fetch: implement support for atomic reference updates

Hi,

this is a short patch series to implement support for atomic reference updates for git-fetch(1). It's similar to git push --atomic, only that it applies to the local side. That is the fetch will either succeed and update all remote references or it will fail and update none.

Patrick

Patrick Steinhardt (2):
fetch: allow passing a transaction to s_update_ref()
fetch: implement support for atomic reference updates

 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 |  72 ++++++++++++-----
 t/t5510-fetch.sh                | 139 ++++++++++++++++++++++++++++++++
 3 files changed, 197 insertions(+), 18 deletions(-)

2.30.0

The handling of ref updates is completely handled by s_update_ref(), which will manage the complete lifecycle of the reference transaction. This is fine right now given that git-fetch(1) does not support atomic fetches, so each reference gets its own transaction. It is quite inflexible though, as s_update_ref() only knows about a single reference update at a time, so it doesn't allow us to alter the strategy.

This commit prepares s_update_ref() and its only caller update_local_ref() to allow passing an external transaction. If none is given, then the existing behaviour is triggered which creates a new transaction and directly commits it. Otherwise, if the caller provides a transaction, then we only queue the update but don't commit it. This optionally allows the caller to manage when a transaction will be committed.

Given that update_local_ref() is always called with a NULL transaction for now, no change in behaviour is expected from this change.

---
 builtin/fetch.c | 48 +++++++++++++++++++++++++++++++-----------------
 1 file changed, 31 insertions(+), 17 deletions(-)

index ecf8537605..020a977bc7 100644
@@ -583,13 +583,14 @@ static struct ref *get_ref_map(struct remote *remote,
 
 static int s_update_ref(const char *action,
 			struct ref *ref,
+			struct ref_transaction *transaction,
 			int check_old)
 {
 	char *msg;
 	char *rla = getenv("GIT_REFLOG_ACTION");
-	struct ref_transaction *transaction;
+	struct ref_transaction *transaction_to_free = NULL;
 	struct strbuf err = STRBUF_INIT;
-	int ret, df_conflict = 0;
+	int ret, df_conflict = 0, commit = 0;
 
 	if (dry_run)
 		return 0;

@@ -597,26 +598,38 @@ static int s_update_ref(const char *action,
 		rla = default_rla.buf;
 	msg = xstrfmt("%s: %s", rla, action);
 
-	transaction = ref_transaction_begin(&err);
-	if (!transaction ||
-	    ref_transaction_update(transaction, ref->name,
+	/*
+	 * If no transaction was passed to us, we manage the transaction
+	 * ourselves. Otherwise, we trust the caller to handle the transaction
+	 * lifecycle.
+	 */
+	if (!transaction) {
+		transaction = transaction_to_free = ref_transaction_begin(&err);
+		if (!transaction)
+			goto fail;
+		commit = 1;
+	}
+
+	if (ref_transaction_update(transaction, ref->name,
 				   &ref->new_oid,
 				   check_old ? &ref->old_oid : NULL,
 				   0, msg, &err))
 		goto fail;
 
-	ret = ref_transaction_commit(transaction, &err);
-	if (ret) {
-		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
-		goto fail;
+	if (commit) {
+		ret = ref_transaction_commit(transaction, &err);
+		if (ret) {
+			df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
+			goto fail;
+		}
 	}
 
-	ref_transaction_free(transaction);
+	ref_transaction_free(transaction_to_free);
 	strbuf_release(&err);
 	free(msg);
 	return 0;
 fail:
-	ref_transaction_free(transaction);
+	ref_transaction_free(transaction_to_free);
 	error("%s", err.buf);
 	strbuf_release(&err);
 	free(msg);

@@ -759,6 +772,7 @@ static void format_display(struct strbuf *display, char code,
 }
 
 static int update_local_ref(struct ref *ref,
+			    struct ref_transaction *transaction,
 			    const char *remote,
 			    const struct ref *remote_ref,
 			    struct strbuf *display,

@@ -799,7 +813,7 @@ static int update_local_ref(struct ref *ref,
 	    starts_with(ref->name, "refs/tags/")) {
 		if (force || ref->force) {
 			int r;
-			r = s_update_ref("updating tag", ref, 0);
+			r = s_update_ref("updating tag", ref, transaction, 0);
 			format_display(display, r ? '!' : 't', _("[tag update]"),
 				       r ? _("unable to update local ref") : NULL,
 				       remote, pretty_ref, summary_width);

@@ -836,7 +850,7 @@ static int update_local_ref(struct ref *ref,
 			what = _("[new ref]");
 		}
 
-		r = s_update_ref(msg, ref, 0);
+		r = s_update_ref(msg, ref, transaction, 0);
 		format_display(display, r ? '!' : '*', what,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -858,7 +872,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "..");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("fast-forward", ref, 1);
+		r = s_update_ref("fast-forward", ref, transaction, 1);
 		format_display(display, r ? '!' : ' ', quickref.buf,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -870,7 +884,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "...");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("forced-update", ref, 1);
+		r = s_update_ref("forced-update", ref, transaction, 1);
 		format_display(display, r ? '!' : '+', quickref.buf,
 			       r ? _("unable to update local ref") : _("forced update"),
 			       remote, pretty_ref, summary_width);

@@ -1034,8 +1048,8 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 
 			strbuf_reset(&note);
 			if (ref) {
-				rc |= update_local_ref(ref, what, rm, &note,
-						       summary_width);
+				rc |= update_local_ref(ref, NULL, what,
+						       rm, &note, summary_width);
 				free(ref);
 			} else if (write_fetch_head || dry_run) {
 				/*

2.30.0

Yeah, I think we also don't need the df_conflict variable, and I don't like the duplication of the following clean up code:

        ref_transaction_free(transaction_to_free);
        strbuf_release(&err);
        free(msg);

Patrick's patch didn't introduce them, but we might still want to get rid of them (see below).

"our_transaction" looks fine.

Here is a suggested refactoring patch that could come before Patrick's patch:

-------------------------------------------------------------------------

index ecf8537605..8017fc19da 100644
@@ -581,6 +581,22 @@ static struct ref *get_ref_map(struct remote *remote,

#define STORE_REF_ERROR_OTHER 1
#define STORE_REF_ERROR_DF_CONFLICT 2

+static int do_s_update_ref(struct ref_transaction *transaction,

+                          struct ref *ref,
+                          int check_old,
+                          struct strbuf *err,
+                          char *msg)
+{
+       if (!transaction ||
+           ref_transaction_update(transaction, ref->name,
+                                  &ref->new_oid,
+                                  check_old ? &ref->old_oid : NULL,
+                                  0, msg, err))

+ return TRANSACTION_GENERIC_ERROR;
+
+ return ref_transaction_commit(transaction, err);
+}
+
static int s_update_ref(const char *action,
struct ref *ref,
int check_old)
@@ -589,7 +605,7 @@ static int s_update_ref(const char *action,

       char *rla = getenv("GIT_REFLOG_ACTION");
       struct ref_transaction *transaction;
       struct strbuf err = STRBUF_INIT;

- int ret, df_conflict = 0;
+ int ret = 0;

       if (dry_run)
               return 0;
@@ -598,30 +614,19 @@ static int s_update_ref(const char *action,
       msg = xstrfmt("%s: %s", rla, action);

       transaction = ref_transaction_begin(&err);
-       if (!transaction ||
-           ref_transaction_update(transaction, ref->name,
-                                  &ref->new_oid,
-                                  check_old ? &ref->old_oid : NULL,
-                                  0, msg, &err))

- goto fail;

- ret = ref_transaction_commit(transaction, &err);

+       ret = do_s_update_ref(transaction, ref, check_old, &err, msg);
       if (ret) {
-               df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
-               goto fail;
+               error("%s", err.buf);
+               ret = (ret == TRANSACTION_NAME_CONFLICT)
+                       ? STORE_REF_ERROR_DF_CONFLICT
+                       : STORE_REF_ERROR_OTHER;
       }

       ref_transaction_free(transaction);
       strbuf_release(&err);
       free(msg);
-       return 0;
-fail:
-       ref_transaction_free(transaction);
-       error("%s", err.buf);
-       strbuf_release(&err);
-       free(msg);
-       return df_conflict ? STORE_REF_ERROR_DF_CONFLICT

- : STORE_REF_ERROR_OTHER;
+ return ret;
}

static int refcol_width = 10;
-------------------------------------------------------------------------

After the above patch, Patrick's patch would become:

-------------------------------------------------------------------------

index 8017fc19da..d58805c03d 100644
@@ -584,6 +584,7 @@ static struct ref *get_ref_map(struct remote *remote,

static int do_s_update_ref(struct ref_transaction *transaction,
struct ref *ref,
int check_old,
+ int commit,
struct strbuf *err,
char *msg)
{
@@ -594,16 +595,17 @@ static int do_s_update_ref(struct ref_transaction *transaction,

                                  0, msg, err))
               return TRANSACTION_GENERIC_ERROR;

- return ref_transaction_commit(transaction, err);
+ return (commit) ? ref_transaction_commit(transaction, err) : 0;
}

static int s_update_ref(const char *action,
+ struct ref_transaction *transaction,
struct ref *ref,
int check_old)

{
       char *msg;
       char *rla = getenv("GIT_REFLOG_ACTION");
-       struct ref_transaction *transaction;
+       struct ref_transaction *our_transaction = NULL;
       struct strbuf err = STRBUF_INIT;
       int ret = 0;

@@ -613,9 +615,16 @@ static int s_update_ref(const char *action,

               rla = default_rla.buf;
       msg = xstrfmt("%s: %s", rla, action);

- transaction = ref_transaction_begin(&err);
+ /*

+        * If no transaction was passed to us, we manage the
+        * transaction ourselves. Otherwise, we trust the caller to
+        * handle the transaction lifecycle.
+        */

+ if (!transaction)
+ transaction = our_transaction = ref_transaction_begin(&err);

- ret = do_s_update_ref(transaction, ref, check_old, &err, msg);
+ ret = do_s_update_ref(transaction, ref, check_old, !!our_transaction,

+                             &err, msg);
       if (ret) {
               error("%s", err.buf);
               ret = (ret == TRANSACTION_NAME_CONFLICT)
@@ -623,7 +632,7 @@ static int s_update_ref(const char *action,
                       : STORE_REF_ERROR_OTHER;
       }

- ref_transaction_free(transaction);

+       ref_transaction_free(our_transaction);
       strbuf_release(&err);
       free(msg);
       return ret;

...
-------------------------------------------------------------------------

You can find these patches, with Patrick as the author, there:

https://gitlab.com/gitlab-org/gitlab-git/-/commits/cc-improve-s-update-ref/

OK, I think the above as a preliminary clean-up would not hurt, but the "if the caller gave us NULL as the transaction, we are responsible for handling that bogosity" bit feels a bit too magical to my taste. I do understand that your intention is that it releaves the caller ...

When executing a fetch, then git will currently allocate one reference transaction per reference update and directly commit it. This means that fetches are non-atomic: even if some of the reference updates fail, others may still succeed and modify local references.

This is fine in many scenarios, but this strategy has its downsides.

- The view of remote references may be inconsistent and may show a bastardized state of the remote repository.

- Batching together updates may improve performance in certain scenarios. While the impact probably isn't as pronounced with loose references, the upcoming reftable backend may benefit as it needs to write less files in case the update is batched.

- The reference-update hook is currently being executed twice per updated reference. While this doesn't matter when there is no such hook, we have seen severe performance regressions when doing a git-fetch(1) with reference-transaction hook when the remote repository has hundreds of thousands of references.

Similar to git push --atomic, this commit thus introduces atomic fetches. Instead of allocating one reference transaction per updated reference, it causes us to only allocate a single transaction and commit it as soon as all updates were received. If locking of any reference fails, then we abort the complete transaction and don't update any reference, which gives us an all-or-nothing fetch.

Note that this may not completely fix the first of above downsides, as the consistent view also depends on the server-side. If the server doesn't have a consistent view of its own references during the reference negotiation phase, then the client would get the same inconsistent view the server has. This is a separate problem though and, if it actually exists, can be fixed at a later point.

---
 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 |  26 +++++-
 t/t5510-fetch.sh                | 139 ++++++++++++++++++++++++++++++++
 3 files changed, 167 insertions(+), 2 deletions(-)

index 2bf77b46fd..07783deee3 100644
@@ -7,6 +7,10 @@
 	existing contents of `.git/FETCH_HEAD`.  Without this
 	option old data in `.git/FETCH_HEAD` will be overwritten.
 
