Hi Sekhar,
On 23/08/2013 10:50, Sekhar Nori wrote:
Hi Benoit,
Did you get a chance to think about this, I have provided some
replies below.
On Friday 16 August 2013 10:03 PM, Benoit Cousson wrote:
Hi Sekhar,
On 16/08/2013 17:41, Sekhar Nori wrote:
On 8/16/2013 7:45 PM, Benoit Cousson wrote:
Hi Gururaja,
On 16/08/2013 13:36, Hebbar, Gururaja wrote:
The syntax of compatible property in DT is to mention the
Most specific match to most generic match.
Since AM335x is the platform with latest IP revision, add it
1st in the device id table.
I don't understand why? The order should not matter at all.
Yes, it should not. We are trying to work around a bug in the
kernel where the compatible order is not honored (instead the
order in of_match[] array matters). There were patches being
discussed to fix this.
I've tried to follow the thread you had with Mark on the v2,
but AFAIK, you've never answered to his latest question.
Moreover, checking the differences between the Davinci and the
am3352 RTC IP, I would not claim that both are compatible.
Sure you can use the am3352 with the Davinci driver, but you
will lose the wakeup functionality without even being notify
about that.
When the kernel is fixed for the bug pointed out above, this
should not happen with properly defined compatible property.
For my point of view, compatible mean that the HW will still be
fully functional with both versions of the driver, which is not
the case here.
I do not think that's the interpretation of compatible. Its goes
from most specific to most generic per the ePAPR spec. That in
itself says that 100% functionality is not expected if you don't
find a match for the more specific property.
Well, to be honest I checked as well the official documentation,
and this is far from being clear:
page 21 of the ePAPR:
" Property: compatible Value type: <stringlist>
Description: The compatible property value consists of one or more
strings that define the specific programming model for the device.
This list of strings should be used by a client program for device
driver selection. The property value consists of a concatenated
list of null terminated strings, from most specific to most
general. They allow a device to express its compatibility with a
family of similar devices, potentially allowing a single device
driver to match against several devices.
The recommended format is “manufacturer,model”, where manufacturer
is a string describing the name of the manufacturer (such as a
stock ticker symbol), and model specifies the model number.
Example: compatible = “fsl,mpc8641-uart”, “ns16550"; In this
example, an operating system would first try to locate a device
driver that supported fsl,mpc8641-uart. If a driver was not found,
it would then try to locate a driver that supported the more
general ns16550 device type. "
In this example, the generic vs specific is just about the driver.
You could have one driver that manage 10 different UARTs or a
specific one that manage only your HW.
Right, the assumption is that a specific driver is written because
there are parts of the hardware that cannot be serviced by generic
driver. So ultimately its all driven by changes in hardware.
Yeah, that example remind me the omap-serial driver we had done to
handle the DMA mode that was not supported by the generic one.
There is no assumption about the lost of functionality by using the
generic version of the driver. How the user is supposed to know the
amount of functionality he will lose, and if this is acceptable to
him.
I suppose the generic driver would return some error code like
-ENOTSUP for the functionality it cannot provide.
Well, I guess it depends of the added functionality add how far the IP
version is forward compatible with the old driver. If the new IP version
is just improving the performance by adding a DMA mode instead of the
PIO, then the driver API can remain the same.
But if you add a new functionality that will require an extended API,
there is no way you can report such error. Except if the API itself is
done with a good forward compatibility support.
And if the new functionality is mandatory to make the new system
operational, returning an error might just make the system not working
at all.
I'd rather have an error at boot saying that the driver is not
compatible with the HW and thus I will have to upgrade the kernel,
than booting a kernel that will not work as expected because some
functionality are missing for that specific HW.
If you have an update available, you can always upgrade to it. But
what if there is no update available? Would you rather not have the
user use the available functionality in the meanwhile while he waits
for the kernel support to be developed.
It depends... In the UART example, that's just a matter of performance
improvement, so keeping the old version is fine.
As soon as the driver is usable, and the new feature is not mandatory to
ensure the system will work properly, I guess it is fine.
Claiming that a piece of HW is compatible with an other one that is
just a subset in term of functionality is wrong. You can claim that
the ti,da830-rtc is compatible with the ti,am3352-rtc driver, but
not the opposite.
If this is wrong, then what is the use of compatible property?
That's a good question :-) And that's why I was always confused with
that property.
Why
would someone write a driver supporting “fsl,mpc8641-uart” if 100% of
its hardware features are also supported by “ns16550" driver?
