2022-03-23 02:27:26

by Arnaldo Carvalho de Melo

[permalink] [raw]
Subject: Re: perf version issues

Re-adding lkml and Ian, Jiri, as I found some strange behaviour with
'git cherry-pick'.

Em Mon, Mar 21, 2022 at 08:43:26AM +0000, John Garry escreveu:
> Hi Arnaldo,
>
> I have been finding that during my development that the output from "perf
> -v" may be incorrect in certain circumstances. I sent some patches to fix it
> up:
> https://lore.kernel.org/linux-perf-users/[email protected]/T/#mf7234d2e331a53370c148eb4a0633d405a5c6832
>
> Is there some issue that these cannot be processed?

So, while testing it I noticed that if we use 'git cherry-pick' it
doesn't trigger a rebuild, even having a new head :-\

I.e. with the two patches in at the HEAD, try:

$ git log --oneline -2
cb66befccba18fac (HEAD -> perf/core) perf tools: Fix version kernel tag
4e666cdb06eede20 perf tools: Fix dependency for version file creation
$ make -C tools/perf O=/tmp/build/perf install-bin
$ perf -v
perf version 5.17.rc8.gcb66befccba1

Ok so far, now lets remove HEAD and leave just the first patch, that has
the ORIG_HEAD dependency to trigger a rebuild:

$ git reset --hard HEAD~
HEAD is now at 4e666cdb06eede20 perf tools: Fix dependency for version file creation
$ make -C tools/perf O=/tmp/build/perf install-bin
$ perf -v
perf version 5.17.rc8.g4e666cdb06ee

Ok, detected and rebuilt, perf -v works as expected.

Now lets try cherry picking your second patch:

$ git cherry-pick cb66befccba18fac
[perf/core 8ff6a6c06a90a362] perf tools: Fix version kernel tag
Author: John Garry <[email protected]>
Date: Mon Feb 21 21:16:49 2022 +0800
1 file changed, 4 insertions(+), 9 deletions(-)
$ make -C tools/perf O=/tmp/build/perf install-bin
$ perf -v
perf version 5.17.rc8.g4e666cdb06ee

Now it doesn?t notice it and there is no automatic rebuild triggered, we
stay with the cset from before the cherry-pick :-\

$ git log --oneline -2
8ff6a6c06a90a362 (HEAD -> perf/core) perf tools: Fix version kernel tag
4e666cdb06eede20 perf tools: Fix dependency for version file creation
$

Anyway, your patch works for some cases, so its an improvement and I'll
apply it, we can continue from there.

- Arnaldo


2022-03-23 23:33:52

by John Garry

[permalink] [raw]
Subject: Re: perf version issues

On 22/03/2022 20:11, Arnaldo Carvalho de Melo wrote:
>> I have been finding that during my development that the output from "perf
>> -v" may be incorrect in certain circumstances. I sent some patches to fix it
>> up:
>> https://lore.kernel.org/linux-perf-users/[email protected]/T/#mf7234d2e331a53370c148eb4a0633d405a5c6832
>>
>> Is there some issue that these cannot be processed?
> So, while testing it I noticed that if we use 'git cherry-pick' it
> doesn't trigger a rebuild, even having a new head:-\
>
> I.e. with the two patches in at the HEAD, try:
>
> $ git log --oneline -2
> cb66befccba18fac (HEAD -> perf/core) perf tools: Fix version kernel tag
> 4e666cdb06eede20 perf tools: Fix dependency for version file creation
> $ make -C tools/perf O=/tmp/build/perf install-bin
> $ perf -v
> perf version 5.17.rc8.gcb66befccba1
>
> Ok so far, now lets remove HEAD and leave just the first patch, that has
> the ORIG_HEAD dependency to trigger a rebuild:
>
> $ git reset --hard HEAD~
> HEAD is now at 4e666cdb06eede20 perf tools: Fix dependency for version file creation
> $ make -C tools/perf O=/tmp/build/perf install-bin
> $ perf -v
> perf version 5.17.rc8.g4e666cdb06ee
>
> Ok, detected and rebuilt, perf -v works as expected.

ok, good

>
> Now lets try cherry picking your second patch:
>
> $ git cherry-pick cb66befccba18fac
> [perf/core 8ff6a6c06a90a362] perf tools: Fix version kernel tag
> Author: John Garry<[email protected]>
> Date: Mon Feb 21 21:16:49 2022 +0800
> 1 file changed, 4 insertions(+), 9 deletions(-)
> $ make -C tools/perf O=/tmp/build/perf install-bin
> $ perf -v
> perf version 5.17.rc8.g4e666cdb06ee
>
> Now it doesn´t notice it and there is no automatic rebuild triggered, we
> stay with the cset from before the cherry-pick:-\
>
> $ git log --oneline -2
> 8ff6a6c06a90a362 (HEAD -> perf/core) perf tools: Fix version kernel tag
> 4e666cdb06eede20 perf tools: Fix dependency for version file creation
> $
>
> Anyway, your patch works for some cases, so its an improvement and I'll
> apply it, we can continue from there.

