double dereferences and because temp_config could change (resistance correction enabled/disabled should be a devicetree property, conversion rate as well as channel enable should be sysfs attributes, and channel enable/disable should also be devicetree properties). The value could also be used to support suspend/resume in the driver if someone wants to add that at some point. In this context, if (data->id == tmp512 && channel == 4) return 0; is wrong because there are 3 or 4 channels, meaning the channel numbers are 0..3 and there is no channel 4. This should be "channel == 3" to disable the 4th channel on tmp512. Interesting that no one (including me ;-) noticed this. > > > > > This patch series is only compile tested. > > > > I assume that means you don't have access to the chip. Is this correct ? > > Just asking, because it would be great to have a register dump which would > > enable me to write unit test code. > > That is correct, I don't have access to the chip. > Too bad. Eric, if you are still around, would it be possible to send me register dumps ? Thanks, Guenter > Cheers, > Biju > > > > > > > > > Biju Das (2): > > > hwmon: tmp513: Convert enum->pointer for data in the match tables > > > hwmon: tmp513: Add temp_config to struct tmp51x_info > > > > > > drivers/hwmon/tmp513.c | 51 > > > ++++++++++++++++++++++-------------------- > > > 1 file changed, 27 insertions(+), 24 deletions(-) > > > > > > -- > > > 2.25.1 > > >
Hi Guenter Roeck, Thanks for the feedback. > Subject: Re: [PATCH 0/2] Convert enum->pointer for data in the tmp51x match > tables > > On Sun, Aug 20, 2023 at 02:55:08PM +0000, Biju Das wrote: > > Hi Guenter Roeck, > > > > Thanks for the feedback. > > > > > Subject: Re: [PATCH 0/2] Convert enum->pointer for data in the > > > tmp51x match tables > > > > > > On Sun, Aug 20, 2023 at 01:49:08PM +0100, Biju Das wrote: > > > > Convert enum->pointer for data in the match tables, so that > > > > device_get_match_data() can do match against OF/ACPI/I2C tables, > > > > once i2c bus type match support added to it. > > > > > > > > > > I don't see why this would be necessary. You don't explain why the > > > current implementation would no longer work. Various other drivers > > > implement the same mechanism as this driver, i.e., type cast the > > > return value of > > > device_get_match_data() to a non-pointer. I'd argue that changing > > > the functionality of device_get_match_data() such that this is no > > > longer possible would be inherently flawed and would introduce > > > unnecessary complexity to drivers using that mechanism today. If > > > device_get_match_data() is enhanced to include the functionality of > > > i2c_match_id(), it should be done in a way that doesn't mandate such > > > an API change. > > > > Currently nothing is broken. There is a helper function to do > > OF/ACPI/ID match(i2c_get_match_data). But Dmitry proposed to extend > > device_get_match_data() to buses so that we can get rid of > > i2c_get_match_data and future it can be extended to SPI aswell. see [1]. > > > > While doing this Jonathan found an issue where it won't work if OF > > table is using pointer and ID is using enum in the match data. > Moreover,the proposed API returns NULL for non-match. > > I think that is may be problem because many drivers don't have a value in > the driver_data field. Maybe that doesn't matter because such drivers would > normally not call device_get_match_data() or i2c_match_id(), Yes, that is correct. but it would > create some ambiguity because those functions would no longer work for all > cases. > > > > > So Andy proposed to convert the valid enums to non-zero or use a pointer. > > > There are _lots_ of drivers where the chip ID is in driver_data and starts > with 0, or with the field not used. It is not my call to make, but I really > think that demanding that this field is always != 0 is just wrong. For hwmon subsystem, this is the only i2c client driver using device_match_data(), other hwmon drivers are using of_device_get_match_data() and ID lookup. > > > In this case, pointer makes sense as the hardware differences between > > the chips are compared against type rather than using feature and > > driver data. In the second patch, I just used a driver data to avoid > > one such case. > > > > [1] > > > > The suggested changes do make some sense independently of the API change, > but not as suggested. I'd be inclined to accept the changes if the number > of channels is kept in the configuration data. With that, the chip ID would > no longer be needed. TMP513_TEMP_CONFIG_DEFAULT and > TMP512_TEMP_CONFIG_DEFAULT are not really needed in the configuration data > since the value can be derived from the number of channels instead of the > chip type. Yep, instead of using chip type for HW differences, we can use a macro to get default value based on the number of channels supported by the chip. > > According to what you said earlier, that suggests that the only necessary > change would be to report the number of channels in driver_data / data > because that would