That's still doable, if you want to reduce the size of a big generic
driver into a smaller specific driver. The point is that the compatible
value does not have any assumption about the driver that will handle it.
The other issue is that we are supposed to always use the latest SoC
version even if the IP is 100% the same. Like
omap5-timer {
compatible = "omap5-timer", "omap4-timer";
}
So how are you suppose to differentiate the same IP, and a compatible IP???
That's what I don't like in that compatible string definition. You do
not necessarily know the amount of compatibility you are talking about.
am3352 DTS must use the ti,am3352-rtc to have the expected
behavior.
Yes, that's what patch 2/2 does.
Using the ti,da830-rtc version will not make the board working
as expected. So we cannot claim the compatibility.
Ideally, the DT file was *always* written as
compatible = "ti,am3352-rtc", "ti,da830-rtc";
even when there was no kernel support for AM3352 RTC.
You could do that only for the davinci DTS, because it is a subset
of the am3352 but you cannot do that for the am3352 as explained
before.
DaVinci DTS will use compatible = "ti,da830-rtc"; Thats not changing
with this patchset.
That way, there is no need to update the .dts[i] file. As kernel
gains functionality, more features (like rtc wake) is available
to users.
Well, you cannot anticipate the HW evolution anyway and you did not
know when Davinci DTS was done that an am3352 will exist in the
future and reuse the same IP.
Yeah, I am not claiming DaVinci DTS should be updated. But when
am335x.dtsi was written, there was knowledge that a (potentially)
new version of RTC IP is available and therefore the .dtsi should
have been written as
compatible = "ti,am3352-rtc", "ti,da830-rtc";
instead of
compatible = "ti,da830-rtc";
that it is today. Thing is, you can never know if the IP needs any
additional handling even after reading the spec. We keep discovering
new features/quirks. So, when writing <new-soc>.dtsi its safer to
always use
compatible = "ti,<new-soc>-rtc", "ti,<compatible-soc>-rtc";
even though _today_ you may not have code that specifically handles
"ti,<new-soc>-rtc".
Well, we can do that, but honestly, I do not see the need. You can
always update the dts later. Why would we add hundreds more compatible
strings just in case few devices will need special handling. That's
overkill.
If was maybe easy and harmless in the good old PPC time when the DTS
files were relatively small, but the ARM DTS files are much bigger.
In the name of the Keep It Simple Principle, I would just avoid adding
something just because it might be useful in 5 % of the cases.
Moreover DT is a ABI, so extending it according to the HW feature
evolution is absolutely normal.
You mean the DT bindings and not specifically the .dts[i], right?
Yes, you're right, sorry for the confusing sentence.
Then I agree. Ideally the .dts[i] is written once per hardware. I
know its almost impossible to get that right.
Every SoC that are using a RTC with the same programing model than
the da830 can claim the compatibility. A SoC that is using a newer
version of the IP with an extended programming model cannot claim
that.
We differ here. As long as programming model is _extended_, new IP
is still compatible with old. If the programming model is _changed_
then its not.
Yes, I was assuming compatible as a 100% match which is not really
compatible but identical in that case. But the current syntax, does not
allow you anyway to differentiate a new SoC with the exact same IP
version to a different for compatible IP.
Otherwise they get plain RTC functionality - but at least they
get something instead of no RTC.
Because you assume that this feature is not important and thus you
can use the plain RTC, but what if someone is buying that HW
because of this new feature. Without that feature his system will
just not work properly.
Right, but thats not a problem DT can solve. He/she needs to hassle
TI for updated drivers. But there could be 10 other customers who are
just okay with plain RTC. So till the time driver is updated to use
ti,am3352-rtc", are you saying no one should be able to use the RTC
on the SoC at all even though 90% features are available?
Again, if it will not prevent the system to work properly, then it is
fine. But let's assume that without the wakeup RTC functionality, you
just cannot wakeup from suspend an am3352 board. Then you end up with a
non functional system for the PM point of view.
Someone who is not aware that the compatible RTC is not supporting that
feature might spend some time figuring out why he cannot wakeup from
suspend on a RTC alarm.
In general, I'd rather have nothing than something that is working at 50%.
But maybe that's just me :-)
Saying that this is compatible whereas you lost functionality is
lying to the customer for my point of view.
If 100% functionality is required for compatibility then I am afraid
there is nothing like "compatibility". There are just different
isolated versions.
I guess you are right.
Bottom-line, I'm really disappointed but that lack of accuracy in the
compatible string, but I guess, it was done for what you guys are doing.
And maybe, it is something that should just be well documented in the
bindings to avoid confusing the users.
Regards,
Benoit
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
