[PATCH 2/2] sunxi: a20-LIME2 define board specific operating points

Iain Paton ipaton0 at gmail.com
Sat Apr 4 01:19:37 PDT 2015


On 31/03/15 00:17, Maxime Ripard wrote:
> On Thu, Mar 26, 2015 at 05:08:29PM +0000, Iain Paton wrote:
>> these settings override the inappropriate ones in the dtsi with the
>> manufacturers provided settings from their 3.4.x fex file for this board
> 
> Inappropriate how?

inappropriate in that the dtsi sets values that cause the board to lockup 
or to be unstable

> I'd rather have a better common operating points set and change that
> to accomodate all boards, rather than duplicating that information
> everywhere.

then anything that you have common needs to cope with the worst cases 
for *all* boards, to the detriment of anything that's capable of doing
better. i.e. the smallest most conservative subset.

Whether we like it or not, stuff like this is simply going to end up 
being board specific. You can't easily deal with cases like a particular 
PCB design requiring an extra 0.05v due to a too-thin trace somewhere 
without having board-specific knowledge.

So rather than us, as software people, trying to second guess the board 
designers, I believe we should simply use the settings they provide if 
we have them in the fex files. Likely the manufacturer has tested those 
settings on a much larger number of boards than we'll ever be able to do.

The default settings for a board need to be reliable for everyone.

>> +	#cooling-cells = <2>;
> 
> This is also defined in the DTSI.

Yes, exactly. The point being to completely isolate the board from changes 
to this section of the dtsi.

To further illustrate, let's take a look at the sun7i opp table in your 
sunxi/for-next tree

/* kHz    uV */
960000  1400000
912000  1400000
864000  1300000
720000  1200000
528000  1100000
312000  1000000
144000  900000

Comparing to the 38 A20 fex files in https://github.com/linux-sunxi/sunxi-boards 
as of today

boot_freq: 912=37, 1008=1
max_freq: 1200=1, 1008=10, 912=27
min_freq: 720MHz=7, 408MHz=3, 60MHz=28

The dvfs section doesn't define 960MHz in any of the 38 files, so that entry 
is pure conjecture and likely not properly tested by any of the manufacturers.

On the voltage front:
912MHz: 1425=7, 1400=27, 1350=4
864MHz: 1350=7, 1300=27, 1250=4
720MHz: 1300=1, 1250=7,  1200=30
528MHz: 1150=8, 1100=30
312MHz: 1100=8, 1050=27, 1000=3

144MHz is only defined for 12 out of 38 boards
144MHz: 1050=9, 1000=3

There are additional settings at 792MHz and 624MHz in the fex files, both
defined by 26 out of 38 boards.

1008MHz @ 1450 is defined by all 38.

To come to a generic subset, we need to exclude things as follows
1. anything greater than the lowest max_freq
2. anything less than the highest min_freq
3. anything less than the highest required voltage at a given frequency

That leaves us with this:

/* kHz    uV */
912000  1425000
864000  1350000
720000  1300000

Even then, we can't guarantee those three will be safe or reliable for
everything. The A20 datasheet gives 1.4v as the absolute maximum which 
means we need to remove the 912MHz completely as some devices require 
overvolting to reach it and that may be unsafe for the rest.

So a generic dtsi subset is limited to

/* kHz    uV */
864000  1350000
720000  1300000

as that's the subset that covers all devices we have data on and doesn't
involve overclocking, overvolting, undervolting, operating outside AllWinner
SoC ratings, outside device manufacturer specified settings, and doesn't
require any board specific knowledge.

So in the case of a generic, not board-specific, dtsi file most of what's 
currently in the dtsi is inappropriate for a large proportion of devices.
The commonality is tiny, without board specific knowledge we can't do any 
better.

While we can be relatively confident the remaining two settings will work on 
almost all boards, confidence in the other settings is at best uncertain.
I certainly have little interest in 7.9% at 312MHz, or 0% at 144MHz
possibility of success. 
While the faster speeds aren't quite so bad, you're still suggesting that 
18-21% of boards having problems is ok. 
Sorry, thats simply not acceptable.


I hope this explains my point of view and why I believe this data to be 
board specific stuff that belongs in the boards dts file and not in the dtsi.

Rgds,
Iain






More information about the linux-arm-kernel mailing list