+--atomic::
+	Use an atomic transaction to update local refs. Either all refs are
+	updated, or on error, no refs are updated.
+
 --depth=<depth>::
 	Limit fetching to the specified number of commits from the tip of
 	each remote branch history. If fetching to a 'shallow' repository

index 020a977bc7..5675ae4ec3 100644
@@ -63,6 +63,7 @@ static int enable_auto_gc = 1;
 static int tags = TAGS_DEFAULT, unshallow, update_shallow, deepen;
 static int max_jobs = -1, submodule_fetch_jobs_config = -1;
 static int fetch_parallel_config = 1;
+static int atomic_fetch;
 static enum transport_family family;
 static const char *depth;
 static const char *deepen_since;

@@ -144,6 +145,8 @@ static struct option builtin_fetch_options[] = {
 		 N_("set upstream for git pull/fetch")),
 	OPT_BOOL('a', "append", &append,
 		 N_("append to .git/FETCH_HEAD instead of overwriting")),
+	OPT_BOOL(0, "atomic", &atomic_fetch,
+		 N_("use atomic transaction to update references")),
 	OPT_STRING(0, "upload-pack", &upload_pack, N_("path"),
 		   N_("path to upload pack on remote end")),
 	OPT__FORCE(&force, N_("force overwrite of local reference"), 0),

@@ -926,7 +929,8 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 	FILE *fp;
 	struct commit *commit;
 	int url_len, i, rc = 0;
-	struct strbuf note = STRBUF_INIT;
+	struct strbuf note = STRBUF_INIT, err = STRBUF_INIT;
+	struct ref_transaction *transaction = NULL;
 	const char *what, *kind;
 	struct ref *rm;
 	char *url;

@@ -955,6 +959,14 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 		}
 	}
 
+	if (atomic_fetch) {
+		transaction = ref_transaction_begin(&err);
+		if (!transaction) {
+			error("%s", err.buf);
+			goto abort;
+		}
+	}
+
 	prepare_format_display(ref_map);
 
 	/*

@@ -1048,7 +1060,7 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 
 			strbuf_reset(&note);
 			if (ref) {
-				rc |= update_local_ref(ref, NULL, what,
+				rc |= update_local_ref(ref, transaction, what,
 						       rm, &note, summary_width);
 				free(ref);
 			} else if (write_fetch_head || dry_run) {

@@ -1074,6 +1086,14 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 		}
 	}
 
+	if (!rc && transaction) {
+		rc = ref_transaction_commit(transaction, &err);
+		if (rc) {
+			error("%s", err.buf);
+			goto abort;
+		}
+	}
+
 	if (rc & STORE_REF_ERROR_DF_CONFLICT)
 		error(_("some local refs could not be updated; try running\n"
 		      " 'git remote prune %s' to remove any old, conflicting "

@@ -1090,6 +1110,8 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 
  abort:
 	strbuf_release(&note);
+	strbuf_release(&err);
+	ref_transaction_free(transaction);
 	free(url);
 	fclose(fp);
 	return rc;

2.30.0

With the help of the previous step, it is trivial to see that the single transaction around store_updated_refs() would be sufficient to achieve the atomicity.

But what about FETCH_HEAD? Do we refrain from writing and/or appending to it when there is any failure?

Thanks.

Hi,

this is the second version of my patch series to implement support for atomic reference updates for git-fetch(1). It's similar to git push --atomic, only that it applies to the local side. That is the fetch will either succeed and update all remote references or it will fail and update none.

Changes compared to v1:

- In v1, we still wrote to FETCH_HEAD even if the fetch failed. I've fixed this now by pulling out logic to write to FETCH_HEAD in 1/4 so that we can now easily buffer updates and commit them only if the reference transaction succeeds. There's some additional tests in 4/4 to test it works as expected.

- As suggested by Christian, I've unified the exit path in s_update_ref() in 2/4.

- I've dropped the commit variable and renamed transaction_to_free to our_transaction as suggested by Junio.

Patrick

Patrick Steinhardt (4):
fetch: extract writing to FETCH_HEAD
fetch: refactor s_update_ref to use common exit path
fetch: allow passing a transaction to s_update_ref()
fetch: implement support for atomic reference updates

 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 | 198 ++++++++++++++++++++++++--------
 t/t5510-fetch.sh                | 168 +++++++++++++++++++++++++++
 3 files changed, 322 insertions(+), 48 deletions(-)

Range-diff against v1:

-:  ---------- > 1:  d80dbc5a9c fetch: extract writing to FETCH_HEAD
-:  ---------- > 2:  718a8bf5d7 fetch: refactor `s_update_ref` to use common exit path
1:  e627e729e5 ! 3:  4162d10fcb fetch: allow passing a transaction to `s_update_ref()`
    @@ builtin/fetch.c: static struct ref *get_ref_map(struct remote *remote,
      	char *msg;
      	char *rla = getenv("GIT_REFLOG_ACTION");
     -	struct ref_transaction *transaction;
    -+	struct ref_transaction *transaction_to_free = NULL;
    ++	struct ref_transaction *our_transaction = NULL;
      	struct strbuf err = STRBUF_INIT;
    --	int ret, df_conflict = 0;
    -+	int ret, df_conflict = 0, commit = 0;
    + 	int ret;
      
    - 	if (dry_run)
    - 		return 0;
     @@ builtin/fetch.c: static int s_update_ref(const char *action,
      		rla = default_rla.buf;
      	msg = xstrfmt("%s: %s", rla, action);
      
     -	transaction = ref_transaction_begin(&err);
    --	if (!transaction ||
    --	    ref_transaction_update(transaction, ref->name,
     +	/*
     +	 * If no transaction was passed to us, we manage the transaction
     +	 * ourselves. Otherwise, we trust the caller to handle the transaction
     +	 * lifecycle.
     +	 */
    -+	if (!transaction) {
    -+		transaction = transaction_to_free = ref_transaction_begin(&err);
    -+		if (!transaction)
    -+			goto fail;
    -+		commit = 1;
    -+	}
    -+
    -+	if (ref_transaction_update(transaction, ref->name,
    - 				   &ref->new_oid,
    - 				   check_old ? &ref->old_oid : NULL,
    - 				   0, msg, &err))
    - 		goto fail;
    - 
    --	ret = ref_transaction_commit(transaction, &err);
    --	if (ret) {
    --		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
    --		goto fail;
    -+	if (commit) {
    -+		ret = ref_transaction_commit(transaction, &err);
    -+		if (ret) {
    -+			df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
    -+			goto fail;
    + 	if (!transaction) {
    +-		ret = STORE_REF_ERROR_OTHER;
    +-		goto out;
    ++		transaction = our_transaction = ref_transaction_begin(&err);
    ++		if (!transaction) {
    ++			ret = STORE_REF_ERROR_OTHER;
    ++			goto out;
     +		}
      	}
      
    + 	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,
    +@@ builtin/fetch.c: static int s_update_ref(const char *action,
    + 		goto out;
    + 	}
    + 
    +-	ret = ref_transaction_commit(transaction, &err);
    +-	if (ret) {
    +-		ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
    +-							 : STORE_REF_ERROR_OTHER;
    +-		goto out;
    ++	if (our_transaction) {
    ++		ret = ref_transaction_commit(our_transaction, &err);
    ++		if (ret) {
    ++			ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
    ++								 : STORE_REF_ERROR_OTHER;
    ++			goto out;
    ++		}
    + 	}
    + 
    + out:
     -	ref_transaction_free(transaction);
    -+	ref_transaction_free(transaction_to_free);
    ++	ref_transaction_free(our_transaction);
    + 	if (ret)
    + 		error("%s", err.buf);
      	strbuf_release(&err);
    - 	free(msg);
    - 	return 0;
    - fail:
    --	ref_transaction_free(transaction);
    -+	ref_transaction_free(transaction_to_free);
    - 	error("%s", err.buf);
    - 	strbuf_release(&err);
    - 	free(msg);
     @@ builtin/fetch.c: static void format_display(struct strbuf *display, char code,
      }

2: 4807344e92 ! 4: 53705281b6 fetch: implement support for atomic reference updates

    @@ Commit message
         inconsistent view the server has. This is a separate problem though and,
         if it actually exists, can be fixed at a later point.
     
    +    This commit also changes the way we write FETCH_HEAD in case `--atomic`
    +    is passed. Instead of writing changes as we go, we need to accumulate
    +    all changes first and only commit them at the end when we know that all
    +    reference updates succeeded. Ideally, we'd just do so via a temporary
    +    file so that we don't need to carry all updates in-memory. This isn't
    +    trivially doable though considering the `--append` mode, where we do not
    +    truncate the file but simply append to it. And given that we support
    +    concurrent processes appending to FETCH_HEAD at the same time without
    +    any loss of data, seeding the temporary file with current contents of
    +    FETCH_HEAD initially and then doing a rename wouldn't work either. So
    +    this commit implements the simple strategy of buffering all changes and
    +    appending them to the file on commit.
    +
         Signed-off-by: Patrick Steinhardt <[email protected]>
     
      ## Documentation/fetch-options.txt ##
    @@ builtin/fetch.c: static struct option builtin_fetch_options[] = {
      	OPT_STRING(0, "upload-pack", &upload_pack, N_("path"),
      		   N_("path to upload pack on remote end")),
      	OPT__FORCE(&force, N_("force overwrite of local reference"), 0),
    -@@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
    +@@ builtin/fetch.c: static int iterate_ref_map(void *cb_data, struct object_id *oid)
    + 
    + struct fetch_head {
      	FILE *fp;
    ++	struct strbuf buf;
    + };
    + 
    + static int open_fetch_head(struct fetch_head *fetch_head)
    +@@ builtin/fetch.c: static int open_fetch_head(struct fetch_head *fetch_head)
    + 	if (!write_fetch_head)
    + 		return 0;
    + 
    ++	strbuf_init(&fetch_head->buf, 0);
    + 	fetch_head->fp = fopen(filename, "a");
    + 	if (!fetch_head->fp)
    + 		return error_errno(_("cannot open %s"), filename);
    +@@ builtin/fetch.c: static void append_fetch_head(struct fetch_head *fetch_head, const char *old_oid
    + 	if (!write_fetch_head)
    + 		return;
    + 
    +-	fprintf(fetch_head->fp, "%s\t%s\t%s",
    +-		old_oid, merge_status_marker, note);
    +-	for (i = 0; i < url_len; ++i)
    +-		if ('\n' == url[i])
    +-			fputs("\\n", fetch_head->fp);
    +-		else
    +-			fputc(url[i], fetch_head->fp);
    +-	fputc('\n', fetch_head->fp);
    ++	/*
    ++	 * When using an atomic fetch, we do not want to update FETCH_HEAD if
    ++	 * any of the reference updates fails. We thus have to write all
    ++	 * updates to a buffer first and only commit it as soon as all
    ++	 * references have been successfully updated.
    ++	 */
    ++	if (atomic_fetch) {
    ++		strbuf_addf(&fetch_head->buf, "%s\t%s\t%s",
    ++			    old_oid, merge_status_marker, note);
    ++		strbuf_add(&fetch_head->buf, url, url_len);
    ++		strbuf_addch(&fetch_head->buf, '\n');
    ++	} else {
    ++		fprintf(fetch_head->fp, "%s\t%s\t%s",
    ++			old_oid, merge_status_marker, note);
    ++		for (i = 0; i < url_len; ++i)
    ++			if ('\n' == url[i])
    ++				fputs("\\n", fetch_head->fp);
    ++			else
    ++				fputc(url[i], fetch_head->fp);
    ++		fputc('\n', fetch_head->fp);
    ++	}
    + }
    + 
    + static void commit_fetch_head(struct fetch_head *fetch_head)
    + {
    +-	/* Nothing to commit yet. */
    ++	if (!write_fetch_head || !atomic_fetch)
    ++		return;
    ++	strbuf_write(&fetch_head->buf, fetch_head->fp);
    + }
    + 
    + static void close_fetch_head(struct fetch_head *fetch_head)
    +@@ builtin/fetch.c: static void close_fetch_head(struct fetch_head *fetch_head)
    + 		return;
    + 
    + 	fclose(fetch_head->fp);
    ++	strbuf_release(&fetch_head->buf);
    + }
    + 
    + static const char warn_show_forced_updates[] =
    +@@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
    + 	struct fetch_head fetch_head;
      	struct commit *commit;
      	int url_len, i, rc = 0;
     -	struct strbuf note = STRBUF_INIT;
    @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *
     +		}
     +	}
     +
    - 	if (rc & STORE_REF_ERROR_DF_CONFLICT)
    - 		error(_("some local refs could not be updated; try running\n"
    - 		      " 'git remote prune %s' to remove any old, conflicting "
    + 	if (!rc)
    + 		commit_fetch_head(&fetch_head);
    + 
     @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
      
       abort:
    @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *
     +	strbuf_release(&err);
     +	ref_transaction_free(transaction);
      	free(url);
    - 	fclose(fp);
    + 	close_fetch_head(&fetch_head);
      	return rc;
    +@@ builtin/fetch.c: int cmd_fetch(int argc, const char **argv, const char *prefix)
    + 			die(_("--filter can only be used with the remote "
    + 			      "configured in extensions.partialclone"));
    + 
    ++		if (atomic_fetch)
    ++			die(_("--atomic can only be used when fetching "
    ++			      "from one remote"));
    ++
    + 		if (stdin_refspecs)
    + 			die(_("--stdin can only be used when fetching "
    + 			      "from one remote"));
     
