refs/files: use heuristic to decide whether to repack with `--auto`

Hi,

I recently noticed that git maintenance run --auto always ends up repacking loose references. This is because the "files" backend does not yet have any heuristics wired up to decide whether or not packing refs is in order. The consequence is that we always decided to pack them, which is of course quite a waste of time.

This small patch series fixes this by introducing a new heuristic for the "files" backend. The heuristic is rather simple: the bigger the "packed-refs" file, the more loose refs we require before we decide to repack. We have been using this heurisitc successfully for a long time in Gitaly by now.

Thanks!

Patrick

Patrick Steinhardt (2):
t0601: merge tests for auto-packing of refs
refs/files: use heuristic to decide whether to repack with --auto

 refs/files-backend.c          |  75 +++++++++++++++++++++++++
 refs/packed-backend.c         |  18 ++++++
 refs/packed-backend.h         |   7 +++
 t/t0601-reffiles-pack-refs.sh | 101 ++++++++++++++++++++++++++++------
 4 files changed, 185 insertions(+), 16 deletions(-)

2.46.0.421.g159f2d50e7.dirty

We have two tests in t0601 which exercise the same underlying logic, once via git pack-refs --auto and once via git maintenance run --auto. Merge these two tests into one such that it becomes easier to extend test coverage for both commands at the same time.

---
 t/t0601-reffiles-pack-refs.sh | 36 +++++++++++++++++++----------------
 1 file changed, 20 insertions(+), 16 deletions(-)

index 60a544b8ee8..ed9652bb829 100755
@@ -161,13 +161,6 @@ test_expect_success 'test --exclude takes precedence over --include' '
 	git pack-refs --include "refs/heads/pack*" --exclude "refs/heads/pack*" &&
 	test -f .git/refs/heads/dont_pack5'
 
-test_expect_success '--auto packs and prunes refs as usual' '
-	git branch auto &&
-	test_path_is_file .git/refs/heads/auto &&
-	git pack-refs --auto --all &&
-	test_path_is_missing .git/refs/heads/auto
-'
-
 test_expect_success 'see if up-to-date packed refs are preserved' '
 	git branch q &&
 	git pack-refs --all --prune &&

@@ -367,14 +360,25 @@ test_expect_success 'pack-refs does not drop broken refs during deletion' '
 	test_cmp expect actual
 '
 
-test_expect_success 'maintenance --auto unconditionally packs loose refs' '
-	git update-ref refs/heads/something HEAD &&
-	test_path_is_file .git/refs/heads/something &&
-	git rev-parse refs/heads/something >expect &&
-	git maintenance run --task=pack-refs --auto &&
-	test_path_is_missing .git/refs/heads/something &&
-	git rev-parse refs/heads/something >actual &&
-	test_cmp expect actual
-'
+for command in "git pack-refs --all --auto" "git maintenance run --task=pack-refs --auto"
+do
+	test_expect_success "$command unconditionally packs loose refs" '
+		test_when_finished "rm -rf repo" &&
+		git init repo &&
+		(
+			cd repo &&
+			git config set maintenance.auto false &&
+			test_commit initial &&
+
+			git update-ref refs/heads/something HEAD &&
+			test_path_is_file .git/refs/heads/something &&
+			git rev-parse refs/heads/something >expect &&
+			$command &&
+			test_path_is_missing .git/refs/heads/something &&
+			git rev-parse refs/heads/something >actual &&
+			test_cmp expect actual
+		)
+	'
+done
 
 test_done

2.46.0.421.g159f2d50e7.dirty

The --auto flag for git-pack-refs(1) allows the ref backend to decide whether or not a repack is in order. This switch has been introduced mostly with the "reftable" backend in mind, which already knows to auto-compact its tables during normal operations. When the flag is set, then it will use the same auto-compaction mechanism and thus end up doing nothing in most cases.

The "files" backend does not have any such heuristic yet and instead packs any loose references unconditionally. So we rewrite the complete "packed-refs" file even if there's only a single loose reference to be packed.

Even worse, starting with 9f6714ab3e (builtin/gc: pack refs when using git maintenance run --auto, 2024-03-25), git pack-refs --auto is unconditionally executed via our auto maintenance, so we end up repacking references every single time auto maintenance kicks in. And while that commit already mentioned that the "files" backend unconditionally packs refs now, the author obviously didn't quite think about the consequences thereof. So while the idea was sound, we really should have added a heuristic to the "files" backend before implementing it.

Introduce a heuristic that decides whether or not it is worth to pack loose references. The important factors to decide here are the number of loose references in comparison to the overall size of the "packed-refs" file. The bigger the "packed-refs" file, the longer it takes to rewrite it and thus we scale up the limit of allowed loose references before we repack.

As is the nature of heuristics, this mechansim isn't obviously "correct", but should rather be seen as a tradeoff between how much resources we spend packing refs and how inefficient the ref store becomes. For all I can say, we have successfully been using the exact same heuristic in Gitaly for several years by now.

---
 refs/files-backend.c          | 75 +++++++++++++++++++++++++++++++
 refs/packed-backend.c         | 18 ++++++++
 refs/packed-backend.h         |  7 +++
 t/t0601-reffiles-pack-refs.sh | 85 ++++++++++++++++++++++++++++++-----
 4 files changed, 175 insertions(+), 10 deletions(-)

index 1cff65f6ae5..261e2b8570f 100644
@@ -1311,6 +1311,78 @@ static int should_pack_ref(struct files_ref_store *refs,
 	return 0;
 }
 
+static size_t fastlog2(size_t sz)
+{
+	size_t l = 0;
+	if (!sz)
+		return 0;
+	for (; sz; sz /= 2)
+		l++;
+	return l - 1;
+}
+
+static int should_pack_refs(struct files_ref_store *refs,
+			    struct pack_refs_opts *opts)
+{
+	struct ref_iterator *iter;
+	size_t packed_size;
+	size_t refcount = 0;
+	size_t limit;
+	int ret;
+
+	if (!(opts->flags & PACK_REFS_AUTO))
+		return 1;
+
+	ret = packed_refs_size(refs->packed_ref_store, &packed_size);
+	if (ret < 0)
+		die("cannot determine packed-refs size");
+
+	/*
+	 * Packing loose references into the packed-refs file scales with the
+	 * number of references we're about to write. We thus decide whether we
+	 * repack refs by weighing the current size of the packed-refs file
+	 * against the number of loose references. This is done such that we do
+	 * not repack too often on repositories with a huge number of
+	 * references, where we can expect a lot of churn in the number of
+	 * references.
+	 *
+	 * As a heuristic, we repack if the number of loose references in the
+	 * repository exceeds `log2(nr_packed_refs) * 5`, where we estimate
+	 * `nr_packed_refs = packed_size / 100`, which scales as following:
+	 *
+	 * - 1kB ~ 10 packed refs: 16 refs
+	 * - 10kB ~ 100 packed refs: 33 refs
+	 * - 100kB ~ 1k packed refs: 49 refs
+	 * - 1MB ~ 10k packed refs: 66 refs
+	 * - 10MB ~ 100k packed refs: 82 refs
+	 * - 100MB ~ 1m packed refs: 99 refs
+	 *
+	 * We thus allow roughly 16 additional loose refs per factor of ten of
+	 * packed refs. This heuristic may be tweaked in the future, but should
+	 * serve as a sufficiently good first iteration.
+	 */
+	limit = fastlog2(packed_size / 100) * 5;
+	if (limit < 16)
+		limit = 16;
+
+	iter = cache_ref_iterator_begin(get_loose_ref_cache(refs, 0), NULL,
+					refs->base.repo, 0);
+	while ((ret = ref_iterator_advance(iter)) == ITER_OK) {
+		if (should_pack_ref(refs, iter->refname, iter->oid,
+				    iter->flags, opts))
+			refcount++;
+		if (refcount >= limit) {
+			ref_iterator_abort(iter);
+			return 1;
+		}
+	}
+
+	if (ret != ITER_DONE)
+		die("error while iterating over references");
+
+	return 0;
+}
+
 static int files_pack_refs(struct ref_store *ref_store,
 			   struct pack_refs_opts *opts)
 {

