tree aef13bf166389ee84c1dd6ffb08a5b81c7f4c395
parent 26b5d6b0e56898a0284d0f45dd5a69f59257a6ab
author Elijah Newren <newren@gmail.com> 1642175981 +0000
committer Junio C Hamano <gitster@pobox.com> 1642200202 -0800

repo_read_index: clear SKIP_WORKTREE bit from files present in worktree

The fix is short (~30 lines), but the description is not.  Sorry.

There is a set of problems caused by files in what I'll refer to as the
"present-despite-SKIP_WORKTREE" state.  This commit aims to not just fix
these problems, but remove the entire class as a possibility -- for
those using sparse checkouts.  But first, we need to understand the
problems this class presents.  A quick outline:

   * Problems
     * User facing issues
     * Problem space complexity
     * Maintenance and code correctness challenges
   * SKIP_WORKTREE expectations in Git
   * Suggested solution
   * Pros/Cons of suggested solution
   * Notes on testcase modifications

=== User facing issues ===

There are various ways for users to get files to be present in the
working copy despite having the SKIP_WORKTREE bit set for that file in
the index.  This may come from:
  * various git commands not really supporting the SKIP_WORKTREE bit[1,2]
  * users grabbing files from elsewhere and writing them to the worktree
    (perhaps even cached in their editor)
  * users attempting to "abort" a sparse-checkout operation with a
    not-so-early Ctrl+C (updating $GIT_DIR/info/sparse-checkout and the
    working tree is not atomic)[3].

Once users have present-despite-SKIP_WORKTREE files, any modifications
users make to these files will be ignored, possibly to users' confusion.

Further:
  * these files will degrade performance for the sparse-index case due
    to requiring the index to be expanded (see commit 55dfcf9591
    ("sparse-checkout: clear tracked sparse dirs", 2021-09-08) for why
    we try to delete entire directories outside the sparse cone).
  * these files will not be updated by by standard commands
    (switch/checkout/pull/merge/rebase will leave them alone unless
    conflicts happen -- and even then, the conflicted file may be
    written somewhere else to avoid overwriting the SKIP_WORKTREE file
    that is present and in the way)
  * there is nothing in Git that users can use to discover such
    files (status, diff, grep, etc. all ignore it)
  * there is no reasonable mechanism to "recover" from such a condition
    (neither `git sparse-checkout reapply` nor `git reset --hard` will
    correct it).

So, not only are users modifications ignored, but the files get
progressively more stale over time.  At some point in the future, they
may change their sparseness specification or disable sparse-checkouts.
At that time, all present-despite-SKIP_WORKTREE files will show up as
having lots of modifications because they represent a version from a
different branch or commit.  These might include user-made local changes
from days before, but the only way to tell is to have users look through
them all closely.

If these users come to others for help, there will be no logs that
explain the issue; it's just a mysterious list of changes.  Users might
adamantly claim (correctly, as it turns out) that they didn't modify
these files, while others presume they did.

[1] https://lore.kernel.org/git/xmqqbmb1a7ga.fsf@gitster-ct.c.googlers.com/
[2] https://lore.kernel.org/git/CABPp-BH9tju7WVm=QZDOvaMDdZbpNXrVWQdN-jmfN8wC6YVhmw@mail.gmail.com/
[3] https://lore.kernel.org/git/CABPp-BFnFpzwGC11TLoLs8YK5yiisA5D5-fFjXnJsbESVDwZsA@mail.gmail.com/

=== Problem space complexity ===

SKIP_WORKTREE has been part of Git for over a decade.  Duy did lots of
work on it initially, and several others have since come along and put
lots of work into it.  Stolee spent most of 2021 on the sparse-index,
with lots of bugfixes along the way including to non-sparse-index cases
as we are still trying to get sparse checkouts to behave reasonably.
Basically every codepath throughout the treat needs to be aware of an
additional type of file: tracked-but-not-present.  The extra type
results in lots of extra testcases and lots of extra code everywhere.