      ## t/t5510-fetch.sh ##
     @@ t/t5510-fetch.sh: test_expect_success 'fetch --prune --tags with refspec prunes based on refspec'
    @@ t/t5510-fetch.sh: test_expect_success 'fetch --prune --tags with refspec prunes
     +	cd "$D" &&
     +	git clone . atomic &&
     +	git branch atomic-branch &&
    -+	git rev-parse atomic-branch >expected &&
    ++	oid=$(git rev-parse atomic-branch) &&
    ++	echo "$oid" >expected &&
     +
     +	git -C atomic fetch --atomic origin &&
     +	git -C atomic rev-parse origin/atomic-branch >actual &&
    -+	test_cmp expected actual
    ++	test_cmp expected actual &&
    ++	test $oid = "$(git -C atomic rev-parse --verify FETCH_HEAD)"
     +'
     +
     +test_expect_success 'fetch --atomic works with multiple branches' '
    @@ t/t5510-fetch.sh: test_expect_success 'fetch --prune --tags with refspec prunes
     +	test_must_fail git -C atomic fetch --atomic origin refs/heads/*:refs/remotes/origin/* &&
     +	test_must_fail git -C atomic rev-parse refs/remotes/origin/atomic-new-branch &&
     +	git -C atomic rev-parse refs/remotes/origin/atomic-non-ff >expected &&
    -+	test_cmp expected actual
    ++	test_cmp expected actual &&
    ++	test_must_be_empty atomic/.git/FETCH_HEAD
     +'
     +
     +test_expect_success 'fetch --atomic executes a single reference transaction only' '
    @@ t/t5510-fetch.sh: test_expect_success 'fetch --prune --tags with refspec prunes
     +	git -C atomic for-each-ref >expected-refs &&
     +	test_must_fail git -C atomic fetch --tags --atomic origin &&
     +	git -C atomic for-each-ref >actual-refs &&
    -+	test_cmp expected-refs actual-refs
    ++	test_cmp expected-refs actual-refs &&
    ++	test_must_be_empty atomic/.git/FETCH_HEAD
    ++'
    ++
    ++test_expect_success 'fetch --atomic --append appends to FETCH_HEAD' '
    ++	test_when_finished "rm -rf \"$D\"/atomic" &&
    ++
    ++	cd "$D" &&
    ++	git clone . atomic &&
    ++	oid=$(git rev-parse HEAD) &&
    ++
    ++	git branch atomic-fetch-head-1 &&
    ++	git -C atomic fetch --atomic origin atomic-fetch-head-1 &&
    ++	test_line_count = 1 atomic/.git/FETCH_HEAD &&
    ++
    ++	git branch atomic-fetch-head-2 &&
    ++	git -C atomic fetch --atomic --append origin atomic-fetch-head-2 &&
    ++	test_line_count = 2 atomic/.git/FETCH_HEAD &&
    ++	cp atomic/.git/FETCH_HEAD expected &&
    ++
    ++	write_script atomic/.git/hooks/reference-transaction <<-\EOF &&
    ++		exit 1
    ++	EOF
    ++
    ++	git branch atomic-fetch-head-3 &&
    ++	test_must_fail git -C atomic fetch --atomic --append origin atomic-fetch-head-3 &&
    ++	test_cmp expected atomic/.git/FETCH_HEAD
     +'
     +
      test_expect_success '--refmap="" ignores configured refspec' '

2.30.0

When performing a fetch with the default --write-fetch-head option, we write all updated references to FETCH_HEAD while the updates are performed. Given that updates are not performed atomically, it means that we we write to FETCH_HEAD even if some or all of the reference updates fail.

Given that we simply update FETCH_HEAD ad-hoc with each reference, the logic is completely contained in store_update_refs and thus quite hard to extend. This can already be seen by the way we skip writing to the
FETCH_HEAD: instead of having a conditional which simply skips writing, we instead open "/dev/null" and needlessly write all updates there.

We are about to extend git-fetch(1) to accept an --atomic flag which will make the fetch an all-or-nothing operation with regards to the reference updates. This will also require us to make the updates to FETCH_HEAD an all-or-nothing operation, but as explained doing so is not easy with the current layout. This commit thus refactors the wa we write to FETCH_HEAD and pulls out the logic to open, append to, commit and close the file. While this may seem rather over-the top at first, pulling out this logic will make it a lot easier to update the code in a subsequent commit. It also allows us to easily skip writing completely in case --no-write-fetch-head was passed.

---
 builtin/fetch.c | 80 ++++++++++++++++++++++++++++++++++++++-----------
 1 file changed, 62 insertions(+), 18 deletions(-)

index ecf8537605..557ec130db 100644
@@ -897,6 +897,56 @@ static int iterate_ref_map(void *cb_data, struct object_id *oid)
 	return 0;
 }
 
+struct fetch_head {
+	FILE *fp;
+};
+
+static int open_fetch_head(struct fetch_head *fetch_head)
+{
+	const char *filename = git_path_fetch_head(the_repository);
+
+	if (!write_fetch_head)
+		return 0;
+
+	fetch_head->fp = fopen(filename, "a");
+	if (!fetch_head->fp)
+		return error_errno(_("cannot open %s"), filename);
+
+	return 0;
+}
+
+static void append_fetch_head(struct fetch_head *fetch_head, const char *old_oid,
+			      const char *merge_status_marker, const char *note,
+			      const char *url, size_t url_len)
+{
+	size_t i;
+
+	if (!write_fetch_head)
+		return;
+
+	fprintf(fetch_head->fp, "%s\t%s\t%s",
+		old_oid, merge_status_marker, note);
+	for (i = 0; i < url_len; ++i)
+		if ('\n' == url[i])
+			fputs("\\n", fetch_head->fp);
+		else
+			fputc(url[i], fetch_head->fp);
+	fputc('\n', fetch_head->fp);
+}
+
+static void commit_fetch_head(struct fetch_head *fetch_head)
+{
+	/* Nothing to commit yet. */
+}
+
+static void close_fetch_head(struct fetch_head *fetch_head)
+{
+	if (!write_fetch_head)
+		return;
+
+	fclose(fetch_head->fp);
+}
+
 static const char warn_show_forced_updates[] =
 N_("Fetch normally indicates which branches had a forced update,\n"
    "but that check has been disabled. To re-enable, use '--show-forced-updates'\n"

@@ -909,22 +959,19 @@ N_("It took %.2f seconds to check forced updates. You can use\n"
 static int store_updated_refs(const char *raw_url, const char *remote_name,
 			      int connectivity_checked, struct ref *ref_map)
 {
-	FILE *fp;
+	struct fetch_head fetch_head;
 	struct commit *commit;
 	int url_len, i, rc = 0;
 	struct strbuf note = STRBUF_INIT;
 	const char *what, *kind;
 	struct ref *rm;
 	char *url;
-	const char *filename = (!write_fetch_head
-				? "/dev/null"
-				: git_path_fetch_head(the_repository));
 	int want_status;
 	int summary_width = transport_summary_width(ref_map);
 
-	fp = fopen(filename, "a");
-	if (!fp)
-		return error_errno(_("cannot open %s"), filename);
+	rc = open_fetch_head(&fetch_head);
+	if (rc)
+		return -1;
 
 	if (raw_url)
 		url = transport_anonymize_url(raw_url);

@@ -1016,16 +1063,10 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 				merge_status_marker = "not-for-merge";
 				/* fall-through */
 			case FETCH_HEAD_MERGE:
-				fprintf(fp, "%s\t%s\t%s",
-					oid_to_hex(&rm->old_oid),
-					merge_status_marker,
-					note.buf);
-				for (i = 0; i < url_len; ++i)
-					if ('\n' == url[i])
-						fputs("\\n", fp);
-					else
-						fputc(url[i], fp);
-				fputc('\n', fp);
+				append_fetch_head(&fetch_head,
+						  oid_to_hex(&rm->old_oid),
+						  merge_status_marker,
+						  note.buf, url, url_len);
 				break;
 			default:
 				/* do not write anything to FETCH_HEAD */

@@ -1060,6 +1101,9 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 		}
 	}
 
+	if (!rc)
+		commit_fetch_head(&fetch_head);
+
 	if (rc & STORE_REF_ERROR_DF_CONFLICT)
 		error(_("some local refs could not be updated; try running\n"
 		      " 'git remote prune %s' to remove any old, conflicting "

@@ -1077,7 +1121,7 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
  abort:
 	strbuf_release(&note);
 	free(url);
-	fclose(fp);
+	close_fetch_head(&fetch_head);
 	return rc;
 }
 

2.30.0

It is clear from the type these days but variable names like "old_oid" hint the readers that they are not a hexadecimal object name string but either an array of uchar[40] or a struct object_id; perhaps "old_oid_hex" would be less misleading.

If the caller does have struct object_id, then it would be even better to take it as-is as a parameter and use oid_to_hex_r() on it in this function when it is given to fprintf(). [Nit #1]

		switch (rm->fetch_head_status) {
		case FETCH_HEAD_NOT_FOR_MERGE:
		case FETCH_HEAD_MERGE:
			append_fetch_head(&fetch_head, &rm->old_oid,
				rm->fetch_head_status == FETCH_HEAD_MERGE,
                                note.buf, url, url_len);

Agreed.

[snip]

I think it's even nicer to just pass in rm->fetch_head_status directly, which allows us to move below switch into append_fetch_head.

[snip]

I like your proposal more. I wasn't quite happy with leaving fp uninitialized, and using it as a sentinel for whether to write something or not feels natural to me.

[snip]

As said above, I'm just moving this complete switch into append_fetch_head and pass rm->fetch_head_status.

Patrick

OK. That may even be better.

Thanks.

The cleanup code in s_update_ref() is currently duplicated for both succesful and erroneous exit paths. This commit refactors the function to have a shared exit path for both cases to remove the duplication.

---
 builtin/fetch.c | 37 ++++++++++++++++++++-----------------
 1 file changed, 20 insertions(+), 17 deletions(-)

index 557ec130db..5221a9dbed 100644
@@ -589,7 +589,7 @@ static int s_update_ref(const char *action,
 	char *rla = getenv("GIT_REFLOG_ACTION");
 	struct ref_transaction *transaction;
 	struct strbuf err = STRBUF_INIT;
-	int ret, df_conflict = 0;
+	int ret;
 
 	if (dry_run)
 		return 0;

@@ -598,30 +598,33 @@ static int s_update_ref(const char *action,
 	msg = xstrfmt("%s: %s", rla, action);
 
 	transaction = ref_transaction_begin(&err);
-	if (!transaction ||
-	    ref_transaction_update(transaction, ref->name,
-				   &ref->new_oid,
-				   check_old ? &ref->old_oid : NULL,
-				   0, msg, &err))
-		goto fail;
+	if (!transaction) {
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
+	}
+
+	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,
+				     check_old ? &ref->old_oid : NULL,
+				     0, msg, &err);
+	if (ret) {
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
+	}
 
 	ret = ref_transaction_commit(transaction, &err);
 	if (ret) {
-		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
-		goto fail;
+		ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
+							 : STORE_REF_ERROR_OTHER;
+		goto out;
 	}
 
+out:
 	ref_transaction_free(transaction);
+	if (ret)
+		error("%s", err.buf);
 	strbuf_release(&err);
 	free(msg);
-	return 0;
-fail:
-	ref_transaction_free(transaction);
-	error("%s", err.buf);
-	strbuf_release(&err);
-	free(msg);
-	return df_conflict ? STORE_REF_ERROR_DF_CONFLICT
-			   : STORE_REF_ERROR_OTHER;
+	return ret;
 }
 
 static int refcol_width = 10;

2.30.0

[snip]

Agreed, that is easier to read. Thanks!

Patrick

The handling of ref updates is completely handled by s_update_ref(), which will manage the complete lifecycle of the reference transaction. This is fine right now given that git-fetch(1) does not support atomic fetches, so each reference gets its own transaction. It is quite inflexible though, as s_update_ref() only knows about a single reference update at a time, so it doesn't allow us to alter the strategy.

This commit prepares s_update_ref() and its only caller update_local_ref() to allow passing an external transaction. If none is given, then the existing behaviour is triggered which creates a new transaction and directly commits it. Otherwise, if the caller provides a transaction, then we only queue the update but don't commit it. This optionally allows the caller to manage when a transaction will be committed.

Given that update_local_ref() is always called with a NULL transaction for now, no change in behaviour is expected from this change.

