pty: optimize read loop

As it turns out, I can actually send data to the pty faster than the
terminal can read. Therefore, make sure we read as much data as possible
but bail out early enough to not cause starvation.

Kernel TTY buffers are 4k, so reduce the overall buffer size, but read
more than once if possible (up to 8 times sounds reasonable).

diff --git a/src/shared/pty.c b/src/shared/pty.c
index adcb32d0be733af138ceb7bec97821065d6cb016..52a426c0e15fd8314571570bcb94f8291cc3ae85 100644 (file)
--- a/src/shared/pty.c
+++ b/src/shared/pty.c
@@ -67,7 +67,7 @@
 #include "ring.h"
 #include "util.h"
 
-#define PTY_BUFSIZE 16384
+#define PTY_BUFSIZE 4096
 
 enum {
         PTY_ROLE_UNKNOWN,
@@ -305,11 +305,11 @@ static int pty_dispatch_read(Pty *pty) {
         /*
          * We're edge-triggered, means we need to read the whole queue. This,
          * however, might cause us to stall if the writer is faster than we
-         * are. Therefore, we read twice and if the second read still returned
-         * data, we reschedule.
+         * are. Therefore, try reading as much as 8 times (32KiB) and only
+         * bail out then.
          */
 
-        for (i = 0; i < 2; ++i) {
+        for (i = 0; i < 8; ++i) {
                 len = read(pty->fd, pty->in_buf, sizeof(pty->in_buf) - 1);
                 if (len < 0) {
                         if (errno == EINTR)