@@ -1323,6 +1395,9 @@ static int files_pack_refs(struct ref_store *ref_store,
 	struct strbuf err = STRBUF_INIT;
 	struct ref_transaction *transaction;
 
+	if (!should_pack_refs(refs, opts))
+		return 0;
+
 	transaction = ref_store_transaction_begin(refs->packed_ref_store, &err);
 	if (!transaction)
 		return -1;

index 7a0a695ca2f..07c57fd541a 100644
@@ -1250,6 +1250,24 @@ int packed_refs_is_locked(struct ref_store *ref_store)
 	return is_lock_file_locked(&refs->lock);
 }
 
+int packed_refs_size(struct ref_store *ref_store,
+		     size_t *out)
+{
+	struct packed_ref_store *refs = packed_downcast(ref_store, REF_STORE_READ,
+							"packed_refs_size");
+	struct stat st;
+
+	if (stat(refs->path, &st) < 0) {
+		if (errno != ENOENT)
+			return -1;
+		*out = 0;
+		return 0;
+	}
+
+	*out = st.st_size;
+	return 0;
+}
+
 /*
  * The packed-refs header line that we write out. Perhaps other traits
  * will be added later.

index 09437ad13bd..9481d5e7c2e 100644
@@ -27,6 +27,13 @@ int packed_refs_lock(struct ref_store *ref_store, int flags, struct strbuf *err)
 void packed_refs_unlock(struct ref_store *ref_store);
 int packed_refs_is_locked(struct ref_store *ref_store);
 
+/*
+ * Obtain the size of the `packed-refs` file. Reports `0` as size in case there
+ * is no packed-refs file. Returns 0 on success, negative otherwise.
+ */
+int packed_refs_size(struct ref_store *ref_store,
+		     size_t *out);
+
 /*
  * Return true if `transaction` really needs to be carried out against
  * the specified packed_ref_store, or false if it can be skipped

2.46.0.421.g159f2d50e7.dirty

Nit: s/instead/will instead/

Yeah, the version we have here is in fact an (almost exact) copy of the function in our reftable codebase. And as you note, we do not want to deduplicate these such that the reftable library is somewhat isolated from the rest of the codebase.

I also noticed log2i(), but honestly the only reason why I didn't reuse it is that I had no clue where to put it. There isn't any header that would be a good fit for it, and creating a new "math.h" header for a single function felt overblown to me. So I decided to just not bother. I'm happy to adjust though if somebody has a suggestion for where to put it.

Patrick

Given the existing contents of wrapper.h (near the end of the file), I think wrapper.c would be a good place to do so.

Shouldn't this essentially be a call to ffs() with the argument tweaked, by the way?

Indeed. Quite a grab bag of functions, thanks for the hint!

We are looking for the last set bit, not for the first set bit. We can massage things a bit, but wouldn't that require us to be aware of the platforms endianess?

In any case, GCC is clever enough to notice what we're doing:

fastlog2(unsigned long):
xor eax, eax
test rdi, rdi
je .L5
bsr rax, rdi
.L5:
ret

The log2i() function looks a bit less efficient:

log2i(unsigned long):
test rdi, rdi
je .L3
bsr rdi, rdi
lea eax, [rdi+1]
cdqe
ret
.L3:
xor eax, eax
ret

Clang isn't yet clever enough with v18.1, but is with trunk:

log2i(unsigned long):
bsr rax, rdi
inc rax
test rdi, rdi
cmove rax, rdi
ret

fastlog2(unsigned long):
test rdi, rdi
je .LBB1_1
shr rdi
bsr rcx, rdi
mov eax, 127
cmovne rax, rcx
xor eax, -64
add eax, 65
ret
.LBB1_1:
mov eax, -1
ret

So with the following definition we're optimizing both with GCC and Clang:

    size_t fastlog2(size_t sz)
    {
        size_t l = 0;
        if (!sz)
            return 0;
        for (; sz; sz >>= 1)
            l++;
        return l;
    }

I'd thus say we can just pick that function instead of caring about platform endianess with ffs().

