[RFC,v2,0/4] share a config between submodule and superproject

Message

Emily Shaffer April 23, 2021, 12:15 a.m. UTC

Hi all,

With the second iteration, I bothered to make the tests pass (oops) -
and I'm actually fairly unhappy with what I found. Patches 2 and 3 of
this iteration clean up tests which were (accidentally) explicitly
checking that no child processes were invoked besides the ones they
expected, by way of strictly grepping traces (by test_cmp or by line
counting). I don't like those tests as they were - to me, they felt like
the brittle kind of white-box test - but I also got a stronger feeling
that adding an additional child process or two during every Git
invocation is a bad idea. I also saw a pretty significant increase in
test run time:

All tests successful.
Files=927, Tests=24148, 693 wallclock secs ( 8.71 usr  2.05 sys +
3254.41 cusr 1571.78 csys = 4836.95 CPU)
Result: PASS

real    11m33.029s
user    54m23.187s
sys     26m13.857s

vs before:

All tests successful.
Files=926, Tests=24138, 144 wallclock secs ( 8.14 usr  2.03 sys + 882.29
cusr 535.61 csys = 1428.07 CPU)
Result: PASS

real    2m24.116s
user    14m50.499s
sys     8m57.649s

And that's on a Linux machine; as I understand it, invoking child
processes can be even more painful on other operating systems.

If we could be assured that this extra step (finding the parent's gitdir
and checking that config) was only running when we know we're in a
submodule, I'd be less worried, I think. And there are a couple other
pieces in the big picture submodule plan I sent[1] around which would
require repos to answer "am I a submodule?"

So I think this series may need to be shelved pending an answer to that
question - whether we *should* let submodules know they are
submodules[2] to turn on more behavior, and if so, how we should
implement that.

 - Emily

[1] https://lore.kernel.org/git/[e-mail]
[2] https://lore.kernel.org/git/xmqqk0pbm6qt.fsf@gitster.g

Emily Shaffer (4):
  config: rename "submodule" scope to "gitmodules"
  t1510-repo-setup: don't use exact matching on traces
  t7006-pager.sh: more lenient trace checking
  config: add 'config.superproject' file

 Documentation/git-config.txt   |  21 +++++-
 builtin/config.c               |   9 ++-
 config.c                       |  28 +++++++-
 config.h                       |   5 +-
 submodule-config.c             |   2 +-
 submodule.c                    |  29 +++++++++
 submodule.h                    |   8 +++
 t/t1311-superproject-config.sh | 116 +++++++++++++++++++++++++++++++++
 t/t1510-repo-setup.sh          |   2 +-
 t/t7006-pager.sh               |  24 +++++--
 10 files changed, 230 insertions(+), 14 deletions(-)
 create mode 100755 t/t1311-superproject-config.sh