[Bug 1645324] Re: ebtables: Lock file handling has races

[Eric Desrochers]

<slashd> rbasak, morning I see you talked with dragan about this LP:
#1645324. From now on, I'm taking care of this bug, I see that some ppls
suggest Dragan to submit in the upstream project first, but in this case
it seems that ebtables is nearly "dead" last commits were made in 2015.
Nowadays, all the development happen in nft. Do you think Dragan's patch
could be eligilble for SRu anyway ? considering that Dragan already
submitted the same patch to Debian ebtables but not upstream.

<rbasak> slashd: can we get the patch committed/uploaded into Debian?
That would also alleviate Ubuntu's maintenance burden.

-- 
You received this bug notification because you are a member of Ubuntu
Sponsors Team, which is subscribed to the bug report.
https://bugs.launchpad.net/bugs/1645324

Title:
  ebtables: Lock file handling has races

Status in ebtables package in Ubuntu:
  In Progress
Status in ebtables source package in Trusty:
  In Progress
Status in ebtables source package in Zesty:
  Triaged
Status in ebtables source package in Artful:
  In Progress
Status in ebtables package in Debian:
  New

Bug description:
  [Impact]

   * ebtables uses creation of a file with an exclusive flag
     as a lock to synchronize table modification when used
     with --concurrent parameter.

   * If ebtables crashes it will leave a stale lock file.
     This will prevent another copy of ebtables from running,
     and indirectly any other service that depends on ebtables
     will also be affected.

   * This change adds support for real locks that get
     cleaned up if a process exits or crashes.

  [Test Case]

   * Test Case1:
     1. $ sudo touch /var/lib/ebtables/lock"
     2. $ sudo ebtables --concurrent -L
     3. ebtables can't acquire a lock

   * Test Case 2:
     1. $ while true; do /usr/sbin/ebtables --concurrent -L; done
     2. hard reboot VM
     3. likely that the lock file is present under /var/lib/ebtables
     4. libvird hanging, try to connect to qemu:///system

  [Regression Potential]

   * Normal Use:
     There is no regression potential during normal use and
     operation of ebtables.

   * Package Upgrade:
     There is a very very small regression potential during the package
     upgrade to the latest version. Once the package is upgraded that
     potential is gone. It is a very small potential because several
     things have to happen in a very small time frame and in an exact
     order since ebtables is not a resident program like a daemon:
       1. ebtables is launched during package upgrade AND
       2. new ebtables binary has not yet been written to disk AND
       3. it is launched with --concurent switch AND
       4. another ebtables with new binary is launched AND
       5. it is launched with --concurent switch AND
       6. the first ebtables copy hasn't exited yet AND
       7. both copies of ebtables are launched with a WRITE command AND
       8. both copies of ebtables are manipulating the same resource.
     Then one of the binaries could potentially fail, but once the old
     binary exits the potential is gone so subsequent re-runs of
     ebtables will succeed.

  * Dragan's patch has been submitted to Debian via :
    https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=860590

  * Note that the ebtables upstream project is nearly dead. Nowadays,
  all the development is now happening in nft project which is intended
  to be replacement.

  
  [Original Text]
  libvirtd is hanging after startup due to ebtables lock file -from an earlier run- remains intact when the system reboots.
  Same issue is happening than it is reported here: https://bugzilla.redhat.com/show_bug.cgi?id=1290327 when the system boots.

  After booting the system, It's not possible connect to the qemu-service.
  - libvirt daemon tried to obtain a lock:
  [pid 20966] read(24, "Trying to obtain lock /var/lib/e"..., 1024) = 45
  [pid 20966] poll([{fd=22, events=POLLIN}, {fd=24, events=POLLIN}], 2, 4294967295) = 1 ([{fd=24, revents=POLLIN}])
  [pid 20966] read(24, "Trying to obtain lock /var/lib/e"..., 1024) = 45
  [pid 20966] poll([{fd=22, events=POLLIN}, {fd=24, events=POLLIN}], 2, 4294967295) = 1 ([{fd=24, revents=POLLIN}])
  [pid 20966] read(24, "Trying to obtain lock /var/lib/e"..., 1024) = 45
  [pid 20966] poll([{fd=22, events=POLLIN}, {fd=24, events=POLLIN}], 2, 4294967295^CProcess 20916 detached

  - there was a file named 'lock' in /var/lib/ebtables directory with timestamp 14:54
  - ebtables was configured:
  * Ebtables support available, number of installed rules [ OK ]
  (other nodes appeared to be in the same state from ebtables point of view, but without the lock file)
  - I removed the lock file and libvirt started to work instantly - the lock obtain messages have disappeared from the trace and virsh commands are working
  - at 14:54 the host was booting up. According to the logs, there were other reboots after that one, but the lock file remained intact (at least the timestamp was not updated).

  Could you please suggest a solution to be sure that ebtables lock file
  is removed during boot?

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/ebtables/+bug/1645324/+subscriptions