Patrick

Revisiting this: isn't the current version of this sentence correct? Replacing it with "will instead" certainly makes it incorrect without also changing "packs" to "pack".

Patrick

Hi,

this is the second version of my patch series that introduces a new heuristic for packing refs with the "files" backend. This heuristic is designed to avoid needlessly rewriting the "packed-refs" file when there are only a small set of loose refs. The number of loose refs required scales with the size of the "packed-refs" file.

There is only one change compared to v1, namely a deduplication of the log2 functions we have in our tree.

Thanks!

Patrick

Patrick Steinhardt (3):
wrapper: introduce log2u()
t0601: merge tests for auto-packing of refs
refs/files: use heuristic to decide whether to repack with --auto

 bisect.c                      |  12 +---
 refs/files-backend.c          |  65 ++++++++++++++++++++++
 refs/packed-backend.c         |  18 ++++++
 refs/packed-backend.h         |   7 +++
 t/t0601-reffiles-pack-refs.sh | 101 ++++++++++++++++++++++++++++------
 wrapper.h                     |  18 ++++++
 6 files changed, 194 insertions(+), 27 deletions(-)

Range-diff against v1:

-:  ----------- > 1:  df8c5dffffe wrapper: introduce `log2u()`
1:  3a8063e8b2c = 2:  4a59cec205d t0601: merge tests for auto-packing of refs
2:  9a63abfe3b8 ! 3:  49f953142b1 refs/files: use heuristic to decide whether to repack with `--auto`
    @@ refs/files-backend.c: static int should_pack_ref(struct files_ref_store *refs,
      	return 0;
      }
      
    -+static size_t fastlog2(size_t sz)
    -+{
    -+	size_t l = 0;
    -+	if (!sz)
    -+		return 0;
    -+	for (; sz; sz /= 2)
    -+		l++;
    -+	return l - 1;
    -+}
    -+
     +static int should_pack_refs(struct files_ref_store *refs,
     +			    struct pack_refs_opts *opts)
     +{
    @@ refs/files-backend.c: static int should_pack_ref(struct files_ref_store *refs,
     +	 * packed refs. This heuristic may be tweaked in the future, but should
     +	 * serve as a sufficiently good first iteration.
     +	 */
    -+	limit = fastlog2(packed_size / 100) * 5;
    ++	limit = log2u(packed_size / 100) * 5;
     +	if (limit < 16)
     +		limit = 16;
     +

base-commit: 4590f2e9412378c61eac95966709c78766d326ba

2.46.0.519.g2e7b89e038.dirty

We have an implementation of a function that computes the log2 for an integer. While we could instead use log2(3P), that involves floating point numbers and is thus needlessly complex and inefficient.

We're about to add a second caller that wants to compute log2 for a size_t. Let's thus move the function into "wrapper.h" such that it becomes generally available.

While at it, tweak the implementation a bit:

- The parameter is converted from int to uintmax_t. This conversion is safe to do in "bisect.c" because we already check that the argument is positive.

- The return value is an unsigned. It cannot ever be negative, so it is pointless for it to be a signed integer.

- Loop until !n instead of requiring that n > 1 and then subtract 1 from the result and add a special case for !sz. This helps compilers to generate more efficient code.

Compilers recognize the pattern of this function and optimize accordingly. On GCC 14.2 x86_64:

log2u(unsigned long):
test rdi, rdi
je .L3
bsr rax, rdi
ret
.L3:
mov eax, -1
ret

Clang 18.1 does not yet recognize the pattern, but starts to do so on Clang trunk x86_64. The code isn't quite as efficient as the one generated by GCC, but still manages to optimize away the loop:

log2u(unsigned long):
test rdi, rdi
je .LBB0_1
shr rdi
bsr rcx, rdi
mov eax, 127
cmovne rax, rcx
xor eax, -64
add eax, 65
ret
.LBB0_1:
mov eax, -1
ret

The pattern is also recognized on other platforms like ARM64 GCC 14.2.0, where we end up using clz:

log2u(unsigned long):
clz x2, x0
cmp x0, 0
mov w1, 63
sub w0, w1, w2
csinv w0, w0, wzr, ne
ret

Note that we have a similar function fastlog2() in the reftable code. As that codebase is separate from the Git codebase we do not adapt it to use the new function.

