files-backend: don't rewrite the `packed-refs` file unnecessarily

Even when we are deleting references, we needn't overwrite the
`packed-refs` file if the references that we are deleting only exist
as loose references. Implement this optimization as follows:

* Add a function `is_packed_transaction_needed()`, which checks
  whether a given packed-refs transaction actually needs to be carried
  out (i.e., it returns false if the transaction obviously wouldn't
  have any effect). This function must be called while holding the
  `packed-refs` lock to avoid races.

* Change `files_transaction_prepare()` to check whether the
  packed-refs transaction is actually needed. If not, squelch it, but
  continue holding the `packed-refs` lock until the end of the
  transaction to avoid races.

This fixes a mild regression caused by dc39e09942 (files_ref_store:
use a transaction to update packed refs, 2017-09-08). Before that
commit, unnecessary rewrites of `packed-refs` were suppressed by
`repack_without_refs()`. But the transaction-based writing introduced
by that commit didn't perform that optimization.

Note that the pre-dc39e09942 code still had to *read* the whole
`packed-refs` file to determine that the rewrite could be skipped, so
the performance for the cases that the write could be elided was
`O(N)` in the number of packed references both before and after
dc39e09942. But after that commit the constant factor increased.

This commit reimplements the optimization of eliding unnecessary
`packed-refs` rewrites. That, plus the fact that since
cfa2e29c34 (packed_ref_store: get rid of the `ref_cache` entirely,
2017-03-17) we don't necessarily have to read the whole `packed-refs`
file at all, means that deletes of one or a few loose references can
now be done with `O(n lg N)` effort, where `n` is the number of loose
references being deleted and `N` is the total number of packed
references.

This commit fixes two tests in t1409.

Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

1 files changed, 94 insertions, 0 deletions

diff --git a/refs/packed-backend.c b/refs/packed-backend.c
index 0279aeceea..0b0a17ca8e 100644
--- a/refs/packed-backend.c
+++ b/refs/packed-backend.c
@@ -754,6 +754,100 @@ error:
 	return -1;
 }
 
+int is_packed_transaction_needed(struct ref_store *ref_store,
+				 struct ref_transaction *transaction)
+{
+	struct packed_ref_store *refs = packed_downcast(
+			ref_store,
+			REF_STORE_READ,
+			"is_packed_transaction_needed");
+	struct strbuf referent = STRBUF_INIT;
+	size_t i;
+	int ret;
+
+	if (!is_lock_file_locked(&refs->lock))
+		BUG("is_packed_transaction_needed() called while unlocked");
+
+	/*
+	 * We're only going to bother returning false for the common,
+	 * trivial case that references are only being deleted, their
+	 * old values are not being checked, and the old `packed-refs`
+	 * file doesn't contain any of those reference(s). This gives
+	 * false positives for some other cases that could
+	 * theoretically be optimized away:
+	 *
+	 * 1. It could be that the old value is being verified without
+	 *    setting a new value. In this case, we could verify the
+	 *    old value here and skip the update if it agrees. If it
+	 *    disagrees, we could either let the update go through
+	 *    (the actual commit would re-detect and report the
+	 *    problem), or come up with a way of reporting such an
+	 *    error to *our* caller.
+	 *
+	 * 2. It could be that a new value is being set, but that it
+	 *    is identical to the current packed value of the
+	 *    reference.
+	 *
+	 * Neither of these cases will come up in the current code,
+	 * because the only caller of this function passes to it a
+	 * transaction that only includes `delete` updates with no
+	 * `old_id`. Even if that ever changes, false positives only
+	 * cause an optimization to be missed; they do not affect
+	 * correctness.
+	 */
+
+	/*
+	 * Start with the cheap checks that don't require old
+	 * reference values to be read:
+	 */
+	for (i = 0; i < transaction->nr; i++) {
+		struct ref_update *update = transaction->updates[i];
+
+		if (update->flags & REF_HAVE_OLD)
+			/* Have to check the old value -> needed. */
+			return 1;
+
+		if ((update->flags & REF_HAVE_NEW) && !is_null_oid(&update->new_oid))
+			/* Have to set a new value -> needed. */
+			return 1;
+	}
+
+	/*
+	 * The transaction isn't checking any old values nor is it
+	 * setting any nonzero new values, so it still might be able
+	 * to be skipped. Now do the more expensive check: the update
+	 * is needed if any of the updates is a delete, and the old
+	 * `packed-refs` file contains a value for that reference.
+	 */
+	ret = 0;
+	for (i = 0; i < transaction->nr; i++) {
+		struct ref_update *update = transaction->updates[i];
+		unsigned int type;
+		struct object_id oid;
+
+		if (!(update->flags & REF_HAVE_NEW))
+			/*
+			 * This reference isn't being deleted -> not
+			 * needed.
+			 */
+			continue;
+
+		if (!refs_read_raw_ref(ref_store, update->refname,
+				       oid.hash, &referent, &type) ||
+		    errno != ENOENT) {
+			/*
+			 * We have to actually delete that reference
+			 * -> this transaction is needed.
+			 */
+			ret = 1;
+			break;
+		}
+	}
+
+	strbuf_release(&referent);
+	return ret;
+}
+
 struct packed_transaction_backend_data {
 	/* True iff the transaction owns the packed-refs lock. */
 	int own_lock;