---
 builtin/fetch.c | 43 +++++++++++++++++++++++++++----------------
 1 file changed, 27 insertions(+), 16 deletions(-)

index 5221a9dbed..654e0bb520 100644
@@ -583,11 +583,12 @@ static struct ref *get_ref_map(struct remote *remote,
 
 static int s_update_ref(const char *action,
 			struct ref *ref,
+			struct ref_transaction *transaction,
 			int check_old)
 {
 	char *msg;
 	char *rla = getenv("GIT_REFLOG_ACTION");
-	struct ref_transaction *transaction;
+	struct ref_transaction *our_transaction = NULL;
 	struct strbuf err = STRBUF_INIT;
 	int ret;
 

@@ -597,10 +598,17 @@ static int s_update_ref(const char *action,
 		rla = default_rla.buf;
 	msg = xstrfmt("%s: %s", rla, action);
 
-	transaction = ref_transaction_begin(&err);
+	/*
+	 * If no transaction was passed to us, we manage the transaction
+	 * ourselves. Otherwise, we trust the caller to handle the transaction
+	 * lifecycle.
+	 */
 	if (!transaction) {
-		ret = STORE_REF_ERROR_OTHER;
-		goto out;
+		transaction = our_transaction = ref_transaction_begin(&err);
+		if (!transaction) {
+			ret = STORE_REF_ERROR_OTHER;
+			goto out;
+		}
 	}
 
 	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,

@@ -611,15 +619,17 @@ static int s_update_ref(const char *action,
 		goto out;
 	}
 
-	ret = ref_transaction_commit(transaction, &err);
-	if (ret) {
-		ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
-							 : STORE_REF_ERROR_OTHER;
-		goto out;
+	if (our_transaction) {
+		ret = ref_transaction_commit(our_transaction, &err);
+		if (ret) {
+			ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
+								 : STORE_REF_ERROR_OTHER;
+			goto out;
+		}
 	}
 
 out:
-	ref_transaction_free(transaction);
+	ref_transaction_free(our_transaction);
 	if (ret)
 		error("%s", err.buf);
 	strbuf_release(&err);

@@ -762,6 +772,7 @@ static void format_display(struct strbuf *display, char code,
 }
 
 static int update_local_ref(struct ref *ref,
+			    struct ref_transaction *transaction,
 			    const char *remote,
 			    const struct ref *remote_ref,
 			    struct strbuf *display,

@@ -802,7 +813,7 @@ static int update_local_ref(struct ref *ref,
 	    starts_with(ref->name, "refs/tags/")) {
 		if (force || ref->force) {
 			int r;
-			r = s_update_ref("updating tag", ref, 0);
+			r = s_update_ref("updating tag", ref, transaction, 0);
 			format_display(display, r ? '!' : 't', _("[tag update]"),
 				       r ? _("unable to update local ref") : NULL,
 				       remote, pretty_ref, summary_width);

@@ -839,7 +850,7 @@ static int update_local_ref(struct ref *ref,
 			what = _("[new ref]");
 		}
 
-		r = s_update_ref(msg, ref, 0);
+		r = s_update_ref(msg, ref, transaction, 0);
 		format_display(display, r ? '!' : '*', what,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -861,7 +872,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "..");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("fast-forward", ref, 1);
+		r = s_update_ref("fast-forward", ref, transaction, 1);
 		format_display(display, r ? '!' : ' ', quickref.buf,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -873,7 +884,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "...");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("forced-update", ref, 1);
+		r = s_update_ref("forced-update", ref, transaction, 1);
 		format_display(display, r ? '!' : '+', quickref.buf,
 			       r ? _("unable to update local ref") : _("forced update"),
 			       remote, pretty_ref, summary_width);

@@ -1078,8 +1089,8 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 
 			strbuf_reset(&note);
 			if (ref) {
-				rc |= update_local_ref(ref, what, rm, &note,
-						       summary_width);
+				rc |= update_local_ref(ref, NULL, what,
+						       rm, &note, summary_width);
 				free(ref);
 			} else if (write_fetch_head || dry_run) {
 				/*

2.30.0

When executing a fetch, then git will currently allocate one reference transaction per reference update and directly commit it. This means that fetches are non-atomic: even if some of the reference updates fail, others may still succeed and modify local references.

This is fine in many scenarios, but this strategy has its downsides.

- The view of remote references may be inconsistent and may show a bastardized state of the remote repository.

- Batching together updates may improve performance in certain scenarios. While the impact probably isn't as pronounced with loose references, the upcoming reftable backend may benefit as it needs to write less files in case the update is batched.

- The reference-update hook is currently being executed twice per updated reference. While this doesn't matter when there is no such hook, we have seen severe performance regressions when doing a git-fetch(1) with reference-transaction hook when the remote repository has hundreds of thousands of references.

Similar to git push --atomic, this commit thus introduces atomic fetches. Instead of allocating one reference transaction per updated reference, it causes us to only allocate a single transaction and commit it as soon as all updates were received. If locking of any reference fails, then we abort the complete transaction and don't update any reference, which gives us an all-or-nothing fetch.

Note that this may not completely fix the first of above downsides, as the consistent view also depends on the server-side. If the server doesn't have a consistent view of its own references during the reference negotiation phase, then the client would get the same inconsistent view the server has. This is a separate problem though and, if it actually exists, can be fixed at a later point.

This commit also changes the way we write FETCH_HEAD in case --atomic is passed. Instead of writing changes as we go, we need to accumulate all changes first and only commit them at the end when we know that all reference updates succeeded. Ideally, we'd just do so via a temporary file so that we don't need to carry all updates in-memory. This isn't trivially doable though considering the --append mode, where we do not truncate the file but simply append to it. And given that we support concurrent processes appending to FETCH_HEAD at the same time without any loss of data, seeding the temporary file with current contents of FETCH_HEAD initially and then doing a rename wouldn't work either. So this commit implements the simple strategy of buffering all changes and appending them to the file on commit.

---
 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 |  66 ++++++++++---
 t/t5510-fetch.sh                | 168 ++++++++++++++++++++++++++++++++
 3 files changed, 227 insertions(+), 11 deletions(-)

index 2bf77b46fd..07783deee3 100644
@@ -7,6 +7,10 @@
 	existing contents of `.git/FETCH_HEAD`.  Without this
 	option old data in `.git/FETCH_HEAD` will be overwritten.
 
+--atomic::
+	Use an atomic transaction to update local refs. Either all refs are
+	updated, or on error, no refs are updated.
+
 --depth=<depth>::
 	Limit fetching to the specified number of commits from the tip of
 	each remote branch history. If fetching to a 'shallow' repository

2.30.0

Good point. I'll unify those code paths.

This is a change I'm hesitant to make. The thing is that FETCH_HEAD is quite weird as the current design allows different git fetch --append processes to write to FETCH_HEAD at the same time. If we change to always accumulate first and flush once we're done, then we essentially have a change in behaviour as FETCH_HEADs aren't written in an interleaving fashion anymore, but only at the end. Also, if there is any unexpected error in between updating references which causes us to die, then we wouldn't have written the already updated references to FETCH_HEAD now.

So I'm all in when it comes to merging formatting directives, even more so as I have missed the "\\n" weirdness. But for now, I'll keep the current flushing semantics in the non-atomic case. Please let me know if you're not convinced and I'll have another look for v4.

Patrick

Hmph, do you mean

	git fetch --append remote-a & git fetch --append remote-b

or something else [*1*]?

With the current write-out codepath, there is no guarantee that ...

... these stdio operations for a single record would result in a single atomic write() that would not compete with writes by other processes. So I wouldn't call "the current design allows ... at the same time." It has been broken for years and nobody complained.

I view it a good thing, actually, for another reason [*2*], but that would take a follow-up fix that is much more involved, so let's not go there (yet). At least buffering a single line's worth of output in a buffer and writing it out in a single write() would get us much closer to solving the "mixed up output" from multiple processes problem the current code seems to already have.

That one may be a valid concern.

Thanks.

[Footnote]

*1* "git fetch --all/--multiple" was strictly serial operation to
avoid such a mixed-up output problem, but perhaps we weren't careful enough when we introduced parallel fetch and broke it?

*2* When fetching from a single remote, the code makes sure that a
record that is not marked as 'not-for-merge' is listed first by reordering the records, but it does not work well when fetching from multiple remotes. Queuing all in the buffer before showing them would give us an opportunity to sort, but as I said, it takes coordination among the processes---instead of each process writing to FETCH_HEAD on their own, somebody has to collect the records from them, reorder as needed and write them all out.

cf. https://lore.kernel.org/git/X8oL190Vl03B0cQ%[email protected]/

That's what I mean.

Fair enough. I somehow blindly assumed that git fetch --all, which does invoke git fetch --append, did perform the fetch in parallel. If that's not the case: all the better.

The buffering of a single line is exactly what we're doing now in the non-atomic case. Previously there had been multiple writes, now we only call strbuf_write() once on the buffer for each reference.

Just to be explicit: are you fine with changes in this patch as-is? They don't solve concurrent writes to FETCH_HEAD, but they should make it easier to solve that if we ever need to.

Patrick

[snip]

Never mind, I forgot that I'm still replying on v2 of this patch series.

Patrick

The "--all" option started as "one after another, one at a time", but these days goes thru fetch_multiple() where we started to use run_processes_parallel() API without giving it much thought what it would do to the writing of FETCH_HEAD; since around d54dea77 (fetch:
let --jobs=<n> parallelize --multiple, too, 2019-10-05), this codepath wrt FETCH_HEAD is utterly broken, I would have to say.

Exactly.

Hi,

this is the third version of my patch series to implement support for atomic reference updates for git-fetch(1). It's similar to git push --atomic, only that it applies to the local side. That is the fetch will either succeed and update all remote references or it will fail and update none.

Changes compared to v2:

- append_fetch_head now takes a struct object_id *old_oid instead of a char *old_oid.

- Handling of the merge status marker was moved into append_fetch_head.

- I've unified code paths to format the line we're about to write to FETCH_HEAD in atomic and non-atomic cases, which is the new 2/5 patch.

- We now always initialize fetch_head->fp and use it to derive whether anything should be written at all instead of using the global write_fetch_head variable.

I think this should address all of Junio's feedback. Thanks for your input!

Patrick

Patrick Steinhardt (5):
fetch: extract writing to FETCH_HEAD
fetch: use strbuf to format FETCH_HEAD updates
fetch: refactor s_update_ref to use common exit path
fetch: allow passing a transaction to s_update_ref()
fetch: implement support for atomic reference updates

 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 | 228 +++++++++++++++++++++++---------
 remote.h                        |   2 +-
 t/t5510-fetch.sh                | 168 +++++++++++++++++++++++
 4 files changed, 342 insertions(+), 60 deletions(-)

Range-diff against v2:
1: d80dbc5a9c ! 1: 61dc19a1ca fetch: extract writing to FETCH_HEAD

    @@ builtin/fetch.c: static int iterate_ref_map(void *cb_data, struct object_id *oid
     +{
     +	const char *filename = git_path_fetch_head(the_repository);
     +
    -+	if (!write_fetch_head)
    -+		return 0;
    -+
    -+	fetch_head->fp = fopen(filename, "a");
    -+	if (!fetch_head->fp)
    -+		return error_errno(_("cannot open %s"), filename);
    ++	if (write_fetch_head) {
    ++		fetch_head->fp = fopen(filename, "a");
    ++		if (!fetch_head->fp)
    ++			return error_errno(_("cannot open %s"), filename);
    ++	} else {
    ++		fetch_head->fp = NULL;
    ++	}
     +
     +	return 0;
     +}
     +
    -+static void append_fetch_head(struct fetch_head *fetch_head, const char *old_oid,
    -+			      const char *merge_status_marker, const char *note,
    ++static void append_fetch_head(struct fetch_head *fetch_head,
    ++			      const struct object_id *old_oid,
    ++			      enum fetch_head_status fetch_head_status,
    ++			      const char *note,
     +			      const char *url, size_t url_len)
     +{
    ++	char old_oid_hex[GIT_MAX_HEXSZ + 1];
    ++	const char *merge_status_marker;
     +	size_t i;
     +
    -+	if (!write_fetch_head)
    ++	if (!fetch_head->fp)
     +		return;
     +
    ++	switch (fetch_head_status) {
    ++		case FETCH_HEAD_NOT_FOR_MERGE:
    ++			merge_status_marker = "not-for-merge";
    ++			break;
    ++		case FETCH_HEAD_MERGE:
    ++			merge_status_marker = "";
    ++			break;
    ++		default:
    ++			/* do not write anything to FETCH_HEAD */
    ++			return;
    ++	}
    ++
     +	fprintf(fetch_head->fp, "%s\t%s\t%s",
    -+		old_oid, merge_status_marker, note);
    ++		oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
     +	for (i = 0; i < url_len; ++i)
     +		if ('\n' == url[i])
     +			fputs("\\n", fetch_head->fp);
    @@ builtin/fetch.c: static int iterate_ref_map(void *cb_data, struct object_id *oid
     +
     +static void close_fetch_head(struct fetch_head *fetch_head)
     +{
    -+	if (!write_fetch_head)
    ++	if (!fetch_head->fp)
     +		return;
     +
     +	fclose(fetch_head->fp);
    @@ builtin/fetch.c: N_("It took %.2f seconds to check forced updates. You can use\n
      	if (raw_url)
      		url = transport_anonymize_url(raw_url);
     @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
    - 				merge_status_marker = "not-for-merge";
    - 				/* fall-through */
    - 			case FETCH_HEAD_MERGE:
    + 	     want_status++) {
    + 		for (rm = ref_map; rm; rm = rm->next) {
    + 			struct ref *ref = NULL;
    +-			const char *merge_status_marker = "";
    + 
    + 			if (rm->status == REF_STATUS_REJECT_SHALLOW) {
    + 				if (want_status == FETCH_HEAD_MERGE)
    +@@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
    + 					strbuf_addf(&note, "%s ", kind);
    + 				strbuf_addf(&note, "'%s' of ", what);
    + 			}
    +-			switch (rm->fetch_head_status) {
    +-			case FETCH_HEAD_NOT_FOR_MERGE:
    +-				merge_status_marker = "not-for-merge";
    +-				/* fall-through */
    +-			case FETCH_HEAD_MERGE:
     -				fprintf(fp, "%s\t%s\t%s",
     -					oid_to_hex(&rm->old_oid),
     -					merge_status_marker,
    @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *
     -					else
     -						fputc(url[i], fp);
     -				fputc('\n', fp);
    -+				append_fetch_head(&fetch_head,
    -+						  oid_to_hex(&rm->old_oid),
    -+						  merge_status_marker,
    -+						  note.buf, url, url_len);
    - 				break;
    - 			default:
    - 				/* do not write anything to FETCH_HEAD */
    +-				break;
    +-			default:
    +-				/* do not write anything to FETCH_HEAD */
    +-				break;
    +-			}
    ++
    ++			append_fetch_head(&fetch_head, &rm->old_oid,
    ++					  rm->fetch_head_status,
    ++					  note.buf, url, url_len);
    + 
    + 			strbuf_reset(&note);
    + 			if (ref) {
     @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
      		}
      	}
    @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *
      	return rc;
      }
      