---
 bisect.c  | 12 +-----------
 wrapper.h | 18 ++++++++++++++++++
 2 files changed, 19 insertions(+), 11 deletions(-)

index 4406fc44cf9..4713226fad4 100644
@@ -1130,16 +1130,6 @@ enum bisect_error bisect_next_all(struct repository *r, const char *prefix)
 	return res;
 }
 
-static inline int log2i(int n)
-{
-	int log2 = 0;
-
-	for (; n > 1; n >>= 1)
-		log2++;
-
-	return log2;
-}
-
 static inline int exp2i(int n)
 {
 	return 1 << n;

@@ -1162,7 +1152,7 @@ int estimate_bisect_steps(int all)
 	if (all < 3)
 		return 0;
 
-	n = log2i(all);
+	n = log2u(all);
 	e = exp2i(n);
 	x = all - e;
 

index 1b2b047ea06..a6b3e1f09ec 100644
@@ -140,4 +140,22 @@ int csprng_bytes(void *buf, size_t len);
  */
 uint32_t git_rand(void);
 
+/* Provide log2 of the given `size_t`. */
+static inline unsigned log2u(uintmax_t sz)
+{
+	unsigned l = 0;
+
+	/*
+	 * Technically this isn't required, but it helps the compiler optimize
+	 * this to a `bsr` instruction.
+	 */
+	if (!sz)
+		return 0;
+
+	for (; sz; sz >>= 1)
+		l++;
+
+	return l - 1;
+}
+
 #endif /* WRAPPER_H */

2.46.0.519.g2e7b89e038.dirty

We have two tests in t0601 which exercise the same underlying logic, once via git pack-refs --auto and once via git maintenance run --auto. Merge these two tests into one such that it becomes easier to extend test coverage for both commands at the same time.

---
 t/t0601-reffiles-pack-refs.sh | 36 +++++++++++++++++++----------------
 1 file changed, 20 insertions(+), 16 deletions(-)

index 60a544b8ee8..ed9652bb829 100755
@@ -161,13 +161,6 @@ test_expect_success 'test --exclude takes precedence over --include' '
 	git pack-refs --include "refs/heads/pack*" --exclude "refs/heads/pack*" &&
 	test -f .git/refs/heads/dont_pack5'
 
-test_expect_success '--auto packs and prunes refs as usual' '
-	git branch auto &&
-	test_path_is_file .git/refs/heads/auto &&
-	git pack-refs --auto --all &&
-	test_path_is_missing .git/refs/heads/auto
-'
-
 test_expect_success 'see if up-to-date packed refs are preserved' '
 	git branch q &&
 	git pack-refs --all --prune &&

@@ -367,14 +360,25 @@ test_expect_success 'pack-refs does not drop broken refs during deletion' '
 	test_cmp expect actual
 '
 
-test_expect_success 'maintenance --auto unconditionally packs loose refs' '
-	git update-ref refs/heads/something HEAD &&
-	test_path_is_file .git/refs/heads/something &&
-	git rev-parse refs/heads/something >expect &&
-	git maintenance run --task=pack-refs --auto &&
-	test_path_is_missing .git/refs/heads/something &&
-	git rev-parse refs/heads/something >actual &&
-	test_cmp expect actual
-'
+for command in "git pack-refs --all --auto" "git maintenance run --task=pack-refs --auto"
+do
+	test_expect_success "$command unconditionally packs loose refs" '
+		test_when_finished "rm -rf repo" &&
+		git init repo &&
+		(
+			cd repo &&
+			git config set maintenance.auto false &&
+			test_commit initial &&
+
+			git update-ref refs/heads/something HEAD &&
+			test_path_is_file .git/refs/heads/something &&
+			git rev-parse refs/heads/something >expect &&
+			$command &&
+			test_path_is_missing .git/refs/heads/something &&
+			git rev-parse refs/heads/something >actual &&
+			test_cmp expect actual
+		)
+	'
+done
 
