Process vs.Atomic Context (Slideshow)

Process Context

Process context: everything that can be identified by a process ID

• Processes (and threads) that execute in user mode ⟶ process address space

• Processes (and threads) that execute in kernel mode ⟶ kernel address space

• Kernel threads ⟶ kernel address space

Preemption …

• Process context is subject to scheduling

• Fair scheduling: preemption at end of time slice

• Realtime: preemption when higher priority process/thread is runnable

Race Conditions

When do race conditions occur?

• Two processes/threads share the same address space

• Manipulate the same data structure

In kernel address space?

• Userspace processes executing a system call (“switch to kernel mode”)

• Kernel threads

Protection through locking

• Mutexes: locker has to wait until unlocked

• Spinlocks: locker loops until unlocked (on different CPU obviously)

  • Atomic context

Atomic Context

Atomic context is where code must not sleep!

• Interrupt service routine

  • Interrupts disabled

  • No preemption, no scheduling, no nothing

  • ⟶ primary source of latency

• Bottom half - code that runs in interrupt context (not subject to scheduling), but interrupts are already enabled

  • Deferred work ⟶ “tasklet”, “soft-IRQ”

  • Best avoided because not easily controllable, realtime-wise

• All code that holds a spinlock

Atomic vs. Process Context

Atomic context must not sleep

• Preemption disabled ⟶ prioritization impossible

• High latency if atomic code runs for too long

• Severe restrictions

  • Paging

  • Locking is difficult

  • …

Process context …

• Subject to scheduling ⟶ easily prioritized (be it realtime or not)

• Easy locking

Conclusion

• Atomic context best avoided

• … at least when absolute control is desired

“Sleep While Atomic” Debugging

⟶ CONFIG_DEBUG_ATOMIC_SLEEP