Re: [PATCH v1] docs: describe how to quickly build Linux

[Thorsten Leemhuis]

On 03.02.23 10:44, Jani Nikula wrote:
> On Thu, 02 Feb 2023, Thorsten Leemhuis <linux@xxxxxxxxxxxxx> wrote:
>> On 02.02.23 16:08, Konstantin Ryabitsev wrote:
>>> On Thu, Feb 02, 2023 at 12:15:36PM +0100, Linux kernel regression tracking (Thorsten Leemhuis) wrote:
>>>> Then I tried creating a shallow clone like this:
>>>>
>>>> git clone
>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
>>>> --depth 1 -b v6.1
>>>> git remote set-branches --add origin master
>>>> git fetch --all --shallow-exclude=v6.1
>>>> git remote add -t linux-6.1.y linux-stable
>>>> https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
>>>> git fetch --all --shallow-exclude=v6.1
>>>>
>>>> This took only roundabout 2 minutes and downloads & stores ~512 MByte
>>>> data (without checkout).
>>>
>>> Can we also include the option of just downloading the tarball, if it's a
>>> released version? That's the fastest and most lightweight option 100% of the
>>> time. :)
>>
>> Don't worry, that was in there and will stay in there:
>>
>> +   If you plan to only build one particular kernel version, download
>> its source
>> +   archive from https://kernel.org; afterwards extract its content to
>> '~/linux/'
>> +   and change into the directory created during extraction.
> 
> The trouble is, if this is for someone who needs to try kernels for
> debugging, a typical idea is to ask them to revert something or apply a
> patch. All the guides for that will be 'git revert' and 'git am'. Bisect
> is right up there on the list too. And then they'll first grab a tarball
> and fail,

Yeah, those are the reasons why I don't like the tarball approach too
much myself. Guess I should point them out in the text to make readers
aware of them...

> then do a shallow copy and fail,

The new test I wrote (still a draft) will suggest to use a recent
release as base, hence bisection or reverting a patch will be possible.
And if the range turns out to be to shallow, there is still "git fetch
--shallow-exclude=v6.1" to deepen it, which should avoid...

> and then finally get a full one... :p

...this scenario -- at least unless I missed anything.

Ciao, Thorsten

>>>> Not totally sure, but the shallow clone somehow feels more appropriate
>>>> for the use case (reminder, there is a "quickly" in the document title),
>>>> even if such a clone is less flexible (e.g. users have to manually add
>>>> stable branches they are interested it; and they need to be careful when
>>>> using git fetch).
>>>>
>>>> That's why I now strongly consider using the shallow clone method by
>>>> default in v2 of this text. Or does that also create a lot of load on
>>>> the servers? Or are there other strong reason why using a shallow clone
>>>> might be a bad idea for this use case?
>>>
>>> As I mentioned elsewhere, this is only a problem when it's done in batch mode
>>> by CI systems. A full clone uses pregenerated pack files and is very cheap,
>>> because it's effectively a sendfile operation. A shallow clone requires
>>> generating a brand new pack, compressing it, and then keeping it around in
>>> memory for the duration of the clone process. Not a big deal when a few humans
>>> here and there do it, but when 50 CI nodes do it all at once, it effectively
>>> becomes a DDoS. :)
>>
>> Thx again for your insights, much appreciated.
>>
>> Ciao, Thorsten
>