SMP

This section explains how SMP (Symetric Multi Processing) is implemented within LSE/OS.

Some SMP Specific terms
Here are some terms needed for understanding this document:

  • AP: Application Processor - The slaves processors.
  • APIC: Advanced Programmable Interrupt Controller
  • APICID ou APID: Identifier of the SMP device (IOAPIC ou BSP ou AP) on the APIC bus.
  • BIPI: Boot IPI - The mainboard send this request to find the BSP.
  • BIST: Built-in Self Test.
  • BSP: Bootstrap Processor - The "master" processor that boots first.
  • FIPI: Final boot IPI - The BSP executes the BIOS
  • IPI: Inter Processor interrupt
  • MP: Multiprocessor
  • SIPI: Start IPI - The BSP sends this request to initialize APs.
  • SMP: Symetric Multiprocessing

    SMP Bootstrap
    Just like the BSP, the AP's need to be boostraped. This section explains how:

    The SMP program starts:

    1

    The SMP program installs IPI interrupts and SMP services for each processor. It also make AP's bootstraping to protected mode (see the smplo.S assembly file) and make them executing the smpkern task (they are in fact context switchers):

    1

    Once they are booted, AP's send a smp_hello() request to their corresponding services. They call then runq() for determining the next task to execute:

    1

    Services and interrupts in SMP
    There is one IDT (Interrupt Descriptor Table) per processor: this implies that drivers must be explicitely programmed to support the SMP mode.

    E.g. the timer service has an entry point per cpu:

    The timer service

    But it also has a tick interrupt per cpu:

    The timer service

    Scheduling policy then could be done according to different ways:

  • The APIC could be programmed to send tick interrupt to a dedicated processor that would interrupt other processors with an IPI interrupt.
  • The APIC could be programmed to send tick interrupts to all the processors: in such case synchronization could be maybe more difficult.

    • For now, SMP was only tested with simple tasks (e.g. cnt) that use only coresrv services.

    Spinlocking in SMP
    Altough in monoprocessor mode, servicing guarantees non-interruptibility, in SMP mode, two processors could execute the same code twice. For avoiding this we need to spinlock critical sections. There are two ways for doing this:

    Spinlock are disseminated into the code to protect critical sections:

    Fine grained locking

    Spinlock are done at the door of the service, as in coresrv, for simplifying the code:

    Majored locking