[PATCH] Parallelize pulling by ssh
This causes ssh-pull to request objects in prefetch() and read then in
fetch(), such that it reduces the unpipelined round-trip time.
This also makes sha1_write_from_fd() support having a buffer of data
which it accidentally read from the fd after the object; this was
formerly not a problem, because it would always get a short read at
the end of an object, because the next object had not been
requested. This is no longer true.
Signed-off-by: Daniel Barkalow <barkalow@iabervon.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
diff --git a/ssh-pull.c b/ssh-pull.c
index 4cf9b6a..bdc99df 100644
--- a/ssh-pull.c
+++ b/ssh-pull.c
@@ -10,24 +10,49 @@
static unsigned char remote_version = 0;
static unsigned char local_version = 1;
+ssize_t force_write(int fd, void *buffer, size_t length)
+{
+ ssize_t ret = 0;
+ while (ret < length) {
+ ssize_t size = write(fd, buffer + ret, length - ret);
+ if (size < 0) {
+ return size;
+ }
+ if (size == 0) {
+ return ret;
+ }
+ ret += size;
+ }
+ return ret;
+}
+
void prefetch(unsigned char *sha1)
{
+ char type = 'o';
+ force_write(fd_out, &type, 1);
+ force_write(fd_out, sha1, 20);
+ //memcpy(requested + 20 * prefetches++, sha1, 20);
}
+static char conn_buf[4096];
+static size_t conn_buf_posn = 0;
+
int fetch(unsigned char *sha1)
{
int ret;
signed char remote;
- char type = 'o';
- if (has_sha1_file(sha1))
- return 0;
- write(fd_out, &type, 1);
- write(fd_out, sha1, 20);
- if (read(fd_in, &remote, 1) < 1)
- return -1;
+
+ if (conn_buf_posn) {
+ remote = conn_buf[0];
+ memmove(conn_buf, conn_buf + 1, --conn_buf_posn);
+ } else {
+ if (read(fd_in, &remote, 1) < 1)
+ return -1;
+ }
+ //fprintf(stderr, "Got %d\n", remote);
if (remote < 0)
return remote;
- ret = write_sha1_from_fd(sha1, fd_in);
+ ret = write_sha1_from_fd(sha1, fd_in, conn_buf, 4096, &conn_buf_posn);
if (!ret)
pull_say("got %s\n", sha1_to_hex(sha1));
return ret;
