[Cluster-devel] [PATCHv2 dlm-next 0/3] fs: dlm: receive handling changes

[Alexander Aring <aahringo@redhat.com>]

Hi,

this patch series will do some changes according to the receive handling
in dlm. My goal is to survive tcpkill [0] which is my main testcase while
doing some heavy load on dlm. The code is still in a situation where dlm
can end in a deadlock state while e.g. using gfs2. For generating a lot
of dlm traffic I use gfs2 and running make_panic [1].

I mainly just run make_panic on two nodes and let run:

tcpkill -9 -i $IFACE port 21064

on both nodes. The first two patches changes the behaviour how to react
on invalid dlm messages (which does not occur on my test case). The third
patch increase in my test scenario the stability of dlm connection and
running tcpkill. The last patch will ensure that the other end agrees
to close the socket if we want to do a clean tcp termination.

FUTURE WORK:

I still get deadlocks, as PATCH 2/3 shows that I think the TCP stack
will drop some packets when receiving a tcp reset (and the application
layer think it's successful send). This is one of my future work to
solve this behaviour. I also think about to extend the test case to e.g.
umount/mount gfs2 while my testcase is running and doing randomly
drops/delaying/reordering of dlm tcp by using the introduced mark
functionality and traffic control.

- Alex

[0] https://salsa.debian.org/pkg-security-team/dsniff
[1] https://fedorapeople.org/cgit/teigland/public_git/dct-stuff.git/tree/fs/make_panic

changes since v2:
 - fixup patch 2/4 with 1/4, otherwise we have some broken state
 - correct cbuf_eat() to "eat" full received buffer size in case of
   receiption of invalid dlm message.

Alexander Aring (3):
  fs: dlm: don't close socket on invalid message
  fs: dlm: change handling of reconnects
  fs: dlm: implement tcp graceful shutdown

 fs/dlm/lowcomms.c | 112 +++++++++++++++++++++++++++++++++++-----------
 1 file changed, 87 insertions(+), 25 deletions(-)

-- 
2.26.2