[RFC] Moving Intrepid to 2.6.27 kernel

[Ben Collins]

Matt Zimmerman wrote:
> On Wed, Aug 20, 2008 at 04:02:47PM -0400, Ben Collins wrote:
>> As many of you know, we've continued a tree called ubuntu-next which is
>> currently following 2.6.27 upstream kernel. It has been kept in sync
>> with intrepid's 2.6.26 based kernel tree.
> 
> I have been in all-day meetings this past week and did not see this thread
> until now.  I think this change requires stronger justification than I've
> seen so far.
> 
>> In recent weeks it has become apparent that we will need several updates
>> to 2.6.26 in order to have a stable kernel for 8.10. These include:
>>
>>  * Updated mac80211, which will require updated wlan drivers in order 
>>    to fix major regressions with suspend/resume and wlan devices.
>>  * Updated alsa-1.0.17 drivers for new codecs
>>  * xen64 paravirt
>>  * Updated KVM
>>
>> These are just the major things we need to do.
> 
> Why are these needed in order to have a stable kernel for 8.10?  Major
> updates are the opposite of stability.  What we want is a kernel which fixes
> more bugs than it introduces while introducing the most important driver
> updates we need for the release.
> 
>> The good thing is, all of this is already in 2.6.27, and is ready to be
>> uploaded on a moments notice. It has been tested by myself for quite some
>> time, and others on the team have also tested it.
> 
> The amount of testing it has received is trivial compared to what 2.6.26 has
> seen.  It's been tested on our certified systems and by many thousands of
> community members on their own hardware.  The fact that it's been smoke
> tested by a few people doesn't give me much confidence that it's free of
> serious regressions.
> 
>> Yesterday, we discussed the possibility of moving to 2.6.27 before
>> feature freeze on #ubuntu-kernel during out weekly IRC meeting. In
>> attendance were the kernel team, Steve Langasek, Chuck Short and Soren
>> Hansen. Soren and Chuck gave particularly good reasons for moving to
>> 2.6.27, which included xen64 and updated KVM.
> 
> Where can I find the summary and log of the meeting?

I can put one together.

>> Steve had questions about the particulars of this move, and I believe we
>> answered all of his questions satisfyingly, but we agreed to take this
>> to a wider audience before making a commitment.
>>
>> So fire away with any concerns or issues that this might cause. If we
>> are going to do it, it needs to be decided by early next week.
> 
> I was in meetings all week and didn't see this until now.  Mario has raised
> a concern in this thread which I think needs to be addressed.  I also have
> the following questions:

I've tested LRM, and it compiles fine. I've tested at least broadcom's 
driver.

> What are the known regressions from 2.6.26 to 2.6.27?

I've looked at the list of known regressions from upstream, and the 
critical/important ones have been addressed or don't affect us.

> Which systems has 2.6.27 been tested on?  What test plan was followed?

I've tested 32-bit and 64-bit on Xeon, 4+ laptops and a few of Intel's 
SDP's. Tim has done further testing. I know the server team has been 
checking the 2.6.27 kernels as well. In addition, I've had feedback from 
more than a dozen others that have been using the 2.6.27 kernels on 
various platforms because of fixes major problems for them over 2.6.26.

> In which areas are further regressions to be expected in 2.6.27 (which parts
> of the kernel have changed in the most significant ways)?

Alsa and wireless are the major areas. In 2.6.26 there were major 
updates to RTC, which seem to have settled now in 2.6.27. Alsa and 
wireless are major improvements over 2.6.26.

The problem with 2.6.26 is that wireless updates there are causing major 
regressions in suspend/resume, which as far as I'm concerned is a 
showstopper in itself. 2.6.27 fixes these issues.

Also included in the wireless update is ath5k/ath9k updates, which could 
make madwifi unneeded, stomping out one lrm driver. There's also a new 
iwlagn driver and iwl5k driver, bringing better support for newer Intel 
wireless cards, including 11n support for 4965 (these new drivers 
require the new wireless stack, so we can't just grab the drivers).

Updating to Alsa 1.0.17 is a must-have for intrepid due to various 
driver updates. If we release with 1.0.16 (which is in 2.6.26) we wont 
be supporting a lot of new hardware.

> What were the results of certification testing with 2.6.27?

No results that I know of. We have no way to do one-off certification 
testing that I'm aware of.

We aren't getting rid of the 2.6.26 tree, so worse case, we revert back 
to it. But I have strong indications right now that we wont need to 
consider that.

Our main reason to do this is that 2.6.26 as-is isn't good enough to 
release intrepid. Updating the various stacks to improve the state will 
require more effort on our part than updating the whole kernel 
(especially in the long term for security and such).