But, the sad thing is that we actually have more than one extra type.
We have tracked, tracked-but-not-present (SKIP_WORKTREE), and
tracked-but-promised-to-not-be-present-but-is-present-anyway
(present-despite-SKIP_WORKTREE).  Two types is a monumental amount of
effort to support, and adding a third feels a bit like insanity[4].

[4] Some examples of which can be seen at
    https://lore.kernel.org/git/CABPp-BGJ_Nvi5TmgriD9Bh6eNXE2EDq2f8e8QKXAeYG3BxZafA@mail.gmail.com/

=== Maintenance and code correctness challenges ===

Matheus' patches to grep stalled for nearly a year, in part because of
complications of how to handle sparse-checkouts appropriately in all
cases[5][6] (with trying to sanely figure out how to sanely handle
present-despite-SKIP_WORKTREE files being one of the complications).
His rm/add follow-ups also took months because of those kinds of
issues[7].  The corner cases with things like submodules and
SKIP_WORKTREE with the addition of present-despite-SKIP_WORKTREE start
becoming really complex[8].

We've had to add ugly logic to merge-ort to attempt to handle
present-despite-SKIP_WORKTREE files[9], and basically just been forced
to give up in merge-recursive knowing full well that we'll sometimes
silently discard user modifications.  Despite stash essentially being a
merge, it needed extra code (beyond what was in merge-ort and
merge-recursive) to manually tweak SKIP_WORKTREE bits in order to avoid
a few different bugs that'd result in an early abort with a partial
stash application[10].

[5] See https://lore.kernel.org/git/5f3f7ac77039d41d1692ceae4b0c5df3bb45b74a.1612901326.git.matheus.bernardino@usp.br/#t
    and the dates on the thread; also Matheus and I had several
    conversations off-list trying to resolve the issues over that time
[6] ...it finally kind of got unstuck after
    https://lore.kernel.org/git/CABPp-BGJ_Nvi5TmgriD9Bh6eNXE2EDq2f8e8QKXAeYG3BxZafA@mail.gmail.com/
[7] See for example
    https://lore.kernel.org/git/CABPp-BHwNoVnooqDFPAsZxBT9aR5Dwk5D9sDRCvYSb8akxAJgA@mail.gmail.com/#t
    and quotes like "The core functionality of sparse-checkout has always
    been only partially implemented", a statement I still believe is true
    today.
[8] https://lore.kernel.org/git/pull.809.git.git.1592356884310.gitgitgadget@gmail.com/
[9] See commit 66b209b86a ("merge-ort: implement CE_SKIP_WORKTREE
    handling with conflicted entries", 2021-03-20)
[10] See commit ba359fd507 ("stash: fix stash application in
     sparse-checkouts", 2020-12-01)

=== SKIP_WORKTREE expectations in Git ===

A couple quotes:

 * From [11] (before the "sparse-checkout" command existed):

   If it needs too many special cases, hacks, and conditionals, then it
   is not worth the complexity---if it is easier to write a correct code
   by allowing Git to populate working tree files, it is perfectly fine
   to do so.

   In a sense, the sparse checkout "feature" itself is a hack by itself,
   and that is why I think this part should be "best effort" as well.

 * From the git-sparse-checkout manual (still present today):

   THIS COMMAND IS EXPERIMENTAL. ITS BEHAVIOR, AND THE BEHAVIOR OF OTHER
   COMMANDS IN THE PRESENCE OF SPARSE-CHECKOUTS, WILL LIKELY CHANGE IN
   THE FUTURE.

[11] https://lore.kernel.org/git/xmqqbmb1a7ga.fsf@gitster-ct.c.googlers.com/

=== Suggested solution ===

SKIP_WORKTREE was written to allow sparse-checkouts, in particular, as
the name of the option implies, to allow the file to NOT be in the
worktree but consider it to be unchanged rather than deleted.

