From: Jeff King <peff@peff.net>
Date: Sun, 05 Apr 2026 06:46:51 GMT
Subject: Re: [PATCH] object-file: don't use object database without a repository
Message-ID: <20260405064651.GA1452907@coredump.intra.peff.net>
In-Reply-To: <20260404172817.2995133-1-luca.stefani.ge1@gmail.com>

On Sat, Apr 04, 2026 at 07:28:17PM +0200, Luca Stefani wrote:

> When running `git diff -- $file1 $file2' on large enough files,
> index_fd() attempts to use 'the_repository->objects', assuming it
> is initialized, but that's not the case for non-repository usecases.
> 
> When git diff is invoked without a backing repository,
> INDEX_WRITE_OBJECT is never set in flags, meaning only the hash is
> needed and nothing should be written to the object store.
> 
> Enforce the use of index_core() in this case.

I don't think we want to use index_core() for a large file, though. A
test like this:

diff --git a/t/t4053-diff-no-index.sh b/t/t4053-diff-no-index.sh
index 15076dfe0d..7ef5604430 100755
--- a/t/t4053-diff-no-index.sh
+++ b/t/t4053-diff-no-index.sh
@@ -413,4 +413,10 @@ test_expect_success 'diff --no-index with pathspec glob and exclude' '
 	test_cmp expect actual
 '
 
+test_expect_success 'diff --no-index on a huge file' '
+	dd if=/dev/zero bs=1M count=4000 >big.file &&
+	echo whatever >small.file &&
+	test_expect_code 1 git diff --no-index big.file small.file
+'
+
 test_done


will now fail on a 32-bit system, because we try to mmap the whole file,
which will fail.  We really do want to follow the streaming code path
(which knows to respect the lack of a WRITE_OBJECT flag and works
without an odb in that case).

It's kind of an expensive test, though, so we probably don't want to
actually include it in the test suite.

-Peff

PS I'd expect a 4GB+ file to work, too, but it looks like the diff code
   barfs when trying to stuff the file into a diff_filespec. A simpler
   example is:

     dd if=/dev/zero bs=1G count=5 >big.file
     git hash-object big.file

   but that dies, too! It looks like the streaming helper uses a size_t
   to take the size, which is wrong. It really should be an off_t. So I
   dunno, maybe nobody cares about ever working with 4GB files on 32-bit
   systems these days. It still feels like we should avoid a large mmap,
   though.