    +
    + ## remote.h ##
    +@@ remote.h: struct ref {
    + 	 * should be 0, so that xcalloc'd structures get it
    + 	 * by default.
    + 	 */
    +-	enum {
    ++	enum fetch_head_status {
    + 		FETCH_HEAD_MERGE = -1,
    + 		FETCH_HEAD_NOT_FOR_MERGE = 0,
    + 		FETCH_HEAD_IGNORE = 1
-:  ---------- > 2:  a19762690e fetch: use strbuf to format FETCH_HEAD updates
2:  718a8bf5d7 ! 3:  c411f30e09 fetch: refactor `s_update_ref` to use common exit path
    @@ builtin/fetch.c: static int s_update_ref(const char *action,
     -				   check_old ? &ref->old_oid : NULL,
     -				   0, msg, &err))
     -		goto fail;
    +-
    +-	ret = ref_transaction_commit(transaction, &err);
    +-	if (ret) {
    +-		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
    +-		goto fail;
     +	if (!transaction) {
     +		ret = STORE_REF_ERROR_OTHER;
     +		goto out;
    -+	}
    -+
    + 	}
    + 
     +	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,
     +				     check_old ? &ref->old_oid : NULL,
     +				     0, msg, &err);
    @@ builtin/fetch.c: static int s_update_ref(const char *action,
     +		ret = STORE_REF_ERROR_OTHER;
     +		goto out;
     +	}
    - 
    - 	ret = ref_transaction_commit(transaction, &err);
    - 	if (ret) {
    --		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
    --		goto fail;
    -+		ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
    -+							 : STORE_REF_ERROR_OTHER;
    ++
    ++	switch (ref_transaction_commit(transaction, &err)) {
    ++	case 0:
    ++		break;
    ++	case TRANSACTION_NAME_CONFLICT:
    ++		ret = STORE_REF_ERROR_DF_CONFLICT;
     +		goto out;
    - 	}
    - 
    ++	default:
    ++		ret = STORE_REF_ERROR_OTHER;
    ++		goto out;
    ++	}
    ++
     +out:
      	ref_transaction_free(transaction);
     +	if (ret)
3:  4162d10fcb ! 4:  865d357ba7 fetch: allow passing a transaction to `s_update_ref()`
    @@ builtin/fetch.c: static int s_update_ref(const char *action,
      		goto out;
      	}
      
    --	ret = ref_transaction_commit(transaction, &err);
    --	if (ret) {
    --		ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
    --							 : STORE_REF_ERROR_OTHER;
    +-	switch (ref_transaction_commit(transaction, &err)) {
    +-	case 0:
    +-		break;
    +-	case TRANSACTION_NAME_CONFLICT:
    +-		ret = STORE_REF_ERROR_DF_CONFLICT;
    +-		goto out;
    +-	default:
    +-		ret = STORE_REF_ERROR_OTHER;
     -		goto out;
     +	if (our_transaction) {
    -+		ret = ref_transaction_commit(our_transaction, &err);
    -+		if (ret) {
    -+			ret = (ret == TRANSACTION_NAME_CONFLICT) ? STORE_REF_ERROR_DF_CONFLICT
    -+								 : STORE_REF_ERROR_OTHER;
    ++		switch (ref_transaction_commit(our_transaction, &err)) {
    ++		case 0:
    ++			break;
    ++		case TRANSACTION_NAME_CONFLICT:
    ++			ret = STORE_REF_ERROR_DF_CONFLICT;
    ++			goto out;
    ++		default:
    ++			ret = STORE_REF_ERROR_OTHER;
     +			goto out;
     +		}
      	}
4:  53705281b6 ! 5:  6a79e7adcc fetch: implement support for atomic reference updates
    @@ builtin/fetch.c: static struct option builtin_fetch_options[] = {
      	OPT_STRING(0, "upload-pack", &upload_pack, N_("path"),
      		   N_("path to upload pack on remote end")),
      	OPT__FORCE(&force, N_("force overwrite of local reference"), 0),
    -@@ builtin/fetch.c: static int iterate_ref_map(void *cb_data, struct object_id *oid)
    +@@ builtin/fetch.c: static void append_fetch_head(struct fetch_head *fetch_head,
    + 			strbuf_addch(&fetch_head->buf, url[i]);
    + 	strbuf_addch(&fetch_head->buf, '\n');
      
    - struct fetch_head {
    - 	FILE *fp;
    -+	struct strbuf buf;
    - };
    - 
    - static int open_fetch_head(struct fetch_head *fetch_head)
    -@@ builtin/fetch.c: static int open_fetch_head(struct fetch_head *fetch_head)
    - 	if (!write_fetch_head)
    - 		return 0;
    - 
    -+	strbuf_init(&fetch_head->buf, 0);
    - 	fetch_head->fp = fopen(filename, "a");
    - 	if (!fetch_head->fp)
    - 		return error_errno(_("cannot open %s"), filename);
    -@@ builtin/fetch.c: static void append_fetch_head(struct fetch_head *fetch_head, const char *old_oid
    - 	if (!write_fetch_head)
    - 		return;
    - 
    --	fprintf(fetch_head->fp, "%s\t%s\t%s",
    --		old_oid, merge_status_marker, note);
    --	for (i = 0; i < url_len; ++i)
    --		if ('\n' == url[i])
    --			fputs("\\n", fetch_head->fp);
    --		else
    --			fputc(url[i], fetch_head->fp);
    --	fputc('\n', fetch_head->fp);
    +-	strbuf_write(&fetch_head->buf, fetch_head->fp);
    +-	strbuf_reset(&fetch_head->buf);
     +	/*
     +	 * When using an atomic fetch, we do not want to update FETCH_HEAD if
     +	 * any of the reference updates fails. We thus have to write all
     +	 * updates to a buffer first and only commit it as soon as all
     +	 * references have been successfully updated.
     +	 */
    -+	if (atomic_fetch) {
    -+		strbuf_addf(&fetch_head->buf, "%s\t%s\t%s",
    -+			    old_oid, merge_status_marker, note);
    -+		strbuf_add(&fetch_head->buf, url, url_len);
    -+		strbuf_addch(&fetch_head->buf, '\n');
    -+	} else {
    -+		fprintf(fetch_head->fp, "%s\t%s\t%s",
    -+			old_oid, merge_status_marker, note);
    -+		for (i = 0; i < url_len; ++i)
    -+			if ('\n' == url[i])
    -+				fputs("\\n", fetch_head->fp);
    -+			else
    -+				fputc(url[i], fetch_head->fp);
    -+		fputc('\n', fetch_head->fp);
    ++	if (!atomic_fetch) {
    ++		strbuf_write(&fetch_head->buf, fetch_head->fp);
    ++		strbuf_reset(&fetch_head->buf);
     +	}
      }
      
      static void commit_fetch_head(struct fetch_head *fetch_head)
      {
     -	/* Nothing to commit yet. */
    -+	if (!write_fetch_head || !atomic_fetch)
    ++	if (!fetch_head->fp || !atomic_fetch)
     +		return;
     +	strbuf_write(&fetch_head->buf, fetch_head->fp);
      }
      
      static void close_fetch_head(struct fetch_head *fetch_head)
    -@@ builtin/fetch.c: static void close_fetch_head(struct fetch_head *fetch_head)
    - 		return;
    - 
    - 	fclose(fetch_head->fp);
    -+	strbuf_release(&fetch_head->buf);
    - }
    - 
    - static const char warn_show_forced_updates[] =
     @@ builtin/fetch.c: static int store_updated_refs(const char *raw_url, const char *remote_name,
      	struct fetch_head fetch_head;
      	struct commit *commit;

2.30.0

When performing a fetch with the default --write-fetch-head option, we write all updated references to FETCH_HEAD while the updates are performed. Given that updates are not performed atomically, it means that we we write to FETCH_HEAD even if some or all of the reference updates fail.

Given that we simply update FETCH_HEAD ad-hoc with each reference, the logic is completely contained in store_update_refs and thus quite hard to extend. This can already be seen by the way we skip writing to the
FETCH_HEAD: instead of having a conditional which simply skips writing, we instead open "/dev/null" and needlessly write all updates there.

We are about to extend git-fetch(1) to accept an --atomic flag which will make the fetch an all-or-nothing operation with regards to the reference updates. This will also require us to make the updates to FETCH_HEAD an all-or-nothing operation, but as explained doing so is not easy with the current layout. This commit thus refactors the wa we write to FETCH_HEAD and pulls out the logic to open, append to, commit and close the file. While this may seem rather over-the top at first, pulling out this logic will make it a lot easier to update the code in a subsequent commit. It also allows us to easily skip writing completely in case --no-write-fetch-head was passed.

---
 builtin/fetch.c | 108 +++++++++++++++++++++++++++++++++++-------------
 remote.h        |   2 +-
 2 files changed, 80 insertions(+), 30 deletions(-)

index ecf8537605..50f0306a92 100644
@@ -897,6 +897,73 @@ static int iterate_ref_map(void *cb_data, struct object_id *oid)
 	return 0;
 }
 
+struct fetch_head {
+	FILE *fp;
+};
+
+static int open_fetch_head(struct fetch_head *fetch_head)
+{
+	const char *filename = git_path_fetch_head(the_repository);
+
+	if (write_fetch_head) {
+		fetch_head->fp = fopen(filename, "a");
+		if (!fetch_head->fp)
+			return error_errno(_("cannot open %s"), filename);
+	} else {
+		fetch_head->fp = NULL;
+	}
+
+	return 0;
+}
+
+static void append_fetch_head(struct fetch_head *fetch_head,
+			      const struct object_id *old_oid,
+			      enum fetch_head_status fetch_head_status,
+			      const char *note,
+			      const char *url, size_t url_len)
+{
+	char old_oid_hex[GIT_MAX_HEXSZ + 1];
+	const char *merge_status_marker;
+	size_t i;
+
+	if (!fetch_head->fp)
+		return;
+
+	switch (fetch_head_status) {
+		case FETCH_HEAD_NOT_FOR_MERGE:
+			merge_status_marker = "not-for-merge";
+			break;
+		case FETCH_HEAD_MERGE:
+			merge_status_marker = "";
+			break;
+		default:
+			/* do not write anything to FETCH_HEAD */
+			return;
+	}
+
+	fprintf(fetch_head->fp, "%s\t%s\t%s",
+		oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
+	for (i = 0; i < url_len; ++i)
+		if ('\n' == url[i])
+			fputs("\\n", fetch_head->fp);
+		else
+			fputc(url[i], fetch_head->fp);
+	fputc('\n', fetch_head->fp);
+}
+
+static void commit_fetch_head(struct fetch_head *fetch_head)
+{
+	/* Nothing to commit yet. */
+}
+
+static void close_fetch_head(struct fetch_head *fetch_head)
+{
+	if (!fetch_head->fp)
+		return;
+
+	fclose(fetch_head->fp);
+}
+
 static const char warn_show_forced_updates[] =
 N_("Fetch normally indicates which branches had a forced update,\n"
    "but that check has been disabled. To re-enable, use '--show-forced-updates'\n"