The suggests a simple solution: present-despite-SKIP_WORKTREE files
should not exist, for those using sparse-checkouts.

Enforce this at index loading time by checking if core.sparseCheckout is
true; if so, check files in the index with the SKIP_WORKTREE bit set to
verify that they are absent from the working tree.  If they are present,
unset the bit (in memory, though any commands that write to the index
will record the update).

Users can, of course, can get the SKIP_WORKTREE bit back such as by
running `git sparse-checkout reapply` (if they have ensured the file is
unmodified and doesn't match the specified sparsity patterns).

=== Pros/Cons of suggested solution ===

Pros:

  * Solves the user visible problems reported above, which I've been
    complaining about for nearly a year but couldn't find a solution to.
  * Helps prevent slow performance degradation with a sparse-index.
  * Much easier behavior in sparse-checkouts for users to reason about
  * Very simple, ~30 lines of code.
  * Significantly simplifies some ugly testcases, and obviates the need
    to test an entire class of potential issues.
  * Reduces code complexity, reasoning, and maintenance.  Avoids
    disagreements about weird corner cases[12].
  * It has been reported that some users might be (ab)using
    SKIP_WORKTREE as a let-me-modify-but-keep-the-file-in-the-worktree
    mechanism[13, and a few other similar references].  These users know
    of multiple caveats and shortcomings in doing so; perhaps not
    surprising given the "SKIP_WORKTREE expecations" section above.
    However, these users use `git update-index --skip-worktree`, and not
    `git sparse-checkout` or core.sparseCheckout=true.  As such, these
    users would be unaffected by this change and can continue abusing
    the system as before.

[12] https://lore.kernel.org/git/CABPp-BH9tju7WVm=QZDOvaMDdZbpNXrVWQdN-jmfN8wC6YVhmw@mail.gmail.com/
[13] https://stackoverflow.com/questions/13630849/git-difference-between-assume-unchanged-and-skip-worktree

Cons:

  * When core.sparseCheckout is enabled, this adds a performance cost to
    reading the index.  I'll defer discussion of this cost to a subsequent
    patch, since I have some optimizations to add.

=== Notes on testcase modifications ===

The good:
  * t1011: Compare to two cases above it ('read-tree will not throw away
    dirty changes, non-sparse'); since the file is present, it should
    match the non-sparse case now
  * t1092: sparse-index & sparse-checkout now match full-worktree
    behavior in more cases!  Yaay for consistency!
  * t6428, t7012: look at how much simpler the tests become!  Merge and
    stash can just fail early telling the user there's a file in the
    way, instead of not noticing until it's about to write a file and
    then have to implement sudden crash avoidance.  Hurray for sanity!
  * t7817: sparse behavior better matches full tree behavior.  Hurray
    for sanity!

The confusing:
  * t3705: These changes were ONLY needed on Windows, but they don't
    hurt other platforms.  Let's discuss each individually:

    * core.sparseCheckout should be false by default.  Nothing in this
      testcase toggles that until many, many tests later.  However,
      early tests (#5 in particular) were testing `update-index
      --skip-worktree` behavior in a non-sparse-checkout, but the
      Windows tests in CI were behaving as if core.sparseCheckout=true
      had been specified somewhere.  I do not have access to a Windows
      machine.  But I just manually did what should have been a no-op
      and turned the config off.  And it fixed the test.
    * I have no idea why the leftover .gitattributes file from this
      test was causing failures for test #18 on Windows, but only with
      these changes of mine.  Test #18 was checking for empty stderr,
      and specifically wanted to know that some error completely
      unrelated to file endings did not appear.  The leftover
      .gitattributes file thus caused some spurious stderr unrelated to
      the thing being checked.  Since other tests did not intend to
      test normalization, just proactively remove the .gitattributes
      file.  I'm certain this is cleaner and better, I'm just unsure
      why/how this didn't trigger problems before.

Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>