[PATCH] connected: use buffered I/O to talk to rev-list

[PATCH] connected: use buffered I/O to talk to rev-list

[René Scharfe]

Like f0bca72dc77 (send-pack: use buffered I/O to talk to pack-objects,
2016-06-08), significantly reduce the number of system calls and
simplify the code for sending object IDs to rev-list by using stdio's
buffering and handling errors after the loop.

Signed-off-by: René Scharfe <l.s.r@web.de>
---
 connected.c | 16 +++++-----------
 1 file changed, 5 insertions(+), 11 deletions(-)

diff --git a/connected.c b/connected.c
index 937b4bae387..05c2916f38d 100644
--- a/connected.c
+++ b/connected.c
@@ -22,14 +22,13 @@ int check_connected(oid_iterate_fn fn, void *cb_data,
 		    struct check_connected_options *opt)
 {
 	struct child_process rev_list = CHILD_PROCESS_INIT;
+	FILE *rev_list_in;
 	struct check_connected_options defaults = CHECK_CONNECTED_INIT;
-	char commit[GIT_MAX_HEXSZ + 1];
 	struct object_id oid;
 	int err = 0;
 	struct packed_git *new_pack = NULL;
 	struct transport *transport;
 	size_t base_len;
-	const unsigned hexsz = the_hash_algo->hexsz;

 	if (!opt)
 		opt = &defaults;
@@ -122,7 +121,8 @@ int check_connected(oid_iterate_fn fn, void *cb_data,

 	sigchain_push(SIGPIPE, SIG_IGN);

-	commit[hexsz] = '\n';
+	rev_list_in = xfdopen(rev_list.in, "w");
+
 	do {
 		/*
 		 * If index-pack already checked that:
@@ -135,16 +135,10 @@ int check_connected(oid_iterate_fn fn, void *cb_data,
 		if (new_pack && find_pack_entry_one(oid.hash, new_pack))
 			continue;

-		memcpy(commit, oid_to_hex(&oid), hexsz);
-		if (write_in_full(rev_list.in, commit, hexsz + 1) < 0) {
-			if (errno != EPIPE && errno != EINVAL)
-				error_errno(_("failed write to rev-list"));
-			err = -1;
-			break;
-		}
+		fprintf(rev_list_in, "%s\n", oid_to_hex(&oid));
 	} while (!fn(cb_data, &oid));

-	if (close(rev_list.in))
+	if (fclose(rev_list_in))
 		err = error_errno(_("failed to close rev-list's stdin"));

 	sigchain_pop(SIGPIPE);
--
2.28.0

Re: [PATCH] connected: use buffered I/O to talk to rev-list

[Chris Torek]

On Sun, Aug 2, 2020 at 7:39 AM René Scharfe <l.s.r@web.de> wrote:
> @@ -135,16 +135,10 @@ int check_connected(oid_iterate_fn fn, void *cb_data,
>                 if (new_pack && find_pack_entry_one(oid.hash, new_pack))
>                         continue;
>
> -               memcpy(commit, oid_to_hex(&oid), hexsz);
> -               if (write_in_full(rev_list.in, commit, hexsz + 1) < 0) {
> -                       if (errno != EPIPE && errno != EINVAL)
> -                               error_errno(_("failed write to rev-list"));
> -                       err = -1;
> -                       break;
> -               }
> +               fprintf(rev_list_in, "%s\n", oid_to_hex(&oid));
>         } while (!fn(cb_data, &oid));
>
> -       if (close(rev_list.in))
> +       if (fclose(rev_list_in))
>                 err = error_errno(_("failed to close rev-list's stdin"));
>
>         sigchain_pop(SIGPIPE);
> --
> 2.28.0

The same ferror()-before-fclose() remarks apply here too,
but this time the explicit errno checking (EPIPE) cannot
be done -- it's too late, errno is probably overwritten.  I'm
not sure how valuable the explicit errno tests are in the first
place so I will leave that to others, but if we want to keep
the explicit tests, use:

    if (fprintf(...) < 0)

to check each fprintf(), and add a final fflush() call (with
another check) before the fclose().

