Re: Keep bpf-next always open

[Song Liu <songliubraving@fb.com>]

> On Sep 23, 2020, at 3:28 PM, Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
> 
> On Wed, Sep 23, 2020 at 10:23:51PM +0000, Song Liu wrote:
>> 
>> 
>>> On Sep 23, 2020, at 3:14 PM, Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
>>> 
>>> On Wed, Sep 23, 2020 at 02:48:24PM -0700, Andrii Nakryiko wrote:
>>>> On Wed, Sep 23, 2020 at 2:20 PM Alexei Starovoitov
>>>> <alexei.starovoitov@gmail.com> wrote:
>>>>> 
>>>>> BPF developers,
>>>>> 
>>>>> The merge window is 1.5 weeks away or 2.5 weeks if rc8 happens. In the past we
>>>>> observed a rush of patches to get in before bpf-next closes for the duration of
>>>>> the merge window. Then there is a flood of patches right after bpf-next
>>>>> reopens. Both periods create unnecessary tension for developers and maintainers.
>>>>> In order to mitigate these issues we're planning to keep bpf-next open
>>>>> during upcoming merge window and if this experiment works out we will keep
>>>>> doing it in the future. The problem that bpf-next cannot be fully open, since
>>>>> during the merge window lots of trees get pulled by Linus with inevitable bugs
>>>>> and conflicts. The merge window is the time to fix bugs that got exposed
>>>>> because of merges and because more people test torvalds/linux.git than
>>>>> bpf/bpf-next.git.
>>>>> 
>>>>> Hence starting roughly one week before the merge window few risky patches will
>>>>> be applied to the 'next' branch in the bpf-next tree instead of
>>>> 
>>>> Riskiness would be up to maintainers to determine or should we mark
>>>> patches with a different tag (bpf-next-next?) explicitly?
>>> 
>>> "Risky" in a sense of needing a revert. The bpf tree and two plus -rc1 to -rc7
>>> weeks should be enough to address any issues except the most fundamental ones.
>>> Something like the recent bpf_tail_call support in subprograms I would consider
>>> for the "next" branch if it was posted a day before the merge window.
>>> In practice, I suspect, such cases will be rare.
>>> 
>>> I think bpf-next-next tag should not be used. All features are for [bpf-next].
>>> Fixes are for [bpf]. The bpf-next/next is a temporary parking place for patches
>>> while the merge window is ongoing.
>> 
>> I wonder whether we can move/rename the branch around so that the developers can 
>> always base their work on bpf-next/master. Something like:
>> 
>> Long before merge window for 5.15:	
>> We only have bpf-next/master
>> 
>> 1 week before merge window for 5.15:	
>> Clone bpf-next/master as bpf-next/for-5.15
>> 
>> From -1 week to the end of merge window
>> Risky features only goes to bpf-next/master, bug fix goes in both master and for-5.15
>> 
>> After merge window:
>> Fast forward bpf-next/master based on net-next. Deprecate for-5.15.
>> 
>> Would this work? 
> 
> It will create headaches for linux-next that merges bpf-next/master.
> All linux-next trees should not add patches to those trees that are not going
> into the merge window.

I see. Keeping bpf-next/master for linux-next/master does make sense. 

How about we keep bpf-next/next always open, or maybe rename it as bpf-next/dev?
Developers could always base their work on bpf-next/dev. When the maintainer 
applies the patch, he can decide whether to apply it to both master and dev, or
just dev. 

Thanks,
Song