@@ -909,22 +976,19 @@ N_("It took %.2f seconds to check forced updates. You can use\n"
 static int store_updated_refs(const char *raw_url, const char *remote_name,
 			      int connectivity_checked, struct ref *ref_map)
 {
-	FILE *fp;
+	struct fetch_head fetch_head;
 	struct commit *commit;
 	int url_len, i, rc = 0;
 	struct strbuf note = STRBUF_INIT;
 	const char *what, *kind;
 	struct ref *rm;
 	char *url;
-	const char *filename = (!write_fetch_head
-				? "/dev/null"
-				: git_path_fetch_head(the_repository));
 	int want_status;
 	int summary_width = transport_summary_width(ref_map);
 
-	fp = fopen(filename, "a");
-	if (!fp)
-		return error_errno(_("cannot open %s"), filename);
+	rc = open_fetch_head(&fetch_head);
+	if (rc)
+		return -1;
 
 	if (raw_url)
 		url = transport_anonymize_url(raw_url);

@@ -953,7 +1017,6 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 	     want_status++) {
 		for (rm = ref_map; rm; rm = rm->next) {
 			struct ref *ref = NULL;
-			const char *merge_status_marker = "";
 
 			if (rm->status == REF_STATUS_REJECT_SHALLOW) {
 				if (want_status == FETCH_HEAD_MERGE)

@@ -1011,26 +1074,10 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 					strbuf_addf(&note, "%s ", kind);
 				strbuf_addf(&note, "'%s' of ", what);
 			}
-			switch (rm->fetch_head_status) {
-			case FETCH_HEAD_NOT_FOR_MERGE:
-				merge_status_marker = "not-for-merge";
-				/* fall-through */
-			case FETCH_HEAD_MERGE:
-				fprintf(fp, "%s\t%s\t%s",
-					oid_to_hex(&rm->old_oid),
-					merge_status_marker,
-					note.buf);
-				for (i = 0; i < url_len; ++i)
-					if ('\n' == url[i])
-						fputs("\\n", fp);
-					else
-						fputc(url[i], fp);
-				fputc('\n', fp);
-				break;
-			default:
-				/* do not write anything to FETCH_HEAD */
-				break;
-			}
+
+			append_fetch_head(&fetch_head, &rm->old_oid,
+					  rm->fetch_head_status,
+					  note.buf, url, url_len);
 
 			strbuf_reset(&note);
 			if (ref) {

@@ -1060,6 +1107,9 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 		}
 	}
 
+	if (!rc)
+		commit_fetch_head(&fetch_head);
+
 	if (rc & STORE_REF_ERROR_DF_CONFLICT)
 		error(_("some local refs could not be updated; try running\n"
 		      " 'git remote prune %s' to remove any old, conflicting "

@@ -1077,7 +1127,7 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
  abort:
 	strbuf_release(&note);
 	free(url);
-	fclose(fp);
+	close_fetch_head(&fetch_head);
 	return rc;
 }
 

index 3211abdf05..aad1a0f080 100644
@@ -134,7 +134,7 @@ struct ref {
 	 * should be 0, so that xcalloc'd structures get it
 	 * by default.
 	 */
-	enum {
+	enum fetch_head_status {
 		FETCH_HEAD_MERGE = -1,
 		FETCH_HEAD_NOT_FOR_MERGE = 0,
 		FETCH_HEAD_IGNORE = 1

2.30.0

This commit refactors append_fetch_head() to use a struct strbuf for formatting the update which we're about to append to the FETCH_HEAD file. While the refactoring doesn't have much of a benefit right now, it servers as a preparatory step to implement atomic fetches where we need to buffer all updates to FETCH_HEAD and only flush them out if all reference updates succeeded.

No change in behaviour is expected from this commit.

---
 builtin/fetch.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

index 50f0306a92..1252f37493 100644
@@ -899,6 +899,7 @@ static int iterate_ref_map(void *cb_data, struct object_id *oid)
 
 struct fetch_head {
 	FILE *fp;
+	struct strbuf buf;
 };
 
 static int open_fetch_head(struct fetch_head *fetch_head)

@@ -909,6 +910,7 @@ static int open_fetch_head(struct fetch_head *fetch_head)
 		fetch_head->fp = fopen(filename, "a");
 		if (!fetch_head->fp)
 			return error_errno(_("cannot open %s"), filename);
+		strbuf_init(&fetch_head->buf, 0);
 	} else {
 		fetch_head->fp = NULL;
 	}

@@ -941,14 +943,17 @@ static void append_fetch_head(struct fetch_head *fetch_head,
 			return;
 	}
 
-	fprintf(fetch_head->fp, "%s\t%s\t%s",
-		oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
+	strbuf_addf(&fetch_head->buf, "%s\t%s\t%s",
+		    oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
 	for (i = 0; i < url_len; ++i)
 		if ('\n' == url[i])
-			fputs("\\n", fetch_head->fp);
+			strbuf_addstr(&fetch_head->buf, "\\n");
 		else
-			fputc(url[i], fetch_head->fp);
-	fputc('\n', fetch_head->fp);
+			strbuf_addch(&fetch_head->buf, url[i]);
+	strbuf_addch(&fetch_head->buf, '\n');
+
+	strbuf_write(&fetch_head->buf, fetch_head->fp);
+	strbuf_reset(&fetch_head->buf);
 }
 
 static void commit_fetch_head(struct fetch_head *fetch_head)

@@ -962,6 +967,7 @@ static void close_fetch_head(struct fetch_head *fetch_head)
 		return;
 
 	fclose(fetch_head->fp);
+	strbuf_release(&fetch_head->buf);
 }
 
 static const char warn_show_forced_updates[] =

2.30.0

Forgot to add my

Will amend in v4.

Patrick

s/servers/serves/

The cleanup code in s_update_ref() is currently duplicated for both succesful and erroneous exit paths. This commit refactors the function to have a shared exit path for both cases to remove the duplication.

---
 builtin/fetch.c | 47 +++++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 20 deletions(-)

index 1252f37493..991771f8eb 100644
@@ -589,7 +589,7 @@ static int s_update_ref(const char *action,
 	char *rla = getenv("GIT_REFLOG_ACTION");
 	struct ref_transaction *transaction;
 	struct strbuf err = STRBUF_INIT;
-	int ret, df_conflict = 0;
+	int ret;
 
 	if (dry_run)
 		return 0;

@@ -598,30 +598,37 @@ static int s_update_ref(const char *action,
 	msg = xstrfmt("%s: %s", rla, action);
 
 	transaction = ref_transaction_begin(&err);
-	if (!transaction ||
-	    ref_transaction_update(transaction, ref->name,
-				   &ref->new_oid,
-				   check_old ? &ref->old_oid : NULL,
-				   0, msg, &err))
-		goto fail;
-
-	ret = ref_transaction_commit(transaction, &err);
-	if (ret) {
-		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
-		goto fail;
+	if (!transaction) {
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
 	}
 
+	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,
+				     check_old ? &ref->old_oid : NULL,
+				     0, msg, &err);
+	if (ret) {
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
+	}
+
+	switch (ref_transaction_commit(transaction, &err)) {
+	case 0:
+		break;
+	case TRANSACTION_NAME_CONFLICT:
+		ret = STORE_REF_ERROR_DF_CONFLICT;
+		goto out;
+	default:
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
+	}
+
+out:
 	ref_transaction_free(transaction);
+	if (ret)
+		error("%s", err.buf);
 	strbuf_release(&err);
 	free(msg);
-	return 0;
-fail:
-	ref_transaction_free(transaction);
-	error("%s", err.buf);
-	strbuf_release(&err);
-	free(msg);
-	return df_conflict ? STORE_REF_ERROR_DF_CONFLICT
-			   : STORE_REF_ERROR_OTHER;
+	return ret;
 }
 
 static int refcol_width = 10;

2.30.0

The handling of ref updates is completely handled by s_update_ref(), which will manage the complete lifecycle of the reference transaction. This is fine right now given that git-fetch(1) does not support atomic fetches, so each reference gets its own transaction. It is quite inflexible though, as s_update_ref() only knows about a single reference update at a time, so it doesn't allow us to alter the strategy.

This commit prepares s_update_ref() and its only caller update_local_ref() to allow passing an external transaction. If none is given, then the existing behaviour is triggered which creates a new transaction and directly commits it. Otherwise, if the caller provides a transaction, then we only queue the update but don't commit it. This optionally allows the caller to manage when a transaction will be committed.

Given that update_local_ref() is always called with a NULL transaction for now, no change in behaviour is expected from this change.

---
 builtin/fetch.c | 51 ++++++++++++++++++++++++++++++-------------------
 1 file changed, 31 insertions(+), 20 deletions(-)

index 991771f8eb..654c7a7eed 100644
@@ -583,11 +583,12 @@ static struct ref *get_ref_map(struct remote *remote,
 
 static int s_update_ref(const char *action,
 			struct ref *ref,
+			struct ref_transaction *transaction,
 			int check_old)
 {
 	char *msg;
 	char *rla = getenv("GIT_REFLOG_ACTION");
-	struct ref_transaction *transaction;
+	struct ref_transaction *our_transaction = NULL;
 	struct strbuf err = STRBUF_INIT;
 	int ret;
 

@@ -597,10 +598,17 @@ static int s_update_ref(const char *action,
 		rla = default_rla.buf;
 	msg = xstrfmt("%s: %s", rla, action);
 
-	transaction = ref_transaction_begin(&err);
+	/*
+	 * If no transaction was passed to us, we manage the transaction
+	 * ourselves. Otherwise, we trust the caller to handle the transaction
+	 * lifecycle.
+	 */
 	if (!transaction) {
-		ret = STORE_REF_ERROR_OTHER;
-		goto out;
+		transaction = our_transaction = ref_transaction_begin(&err);
+		if (!transaction) {
+			ret = STORE_REF_ERROR_OTHER;
+			goto out;
+		}
 	}
 
 	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,

@@ -611,19 +619,21 @@ static int s_update_ref(const char *action,
 		goto out;
 	}
 
-	switch (ref_transaction_commit(transaction, &err)) {
-	case 0:
-		break;
-	case TRANSACTION_NAME_CONFLICT:
-		ret = STORE_REF_ERROR_DF_CONFLICT;
-		goto out;
-	default:
-		ret = STORE_REF_ERROR_OTHER;
-		goto out;
+	if (our_transaction) {
+		switch (ref_transaction_commit(our_transaction, &err)) {
+		case 0:
+			break;
+		case TRANSACTION_NAME_CONFLICT:
+			ret = STORE_REF_ERROR_DF_CONFLICT;
+			goto out;
+		default:
+			ret = STORE_REF_ERROR_OTHER;
+			goto out;
+		}
 	}
 
 out:
-	ref_transaction_free(transaction);
+	ref_transaction_free(our_transaction);
 	if (ret)
 		error("%s", err.buf);
 	strbuf_release(&err);

@@ -766,6 +776,7 @@ static void format_display(struct strbuf *display, char code,
 }
 
 static int update_local_ref(struct ref *ref,
+			    struct ref_transaction *transaction,
 			    const char *remote,
 			    const struct ref *remote_ref,
 			    struct strbuf *display,

@@ -806,7 +817,7 @@ static int update_local_ref(struct ref *ref,
 	    starts_with(ref->name, "refs/tags/")) {
 		if (force || ref->force) {
 			int r;
-			r = s_update_ref("updating tag", ref, 0);
+			r = s_update_ref("updating tag", ref, transaction, 0);
 			format_display(display, r ? '!' : 't', _("[tag update]"),
 				       r ? _("unable to update local ref") : NULL,
 				       remote, pretty_ref, summary_width);

@@ -843,7 +854,7 @@ static int update_local_ref(struct ref *ref,
 			what = _("[new ref]");
 		}
 
-		r = s_update_ref(msg, ref, 0);
+		r = s_update_ref(msg, ref, transaction, 0);
 		format_display(display, r ? '!' : '*', what,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -865,7 +876,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "..");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("fast-forward", ref, 1);
+		r = s_update_ref("fast-forward", ref, transaction, 1);
 		format_display(display, r ? '!' : ' ', quickref.buf,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -877,7 +888,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "...");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("forced-update", ref, 1);
+		r = s_update_ref("forced-update", ref, transaction, 1);
 		format_display(display, r ? '!' : '+', quickref.buf,
 			       r ? _("unable to update local ref") : _("forced update"),
 			       remote, pretty_ref, summary_width);

@@ -1094,8 +1105,8 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 
 			strbuf_reset(&note);
 			if (ref) {
-				rc |= update_local_ref(ref, what, rm, &note,
-						       summary_width);
+				rc |= update_local_ref(ref, NULL, what,
+						       rm, &note, summary_width);
 				free(ref);
 			} else if (write_fetch_head || dry_run) {
 				/*

2.30.0

When executing a fetch, then git will currently allocate one reference transaction per reference update and directly commit it. This means that fetches are non-atomic: even if some of the reference updates fail, others may still succeed and modify local references.

