diff options
Diffstat (limited to 'results/classifier/gemma3:12b/debug/1725')
| -rw-r--r-- | results/classifier/gemma3:12b/debug/1725 | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/results/classifier/gemma3:12b/debug/1725 b/results/classifier/gemma3:12b/debug/1725 new file mode 100644 index 000000000..8679a786e --- /dev/null +++ b/results/classifier/gemma3:12b/debug/1725 @@ -0,0 +1,22 @@ + +qemu-system-x86_64 reports wrong thread to GDB on SIGINT +Description of problem: +Upon interruption of a thread by GDB, QEMU in some circumstances will send a stop reply with the ID of a thread that had not been resumed. + +This happens for the following reasons: +1. GDB uses `vCont` exclusively to resume and step through threads. +2. When a thread is interrupted by GDB, QEMU runs `vm_stop(RUN_STATE_PAUSED)`, which triggers `gdb_vm_state_change`, which, in turn, uses whatever CPU is pointed to by `gdbserver_state.c_cpu` at that time to construct the stop reply. +3. The `vCont` handler in QEMU doesn't set `gdbserver_state.c_cpu` before resuming any CPUs. + +Important to note is that stepping is not affected by this issue because the `EXCP_DEBUG` handler sets `gdbserver_state.c_cpu` to the CPU the exception happened in before `gdb_vm_state_change` runs. Which also means single stepping before continuing is an effective way to work around this bug. +Steps to reproduce: +1. Run QEMU with at least two threads and the GDB stub enabled. +2. Run `gdb --nx --ex 'target remote :1234' --ex 'set scheduler-locking on'` +3. Switch to Thread 1.2 in GDB with `thr 2` +4. Resume Thread 1.2 in GDB with `c` +5. Press Ctrl+C to interrupt the VM +6. Notice that the event is reported as having happened in Thread 1.1, which has not been resumed. +Additional information: +Note that, while this bug happens no matter the state of `scheduler-locking`, it only becomes a problem when it is enabled. This is because, when it is disabled, GDB will always resume all threads on `continue`, so it doesn't matter what thread ID QEMU says the interrupt happened in, as it is guaranteed to have been resumed anyway. That, however, is not the case when `scheduler-locking` is enabled. + +Regardless, I don't think it makes sense for QEMU to be reporting events happening in threads that weren't resumed through either `s/S/c/C` or `vCont`, which is what it's doing here. |