FTR I messed up the descriptions when updating this and bug 2172863
(similar titles, sorry). I've updated both descriptions before starting
the SRU verification.

** Description changed:

- [ Impact ]
+   [ Impact ]
  
-  * Running kill-session on a sesion whose window is zoomed crashes the tmux
-    server with SIGSEGV.
+    * Tmux server crashes with SIGSEGV when a control mode client that has
+      subscriptions and the wait-exit flag detaches.
  
-  * This is fixed upstream in tmux 3.5, and the issue does not manifest in
-    versions before 3.4. Therefore, only Noble is affected.
+    * This affects iTerm2 with the tmux integration active. Closing an iTerm2 
tmux
+      window over ssh can bring down the remote ssh server.
  
-  * This bug is fixed by cherry-picking the upstream patch.
+    * This bug is fixed by backporting an upstream patch that checks for a null
+      session.
  
- [ Test Plan ]
+   [ Test Plan ]
  
-  * Run the following script:
+    * Run the following script:
  
-   unset TMUX
-   export TMUX_TMPDIR=$(mktemp -d)
-   T="tmux -f /dev/null -L sru"
-   $T new-session -d -s anchor -x 80 -y 24
-   $T new-session -d -s a -x 80 -y 24
-   $T split-window -t =a:
-   $T resize-pane -Z -t =a:
-   $T kill-session -t =a
-   sleep 1
-   $T list-sessions
+   #!/bin/sh
+   # Reproducer: tmux server SIGSEGV in control_check_subs_timer after a
+   # control client with subscriptions and wait-exit detaches.
+   TMUX_BIN=${TMUX_BIN:-tmux}
+   T="$TMUX_BIN -L subs-repro -f /dev/null"
  
-  * On an unpatched tmux, this outputs `server exited unexpectedly`
+   $T kill-server 2>/dev/null
+   $T new-session -d -s keep          # a second session, so the server has no 
reason to exit
+   $T new-session -d -s a
+   PID=$($T display-message -p '#{pid}')
  
-  * On the fixed version, the server does not exit and the script outputs the
-    anchor session.
+   # Control client: add a subscription, enable wait-exit (as iTerm2 does),
+   # then detach. After %exit the client waits for an empty line on stdin,
+   # which we never send - like a terminal that is slow to answer.
+   {
+     printf '%s\n' "refresh-client -B 'foo:%*:#{pane_id}'"
+     printf '%s\n' 'refresh-client -f wait-exit'
+     sleep 2
+     printf '%s\n' 'detach-client'
+     sleep 5
+   } | $T -C attach-session -t a >/dev/null &
  
- [ Where problems could occur ]
+   sleep 5
+   case "$(ps -o stat= -p "$PID")" in
+     ""|Z*) echo "server $PID died (crash)" ;;
+     *)     echo "server $PID alive" ;;
+   esac
+   wait
+   $T kill-server 2>/dev/null
  
-  * Regressions could appear in other instances of window teardown: 
kill-window,
-    kill-pane, etc.
+    * On an unpatched tmux, this shows `server <pid> died (crash)`
  
- [ Other Info ]
+    * On the fixed version, this shows `server <pid> alive`
  
-  * The original report (copied below) was itself a SRU bug template. I decided
-    to not edit it directly. Hope this is OK.
+   [ Where problems could occur ]
  
- [ Original Report ]
+    * The backported patch only touches the subscription timer callback code 
path.
+      Regressions would manifest in that area, possibly also related to the 
iTerm2
+      integration.
  
- [Impact]
+   [ Other Info ]
  
- On tmux 3.4-1ubuntu0.2 (noble), running kill-session on a session whose
- window is zoomed crashes the tmux server with SIGSEGV. One server
- normally holds all of a user's sessions, so every session on it ends,
- along with every process running in them. There is no warning and
- nothing can be recovered. It reproduces with no configuration file.
+    * The original report (copied below) was itself a SRU bug template. I 
decided
+      to not edit it directly. Hope this is OK.
  
- Zooming a pane (C-b z) is a common action, and ending a session is too.
- Scripts that manage sessions can reach this without the user doing
- anything unusual.
+   [Original Report]
  
- Upstream fixed this in tmux 3.5. The bug report is
- https://github.com/tmux/tmux/issues/3857 ("Zoomed windows crash tmux
- after killing session with `x` on tree mode"), which reproduces it from
- the keyboard: split with C-b %, zoom with C-b z, open tree mode with C-b
- s, press x on the session, confirm with y. The fix is
- https://github.com/tmux/tmux/pull/3860 ("Don't notify when unzooming
- windows from window_destroy()", by Romain Francoise), whose branch
- commit is 5363a5475a1dd4158c1ddb23de2da34e88738673. The maintainer
- applied it through OpenBSD, so the hash on tmux's master may differ; I
- haven't confirmed it. 3.5's CHANGES lists it as "Do not notify window-
- layout-changed if the window is about to be destroyed". The pull request
- says the bug came in with commit 36e1ac6. I've read the 3.5a source but
- haven't built or run it.
+   Release and package (output of `lsb_release -rd` and `apt-cache policy tmux`
+   on the affected server):
  
- Backtrace, from a core dump with tmux-dbgsym 3.4-1ubuntu0.2:
+       No LSB modules are available.
+       Description:    Ubuntu 24.04.5 LTS
+       Release:        24.04
+       tmux:
+         Installed: 3.4-1ubuntu0.2
+         Candidate: 3.4-1ubuntu0.2
+         Version table:
+        *** 3.4-1ubuntu0.2 500
+               500 http://se.archive.ubuntu.com/ubuntu noble-updates/main 
amd64 Packages
+               100 /var/lib/dpkg/status
+            3.4-1build1 500
+               500 http://se.archive.ubuntu.com/ubuntu noble/main amd64 
Packages
  
-   #0 options_tree_RB_MINMAX   options.c:90
-   #1 options_free             options.c:186
-   #2 window_destroy           window.c:360
-   #3 window_remove_ref        window.c:398   (references now 0)
-   #4 notify_callback          notify.c:165   (ne->name 
"window-layout-changed", ne->window the same window)
-   #5 cmdq_next ... server_loop
+   Expected: closing an iTerm2 tmux -CC window detaches the client and the
+   tmux server keeps running.
+   Actual: the tmux server crashes with SIGSEGV and all sessions are lost.
  
- How it happens, from reading 3.4's window.c and notify.c, and consistent with 
the core:
- 1. The last reference to a zoomed window goes, so window_destroy starts.
- 2. window_destroy calls window_unzoom, which calls 
notify_window("window-layout-changed", w). That takes a new reference and 
queues a callback.
- 3. window_destroy carries on and frees w.
- 4. The queued callback drops its reference on freed memory, reaches 0, and 
calls window_destroy again.
+   [Impact]
  
- [Test Plan]
+   On noble, the tmux server crashes with SIGSEGV when a control mode client 
that
+   has subscriptions and the wait-exit flag detaches. Every session and every
+   process running in them is lost.
  
- As an ordinary user, on a private socket so no existing tmux server is
- touched:
+   iTerm2's tmux integration (`tmux -CC`) sets both. It sends
+   `refresh-client -fpause-after=0,wait-exit` and subscribes with
+   `refresh-client -B`. So closing an iTerm2 tmux window over ssh can bring 
down
+   the whole server. We hit it three times in three days on a noble server.
  
-   unset TMUX
-   export TMUX_TMPDIR=$(mktemp -d)
-   T="tmux -f /dev/null -L sru"
-   $T new-session -d -s anchor -x 80 -y 24
-   $T new-session -d -s a -x 80 -y 24
-   $T split-window -t =a:
-   $T resize-pane -Z -t =a:
-   $T kill-session -t =a
-   sleep 1
-   $T list-sessions
+   Kernel log, the same instruction every time (offset 0xf0 is
+   `struct session` member `windows` and 0x9f8 is `struct client` member
+   `session` in the 3.4-1ubuntu0.2 debug symbols, so this is the
+   `%*`/`@*` subscription path that iTerm2 uses):
  
- Broken: "server exited unexpectedly" from kill-session, then "no server 
running on ..." from list-sessions.
- Fixed: list-sessions prints the "anchor" session.
+       tmux: server[1057941]: segfault at f0 ip 0000593fe1c9db25 sp
+   00007ffe794273f0 error 4 in tmux[593fe1c6e000+af000]
  
- To clean up: $T kill-server; rm -rf "$TMUX_TMPDIR"
+   Cause: on MSG_EXITING the server clears c->session. With wait-exit the 
client
+   then stays connected until the terminal sends an empty line. If the
+   one-second subscription timer fires in that window, 
control_check_subs_timer()
+   walks s->windows with s == NULL:
  
- The same steps without the resize-pane line leave "anchor" alive, on the
- broken version too.
+       #0  winlinks_RB_MINMAX ()
+       #1  control_check_subs_timer ()
+       #2  ?? () from /lib/x86_64-linux-gnu/libevent_core-2.1.so.7
+       #3  event_base_loop ()
+       #4  proc_loop ()
+       #5  server_start ()
  
- Results on 3.4-1ubuntu0.2, Ubuntu 24.04.5: broken 3 runs of 3 with these
- exact steps, and 8 of 8 in an earlier scripted run; without the zoom, 0
- of 3 and 0 of 8.
+   The same NULL session is also reachable when the attached session is 
destroyed
+   with no replacement (server_destroy_session) and on server exit
+   (server_send_exit).
  
- [Where problems could occur]
+   Upstream: there is no targeted fix. The crash went away in tmux 3.8 as a 
side
+   effect of moving subscriptions into
+   monitor.c (commit 8e03310a, "Move monitor stuff out of control.c into its 
own
+   file for reuse"). The new code checks for a NULL session before evaluating.
+   That rewrite is too large to backport. The attached debdiff adds the
+   equivalent three-line check to control_check_subs_timer() instead.
  
- The fix changes what happens when a zoomed window is destroyed, and whether 
window-layout-changed fires then.
- - A regression could appear as a crash or a leak elsewhere in window 
teardown: kill-window, kill-pane, a pane's last process exiting, or a server 
shutting down with zoomed windows.
- - Hooks and control-mode clients listening for window-layout-changed would no 
longer see that event for a window being destroyed. Anything relying on it 
would notice.
- - Unzooming on ordinary paths (resize-pane -Z, select-pane -Z) must keep 
firing the notification and keep the layout; if the change reaches further than 
destruction, zoom behaviour would be where it shows.
+   [Test Plan]
  
- Regression testing should cover zoom and unzoom, kill-window and kill-
- pane on zoomed windows, and a hook on window-layout-changed.
+   No terminal is needed. The attached `tmux-subs-repro.sh` starts a private
+   server (`-L subs-repro -f /dev/null`), attaches a control client on a pipe,
+   subscribes, sets wait-exit and detaches, and then never sends the empty 
line:
  
- [Other Info]
+       sh tmux-subs-repro.sh
  
- Other routes to the same crash, same setup: an ungrouped session with a
- zoomed window, with destroy-unattached set, when its last client
- disconnects. Killing the zoomed window or pane directly (kill-window,
- kill-pane) did not crash in my runs, 8 of 8.
+   Expected with the fix: `server <pid> alive`. Without it: `server <pid> died
+   (crash)`.
  
- Workaround: unzoom before kill-session.
+   Results (run in an Ubuntu 24.04 container):
  
- I searched for an existing report and didn't find one. I have not
- prepared a debdiff.
+   | Build                                | Crashes                            
           |
+   | ------------------------------------ | 
--------------------------------------------- |
+   | 3.4-1ubuntu0.2                       | 3/3, 1/2 (it is a race against the 
1 s timer) |
+   | 3.4-1ubuntu0.2 + attached patch      | 0/5                                
           |
+   | upstream 3.6a                        | 5/5                                
           |
+   | upstream 3.8 and master (2026-10-30) | 0/5 each                           
           |
+ 
+   Also checked with the patched build: subscriptions still report changes
+   (`%subscription-changed` for a pane title change), and the destroyed-session
+   path (kill-session while a control client with subscriptions has queued
+   output, detach-on-destroy on) no longer crashes (0/3). Without the patch it
+   crashes.
+ 
+   [Where problems could occur]
+ 
+   The change only returns early from the subscription timer callback while the
+   control client has no session. The timer stays armed, so a client that gets 
a
+   session again resumes its subscriptions. A client without a session has
+   nothing its subscriptions could report, since every subscription type is
+   relative to the client's session. A regression would show up as missing
+   `%subscription-changed` notifications in control mode clients such as 
iTerm2.
+ 
+   [Other Info]
+ 
+   - Workaround: none that is reliable. `detach-on-destroy off` covers only the
+     destroyed-session path, not the detach path that iTerm2 triggers.
+   - The same timer code is present in upstream 3.5a, 3.6a and 3.7c. I have
+     tested only noble.

** Tags removed: verification-needed-resolute
** Tags added: verification-done-resolute

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2172659

Title:
  tmux server crashes (SIGSEGV in control_check_subs_timer) when an
  iTerm2 tmux -CC client detaches

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/tmux/+bug/2172659/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to