patcht0021: make sure clean filter runs

In t0021.15 one of the things we are checking is that the clean filter is run when checking out empty-branch. The clean filter needs to be run to make sure there are no modifications on the file system for the test.r file, and thus it isn't dangerous to overwrite it.

However in the current test setup it is not always necessary to run the clean filter, and thus the test sometimes fails, as debug.log isn't written.

This happens when test.r has an older mtime than the index itself. That mtime is also recorded as stat data for test.r in the index, and based on the heuristic we're using for index entries, git correctly assumes this file is up-to-date.

Usually this test succeeds because the mtime of test.r is the same as the mtime of the index. In this case test.r is racily clean, so git actually checks the contents, for which the clean filter is run.

Fix the test by updating the mtime of test.r, so git is forced to check the contents of the file, and the clean filter is run as the test expects.

Signed-off-by: Thomas Gummerer <t.gummerer@gmail.com>
---
 t/t0021-conversion.sh | 1 +
 1 file changed, 1 insertion(+)

Thomas Gummerer <t.gummerer@gmail.com> writes:

Hmph, depending on the timestamp granularity, with this patch, test.r would have mtime that is the same or a bit later than that of the index file. Is it sufficient to really "force" Git to check the contents, or does it just make the likelyhood that it would choose to check a bit bigger (in other words, are we solving the race, or merely making the race window smaller)?

Thanks.

On 08/20, Junio C Hamano wrote:

This test only worked until now because git checks the contents if the mtime of the file and the index are the same. This is because of racy-git. I tried to describe this in the commit message, but looks like it wasn't clear enough. Do you have any suggestions on how to make it clearer?

It will also check the contents if the mtime is greater than the timestamp of the index, so the 'touch' here would also cover that.

So the changes here do solve the race completely.

Thomas Gummerer <t.gummerer@gmail.com> writes:

OK, the explanation makes sense.

Either test.r has been correctly checked out and has an older timestamp or a more recent timestamp. In the former case, the index knows that we did not touch it, so the next "checkout" knows it does not have to ask the clean filter to work on it. In the latter case, the index is unsure if we touched it (or, suspects that it has updated contents in it), so the clean filter needs to read from the working tree to see if we did change it (and we find it is not modified). The outcome at the higher level, the answer to the question "checkout" wanted to ask, is the same: test.r has no local modificaiton and we can switch branches safely.

And that is already validated by seeing what exit status "checkout" gives us, so it sort-of feels to be testing a bit too low level implementation detail to see on which paths the filters are or are not called, but that is not a problem with this fix. If we want to check at that level, we should do so correctly, and making sure that the test.r file has recent timestamp to convince "checkout" that it needs to verify contents is the right thing to do.

Thanks.

Am 20.08.19 um 08:56 schrieb Thomas Gummerer:

		test-tool chmtime +10 test.r

would be more reliable.

-- Hannes

On 08/20, Johannes Sixt wrote:

Hmm, is touch unreliable on some platforms? I didn't think of 'test-tool chmtime', but I'm also not sure it's better than touch in this case.

To me te 'touch' signifies that the timestamp must be updated after the previous checkout, so git thinks it could possibly have been changed, which I think is clearer in this case than setting the mtime to a future time.

But I'm happy to change it if there's something I'm missing why 'test-tool chmtime' is better in this case.

Am 21.08.19 um 16:56 schrieb Thomas Gummerer:

touch does not guarantee that the current time is different from the timestamp that the file already carries, particularly not when the filesystem stores just a resolution of 1 second, and commands are executed quickly.

But when we use test-tool chmtime +10, then the timestamp is definitely different. If you don't like a timestamp in the future, use -10, or anything else that is different from zero.

-- Hannes

On Wed, Aug 21, 2019 at 08:23:23PM +0200, Johannes Sixt wrote:

This 'touch' must ensure that the timestamp of the file is not older than the timestamp of the index, and to achive that it doesn't necessarily have to modify the timestamp.

The file is modified first, then the index is updated, and finally comes this 'touch'. Consequently, if 'touch' doesn't modify the timestamp of the file, then it must have the same timestamp as the index, IOW it's racily clean, and the subsequent 'git checkout' has to look at the file content and has to run the filter, and that's what we want to see here.

