[MESOS-8980] mesos-slave can deadlock with docker pull

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.4.1, 1.5.1, 1.6.0
• Fix Version/s: 1.4.2, 1.5.2, 1.6.1, 1.7.0
• Target Version/s: 1.4.2, 1.5.2, 1.6.1, 1.7.0
• Assignee: Kjetil Joergensen
• Reporter: Kjetil Joergensen
• Created: 05/Jun/18 23:42
• Updated: 08/Jun/18 05:34
• Resolved: 08/Jun/18 05:34

Description

Similar to MESOS-1885.

mesos-slave create pipes for stdout/stderr for docker pull, forks & execs docker pull. When the output of docker pull exceeds the buffer allocated to the pipe, mesos-slave and docker pull will deadlock, where docker pull blocks on writing to stdout, and mesos-slave is waiting for docker pull to exit.

Under "normal" circumstances this seems somewhat rare, although if you have enough jobs running, you'll get to a point where the sum-total of buffer for pipes allocated gets to fs.pipe-max-size, at which point linux will give you a single page of memory for the pipe buffer, at which point a moderate amount of layers will push you over the 4k pipe buffer.

Is stdout for docker pull being used for anything ? (Cursory testing & reading said no immediate observable harm). If not, the following should do the trick

diff --git a/src/docker/docker.cpp b/src/docker/docker.cpp
index d423d56ad..daebb897b 100755
--- a/src/docker/docker.cpp
+++ b/src/docker/docker.cpp
@@ -1413,7 +1413,7 @@ Future<Docker::Image> Docker::__pull(
       path,
       argv,
       Subprocess::PATH("/dev/null"),
-      Subprocess::PIPE(),
+      Subprocess::PATH("/dev/null"),
       Subprocess::PIPE(),
       nullptr,
       environment);