Check out this difference between target remote, and target extended-remote: target remote: >./gdb /home/pedro/gdb/tests/threads32 GNU gdb (GDB) 6.8.50.20081014-cvs [...] (gdb) tar remote :9999 Remote debugging using :9999 0xf7fbb810 in ?? () from /lib/ld-linux.so.2 (gdb) info threads Remote connection closed (gdb) maint print target-stack The current target stack is: - exec (Local exec file) - None (None) (gdb) q [nothing] target extended-remote: >./gdb /home/pedro/gdb/tests/threads32 GNU gdb (GDB) 6.8.50.20081014-cvs [...] (gdb) tar extended-remote :9999 Remote debugging using :9999 0xf7f70810 in ?? () from /lib/ld-linux.so.2 (gdb) c Continuing. [Switching to Thread 22227] Remote connection closed (gdb) info threads putpkt: write failed: Broken pipe. (gdb) (gdb) maint print target-stack The current target stack is: - extended-remote (Extended remote serial target in gdb-specific protocol) - exec (Local exec file) - None (None) (gdb) (gdb) q The program is running. Quit anyway (and kill it)? (y or n) y Quitting: putpkt: write failed: Broken pipe. The issue is again the mixup of "target" as in 'interface'/'debug api'/'connection to system', vs "target" as in "inferior". In the remote target, a target_mourn_inferior unpushes the target_ops, while in the extended-remote target, it doesn't, leaving the user with a useless broken connection. The attached patch makes the extended-remote behave the same as the remote target. Considering an extended-remote connection debugging multi-processes seems to make it clearer that target_mourn_inferior isn't the right call here, me thinks. There are cases in async mode that when the connection was broken, we'd leave the SIGINT signal handler set to handle_remote_sigint or handle_remote_sigint_twice, although we had already poped the target, which would result in later crashes. I'm also making sure in remote_close that that doesn't happen. Any objections to this? No regressions on x86_64-unknown-linux-gnu (sync/async). -- Pedro Alves