tls.c: Better tolerate short TLS writes.

The existing retry policy for SSL_write returning
SSL_ERROR_WANT_WRITE was not very robust in handling
delayed writes from writing a large amount of data.
This improves this by significantly increasing
our tolerance and also resetting the counter when
progress is made.

This is still not perfect, and it would still be
better to do what the comments say.

bbs/tls.c

@@ -569,10 +569,14 @@ static void *ssl_io_thread(void *unused)
 									if (!write_attempts++) { /* Log the first time. Debug, not warning, since this could happen legitimately */
 										bbs_debug(4, "SSL_write returned %ld (%s)\n", wres, ssl_strerror(err));
 									}
-									if (write_attempts < 1000) {
-										usleep(25); /* Don't make the loop super tight, it'll probably take several hunderd/thousand us anyways. */
+									if (write_attempts < 3000) {
+										/* Don't make the loop super tight, it'll probably take several hunderd/thousand us anyways,
+										 * and we don't need to service I/O in realtime.
+										 * Don't make it super loose either, we have other work we need to get on with. */
+										usleep(500);
 										continue;
 									}
+									/* This is more than a second without making any progress, abort. */
 									bbs_error("Max SSL_write retries (%d) exceeded\n", write_attempts);
 									MARK_DEAD(ssl);
 									needcreate = 1;
@@ -596,6 +600,9 @@ static void *ssl_io_thread(void *unused)
 							}
 							bbs_debug(6, "SSL_write returned %ld (%s)\n", wres, ssl_strerror(err));
 							break;
+						} else {
+							/* Reset any time we are able to make progress. */
+							write_attempts = 0;
 						}
 					}
 				} while (wres != ores);