-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/68810/
-----------------------------------------------------------
(Updated Sept. 29, 2018, 12:05 nachm.)
Review request for mesos, Gilbert Song and Jie Yu.
Changes
-------
Addressed @jieyu's comments.
Bugs: MESOS-8978
https://issues.apache.org/jira/browse/MESOS-8978
Repository: mesos
Description (updated)
-------
Previously, we would unconditionally call 'setsid()' in the command
executor whenever we launched a process. This causes the process it
launches to start a new session and NOT inherit the controlling TTY
from it's parent. This obviously causes problems, if the goal of
attaching a TTY to a task is so that we can actually give it control
of that TTY while it is executing.
Interestingly, even if process does not control its TTY, it can still
read and write from the file descriptors associated with it (it just
can't receive any signals associated with it, such as SIGWINCH,
SIGHUP, etc.). This is why things "appeared" to mostly work until this
point because stdin/stdout/stderr were all being redirected properly
to it.
Where we saw problems was with trying to 'attach' to an already
running process launched via the command executor using the
ATTACH_NESTED_CONTAINER_INPUT/OUTPUT calls. If you ran something like
'bash', you would not be able to do job control, and you could not
resize the screen properly when running things like 'vim'.
This commit fixes this issue by checking to see if a TTY has been
attached to a task launched by the command executor, and skipping the
'setsid()' call when forking it. We considered a number of alternative
approaches, but finally settled on this one since it was the least
invasive. I.e. only tasks launched with a TTY (which is a failry new
concept in Mesos) will be affected by this change; the semantics of
all traditional tasks launched via the command executor will go
unchanged.
Note: this problem does not exists for the default executor because
the agent does the job of launching all containers, and there is no
executor sitting between the containerizer and the actual process of a
task intervening to call an extra 'setsid()'. This is why containers
launched via LAUNCH_NESTED_CONTAINER_SESSION have always worked as
expected.
Diffs (updated)
-----
src/launcher/executor.cpp 86fb87dfbaa4daba22ec491ca5fca0e562797521
src/slave/containerizer/mesos/io/switchboard.cpp
498c008c36725b4ed43b4905d3640508beaaae93
Diff: https://reviews.apache.org/r/68810/diff/2/
Changes: https://reviews.apache.org/r/68810/diff/1-2/
Testing
-------
$ make -j 40 check
Thanks,
Kevin Klues
