commit a5a99557dbce7e25efeaf31f37322748fed8da50
from: Omar Polo <op@omarpolo.com>
via: Thomas Adam <thomas@xteddy.org>
date: Fri Jul 29 09:38:22 2022 UTC

gotwebd: handle partial writes

fcgi_send_response can end up trying to send big amount of data, in
particular when serving blobs.  if a write failure occurs, gotwebd and
httpd go out of sync and "bad stuff happens".

debugged with and ok tracey@

commit - 57d4848d969f00db854acc82cd5e920532a9f809
commit + a5a99557dbce7e25efeaf31f37322748fed8da50
blob - 3b6350631e42b62f4a9aa477c0a2aee5464fec6b
blob + dbca0f24e4e64de735cdcf2fbc84ffabe5faa91c
--- gotwebd/fcgi.c
+++ gotwebd/fcgi.c
@@ -353,7 +353,8 @@ fcgi_send_response(struct request *c, struct fcgi_resp
 {
 	struct fcgi_record_header *header;
 	struct timespec ts;
-	size_t padded_len;
+	size_t padded_len, off;
+	ssize_t nw;
 	int err = 0, th = 2000;
 
 	ts.tv_sec = 0;
@@ -379,14 +380,27 @@ fcgi_send_response(struct request *c, struct fcgi_resp
 	 * disconnect. Let's at least try to write the data a few times before
 	 * giving up.
 	 */
-	while ((write(c->fd, resp->data + resp->data_pos,
-	    resp->data_len)) == -1) {
-		nanosleep(&ts, NULL);
-		err++;
-		if (err == th) {
+	for (off = 0; off < resp->data_len; off += nw) {
+		nw = write(c->fd, resp->data + off, resp->data_len - off);
+		if (nw == 0) {
+			c->sock->client_status = CLIENT_DISCONNECT;
+			break;
+		}
+		if (nw == -1) {
+			err++;
+			if (errno == EAGAIN && err < th) {
+				nw = 0;
+				nanosleep(&ts, NULL);
+				continue;
+			}
+			log_warn("write");
 			c->sock->client_status = CLIENT_DISCONNECT;
 			break;
 		}
+
+		if (nw != resp->data_len - off)
+			log_debug("partial write: %zu vs %zu", nw,
+			    resp->data_len - off);
 	}
 
 	free(resp);