never be 0. OK. Since the value is less than 16, there won't be any issue. device_get_match_data() first look for match against firmware nodes, and will fallback to ID lookup(). > > Either case, I would want to keep temp_config and the number of channels in > struct tmp51x_data to avoid repeated double dereferences and because > temp_config could change (resistance correction enabled/disabled should be > a devicetree property, conversion rate as well as channel enable should be > sysfs attributes, and channel enable/disable should also be devicetree > properties). The value could also be used to support suspend/resume in the > driver if someone wants to add that at some point. > > In this context, > if (data->id == tmp512 && channel == 4) > return 0; > is wrong because there are 3 or 4 channels, meaning the channel numbers are > 0..3 and there is no channel 4. This should be "channel == 3" > to disable the 4th channel on tmp512. Interesting that no one (including > me ;-) noticed this. If I am correct, as per above, only max number channels supported by the chip can go to device data. That is the only HW difference between these 2 chips and other chip specific data can be derived from this. Cheers, Biju > > > > > > > > This patch series is only compile tested. > > > > > > I assume that means you don't have access to the chip. Is this > correct ? > > > Just asking, because it would be great to have a register dump which > > > would enable me to write unit test code. > > > > That is correct, I don't have access to the chip. > > > > Too bad. Eric, if you are still around, would it be possible to send me > register dumps ? > > Thanks, > Guenter > > > > Cheers, > > Biju > > > > > > > > > > > > > Biju Das (2): > > > > hwmon: tmp513: Convert enum->pointer for data in the match tables > > > > hwmon: tmp513: Add temp_config to struct tmp51x_info > > > > > > > > drivers/hwmon/tmp513.c | 51 > > > > ++++++++++++++++++++++-------------------- > > > > 1 file changed, 27 insertions(+), 24 deletions(-) > > > > > > > > -- > > > > 2.25.1 > > > >
Hi Guenter Roeck, > Subject: RE: [PATCH 0/2] Convert enum->pointer for data in the tmp51x match > tables > > Hi Guenter Roeck, > > Thanks for the feedback. > > > Subject: Re: [PATCH 0/2] Convert enum->pointer for data in the tmp51x > > match tables > > > > Either case, I would want to keep temp_config and the number of > > channels in struct tmp51x_data to avoid repeated double dereferences > > and because temp_config could change (resistance correction > > enabled/disabled should be a devicetree property, conversion rate as > > well as channel enable should be sysfs attributes, and channel > > enable/disable should also be devicetree properties). The value could > > also be used to support suspend/resume in the driver if someone wants to > add that at some point. > > > > In this context, > > if (data->id == tmp512 && channel == 4) > > return 0; > > is wrong because there are 3 or 4 channels, meaning the channel > > numbers are > > 0..3 and there is no channel 4. This should be "channel == 3" > > to disable the 4th channel on tmp512. Interesting that no one > > (including me ;-) noticed this. > > If I am correct, as per above, only max number channels supported by the > chip can go to device data. That is the only HW difference between these 2 > chips and other chip specific data can be derived from this. I guess, the below initial values for temp_config is chip specific apart from the number of supported channels as we won't be able to derive from number of channels?? #define TMP512_TEMP_CONFIG_DEFAULT 0xBF80 #define TMP513_TEMP_CONFIG_DEFAULT 0xFF80 Cheers, Biju
On Sun, Aug 20, 2023 at 05:36:02PM +0000, Biju Das wrote: > Hi Guenter Roeck, > > > Subject: RE: [PATCH 0/2] Convert enum->pointer for data in the tmp51x match > > tables > > > > Hi Guenter Roeck, > > > > Thanks for the feedback. > > > > > Subject: Re: [PATCH 0/2] Convert enum->pointer for data in the tmp51x > > > match tables > > > > > > Either case, I would want to keep temp_config and the number of > > > channels in struct tmp51x_data to avoid repeated double dereferences > > > and because temp_config could change (resistance correction > > > enabled/disabled should be a devicetree property, conversion rate as > > > well as channel enable should be sysfs attributes, and channel > > > enable/disable should also be devicetree properties). The value could > > > also be used to support suspend/resume in the driver if someone wants to > > add that at some point. > > > > > > In this context, > > > if (data->id == tmp512 && channel == 4) > > > return 0; > > > is wrong because there are 3 or 4 channels, meaning the channel > > > numbers are > > > 0..3 and there is no channel 4. This should be "channel == 3" > > > to disable the 4th channel on tmp512. Interesting that no one > > > (including me ;-) noticed this. > > > > If I am correct, as per above, only max number channels supported by the > > chip can go to device data. That is the only HW difference between these 2 > > chips and other chip specific data can be derived from this. > > I guess, the below initial values for temp_config > is chip specific apart from the number of supported channels > as we won't be able to derive from number of channels?? > > #define TMP512_TEMP_CONFIG_DEFAULT 0xBF80 > #define TMP513_TEMP_CONFIG_DEFAULT 0xFF80 > Bit 3-6 enable temperature sensors. The difference in bit 6 is that the 3rd external temperature sensor (which only exists for tmp513) is enabled if the bit is set. So the base value for TMP513_TEMP_CONFIG_DEFAULT would be 0xBF80 with bit 6 set if the chip has four temperature sensors. Guenter
