Steps towards eliminating FIONREAD.

2 files changed, 12 insertions, 6 deletions

diff --git a/TODO b/TODO
index ec800058..641622b5 100644
--- a/TODO
+++ b/TODO
@@ -78,10 +78,14 @@ Rob Janssen writes:
 > --10, this further increases the load and slows down the backup to a crawl.
 > Killing gpsd makes the backup finish and then it can be started again.
 
-He adds that this occurs with SiRF+Zodiac and SiRF+TSIP.  He suspects
-this may be due to a bug in the SiRF support, interacting with extreme
-conditions in the Linux kernel's memory management.  The bug seems not
-to occur during ext3 backups, only Reiserfs ones.
+He further observers that the FIONREAD ioctl() we use to check for
+input waiting spins returning 0 when this happens.  We suspect that
+memory resource crises do something nasty to the buffering in the
+Linux serial layer.  The bug seems not to occur during ext3 backups,
+only Reiserfs ones.
+
+This probably means we need to get rid of the FIONREAD ioctl(), which
+has always been kind of dodgy anyway.
 
 ** To do:
 
diff --git a/packet.c b/packet.c
index 964441d3..38abc682 100644
--- a/packet.c
+++ b/packet.c
@@ -7,8 +7,8 @@ DESCRIPTION:
 
 Initial conditions of the problem:
 
-1. We have a file descriptor open for read. The device on the other end is 
-   sending packets at us.  
+1. We have a file descriptor open for (possibly non-blockin) read. The device 
+   on the other end is sending packets at us.  
 
 2. It may require more than one read to gather a packet.  Reads may span packet
    boundaries.
@@ -527,6 +527,8 @@ int packet_get(struct gps_device_t *session, size_t waiting)
 
     if (newdata < 0 && errno != EAGAIN)
 	return BAD_PACKET;
+    else if (newdata == 0 || (newdata < 0 && errno == EAGAIN))
+	return 0;
 
 #ifdef STATE_DEBUG
     gpsd_report(6, "Read %d chars to buffer offset %d (total %d): %s\n",