Chris

Re: [PATCH] connected: use buffered I/O to talk to rev-list

[Johannes Sixt]

Am 02.08.20 um 18:08 schrieb Chris Torek:
> On Sun, Aug 2, 2020 at 7:39 AM René Scharfe <l.s.r@web.de> wrote:
>> @@ -135,16 +135,10 @@ int check_connected(oid_iterate_fn fn, void *cb_data,
>>                 if (new_pack && find_pack_entry_one(oid.hash, new_pack))
>>                         continue;
>>
>> -               memcpy(commit, oid_to_hex(&oid), hexsz);
>> -               if (write_in_full(rev_list.in, commit, hexsz + 1) < 0) {
>> -                       if (errno != EPIPE && errno != EINVAL)
>> -                               error_errno(_("failed write to rev-list"));
>> -                       err = -1;
>> -                       break;
>> -               }
>> +               fprintf(rev_list_in, "%s\n", oid_to_hex(&oid));
>>         } while (!fn(cb_data, &oid));
>>
>> -       if (close(rev_list.in))
>> +       if (fclose(rev_list_in))
>>                 err = error_errno(_("failed to close rev-list's stdin"));
>>
>>         sigchain_pop(SIGPIPE);
> 
> The same ferror()-before-fclose() remarks apply here too,
> but this time the explicit errno checking (EPIPE) cannot
> be done -- it's too late, errno is probably overwritten.

The EPIPE check should really remain in the loop above. (Same comment on
EPIPE on Windows applies here.)

-- Hannes

[PATCH v2] connected: use buffered I/O to talk to rev-list

[René Scharfe]

Like f0bca72dc77 (send-pack: use buffered I/O to talk to pack-objects,
2016-06-08), significantly reduce the number of system calls and
simplify the code for sending object IDs to rev-list by using stdio's
buffering.

Take care to handle errors immediately to get the correct error code,
and to flush the buffer explicitly before closing the stream in order to
catch any write errors for these last bytes.

Helped-by: Chris Torek <chris.torek@gmail.com>
Helped-by: Johannes Sixt <j6t@kdbg.org>
---
Changes since v1:
- Handle fprintf() errors immediately.
- Call ferror() and fflush() explicitly before calling fclose().
- Report write errors other than EPIPE and EINVAL, like the original
  code does.

 connected.c | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/connected.c b/connected.c
index 21c1ebe9fbf..b18299fdf0e 100644
--- a/connected.c
+++ b/connected.c
@@ -22,14 +22,13 @@ int check_connected(oid_iterate_fn fn, void *cb_data,
 		    struct check_connected_options *opt)
 {
 	struct child_process rev_list = CHILD_PROCESS_INIT;
+	FILE *rev_list_in;
 	struct check_connected_options defaults = CHECK_CONNECTED_INIT;
-	char commit[GIT_MAX_HEXSZ + 1];
 	struct object_id oid;
 	int err = 0;
 	struct packed_git *new_pack = NULL;
 	struct transport *transport;
 	size_t base_len;
-	const unsigned hexsz = the_hash_algo->hexsz;

 	if (!opt)
 		opt = &defaults;
@@ -122,7 +121,8 @@ int check_connected(oid_iterate_fn fn, void *cb_data,

 	sigchain_push(SIGPIPE, SIG_IGN);

-	commit[hexsz] = '\n';
+	rev_list_in = xfdopen(rev_list.in, "w");
+
 	do {
 		/*
 		 * If index-pack already checked that:
@@ -135,16 +135,17 @@ int check_connected(oid_iterate_fn fn, void *cb_data,
 		if (new_pack && find_pack_entry_one(oid.hash, new_pack))
 			continue;

-		memcpy(commit, oid_to_hex(&oid), hexsz);
-		if (write_in_full(rev_list.in, commit, hexsz + 1) < 0) {
-			if (errno != EPIPE && errno != EINVAL)
-				error_errno(_("failed write to rev-list"));
-			err = -1;
+		if (fprintf(rev_list_in, "%s\n", oid_to_hex(&oid)) < 0)
 			break;
-		}
 	} while (!fn(cb_data, &oid));

-	if (close(rev_list.in))
+	if (ferror(rev_list_in) || fflush(rev_list_in)) {
+		if (errno != EPIPE && errno != EINVAL)
+			error_errno(_("failed write to rev-list"));
+		err = -1;
+	}
+
+	if (fclose(rev_list_in))
 		err = error_errno(_("failed to close rev-list's stdin"));

 	sigchain_pop(SIGPIPE);
--
2.28.0

Re: [PATCH v2] connected: use buffered I/O to talk to rev-list

[Jeff King]

On Wed, Aug 12, 2020 at 06:52:49PM +0200, René Scharfe wrote:

> Like f0bca72dc77 (send-pack: use buffered I/O to talk to pack-objects,
> 2016-06-08), significantly reduce the number of system calls and
> simplify the code for sending object IDs to rev-list by using stdio's
> buffering.
> 
> Take care to handle errors immediately to get the correct error code,
> and to flush the buffer explicitly before closing the stream in order to
> catch any write errors for these last bytes.

FWIW, the error handling in this patch and the other stdio conversions
you sent all look good to me. Thanks (and to Chris and Johannes for
great review on v1).

-Peff