 test_done

2.46.0.519.g2e7b89e038.dirty

The --auto flag for git-pack-refs(1) allows the ref backend to decide whether or not a repack is in order. This switch has been introduced mostly with the "reftable" backend in mind, which already knows to auto-compact its tables during normal operations. When the flag is set, then it will use the same auto-compaction mechanism and thus end up doing nothing in most cases.

The "files" backend does not have any such heuristic yet and instead packs any loose references unconditionally. So we rewrite the complete "packed-refs" file even if there's only a single loose reference to be packed.

Even worse, starting with 9f6714ab3e (builtin/gc: pack refs when using git maintenance run --auto, 2024-03-25), git pack-refs --auto is unconditionally executed via our auto maintenance, so we end up repacking references every single time auto maintenance kicks in. And while that commit already mentioned that the "files" backend unconditionally packs refs now, the author obviously didn't quite think about the consequences thereof. So while the idea was sound, we really should have added a heuristic to the "files" backend before implementing it.

Introduce a heuristic that decides whether or not it is worth to pack loose references. The important factors to decide here are the number of loose references in comparison to the overall size of the "packed-refs" file. The bigger the "packed-refs" file, the longer it takes to rewrite it and thus we scale up the limit of allowed loose references before we repack.

As is the nature of heuristics, this mechansim isn't obviously "correct", but should rather be seen as a tradeoff between how much resources we spend packing refs and how inefficient the ref store becomes. For all I can say, we have successfully been using the exact same heuristic in Gitaly for several years by now.

---
 refs/files-backend.c          | 65 +++++++++++++++++++++++++++
 refs/packed-backend.c         | 18 ++++++++
 refs/packed-backend.h         |  7 +++
 t/t0601-reffiles-pack-refs.sh | 85 ++++++++++++++++++++++++++++++-----
 4 files changed, 165 insertions(+), 10 deletions(-)

index 1cff65f6ae5..51bcd16f821 100644
@@ -1311,6 +1311,68 @@ static int should_pack_ref(struct files_ref_store *refs,
 	return 0;
 }
 
+static int should_pack_refs(struct files_ref_store *refs,
+			    struct pack_refs_opts *opts)
+{
+	struct ref_iterator *iter;
+	size_t packed_size;
+	size_t refcount = 0;
+	size_t limit;
+	int ret;
+
+	if (!(opts->flags & PACK_REFS_AUTO))
+		return 1;
+
+	ret = packed_refs_size(refs->packed_ref_store, &packed_size);
+	if (ret < 0)
+		die("cannot determine packed-refs size");
+
+	/*
+	 * Packing loose references into the packed-refs file scales with the
+	 * number of references we're about to write. We thus decide whether we
+	 * repack refs by weighing the current size of the packed-refs file
+	 * against the number of loose references. This is done such that we do
+	 * not repack too often on repositories with a huge number of
+	 * references, where we can expect a lot of churn in the number of
+	 * references.
+	 *
+	 * As a heuristic, we repack if the number of loose references in the
+	 * repository exceeds `log2(nr_packed_refs) * 5`, where we estimate
+	 * `nr_packed_refs = packed_size / 100`, which scales as following:
+	 *
+	 * - 1kB ~ 10 packed refs: 16 refs
+	 * - 10kB ~ 100 packed refs: 33 refs
+	 * - 100kB ~ 1k packed refs: 49 refs
+	 * - 1MB ~ 10k packed refs: 66 refs
+	 * - 10MB ~ 100k packed refs: 82 refs
+	 * - 100MB ~ 1m packed refs: 99 refs
+	 *
+	 * We thus allow roughly 16 additional loose refs per factor of ten of
+	 * packed refs. This heuristic may be tweaked in the future, but should
+	 * serve as a sufficiently good first iteration.
+	 */
+	limit = log2u(packed_size / 100) * 5;
+	if (limit < 16)
+		limit = 16;
+
+	iter = cache_ref_iterator_begin(get_loose_ref_cache(refs, 0), NULL,
+					refs->base.repo, 0);
+	while ((ret = ref_iterator_advance(iter)) == ITER_OK) {
+		if (should_pack_ref(refs, iter->refname, iter->oid,
+				    iter->flags, opts))
+			refcount++;
+		if (refcount >= limit) {
+			ref_iterator_abort(iter);
+			return 1;
+		}
+	}
+
+	if (ret != ITER_DONE)
+		die("error while iterating over references");
+
+	return 0;
+}
+
 static int files_pack_refs(struct ref_store *ref_store,
 			   struct pack_refs_opts *opts)
 {

