Kernel: Prevent lost wakeups in the idle loop

[spholz]

In the old idle loop implementation, lost wakeups can happen when the processor receives an interrupt immediately before wait_for_interrupt(). In that situation, we still go to sleep, even if the interrupt caused a thread to be runnable or we got notified of a runnable thread.

This new idle loop implementation instead runs with interrupts disabled. The only place in this new idle loop where we can get interrupts is during Processor::idle(). After that, we will only go to sleep if there are still no runnable threads.

With this new idle loop, boot time is reduced from 4 seconds down to 3 seconds on my x86-64 system when running in KVM. If I remove this busy wait, boot time is reduced down to 2 seconds (#26444):

serenity/Userland/Services/SystemServer/main.cpp

Line 135 in cef72c7

sleep(1);

[spholz]

Before vs. after trace generated with the QEMU uftrace plugin and displayed in the Spall profiler:

(The trace snippets don't show the same NVMe interrupt and the first one is from a debug build, but that doesn't matter)

In the new trace, we immediately schedule again after reading the data from disk. Before, we only caused a context switch after receiving a timer interrupt.

[LucasChollet]

Very nice improvement!