Hi Wolfram, On 03/09/19 11:34, Wolfram Sang wrote: > >> Not if you define enough addresses in the pool. E.g. the DS90UB954 >> hardware can have 8 aliases per port, so if you have (n_ports * 8) >> addresses in the pool the problem is solved. > > And then you plug-in somewhere another board with another need for ATR > and you are out of addresses. Now I got what you mean: the aliases in a pool are reserved to an ATR and not available to another ATR. What about hoisting the pool one level up in DT, to the adapter where the ATRs are connected? >>> And another add-on module with >>> non-repogrammable devices may occupy addresses from the defined pool >>> above. >> >> You mean a new device on the local (SoC-to-ATR) bus? Well, it could as >> well occupy a non-described address that the ATR has already picked as >> an alias. > > Nope, I mean a seperate add-on which has a hardcoded I2C address on the > bus of the ATR parent. Then this hardcoded address needs to be removed > from the pool if it is in the wrong range. As I understand it, we are referring to the same use case: .---------------. .-----. .-------------------. | adapter (SoC) |---+---| ATR |---| remote slave 0x10 | '---------------' | '-----' '-------------------' | .----------. | device X | '----------' Here device X with hard-coded address 0x20 is plugged in at runtime. As you say it, if 0x20 is in the ATR pool we have a problem. But even without an explicit pool, initially 0x20 is unused and the ATR can use it for remote slave 0x10. Then at some point device X is connected, and suddenly 0x20 conflicts. To a limited extend the explicit pool could help if the list of possible addons is known: one can put in the pool only address that will never appear in addons. Hey, wait, the no-pool solution could have a way to handle this conflict. When device X is connected, the adapter can look for a new alias (0x21), call the i2c_atr_deconfigure_hw() and i2c_atr_setup_hw() callbacks to stop using 0x20 and start using 0x21. Doesn't look very lovely, but may be worth considering. >>> I am not perfectly happy with the assumption that all undescribed >>> addresses are automatically free. That also might need DTS updates to >>> describe all clients properly. But this change only needs to be done >>> once, and it will improve the description of the hardware. >> >> Right, but I still suspect some users won't do their homework and >> discover address conflicts at runtime, maybe months later, in a painful >> way. Also a chip might be undocumented on a given board, so they could >> do their homework and still have problems. > > Yes, we probably need a binding to mark an address as used even though > we don't know the device or don't have a driver for it. > > Don't get me wrong, I know what you mean. One of my boards has a client > soldered in a way so that it is still in debug mode. That means it > listens to addresses 0x03-0x07 to provide debug information. Took me a > while to find out what is happening there. > > But still, 'i2cdetect' showed all of these. > >> Despite my comments, I'm not strongly against your proposal. To me it >> doesn't seem to solve any problem, while it does introduce some degree >> of risk. Could you elaborate more on but what benefit it introduces? > > I'd think the risk of running out of defined addresses is somewhere > equal to running into (after a while) an unexpectedly used address. > I like the fix for the latter better because describing what is on the > bus is more helpful and generic than updating the pool-property every > time you need it. Plus, as mentioned above, other add-on hardware may > disturb your pool allocation. > > I expect this topic to be one of the discussion points of the BoF. Definitely. And having a list of use cases would help a lot IMO. I had never considered the use case described above, for example. Thanks, -- Luca