You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
gdbstub runs on a separate thread and talks to the vCPU thread through two channels. That needs 12 request and 16 reply message kinds, a startup handshake that sends the interrupt handle, and a loop that watches both the channel and the socket.
Problems this causes on main:
The stop-reason loop spins at 100% CPU while the guest runs. Each pass toggles O_NONBLOCK and peeks the socket (event_loop.rs#L85).
Ctrl-C while the vCPU is not running, for example during a host call, ends the gdb session. on_interrupt returns a fatal error (event_loop.rs#L108).
A sandbox from a running snapshot can be dropped before a client connects. Its gdb thread then stays blocked in accept() and holds the port forever (mod.rs#L305). Later sandboxes on that port only log an error and run without gdb.
An unexpected disconnect leaves debugging enabled and int3 bytes in memory. The next debug exit fails the guest call. The docs say the guest continues (how-to-debug-a-hyperlight-guest.md#L29).
If reporting a crash to gdb fails, the ? returns that debug error instead of the guest's error (hyperlight_vm/mod.rs#L798).
The gdb thread's spawn result and errors are discarded (mod.rs#L301).
Proposal:
The vCPU thread drives gdbstub's GdbStubStateMachine at each debug stop. gdbstub's Target is implemented on HyperlightVm, so requests call the VM directly. The message enums go away.
HyperlightVm owns a GdbSession with the listener and the connection. Dropping the sandbox frees the port.
The client is accepted at the first debug stop. A client may connect before the vCPU stops.
A small watcher thread blocks in peek on a clone of the socket while the guest runs. On data it calls kill_from_debugger. This is the only extra thread.
While stopped, the target holds clones of the HostSharedMemory handles. This avoids lifetimes and unsafe.
Draft result (not yet pushed): about 650 fewer lines. It passes all gdb tests on KVM, plus a new end-to-end Ctrl-C test. It fixes every problem above, #1973, and the rbx read in #1971.
Activity