Thanks,

I just noticed the "reset --hard HEAD^" problem and not cherry-pick'ing.
Let me check this tomorrow.

john

2022-03-24 02:32:01

by John Garry

[permalink] [raw]
Subject: Re: perf version issues

On 22/03/2022 20:11, Arnaldo Carvalho de Melo wrote:

Hi Arnaldo,

> $ git cherry-pick cb66befccba18fac
> [perf/core 8ff6a6c06a90a362] perf tools: Fix version kernel tag
> Author: John Garry<[email protected]>
> Date: Mon Feb 21 21:16:49 2022 +0800
> 1 file changed, 4 insertions(+), 9 deletions(-)
> $ make -C tools/perf O=/tmp/build/perf install-bin
> $ perf -v
> perf version 5.17.rc8.g4e666cdb06ee
>
> Now it doesn´t notice it and there is no automatic rebuild triggered, we
> stay with the cset from before the cherry-pick:-\
>
> $ git log --oneline -2
> 8ff6a6c06a90a362 (HEAD -> perf/core) perf tools: Fix version kernel tag
> 4e666cdb06eede20 perf tools: Fix dependency for version file creation
> $
>
> Anyway, your patch works for some cases, so its an improvement and I'll
> apply it, we can continue from there.

After some experimentation, I find that only .git/HEAD changes for a
cherry-pick depending on whether we're on a branch or not. As such, as
you have seen, we may not rebuild after a cherry-pick, which is no good.

As far as I can see, only .git/index changes in all circumstances, but
we don't have a dependency on that.

TBH, using .git files as Makefile dependencies doesn't seem to work well.

Maybe we should output "git log" and check versus what we have in
PERF-VERSION-FILE and re-build depending on that.

Let me check this further now.

Thanks,
john

2022-03-25 15:26:24

by John Garry

[permalink] [raw]
Subject: Re: perf version issues

+ Ingo

On 23/03/2022 12:44, John Garry wrote:
>> $ git cherry-pick cb66befccba18fac
>>    [perf/core 8ff6a6c06a90a362] perf tools: Fix version kernel tag
>>     Author: John Garry<[email protected]>
>>     Date: Mon Feb 21 21:16:49 2022 +0800
>>     1 file changed, 4 insertions(+), 9 deletions(-)
>>    $ make -C tools/perf O=/tmp/build/perf install-bin
>>    $ perf -v
>>    perf version 5.17.rc8.g4e666cdb06ee
>>
>> Now it doesn´t notice it and there is no automatic rebuild triggered, we
>> stay with the cset from before the cherry-pick:-\
>>
>> $ git log --oneline -2
>> 8ff6a6c06a90a362 (HEAD -> perf/core) perf tools: Fix version kernel tag
>> 4e666cdb06eede20 perf tools: Fix dependency for version file creation
>> $
>>
>> Anyway, your patch works for some cases, so its an improvement and I'll
>> apply it, we can continue from there.
>
> After some experimentation, I find that only .git/HEAD changes for a
> cherry-pick depending on whether we're on a branch or not. As such, as
> you have seen, we may not rebuild after a cherry-pick, which is no good.
>
> As far as I can see, only .git/index changes in all circumstances, but
> we don't have a dependency on that.
>
> TBH, using .git files as Makefile dependencies doesn't seem to work well.
>
> Maybe we should output "git log" and check versus what we have in
> PERF-VERSION-FILE and re-build depending on that.
>
> Let me check this further now.

Just an update on this...

I did some research and found that Ingo introduced the change to make
.git/HEAD a dependency as a build optimisation. See commit commit
c72e3f04b45fb2e50cdd81a50c3778c6a57251d8 ("tools/perf/build: Speed up
git-version test on re-make")

According to the commit log it's quicker to trigger a rebuild of the
version string depending on a change in a git internal file rather than
run "git describe" to see any change in tag/commit every times.

However, I asked the git community about this approach, and their
general idea is that a dependency on .git/HEAD or other git internal
file is just not a good idea.

See
https://lore.kernel.org/git/[email protected]/T/#m4a4dd6de52fdbe21179306cd57b3761eb07f45f8

So to get this working robustly we may need to roll back that named
commit in part.

Thanks,
John