However, I'm not sure what would happens if the system clock were to jump back in between, but since it's only a test I don't think it's worth caring about.

'test-tool chmtime +10' adjusts the timestamp of the file relative to its current timestamp. So yeah, the file's timestamp definitely changes, but that's not enough, because it doesn't ensure that the new timestamp is not older than the timestamp of the index. Just imagine the arguably pathological situation that right after the file was last modified the system miraculously comes to a complete stall, and only manages to resume after 15 seconds to continue with updating the index. This means that the timestamp of the file will be 15s older than the index, and after that 'chmtime +10' it will still be 5s older. Consequently, 'git checkout' will think that the file is clean, it won't run the filter that we expect, and the test will fail. So instead of '+10' it should be '=+10' to set the new timestamp relative to the current time, but I'm not too keen about the timestamp in the future either (though the file is about to be deleted anyway).

I think it would be best to explicitly set the timestamp of the file and the index sort-of relative to each other and add an in-code comment as well, e.g.:

  # Make sure that the file appears dirty, so checkout below has to
  # run the configured filter.
  test-tool chmtime =-10 .git/index &&
  test-tool chmtime =+0 test.r &&

On 08/22, SZEDER Gábor wrote:

Right.

Yeah, I think much of git wouldn't work correctly if that happens, so I think it's fairly safe to ignore in the test suite.

Right, the above is why I think 'touch' is a good idea here. Short of system clocks jumping around, which will most likely break more than this test anyway it guarantees that the timestamp is equal or greater than the timestamp of the index, which is what we need here.

I think the comment is a good idea. I personally still prefer just using 'touch' though, as I find it slightly easier to read (I had to go look up what the =-/=+ in 'test-tool chmtime' does, while I knew what touch would be doing :)

That said that's a minor preference for me, if people have a strong opinion that test-tool chmtime is really better here I'm fine with changing it.

Thomas Gummerer <t.gummerer@gmail.com> writes:

Yup, I do not quite get why people feel 'touch' is a bad idea here.

I also think we should discourage "test-tool chmtime =<anything>" (i.e. set to relative to the timestamp read from time(), as opposed to relative to the timestamp read from the filesystem), when we can avoid it, to allow tests on remote filesystem where the filesystem clock and the system clock may not always be in sync.

Am 22.08.19 um 19:49 schrieb Thomas Gummerer:

Ok, thanks for the clarification. I didn't see the context. It looks like touch is good enough.

-- Hannes

In t0021.15 one of the things we are checking is that the clean filter is run when checking out empty-branch. The clean filter needs to be run to make sure there are no modifications on the file system for the test.r file, and thus it isn't dangerous to overwrite it.

However in the current test setup it is not always necessary to run the clean filter, and thus the test sometimes fails, as debug.log isn't written.

This happens when test.r has an older mtime than the index itself. That mtime is also recorded as stat data for test.r in the index, and based on the heuristic we're using for index entries, git correctly assumes this file is up-to-date.

Usually this test succeeds because the mtime of test.r is the same as the mtime of the index. In this case test.r is racily clean, so git actually checks the contents, for which the clean filter is run.

Fix the test by updating the mtime of test.r, so git is forced to check the contents of the file, and the clean filter is run as the test expects.

Signed-off-by: Thomas Gummerer <t.gummerer@gmail.com>
---

v2 adds the comment as suggested by Szeder.

Junio: I saw this is marked as "merged to 'next'" in the What's
cooking, so if it got merged already I'm fine with just keeping v1,
but otherwise I think adding the comment would be nice.

 t/t0021-conversion.sh | 3 +++
 1 file changed, 3 insertions(+)

Thomas Gummerer <t.gummerer@gmail.com> writes:

I think it was marked with "Will merge to ...".  Replaced.

Thanks.

On Thu, Aug 22, 2019 at 08:22:40PM +0100, Thomas Gummerer wrote:

Yeah, but that comment only really applied when setting both timestamps. With a simple 'touch' it's more like

  # Make sure that the file appears dirty or is at least racily clean,
  # so ...

So the next reader will know that right away, that the author didn't overlook racyness issue.