From d343068e4abc5e43d1ef1d5fed42bf4d7aa8cff4 Mon Sep 17 00:00:00 2001 From: Patrick Steinhardt Date: Wed, 4 Sep 2024 10:53:00 +0200 Subject: [PATCH 1/3] wrapper: introduce `log2u()` 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. Signed-off-by: Patrick Steinhardt Signed-off-by: Junio C Hamano --- bisect.c | 12 +----------- wrapper.h | 18 ++++++++++++++++++ 2 files changed, 19 insertions(+), 11 deletions(-) diff --git a/bisect.c b/bisect.c index 135f94ba09..a7862195a9 100644 --- a/bisect.c +++ b/bisect.c @@ -1130,16 +1130,6 @@ cleanup: 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; diff --git a/wrapper.h b/wrapper.h index 1b2b047ea0..a6b3e1f09e 100644 --- a/wrapper.h +++ b/wrapper.h @@ -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 */ From bd51dca36e0afcff36789bd6eb0135ebe66fcd81 Mon Sep 17 00:00:00 2001 From: Patrick Steinhardt Date: Wed, 4 Sep 2024 10:53:03 +0200 Subject: [PATCH 2/3] t0601: merge tests for auto-packing of refs 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. Signed-off-by: Patrick Steinhardt Signed-off-by: Junio C Hamano --- t/t0601-reffiles-pack-refs.sh | 36 +++++++++++++++++++---------------- 1 file changed, 20 insertions(+), 16 deletions(-) diff --git a/t/t0601-reffiles-pack-refs.sh b/t/t0601-reffiles-pack-refs.sh index 60a544b8ee..ed9652bb82 100755 --- a/t/t0601-reffiles-pack-refs.sh +++ b/t/t0601-reffiles-pack-refs.sh @@ -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 From c3459ae9ef26ee1b49d4de0af128444967910361 Mon Sep 17 00:00:00 2001 From: Patrick Steinhardt Date: Wed, 4 Sep 2024 10:53:08 +0200 Subject: [PATCH 3/3] refs/files: use heuristic to decide whether to repack with `--auto` 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. Signed-off-by: Patrick Steinhardt Signed-off-by: Junio C Hamano --- refs/files-backend.c | 65 +++++++++++++++++++++++++++ refs/packed-backend.c | 18 ++++++++ refs/packed-backend.h | 7 +++ t/t0601-reffiles-pack-refs.sh | 83 +++++++++++++++++++++++++++++++---- 4 files changed, 164 insertions(+), 9 deletions(-) diff --git a/refs/files-backend.c b/refs/files-backend.c index aa52d9be7c..6f047034e7 100644 --- a/refs/files-backend.c +++ b/refs/files-backend.c @@ -1300,6 +1300,68 @@ 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) { @@ -1312,6 +1374,9 @@ 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; diff --git a/refs/packed-backend.c b/refs/packed-backend.c index a0666407cd..a30082cb19 100644 --- a/refs/packed-backend.c +++ b/refs/packed-backend.c @@ -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. diff --git a/refs/packed-backend.h b/refs/packed-backend.h index 09437ad13b..9481d5e7c2 100644 --- a/refs/packed-backend.h +++ b/refs/packed-backend.h @@ -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 diff --git a/t/t0601-reffiles-pack-refs.sh b/t/t0601-reffiles-pack-refs.sh index ed9652bb82..d8cbd3f202 100755 --- a/t/t0601-reffiles-pack-refs.sh +++ b/t/t0601-reffiles-pack-refs.sh @@ -362,21 +362,86 @@ test_expect_success 'pack-refs does not drop broken refs during deletion' ' 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_expect_success "$command does not repack below 16 refs without packed-refs" ' test_when_finished "rm -rf repo" && git init repo && ( cd repo && git config set maintenance.auto false && - test_commit initial && + git commit --allow-empty --message "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 + # Create 14 additional references, which brings us to + # 15 together with the default branch. + printf "create refs/heads/loose-%d HEAD\n" $(test_seq 14) >stdin && + git update-ref --stdin stdin && + git update-ref --stdin stdin && + git update-ref --stdin stdin && + git update-ref --stdin stdin && + git update-ref --stdin