This is fine in many scenarios, but this strategy has its downsides.

- The view of remote references may be inconsistent and may show a bastardized state of the remote repository.

- Batching together updates may improve performance in certain scenarios. While the impact probably isn't as pronounced with loose references, the upcoming reftable backend may benefit as it needs to write less files in case the update is batched.

- The reference-update hook is currently being executed twice per updated reference. While this doesn't matter when there is no such hook, we have seen severe performance regressions when doing a git-fetch(1) with reference-transaction hook when the remote repository has hundreds of thousands of references.

Similar to git push --atomic, this commit thus introduces atomic fetches. Instead of allocating one reference transaction per updated reference, it causes us to only allocate a single transaction and commit it as soon as all updates were received. If locking of any reference fails, then we abort the complete transaction and don't update any reference, which gives us an all-or-nothing fetch.

Note that this may not completely fix the first of above downsides, as the consistent view also depends on the server-side. If the server doesn't have a consistent view of its own references during the reference negotiation phase, then the client would get the same inconsistent view the server has. This is a separate problem though and, if it actually exists, can be fixed at a later point.

This commit also changes the way we write FETCH_HEAD in case --atomic is passed. Instead of writing changes as we go, we need to accumulate all changes first and only commit them at the end when we know that all reference updates succeeded. Ideally, we'd just do so via a temporary file so that we don't need to carry all updates in-memory. This isn't trivially doable though considering the --append mode, where we do not truncate the file but simply append to it. And given that we support concurrent processes appending to FETCH_HEAD at the same time without any loss of data, seeding the temporary file with current contents of FETCH_HEAD initially and then doing a rename wouldn't work either. So this commit implements the simple strategy of buffering all changes and appending them to the file on commit.

---
 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 |  46 ++++++++-
 t/t5510-fetch.sh                | 168 ++++++++++++++++++++++++++++++++
 3 files changed, 213 insertions(+), 5 deletions(-)

index 2bf77b46fd..07783deee3 100644
@@ -7,6 +7,10 @@
 	existing contents of `.git/FETCH_HEAD`.  Without this
 	option old data in `.git/FETCH_HEAD` will be overwritten.
 
+--atomic::
+	Use an atomic transaction to update local refs. Either all refs are
+	updated, or on error, no refs are updated.
+
 --depth=<depth>::
 	Limit fetching to the specified number of commits from the tip of
 	each remote branch history. If fetching to a 'shallow' repository

2.30.0

Looking good. I left a few comments on remaining or newly introduced issues, but IIRC all of them were rather minor.

Thanks, will replace.

Hi,

this is the fourth version of my patch series to implement support for atomic reference updates for git-fetch(1). It's similar to git push --atomic, only that it applies to the local side. That is the fetch will either succeed and update all remote references or it will fail and update none.

Changes compared to v3:

- Fixed indentation of the switch statement in 1/5.

- Added my missing SOB to 2/5 and fixed a typo in the commit message.

Please see the attached range-diff for more details.

Patrick

Patrick Steinhardt (5):
fetch: extract writing to FETCH_HEAD
fetch: use strbuf to format FETCH_HEAD updates
fetch: refactor s_update_ref to use common exit path
fetch: allow passing a transaction to s_update_ref()
fetch: implement support for atomic reference updates

 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 | 228 +++++++++++++++++++++++---------
 remote.h                        |   2 +-
 t/t5510-fetch.sh                | 168 +++++++++++++++++++++++
 4 files changed, 342 insertions(+), 60 deletions(-)

Range-diff against v3:
1: 61dc19a1ca ! 1: 9fcc8b54de fetch: extract writing to FETCH_HEAD

    @@ builtin/fetch.c: static int iterate_ref_map(void *cb_data, struct object_id *oid
     +		return;
     +
     +	switch (fetch_head_status) {
    -+		case FETCH_HEAD_NOT_FOR_MERGE:
    -+			merge_status_marker = "not-for-merge";
    -+			break;
    -+		case FETCH_HEAD_MERGE:
    -+			merge_status_marker = "";
    -+			break;
    -+		default:
    -+			/* do not write anything to FETCH_HEAD */
    -+			return;
    ++	case FETCH_HEAD_NOT_FOR_MERGE:
    ++		merge_status_marker = "not-for-merge";
    ++		break;
    ++	case FETCH_HEAD_MERGE:
    ++		merge_status_marker = "";
    ++		break;
    ++	default:
    ++		/* do not write anything to FETCH_HEAD */
    ++		return;
     +	}
     +
     +	fprintf(fetch_head->fp, "%s\t%s\t%s",

2: a19762690e ! 2: fb8542270a fetch: use strbuf to format FETCH_HEAD updates
@@ Commit message
This commit refactors append_fetch_head() to use a struct strbuf for formatting the update which we're about to append to the FETCH_HEAD file. While the refactoring doesn't have much of a benefit right now, it
- servers as a preparatory step to implement atomic fetches where we need
+ serves as a preparatory step to implement atomic fetches where we need
to buffer all updates to FETCH_HEAD and only flush them out if all reference updates succeeded.

No change in behaviour is expected from this commit.

+ Signed-off-by: Patrick Steinhardt <[email protected]>
+
## builtin/fetch.c ##
@@ builtin/fetch.c: static int iterate_ref_map(void *cb_data, struct object_id *oid)

@@ builtin/fetch.c: static int open_fetch_head(struct fetch_head *fetch_head)
fetch_head->fp = NULL;
}
@@ builtin/fetch.c: static void append_fetch_head(struct fetch_head *fetch_head,
- return;
+ return;
}

- fprintf(fetch_head->fp, "%s\t%s\t%s",

3:  c411f30e09 = 3:  ba6908aa8c fetch: refactor `s_update_ref` to use common exit path
4:  865d357ba7 = 4:  7f820f6f83 fetch: allow passing a transaction to `s_update_ref()`
5:  6a79e7adcc = 5:  0b57d7a651 fetch: implement support for atomic reference updates

2.30.0

When performing a fetch with the default --write-fetch-head option, we write all updated references to FETCH_HEAD while the updates are performed. Given that updates are not performed atomically, it means that we we write to FETCH_HEAD even if some or all of the reference updates fail.

Given that we simply update FETCH_HEAD ad-hoc with each reference, the logic is completely contained in store_update_refs and thus quite hard to extend. This can already be seen by the way we skip writing to the
FETCH_HEAD: instead of having a conditional which simply skips writing, we instead open "/dev/null" and needlessly write all updates there.

We are about to extend git-fetch(1) to accept an --atomic flag which will make the fetch an all-or-nothing operation with regards to the reference updates. This will also require us to make the updates to FETCH_HEAD an all-or-nothing operation, but as explained doing so is not easy with the current layout. This commit thus refactors the wa we write to FETCH_HEAD and pulls out the logic to open, append to, commit and close the file. While this may seem rather over-the top at first, pulling out this logic will make it a lot easier to update the code in a subsequent commit. It also allows us to easily skip writing completely in case --no-write-fetch-head was passed.

---
 builtin/fetch.c | 108 +++++++++++++++++++++++++++++++++++-------------
 remote.h        |   2 +-
 2 files changed, 80 insertions(+), 30 deletions(-)

index ecf8537605..2dd5fcb652 100644
@@ -897,6 +897,73 @@ static int iterate_ref_map(void *cb_data, struct object_id *oid)
 	return 0;
 }
 
+struct fetch_head {
+	FILE *fp;
+};
+
+static int open_fetch_head(struct fetch_head *fetch_head)
+{
+	const char *filename = git_path_fetch_head(the_repository);
+
+	if (write_fetch_head) {
+		fetch_head->fp = fopen(filename, "a");
+		if (!fetch_head->fp)
+			return error_errno(_("cannot open %s"), filename);
+	} else {
+		fetch_head->fp = NULL;
+	}
+
+	return 0;
+}
+
+static void append_fetch_head(struct fetch_head *fetch_head,
+			      const struct object_id *old_oid,
+			      enum fetch_head_status fetch_head_status,
+			      const char *note,
+			      const char *url, size_t url_len)
+{
+	char old_oid_hex[GIT_MAX_HEXSZ + 1];
+	const char *merge_status_marker;
+	size_t i;
+
+	if (!fetch_head->fp)
+		return;
+
+	switch (fetch_head_status) {
+	case FETCH_HEAD_NOT_FOR_MERGE:
+		merge_status_marker = "not-for-merge";
+		break;
+	case FETCH_HEAD_MERGE:
+		merge_status_marker = "";
+		break;
+	default:
+		/* do not write anything to FETCH_HEAD */
+		return;
+	}
+
+	fprintf(fetch_head->fp, "%s\t%s\t%s",
+		oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
+	for (i = 0; i < url_len; ++i)
+		if ('\n' == url[i])
+			fputs("\\n", fetch_head->fp);
+		else
+			fputc(url[i], fetch_head->fp);
+	fputc('\n', fetch_head->fp);
+}
+
+static void commit_fetch_head(struct fetch_head *fetch_head)
+{
+	/* Nothing to commit yet. */
+}
+
+static void close_fetch_head(struct fetch_head *fetch_head)
+{
+	if (!fetch_head->fp)
+		return;
+
+	fclose(fetch_head->fp);
+}
+
 static const char warn_show_forced_updates[] =
 N_("Fetch normally indicates which branches had a forced update,\n"
    "but that check has been disabled. To re-enable, use '--show-forced-updates'\n"

@@ -909,22 +976,19 @@ N_("It took %.2f seconds to check forced updates. You can use\n"
 static int store_updated_refs(const char *raw_url, const char *remote_name,
 			      int connectivity_checked, struct ref *ref_map)
 {
-	FILE *fp;
+	struct fetch_head fetch_head;
 	struct commit *commit;
 	int url_len, i, rc = 0;
 	struct strbuf note = STRBUF_INIT;
 	const char *what, *kind;
 	struct ref *rm;
 	char *url;
-	const char *filename = (!write_fetch_head
-				? "/dev/null"
-				: git_path_fetch_head(the_repository));
 	int want_status;
 	int summary_width = transport_summary_width(ref_map);
 
-	fp = fopen(filename, "a");
-	if (!fp)
-		return error_errno(_("cannot open %s"), filename);
+	rc = open_fetch_head(&fetch_head);
+	if (rc)
+		return -1;
 
 	if (raw_url)
 		url = transport_anonymize_url(raw_url);

@@ -953,7 +1017,6 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 	     want_status++) {
 		for (rm = ref_map; rm; rm = rm->next) {
 			struct ref *ref = NULL;
-			const char *merge_status_marker = "";
 
 			if (rm->status == REF_STATUS_REJECT_SHALLOW) {
 				if (want_status == FETCH_HEAD_MERGE)

@@ -1011,26 +1074,10 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 					strbuf_addf(&note, "%s ", kind);
 				strbuf_addf(&note, "'%s' of ", what);
 			}
-			switch (rm->fetch_head_status) {
-			case FETCH_HEAD_NOT_FOR_MERGE:
-				merge_status_marker = "not-for-merge";
-				/* fall-through */
-			case FETCH_HEAD_MERGE:
-				fprintf(fp, "%s\t%s\t%s",
-					oid_to_hex(&rm->old_oid),
-					merge_status_marker,
-					note.buf);
-				for (i = 0; i < url_len; ++i)
-					if ('\n' == url[i])
-						fputs("\\n", fp);
-					else
-						fputc(url[i], fp);
-				fputc('\n', fp);
-				break;
-			default:
-				/* do not write anything to FETCH_HEAD */
-				break;
-			}
+
+			append_fetch_head(&fetch_head, &rm->old_oid,
+					  rm->fetch_head_status,
+					  note.buf, url, url_len);
 
 			strbuf_reset(&note);
 			if (ref) {

@@ -1060,6 +1107,9 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 		}
 	}
 
+	if (!rc)
+		commit_fetch_head(&fetch_head);
+
 	if (rc & STORE_REF_ERROR_DF_CONFLICT)
 		error(_("some local refs could not be updated; try running\n"
 		      " 'git remote prune %s' to remove any old, conflicting "

@@ -1077,7 +1127,7 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
  abort:
 	strbuf_release(&note);
 	free(url);
-	fclose(fp);
+	close_fetch_head(&fetch_head);
 	return rc;
 }
 

index 3211abdf05..aad1a0f080 100644
@@ -134,7 +134,7 @@ struct ref {
 	 * should be 0, so that xcalloc'd structures get it
 	 * by default.
 	 */
-	enum {
+	enum fetch_head_status {
 		FETCH_HEAD_MERGE = -1,
 		FETCH_HEAD_NOT_FOR_MERGE = 0,
 		FETCH_HEAD_IGNORE = 1

2.30.0

This commit refactors append_fetch_head() to use a struct strbuf for formatting the update which we're about to append to the FETCH_HEAD file. While the refactoring doesn't have much of a benefit right now, it serves as a preparatory step to implement atomic fetches where we need to buffer all updates to FETCH_HEAD and only flush them out if all reference updates succeeded.

