BUG/MINOR: fd: avoid bad tgid assertion in fd_delete() from deinit()

In 2.7, commit 0dc1cc93b ("MAJOR: fd: grab the tgid before manipulating
running") added a check to make sure we never try to delete an FD from
the wrong thread group. It already handles the specific case of an
isolated thread (e.g. stop a listener from the CLI) but forgot to take
into account the deinit() code iterating over all idle server connections
to close them. This results in the crash below during deinit() if thread
groups are enabled and idle connections exist on a thread group higher
than 1.

  [WARNING]  (15711) : Proxy decrypt stopped (cumulated conns: FE: 64, BE: 374511).
  [WARNING]  (15711) : Proxy stats stopped (cumulated conns: FE: 0, BE: 0).
  [WARNING]  (15711) : Proxy GLOBAL stopped (cumulated conns: FE: 0, BE: 0).

  FATAL: bug condition "fd_tgid(fd) != ti->tgid && !thread_isolated()" matched at src/fd.c:369
    call trace(11):
    |       0x4a6060 [c6 04 25 01 00 00 00 00]: main-0x1d60
    |       0x67fcc6 [c7 43 68 fd ad de fd 5b]: sock_conn_ctrl_close+0x16/0x1f
    |       0x59e6f5 [48 89 ef e8 83 65 11 00]: main+0xf6935
    |       0x60ad16 [48 8b 1b 48 81 fb a0 91]: free_proxy+0x716/0xb35
    |       0x62750e [48 85 db 74 35 48 89 dd]: deinit+0xbe/0x87a
    |       0x627ce2 [89 ef e8 97 76 e7 ff 0f]: deinit_and_exit+0x12/0x19
    |       0x4a9694 [bf e6 ff 9d 00 44 89 6c]: main+0x18d4/0x2c1a

There's no harm though since all traffic already ended. This must be
backported to 2.7.

diff --git a/src/fd.c b/src/fd.c
index f4f1bae81e402f1cb024565e2d45db756a89379a..7f1077cd8cdb1784077b9c9af87b2c93aaae688d 100644 (file)
--- a/src/fd.c
+++ b/src/fd.c
@@ -366,7 +366,7 @@ void fd_delete(int fd)
        /* the tgid cannot change before a complete close so we should never
         * face the situation where we try to close an fd that was reassigned.
         */
-       BUG_ON(fd_tgid(fd) != ti->tgid && !thread_isolated());
+       BUG_ON(fd_tgid(fd) != ti->tgid && !thread_isolated() && !(global.mode & MODE_STOPPING));
 
        /* we must postpone removal of an FD that may currently be in use
         * by another thread. This can happen in the following two situations: