From 914d1378b2eaf1dd439ebc562b0baa10127bd4f7 Mon Sep 17 00:00:00 2001
From: "Richard W.M. Jones" <rjones@redhat.com>
Date: Thu, 18 Aug 2011 19:13:32 +0100
Subject: [PATCH] protocol: Fix case where download can fail for small files.

There is another case where downloads of small files could fail if the
library side (writer) fails.  In this case the library would send back
a cancellation, but it would be received after the daemon had finished
sending the whole file (because the file is small enough).  The daemon
would reenter the main loop and immediately get an unexpected cancel
message, causing the daemon to die.

This commit also makes test-cancellation-download-librarycancels.sh
more robust.  We use Monte-Carlo testing with a range of file sizes.
Small file sizes should trigger the error case.
(cherry picked from commit e4cba8f2b1a68e7361ce342ff659cccb0490446e)
---
 daemon/proto.c                                     |  6 +++++
 .../test-cancellation-download-librarycancels.sh   | 26 +++++++++++++++++-----
 2 files changed, 26 insertions(+), 6 deletions(-)

diff --git a/daemon/proto.c b/daemon/proto.c
index 88e2d28..238ee88 100644
--- a/daemon/proto.c
+++ b/daemon/proto.c
@@ -101,6 +101,12 @@ main_loop (int _sock)
 	       "guestfsd: main_loop: new request, len 0x%" PRIx32 "\n",
 	       len);
 
+    /* Cancellation sent from the library and received after the
+     * previous request has finished processing.  Just ignore it.
+     */
+    if (len == GUESTFS_CANCEL_FLAG)
+      continue;
+
     if (len > GUESTFS_MESSAGE_MAX) {
       fprintf (stderr, "guestfsd: incoming message is too long (%u bytes)\n",
                len);
diff --git a/regressions/test-cancellation-download-librarycancels.sh b/regressions/test-cancellation-download-librarycancels.sh
index 20f89d2..5f6ebf1 100755
--- a/regressions/test-cancellation-download-librarycancels.sh
+++ b/regressions/test-cancellation-download-librarycancels.sh
@@ -18,22 +18,36 @@
 
 # Test download where the library cancels.
 #
-#
+# Download big and small files to /dev/full.  This should fail but not
+# kill the appliance.  We test various randomized file sizes because
+# there are many potential race conditions -- for example the daemon
+# may or may not send all of its data because the error condition is
+# detected.
 
 set -e
 
 rm -f test.img
 
-../fish/guestfish <<'EOF'
-add ../images/test.iso
+size=$(awk 'BEGIN{ srand(); print int(16*1024*rand()) }')
+echo "$0: test size $size (bytes)"
+
+../fish/guestfish <<EOF
+# We want the file to be fully allocated.
+alloc test.img 10M
 run
 
-mount-ro /dev/sda /
+part-disk /dev/sda mbr
+mkfs ext2 /dev/sda1
+mount-options "" /dev/sda1 /
 
-# Download a file to /dev/full.
--download /100krandom /dev/full
+fallocate64 /file $size
 
+# Download the file into /dev/full so it fails.
+-download /file /dev/full
+
+# The daemon should still be reachable after the failure.
 ping-daemon
+
 EOF
 
 rm -f test.img
-- 
1.8.3.1