…_) updated the FDT parsing in init(), which now selects the third interrupt triplet of the "arm,armv8-timer" node (Virtual, PPI11 / INTID 27). init_cpu() still selected the second triplet (Non-secure Phys, PPI 14 / INTID 30), so every application processor enabled the physical timer PPI in its redistributor while set_oneshot_timer programs the virtual timer. Timer PPIs are per-core: sleeping tasks on the secondary cores were never woken up again.

This is why thread_test with --smp 4 timed out: of eight threads only the two scheduled on the boot core finished; the six on the secondary cores slept forever.

Skip the same two triplets in init_cpu() so all cores enable the virtual timer interrupt.

test_stdin fed each message to the guest with a one-second pause and
then terminated QEMU immediately, reading the echoed output only
afterwards. On a slow target the last echo had not been written yet
when QEMU was killed, so the collected output ended mid-line and the
assertion failed intermittently. This showed up on aarch64_be, which
is emulated without hardware acceleration.

Collect the guest's stdout on a background thread and wait until every
message has been echoed (bounded by a 60 s timeout) before terminating
QEMU. The check no longer depends on the guest keeping pace with a
fixed schedule.

init() and init_cpu() each parsed the "arm,armv8-timer" device-tree node
on their own and picked a triplet by hand, so the two could disagree about
which interrupt the kernel uses. The kernel arms the physical timer
(CNTP_CVAL_EL0/CNTP_CTL_EL0), whose PPI is the Non-secure Phys triplet
(INTID 30), but init_cpu() selected the Virtual triplet (INTID 27). Timer
PPIs are per-core, so every application processor enabled an interrupt that
never fires and tasks sleeping on the secondary cores were never woken up
again — thread_test hung until the QEMU timeout on aarch64_be with SMP 4.

Parse the timer interrupt once in init(), cache the resolved IntId and
trigger mode in TIMER_IRQ, and let init_cpu() enable that exact interrupt
in its redistributor. This removes the duplicated FDT parsing that let the
two sites drift apart.