On Sun, Aug 20, 2023 at 08:55:18AM -0700, Guenter Roeck wrote: > On Sun, Aug 20, 2023 at 02:55:08PM +0000, Biju Das wrote: > > > On Sun, Aug 20, 2023 at 01:49:08PM +0100, Biju Das wrote: ... > > While doing this Jonathan found an issue where it won't work if OF table is > > using pointer and ID is using enum in the match data. Moreover,the proposed > > API returns NULL for non-match. > > I think that is may be problem because many drivers don't have a value > in the driver_data field. Maybe that doesn't matter because such drivers > would normally not call device_get_match_data() or i2c_match_id(), > but it would create some ambiguity because those functions would no > longer work for all cases. Are you talking here about the cases where 0 / NULL makes things optional? Like when driver data is defined, we use its value, otherwise apply some defaults? If so, where do you see a problem? > > So Andy proposed to convert the valid enums to non-zero or use a pointer. > > > There are _lots_ of drivers where the chip ID is in driver_data and starts > with 0, or with the field not used. It is not my call to make, but I really > think that demanding that this field is always != 0 is just wrong. Those drivers are hackishly abusing OF ID tables. Those, that have _no_ OF ID tables are okay. Moreover this approach allows to have the driver data to be const, and keep it that way which now is problematic and in some cases may even be broken (due to wrong assumptions made by drivers). I.o.w. this makes code more robust against the possible mangling of driver data contents at runtime.
On Mon, Aug 21, 2023 at 12:17:34PM +0300, Andy Shevchenko wrote: > On Sun, Aug 20, 2023 at 08:55:18AM -0700, Guenter Roeck wrote: > > On Sun, Aug 20, 2023 at 02:55:08PM +0000, Biju Das wrote: > > > > On Sun, Aug 20, 2023 at 01:49:08PM +0100, Biju Das wrote: > > ... > > > > While doing this Jonathan found an issue where it won't work if OF table is > > > using pointer and ID is using enum in the match data. Moreover,the proposed > > > API returns NULL for non-match. > > > > I think that is may be problem because many drivers don't have a value > > in the driver_data field. Maybe that doesn't matter because such drivers > > would normally not call device_get_match_data() or i2c_match_id(), > > but it would create some ambiguity because those functions would no > > longer work for all cases. > > Are you talking here about the cases where 0 / NULL makes things optional? > Like when driver data is defined, we use its value, otherwise apply some > defaults? If so, where do you see a problem? > The fact that you have to change lots of drivers to make this work should prove my point. > > > So Andy proposed to convert the valid enums to non-zero or use a pointer. > > > > > There are _lots_ of drivers where the chip ID is in driver_data and starts > > with 0, or with the field not used. It is not my call to make, but I really > > think that demanding that this field is always != 0 is just wrong. > > Those drivers are hackishly abusing OF ID tables. Those, that have _no_ OF ID > tables are okay. This is what you are saying. That doesn't make it true. With the same logic I could claim that drivers providing a pointer in i2c_device_id would hackishly abuse i2c_device_id tables. Guenter > > Moreover this approach allows to have the driver data to be const, and keep it > that way which now is problematic and in some cases may even be broken (due to > wrong assumptions made by drivers). I.o.w. this makes code more robust against > the possible mangling of driver data contents at runtime. > > -- > With Best Regards, > Andy Shevchenko > >
On Mon, Aug 21, 2023 at 05:59:57AM -0700, Guenter Roeck wrote: > On Mon, Aug 21, 2023 at 12:17:34PM +0300, Andy Shevchenko wrote: > > On Sun, Aug 20, 2023 at 08:55:18AM -0700, Guenter Roeck wrote: ... > > Those drivers are hackishly abusing OF ID tables. Those, that have _no_ OF ID > > tables are okay. > > This is what you are saying. That doesn't make it true. With the same logic > I could claim that drivers providing a pointer in i2c_device_id would > hackishly abuse i2c_device_id tables. ...because kernel_ulong_t was poorly chosen type for that to begin with? Abusing pointer is much worse than using long as a pointer holder. Moreover the latter is valid case (okay, not as kernel_ulong_t, but unsigned long) in many places in the kernel, starting from the memory manager, so your argument doesn't look like a good one to me.