libsgxstep: restore APIC before enclave destroy to fix D-state hang

[Hayao0819]

Fixes #90.

Calling sgx_destroy_enclave while SGX-Step still holds /dev/sgx-step open wedges the process in uninterruptible D state, and a reboot is the only way out. As you already suspected in the thread, the culprit is the APIC mode. When the in-kernel driver's sgx_release unregisters its mmu-notifier it waits on synchronize_srcu, but SGX-Step has left the victim core's timer in one-shot mode with no pending tick, so the SRCU grace period never advances and the wait blocks forever.

Closing /dev/sgx-step fixes this, since step_release restores the timer. The catch is the timing. libsgxstep closes the fd from a destructor, which runs after main returns, i.e. long after sgx_destroy_enclave has already hung. So this splits the close out into a new step_close() and calls it just before sgx_destroy_enclave in app/bench, so the victim core is ticking normally again before the driver reaches synchronize_srcu. The destructor still calls step_close(), so the crash-cleanup path is unchanged and the second close is a guarded no-op.

Verification

On a Gemini Lake machine (Debian 12, kernel 6.1) the bench reliably wedged in D state at sgx_destroy_enclave on master. With this change it single-steps the NOP slide and sgx_destroy_enclave returns normally, and the process exits cleanly.

[ all done; counted ... IRQs (AEP/IDT) ]
APP_EXIT=0

[Hayao0819]

The problem recurred even after applying this fix. The cause may lie elsewhere. We are continuing our investigation.