refs: stop re-verifying common prefixes for availability

One of the checks done by `refs_verify_refnames_available()` is whether
any of the prefixes of a reference already exists. For example, given a
reference "refs/heads/main", we'd check whether "refs/heads" or "refs"
already exist, and if so we'd abort the transaction.

When updating multiple references at once, this check is performed for
each of the references individually. Consequently, because references
tend to have common prefixes like "refs/heads/" or refs/tags/", we
evaluate the availability of these prefixes repeatedly. Naturally this
is a waste of compute, as the availability of those prefixes should in
general not change in the middle of a transaction. And if it would,
backends would notice at a later point in time.

Optimize this pattern by storing prefixes in a `strset` so that we can
trivially track those prefixes that we have already checked. This leads
to a significant speedup with the "reftable" backend when creating many
references that all share a common prefix:

    Benchmark 1: update-ref: create many refs (refformat = reftable, preexisting = 100000, new = 10000, revision = HEAD~)
      Time (mean ± σ):      63.1 ms ±   1.8 ms    [User: 41.0 ms, System: 21.6 ms]
      Range (min … max):    60.6 ms …  69.5 ms    38 runs

    Benchmark 2: update-ref: create many refs (refformat = reftable, preexisting = 100000, new = 10000, revision = HEAD)
      Time (mean ± σ):      40.0 ms ±   1.3 ms    [User: 29.3 ms, System: 10.3 ms]
      Range (min … max):    38.1 ms …  47.3 ms    61 runs

    Summary
      update-ref: create many refs (refformat = reftable, preexisting = 100000, new = 10000, revision = HEAD) ran
        1.58 ± 0.07 times faster than update-ref: create many refs (refformat = reftable, preexisting = 100000, new = 10000, revision = HEAD~)

For the "files" backend we see an improvement, but a much smaller one:

    Benchmark 1: update-ref: create many refs (refformat = files, preexisting = 100000, new = 10000, revision = HEAD~)
      Time (mean ± σ):     395.8 ms ±   5.3 ms    [User: 63.6 ms, System: 330.5 ms]
      Range (min … max):   387.0 ms … 404.6 ms    10 runs

    Benchmark 2: update-ref: create many refs (refformat = files, preexisting = 100000, new = 10000, revision = HEAD)
      Time (mean ± σ):     386.0 ms ±   4.0 ms    [User: 51.5 ms, System: 332.8 ms]
      Range (min … max):   380.8 ms … 392.6 ms    10 runs

    Summary
      update-ref: create many refs (refformat = files, preexisting = 100000, new = 10000, revision = HEAD) ran
        1.03 ± 0.02 times faster than update-ref: create many refs (refformat = files, preexisting = 100000, new = 10000, revision = HEAD~)

This change also leads to a modest improvement when writing references
with "initial" semantics, for example when migrating references. The
following benchmarks are migrating 1m references from the "reftable" to
the "files" backend:

    Benchmark 1: migrate reftable:files (refcount = 1000000, revision = HEAD~)
      Time (mean ± σ):     836.6 ms ±   5.6 ms    [User: 645.2 ms, System: 185.2 ms]
      Range (min … max):   829.6 ms … 845.9 ms    10 runs

    Benchmark 2: migrate reftable:files (refcount = 1000000, revision = HEAD)
      Time (mean ± σ):     759.8 ms ±   5.1 ms    [User: 574.9 ms, System: 178.9 ms]
      Range (min … max):   753.1 ms … 768.8 ms    10 runs

    Summary
      migrate reftable:files (refcount = 1000000, revision = HEAD) ran
        1.10 ± 0.01 times faster than migrate reftable:files (refcount = 1000000, revision = HEAD~)

And vice versa:

    Benchmark 1: migrate files:reftable (refcount = 1000000, revision = HEAD~)
      Time (mean ± σ):     870.7 ms ±   5.7 ms    [User: 735.2 ms, System: 127.4 ms]
      Range (min … max):   861.6 ms … 883.2 ms    10 runs

    Benchmark 2: migrate files:reftable (refcount = 1000000, revision = HEAD)
      Time (mean ± σ):     799.1 ms ±   8.5 ms    [User: 661.1 ms, System: 130.2 ms]
      Range (min … max):   787.5 ms … 812.6 ms    10 runs

    Summary
      migrate files:reftable (refcount = 1000000, revision = HEAD) ran
        1.09 ± 0.01 times faster than migrate files:reftable (refcount = 1000000, revision = HEAD~)

The impact here is significantly smaller given that we don't perform any
reference reads with "initial" semantics, so the speedup only comes from
us doing less string list lookups.

Signed-off-by: Patrick Steinhardt <[email protected]>
Signed-off-by: Junio C Hamano <[email protected]>

Author: pks-t

refs.c

@@ -2485,6 +2485,7 @@ int refs_verify_refnames_available(struct ref_store *refs,
 	struct strbuf dirname = STRBUF_INIT;
 	struct strbuf referent = STRBUF_INIT;
 	struct string_list_item *item;
+	struct strset dirnames;
 	int ret = -1;
 
 	/*
@@ -2494,6 +2495,8 @@ int refs_verify_refnames_available(struct ref_store *refs,
 
 	assert(err);
 
+	strset_init(&dirnames);
+
 	for_each_string_list_item(item, refnames) {
 		const char *refname = item->string;
 		const char *extra_refname;
@@ -2523,6 +2526,14 @@ int refs_verify_refnames_available(struct ref_store *refs,
 			if (skip && string_list_has_string(skip, dirname.buf))
 				continue;
 
+			/*
+			 * If we've already seen the directory we don't need to
+			 * process it again. Skip it to avoid checking checking
+			 * common prefixes like "refs/heads/" repeatedly.
+			 */
+			if (!strset_add(&dirnames, dirname.buf))
+				continue;
+
 			if (!initial_transaction &&
 			    !refs_read_raw_ref(refs, dirname.buf, &oid, &referent,
 					       &type, &ignore_errno)) {
@@ -2583,6 +2594,7 @@ int refs_verify_refnames_available(struct ref_store *refs,
 cleanup:
 	strbuf_release(&referent);
 	strbuf_release(&dirname);
+	strset_clear(&dirnames);
 	return ret;
 }