Hi Masahiro,
On 23 January 2015 at 22:04, Masahiro YAMADA yamada.m@jp.panasonic.com wrote:
Hi Simon,
2015-01-24 0:50 GMT+09:00 Simon Glass sjg@chromium.org:
I tried to document the reasoning in the patches, but let me try to expand a bit. Hopefully this can provoke further comments / improvements.
The main motivation for me was that buses want to set up the platform data for their children before they are probed. In fact some children may never be probed. For example a SPI bus wants to know the chip select for each of its children.
At present we have to hunt around in the device tree to figure out which child is the right one, so we can probe it. Worse, the children's drivers (e.g. cros_ec on a SPI bus) have to be involved in setting themselves up. The device_probe_child() function was my attempt to make this fit better, and it did work, but I was not happy with it. You can see that from my unhappy comments in cros_ec, or SPI flash probe, for example.
The new approach makes buses easier to deal with as I hope you can see. The 'bind' step is supposed to set up the entire basic framework of the drivers at start-up. Everything should be visible in the tree (the exception being buses which must be probed at run-time) but nothing should be probed. Now, we are following that approach for buses also.
OK. I understand that this concept makes things much simpler, so
Reviewed-by: Masahiro Yamada yamada.m@jp.panasonic.com
OK I will respin this series.
Perhaps, we can have a flag like DM_FLAG_ALLOC_PLATDATA_LATE: If this flag is set, platdata is allocated in device_probe(), not device_bind(). But, I think it might make things much more complicated and probably make you unhappy. I do not have a strong option about this.
Well I'm keen on anything that makes it more efficient, but yes if we can keep it simple we should do that for now. Let's see how things pan out.
Let's go with your idea! If something does not go well, then we can come back here.
Yes we can.
Re the disadvantages:
- the per-child platform data for a bus is small, and we need not bind
devices which are disabled. So, a board should avoid having a lot of enabled devices which are never used
- malloc() is very fast, it is the platform data setup that takes
time. I agree this slows things down. But currently it only affects bus children, and only their basic information such as chip select.
The use of ofdata_to_platdata() is stil confined to when the device is actually probed. This helps to keep things fast in the common case where we don't need the platform data earlier.
I think we can refine this further, but what I have now feels a lot cleaner to me.
OK!
I have done with a quick review of this series.
I left some comments on some patches,but they are all minor issues.
If you agree to fix them, please send v3. I am OK with the whole concept, so I guess we can get it in during this MW.
Will do. Thanks very much for going through this in detail.
Regards, Simon