[Ksummit-discuss] [MAINTAINERS SUMMIT] Developing across multiple areas of the kernel

[Kees Cook]

On Thu, Jun 29, 2017 at 10:42 AM, James Bottomley
<James.Bottomley at hansenpartnership.com> wrote:
> On Thu, 2017-06-29 at 09:51 -0700, Kees Cook wrote:
>> On Thu, Jun 29, 2017 at 9:36 AM, James Bottomley
>> <James.Bottomley at hansenpartnership.com> wrote:
>> >
>> > On Wed, 2017-06-28 at 16:01 -0700, Kees Cook wrote:
>> > >
>> > > For refcount_t, the conversions have been going per-maintainer,
>> > > and while this is likely the right way to do things, there are
>> > > dependencies that are crossing releases, which seems inefficient.
>> > > For example, obviously doing a refcount_t conversion requires the
>> > > refcount_t implementation first (which landed in v4.11), but then
>> > > later conversions wanted an option for a light implementation
>> > > (expected for v4.13), but in both cases most maintainers wanted
>> > > the implementations entirely landed, not just in -next (vast
>> > > majority of refcount_t conversions currently in the kernel landed
>> > > in v4.12, so the next wave will have to wait until v4.14 it
>> > > seems). This appears mostly to be about avoiding tree
>> > > dependencies, IIUC, but is an awfully slow way to do things.
>> >
>> > Given the performance concerns of the first implementation, this
>> > timetable and the interactions that went with it seem to be pretty
>> > much textbook correct, especially as none of the hot paths seemed
>> > susceptible to overflow attacks.
>> >
>> > Any other way would have produced a lot more friction: imagine if
>> > it had been done tree at once for 4.12 and then performance had
>> > tanked and we'd got reversions all over the place ... you'd be
>> > spending a lot more than a couple of kernel releases trying to
>> > persuade maintainers to take the new improved stuff.
>>
>> Right, I've got no objection to the performance concerns and how that
>> played out, but it's API-to-conversion steps that seem inefficient.
>> E.g., instead of API 1 in v4.11, conversion wave 1 in v4.12, API 2 in
>> v4.13, conversion wave 2 in v4.14, it looks like tree dependencies
>> was the only reason we couldn't have had: API 1 and conversion wave 1
>> in v4.11, API 2 and conversion wave 2 in v4.12 (e.g. btrfs couldn't
>> compile their tree with the API living in tip, so they had to wait
>> until the API was in a release).
>
> Well don't discount tree merge problems, having seen a few caused by
> API plus conversion all at once.  However, by putting them through the
> maintainer trees you got the review that would otherwise have been
> missing which highlighted the performance concerns.  Even this time
> around the affected trees have a whole merge window to run performance
> regressions to verify everything is OK.  Based on this I think the rule
> should be API in release R - 1 and conversion in release R through the
> affected trees with the only exception being changes that are trivial
> enough (for some value of trivial).

I'd argue that's what -next is for, but we lack a way to have people
base their trees on -next sanely. Regardless, I will commence
beer-crying. :)

-Kees

-- 
Kees Cook
Pixel Security