Instead of queuing messages when the control channels are full, xcs just

does nothing (see ctrl_interface.c:ctrl_chan_write_request()).

The following patch throttles the rate in which consoled writes data to
xcs.  With the current values, you get a responsive console but avoid
data corruption in most scenarios.

I'm able to get pretty far in my regression test with this patch.  With
higher throttle values I'm able to get even further (but the console
becomes painfully slow).

I implemented proper control channel queuing in xenctld in VM-Tools and
it's pretty nasty stuff.  This should prevent corruption for most users
until we can get rid of xcs.

Regards,

Anthony Liguori

Signed-off-by: Anthony Liguori <aliguori@us.ibm.com

diff --git a/tools/consoled/io.c b/tools/consoled/io.c
index 52b047fe806e4989a1379bc768996eba3b75fcbd..364245f019c0f042309f8e12c625378df51d406b 100644 (file)
--- a/tools/consoled/io.c
+++ b/tools/consoled/io.c
@@ -288,6 +288,7 @@ void handle_io(void)
        fd_set readfds, writefds;
        int ret;
        int max_fd = -1;
+       int num_of_writes = 0;
 
        do {
                struct domain *d;
@@ -312,6 +313,17 @@ void handle_io(void)
                }
 
                ret = select(max_fd + 1, &readfds, &writefds, 0, &tv);
+               if (tv.tv_sec == 1 && (++num_of_writes % 100) == 0) {
+                       /* FIXME */
+                       /* This is a nasty hack.  xcs does not handle the
+                          control channels filling up well at all.  We'll
+                          throttle ourselves here since we do proper
+                          queueing to give the domains a shot at pulling out
+                          the data.  Fixing xcs is not worth it as it's
+                          going away */
+                       tv.tv_usec = 1000;
+                       select(0, 0, 0, 0, &tv);
+               }
                enum_domains();
 
                if (FD_ISSET(xcs_data_fd, &readfds)) {