On 9/27/23 22:38, Wolfram Sang wrote:
So my next question, is the change to dw_reg_write something that I should
write and submit, or should someone else submit something more generalized,
like option 2 above? I don't own the i2c driver, I'm just trying to fix one
issue on one processor with minimal risk of breaking something. I don't have
the broader view of what's optimal for the whole DesignWare i2c driver. I
also don't have any way to test changes on other models of processors.
Well, I guess this is a question for the designware maintainers: do we
want this one conversion from *_relaxed to non-relaxed. Or are we
playing safe by using non-relaxed all the time. I would suggest the
latter because the drivers I look after hardly write registers in a hot
path (and not many of them at a time). But you guys know your driver
better...
Well I don't have any preference (read enough knowledge) either here and
I hardly think performance becomes issue in any configuration.
Not a showstopper to this fix nor necessarily need to cover either but
one another memory barrier case might be in i2c-slave flows:
1. I2C bus read/write from another host
2. Interrupt to i2c-designware IP
i2c-designware-slave.c: i2c_dw_isr_slave()
i2c-core-slave.c: i2c_slave_event()
-> irq handler goes to slave backend like i2c-slave-eeprom
i2c-slave-eeprom.c: i2c_slave_eeprom_slave_cb()
3. Shared data between irq handler and process context
struct eeprom_data is accessed both from irq handler via
i2c_slave_eeprom_slave_cb() and process context via sysfs node handlers
i2c_slave_eeprom_bin_read() and i2c_slave_eeprom_bin_write()
