Re: [lm-sensors] Hardware monitoring subsystem maintainer position is open

[Hans de Goede <j.w.r.degoede@hhs.nl>]

Rudolf Marek wrote:
> Hello all,
>
>
> Maybe we could try following:
>
> 1) person post the driver
> 2) quick review could be done critical fixes only, driver goes to -mm
> 3) review that goes deeper - check for interface conformity and all the 
> stuff which could break - fixes for non-critical stuff
> 4) after this driver goes to Linus tree at the very beginning for the 
> merge cycle marked as EXPERIMENTAL
> 5) if the person agrees to be in MAINTAINER final review may be done and 
> EXPERIMENTAL could be removed. (This is step which might involve all 
> steps to make the driver really perfect ;)
>
> Well this is just an idea. Neither it does solve the maintainer problem, 
> nor it is perfect solution.

I like the general direction, but I don't really see a clear difference between 
step 2 and 3, IMO the border between these steps is vague. Also this doesn't 
address who can do (which kinda) reviews.

I've been thinking about a solution for the slow path for new driver inclusion 
myself here is what I propose:

1a) We need a set of review guidelines / a review checklist. Here is a start:
* Must follow kernel coding style guidelines
* sysfs interface must be in accordance with Documentation/hwmon/sysfs
* proper locking to avoid 2 simultaneous attempts to communicate with the
  device when for example multiple sysfs files get read at once.
* check the code for any obvious programming errors, like:
  -not freeing resources (in error paths and in general)
  -overrunning an array
  -not checking return values of calls to other parts of the kernel
  -of by one errors / bad loops
  -etc.
1b) We need to decide somehow who can do reviews. For now I say anyone who
  already maintains atleast one hwmon driver. But thats just a wild shot better
  ideas are welcome.

2) We need a review process for new drivers, suggestion:

1.  Person (the submitter) posts driver
2a. Someone (the reviewer) reviews it according to our checklist / guidelines
    and posts a detailed report of his findings and what needs fixing
2b. The submitter posts a fixed version
2c. The reviewer re-reviews the fixed version and either approves the fixed
    version, goto 3 or points out things that (still) need fixing, goto 2b.
3.  The driver gets send to the -mm tree (marked as EXPERIMENTAL)
4.  Another person (the second reviewer) reviews it, same process as with step
    2, once approved goto 5
5.  Send to Linus tree at the very beginning for the merge cycle still marked
    as EXPERIMENTAL
6.  After quite some time and positive usage reports without any negative
    counter reports a patch gets send to Linux to remove EXPERIMENTAL status.

3) We need a review process for fixes to existing drivers, suggestion:

1. Person (the submitter) posts a patch to an existing driver
2. The driver maintainer is responsible for handling this patch and reviews
   it. If necessary he either modifies the patch himself or asks the submitter
   to make modifications.
3. Once the maintainer is happy with the patch, he sends it to Linus tree at
   the beginning for the merge cycle.


The rationale of doing the new driver thing this way is to avoid having a 
single person who bears end responsibilities for all reviews and thus becomes a 
potential bottleneck. The problem with distributing reviews though is that most 
of us aren't as experienced as Jean is. Thus I decided to just do the review 
twice, to compensate (a bit) for this.

The rationale of doing modifications of existing drivers through the maintainer 
is that he knows the code best, and thus can best judge if fixes are really 
fixes or more bandaids / workarounds. And if these fixes have any unwanted side 
effects.

To give proper credit: the above is heavily inspired by how Fedora handles new 
packages, Fedora has thousands of packages written and maintained by 100's of 
contributers. Using a model where any contributer, once he has taken the first 
horde of becoming a contributer and thus proving (some) competence can review 
other contributers new packages.

Problems with the above, I'm not completely happy with my current proposal for 
deciding who can do reviews. Also I guess that Andrew and Linus would like to 
have a single person to take hwmon patches from, thus we still need someone 
todo the actual collecting of approved patches and sending them to Andrew 
and/or Linus.

Last but not least I want to say that we should not focus too much on review, 
review can only get us sofar. In my (abituguru) experience most real life hwmon 
problems do not come from things easily catched by a review, but rather from 
problems with (certain versions of) the hardware responding different then 
expected. No amount of review will catch these cases, thus we also need to 
concentrate on testing.

For both abituguru drivers I've posted to the forums of the abit websites and 
the lm-sensors list as soon as I got something usable and that way managed to 
build a small team of circa 8 testers for each version, a team which I send 
each update to before sending it for upstream inclusion. Especially for the 
abituguru (rev 1 and 2) driver this has helped my much since this is a strange 
beast. The abituguru3 driver is more straight forward but also there the 
testing by others has been invaluable.


> Personally better would be go with Jean and some kind of scheme noted 
> above, which would solve our "takes too long problem" and preserve the 
> high quality of drivers. This would mean some co-maintainers to Jean, 
> rather than to live without Jean at all.

+1

Regards,

Hans


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/