No change in behaviour is expected from this commit.

---
 builtin/fetch.c | 16 +++++++++++-----
 1 file changed, 11 insertions(+), 5 deletions(-)

index 2dd5fcb652..e317e828cd 100644
@@ -899,6 +899,7 @@ static int iterate_ref_map(void *cb_data, struct object_id *oid)
 
 struct fetch_head {
 	FILE *fp;
+	struct strbuf buf;
 };
 
 static int open_fetch_head(struct fetch_head *fetch_head)

@@ -909,6 +910,7 @@ static int open_fetch_head(struct fetch_head *fetch_head)
 		fetch_head->fp = fopen(filename, "a");
 		if (!fetch_head->fp)
 			return error_errno(_("cannot open %s"), filename);
+		strbuf_init(&fetch_head->buf, 0);
 	} else {
 		fetch_head->fp = NULL;
 	}

@@ -941,14 +943,17 @@ static void append_fetch_head(struct fetch_head *fetch_head,
 		return;
 	}
 
-	fprintf(fetch_head->fp, "%s\t%s\t%s",
-		oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
+	strbuf_addf(&fetch_head->buf, "%s\t%s\t%s",
+		    oid_to_hex_r(old_oid_hex, old_oid), merge_status_marker, note);
 	for (i = 0; i < url_len; ++i)
 		if ('\n' == url[i])
-			fputs("\\n", fetch_head->fp);
+			strbuf_addstr(&fetch_head->buf, "\\n");
 		else
-			fputc(url[i], fetch_head->fp);
-	fputc('\n', fetch_head->fp);
+			strbuf_addch(&fetch_head->buf, url[i]);
+	strbuf_addch(&fetch_head->buf, '\n');
+
+	strbuf_write(&fetch_head->buf, fetch_head->fp);
+	strbuf_reset(&fetch_head->buf);
 }
 
 static void commit_fetch_head(struct fetch_head *fetch_head)

@@ -962,6 +967,7 @@ static void close_fetch_head(struct fetch_head *fetch_head)
 		return;
 
 	fclose(fetch_head->fp);
+	strbuf_release(&fetch_head->buf);
 }
 
 static const char warn_show_forced_updates[] =

2.30.0

The cleanup code in s_update_ref() is currently duplicated for both succesful and erroneous exit paths. This commit refactors the function to have a shared exit path for both cases to remove the duplication.

---
 builtin/fetch.c | 47 +++++++++++++++++++++++++++--------------------
 1 file changed, 27 insertions(+), 20 deletions(-)

index e317e828cd..b24a9e09a4 100644
@@ -589,7 +589,7 @@ static int s_update_ref(const char *action,
 	char *rla = getenv("GIT_REFLOG_ACTION");
 	struct ref_transaction *transaction;
 	struct strbuf err = STRBUF_INIT;
-	int ret, df_conflict = 0;
+	int ret;
 
 	if (dry_run)
 		return 0;

@@ -598,30 +598,37 @@ static int s_update_ref(const char *action,
 	msg = xstrfmt("%s: %s", rla, action);
 
 	transaction = ref_transaction_begin(&err);
-	if (!transaction ||
-	    ref_transaction_update(transaction, ref->name,
-				   &ref->new_oid,
-				   check_old ? &ref->old_oid : NULL,
-				   0, msg, &err))
-		goto fail;
-
-	ret = ref_transaction_commit(transaction, &err);
-	if (ret) {
-		df_conflict = (ret == TRANSACTION_NAME_CONFLICT);
-		goto fail;
+	if (!transaction) {
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
 	}
 
+	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,
+				     check_old ? &ref->old_oid : NULL,
+				     0, msg, &err);
+	if (ret) {
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
+	}
+
+	switch (ref_transaction_commit(transaction, &err)) {
+	case 0:
+		break;
+	case TRANSACTION_NAME_CONFLICT:
+		ret = STORE_REF_ERROR_DF_CONFLICT;
+		goto out;
+	default:
+		ret = STORE_REF_ERROR_OTHER;
+		goto out;
+	}
+
+out:
 	ref_transaction_free(transaction);
+	if (ret)
+		error("%s", err.buf);
 	strbuf_release(&err);
 	free(msg);
-	return 0;
-fail:
-	ref_transaction_free(transaction);
-	error("%s", err.buf);
-	strbuf_release(&err);
-	free(msg);
-	return df_conflict ? STORE_REF_ERROR_DF_CONFLICT
-			   : STORE_REF_ERROR_OTHER;
+	return ret;
 }
 
 static int refcol_width = 10;

2.30.0

The handling of ref updates is completely handled by s_update_ref(), which will manage the complete lifecycle of the reference transaction. This is fine right now given that git-fetch(1) does not support atomic fetches, so each reference gets its own transaction. It is quite inflexible though, as s_update_ref() only knows about a single reference update at a time, so it doesn't allow us to alter the strategy.

This commit prepares s_update_ref() and its only caller update_local_ref() to allow passing an external transaction. If none is given, then the existing behaviour is triggered which creates a new transaction and directly commits it. Otherwise, if the caller provides a transaction, then we only queue the update but don't commit it. This optionally allows the caller to manage when a transaction will be committed.

Given that update_local_ref() is always called with a NULL transaction for now, no change in behaviour is expected from this change.

---
 builtin/fetch.c | 51 ++++++++++++++++++++++++++++++-------------------
 1 file changed, 31 insertions(+), 20 deletions(-)

index b24a9e09a4..cada732325 100644
@@ -583,11 +583,12 @@ static struct ref *get_ref_map(struct remote *remote,
 
 static int s_update_ref(const char *action,
 			struct ref *ref,
+			struct ref_transaction *transaction,
 			int check_old)
 {
 	char *msg;
 	char *rla = getenv("GIT_REFLOG_ACTION");
-	struct ref_transaction *transaction;
+	struct ref_transaction *our_transaction = NULL;
 	struct strbuf err = STRBUF_INIT;
 	int ret;
 

@@ -597,10 +598,17 @@ static int s_update_ref(const char *action,
 		rla = default_rla.buf;
 	msg = xstrfmt("%s: %s", rla, action);
 
-	transaction = ref_transaction_begin(&err);
+	/*
+	 * If no transaction was passed to us, we manage the transaction
+	 * ourselves. Otherwise, we trust the caller to handle the transaction
+	 * lifecycle.
+	 */
 	if (!transaction) {
-		ret = STORE_REF_ERROR_OTHER;
-		goto out;
+		transaction = our_transaction = ref_transaction_begin(&err);
+		if (!transaction) {
+			ret = STORE_REF_ERROR_OTHER;
+			goto out;
+		}
 	}
 
 	ret = ref_transaction_update(transaction, ref->name, &ref->new_oid,

@@ -611,19 +619,21 @@ static int s_update_ref(const char *action,
 		goto out;
 	}
 
-	switch (ref_transaction_commit(transaction, &err)) {
-	case 0:
-		break;
-	case TRANSACTION_NAME_CONFLICT:
-		ret = STORE_REF_ERROR_DF_CONFLICT;
-		goto out;
-	default:
-		ret = STORE_REF_ERROR_OTHER;
-		goto out;
+	if (our_transaction) {
+		switch (ref_transaction_commit(our_transaction, &err)) {
+		case 0:
+			break;
+		case TRANSACTION_NAME_CONFLICT:
+			ret = STORE_REF_ERROR_DF_CONFLICT;
+			goto out;
+		default:
+			ret = STORE_REF_ERROR_OTHER;
+			goto out;
+		}
 	}
 
 out:
-	ref_transaction_free(transaction);
+	ref_transaction_free(our_transaction);
 	if (ret)
 		error("%s", err.buf);
 	strbuf_release(&err);

@@ -766,6 +776,7 @@ static void format_display(struct strbuf *display, char code,
 }
 
 static int update_local_ref(struct ref *ref,
+			    struct ref_transaction *transaction,
 			    const char *remote,
 			    const struct ref *remote_ref,
 			    struct strbuf *display,

@@ -806,7 +817,7 @@ static int update_local_ref(struct ref *ref,
 	    starts_with(ref->name, "refs/tags/")) {
 		if (force || ref->force) {
 			int r;
-			r = s_update_ref("updating tag", ref, 0);
+			r = s_update_ref("updating tag", ref, transaction, 0);
 			format_display(display, r ? '!' : 't', _("[tag update]"),
 				       r ? _("unable to update local ref") : NULL,
 				       remote, pretty_ref, summary_width);

@@ -843,7 +854,7 @@ static int update_local_ref(struct ref *ref,
 			what = _("[new ref]");
 		}
 
-		r = s_update_ref(msg, ref, 0);
+		r = s_update_ref(msg, ref, transaction, 0);
 		format_display(display, r ? '!' : '*', what,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -865,7 +876,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "..");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("fast-forward", ref, 1);
+		r = s_update_ref("fast-forward", ref, transaction, 1);
 		format_display(display, r ? '!' : ' ', quickref.buf,
 			       r ? _("unable to update local ref") : NULL,
 			       remote, pretty_ref, summary_width);

@@ -877,7 +888,7 @@ static int update_local_ref(struct ref *ref,
 		strbuf_add_unique_abbrev(&quickref, &current->object.oid, DEFAULT_ABBREV);
 		strbuf_addstr(&quickref, "...");
 		strbuf_add_unique_abbrev(&quickref, &ref->new_oid, DEFAULT_ABBREV);
-		r = s_update_ref("forced-update", ref, 1);
+		r = s_update_ref("forced-update", ref, transaction, 1);
 		format_display(display, r ? '!' : '+', quickref.buf,
 			       r ? _("unable to update local ref") : _("forced update"),
 			       remote, pretty_ref, summary_width);

@@ -1094,8 +1105,8 @@ static int store_updated_refs(const char *raw_url, const char *remote_name,
 
 			strbuf_reset(&note);
 			if (ref) {
-				rc |= update_local_ref(ref, what, rm, &note,
-						       summary_width);
+				rc |= update_local_ref(ref, NULL, what,
+						       rm, &note, summary_width);
 				free(ref);
 			} else if (write_fetch_head || dry_run) {
 				/*

2.30.0

When executing a fetch, then git will currently allocate one reference transaction per reference update and directly commit it. This means that fetches are non-atomic: even if some of the reference updates fail, others may still succeed and modify local references.

This is fine in many scenarios, but this strategy has its downsides.

- The view of remote references may be inconsistent and may show a bastardized state of the remote repository.

- Batching together updates may improve performance in certain scenarios. While the impact probably isn't as pronounced with loose references, the upcoming reftable backend may benefit as it needs to write less files in case the update is batched.

- The reference-update hook is currently being executed twice per updated reference. While this doesn't matter when there is no such hook, we have seen severe performance regressions when doing a git-fetch(1) with reference-transaction hook when the remote repository has hundreds of thousands of references.

Similar to git push --atomic, this commit thus introduces atomic fetches. Instead of allocating one reference transaction per updated reference, it causes us to only allocate a single transaction and commit it as soon as all updates were received. If locking of any reference fails, then we abort the complete transaction and don't update any reference, which gives us an all-or-nothing fetch.

Note that this may not completely fix the first of above downsides, as the consistent view also depends on the server-side. If the server doesn't have a consistent view of its own references during the reference negotiation phase, then the client would get the same inconsistent view the server has. This is a separate problem though and, if it actually exists, can be fixed at a later point.

This commit also changes the way we write FETCH_HEAD in case --atomic is passed. Instead of writing changes as we go, we need to accumulate all changes first and only commit them at the end when we know that all reference updates succeeded. Ideally, we'd just do so via a temporary file so that we don't need to carry all updates in-memory. This isn't trivially doable though considering the --append mode, where we do not truncate the file but simply append to it. And given that we support concurrent processes appending to FETCH_HEAD at the same time without any loss of data, seeding the temporary file with current contents of FETCH_HEAD initially and then doing a rename wouldn't work either. So this commit implements the simple strategy of buffering all changes and appending them to the file on commit.

---
 Documentation/fetch-options.txt |   4 +
 builtin/fetch.c                 |  46 ++++++++-
 t/t5510-fetch.sh                | 168 ++++++++++++++++++++++++++++++++
 3 files changed, 213 insertions(+), 5 deletions(-)

index 2bf77b46fd..07783deee3 100644
@@ -7,6 +7,10 @@
 	existing contents of `.git/FETCH_HEAD`.  Without this
 	option old data in `.git/FETCH_HEAD` will be overwritten.
 
+--atomic::
+	Use an atomic transaction to update local refs. Either all refs are
+	updated, or on error, no refs are updated.
+
 --depth=<depth>::
 	Limit fetching to the specified number of commits from the tip of
 	each remote branch history. If fetching to a 'shallow' repository

2.30.0