@@ -1323,6 +1385,9 @@ static int files_pack_refs(struct ref_store *ref_store,
 	struct strbuf err = STRBUF_INIT;
 	struct ref_transaction *transaction;
 
+	if (!should_pack_refs(refs, opts))
+		return 0;
+
 	transaction = ref_store_transaction_begin(refs->packed_ref_store, &err);
 	if (!transaction)
 		return -1;

index 7a0a695ca2f..07c57fd541a 100644
@@ -1250,6 +1250,24 @@ int packed_refs_is_locked(struct ref_store *ref_store)
 	return is_lock_file_locked(&refs->lock);
 }
 
+int packed_refs_size(struct ref_store *ref_store,
+		     size_t *out)
+{
+	struct packed_ref_store *refs = packed_downcast(ref_store, REF_STORE_READ,
+							"packed_refs_size");
+	struct stat st;
+
+	if (stat(refs->path, &st) < 0) {
+		if (errno != ENOENT)
+			return -1;
+		*out = 0;
+		return 0;
+	}
+
+	*out = st.st_size;
+	return 0;
+}
+
 /*
  * The packed-refs header line that we write out. Perhaps other traits
  * will be added later.

index 09437ad13bd..9481d5e7c2e 100644
@@ -27,6 +27,13 @@ int packed_refs_lock(struct ref_store *ref_store, int flags, struct strbuf *err)
 void packed_refs_unlock(struct ref_store *ref_store);
 int packed_refs_is_locked(struct ref_store *ref_store);
 
+/*
+ * Obtain the size of the `packed-refs` file. Reports `0` as size in case there
+ * is no packed-refs file. Returns 0 on success, negative otherwise.
+ */
+int packed_refs_size(struct ref_store *ref_store,
+		     size_t *out);
+
 /*
  * Return true if `transaction` really needs to be carried out against
  * the specified packed_ref_store, or false if it can be skipped

2.46.0.519.g2e7b89e038.dirty

This seems to hit the balance between the thoroughness of repack (leaving fewer loose refs is good) and the frequencey (doing repack less often is good) in a quite sensible way.

I also have to wonder if it adds a good component to the heuristics to leave younger loose refs (wrt mtime) out of packed-refs, with the expectation that they are more likely to be updated again soon than refs that were written long time ago and stayed the same value.

Thanks.

Nice. Aiming to compile to "bsr" is very good.

The above loop that compilers seem to know to reduce to "bsr" is good.

FWIW, because the definition of "first bit" in ffs() is "least significant bit", you do not need to worry about endianness at all, but what we want is most significant, so it is not directly usable.

You're right, I did missread.

Maybe.

In general, I expect that most users typically only touch a very small set of refs, e.g. the four or five feature branches that they have. So even without such an additional component we would not end up repacking all that often.

That picture changes once you consider remotes, because with bigger teams it's quite likely that you'll get many ref updates there. I'm not sure a time-based heuristic would be a good fit for this usecase, because I'd think that repacking those right away is sensible most of the time.

We also have to consider that an mtime-based component makes the overall system harder to understand and indeterministic. Which isn't to say that it doesn't make sense. But I rather think we should land the simple and stupid solution first, and in case we see that it's insufficient we can iterate and improve it in the future.

Patrick