Calling os::shutdown() doesn't shut down the machine when run with vmrunner because of a missing end-of-transmission (\x04) signal over the wire. To actually shut down the machine, we currently have to call __arch_poweroff().
os::halt() is a confusing name, despite matching the hlt instruction of x86. I feel this should be renamed to something like os::yield(), os::pause(), or os::wait_for_interrupt(). As it stands, without reading the implementation, it's easy to assume it's equivalent to shutting down the machine.
While it would be easy to think os::shutdown() should call __arch_poweroff(), instead of only setting the state of the kernel as non-running, it seems like the semantic meaning of "shutting down" means that there are no more events pending on the unikernel (see src/platform/x86_pc/os.cpp). Is this correct? Feels confusing.
Note that __arch_poweroff() and __arch_reboot() are actually exposed through api/arch.hpp, so services can call this if they need to. That feels like an implementation detail that shouldn't be exposed like that, but rather through os::shutdown() (or perhaps os::poweroff(), if shutdown's name means something else).
Furthermore, it's now expected that services shut down the os themselves. There is a weak Service::stop() function with an empty implementation that is called after the kernel stops running. My first intuition was that this would work, but as explained, this function is actually called after the service has stopped running, which is not equivalent to the service having returned:
#include <os>
void Service::start(){ std::println("Hello :)"); }
void Service::stop(){ os::shutdown(); }We advertise that this should work, which leaves the machine hanging:
#include <os>
int main(){
std::println("Hello :)");
}
Calling
os::shutdown()doesn't shut down the machine when run with vmrunner because of a missing end-of-transmission (\x04) signal over the wire. To actually shut down the machine, we currently have to call__arch_poweroff().os::halt()is a confusing name, despite matching thehltinstruction of x86. I feel this should be renamed to something likeos::yield(),os::pause(), oros::wait_for_interrupt(). As it stands, without reading the implementation, it's easy to assume it's equivalent to shutting down the machine.While it would be easy to think
os::shutdown()should call__arch_poweroff(), instead of only setting the state of the kernel as non-running, it seems like the semantic meaning of "shutting down" means that there are no more events pending on the unikernel (seesrc/platform/x86_pc/os.cpp). Is this correct? Feels confusing.Note that
__arch_poweroff()and__arch_reboot()are actually exposed throughapi/arch.hpp, so services can call this if they need to. That feels like an implementation detail that shouldn't be exposed like that, but rather throughos::shutdown()(or perhapsos::poweroff(), if shutdown's name means something else).Furthermore, it's now expected that services shut down the os themselves. There is a weak
Service::stop()function with an empty implementation that is called after the kernel stops running. My first intuition was that this would work, but as explained, this function is actually called after the service has stopped running, which is not equivalent to the service having returned:We advertise that this should work, which leaves the machine hanging: