Kernel execution contexts

The Linux kernel runs code in four execution contexts. It tells them apart with the bitfields packed into a single per-task word, current->preempt_count, read through the in_*() macros.

#	Context	Macro	May sleep?	Typical sources
1	Process (task)	in_task()	yes	syscall path, kthread, workqueue handler, module init/exit
2	Softirq (bottom half)	in_softirq()	no	network RX/TX, tasklets, timer softirq, RCU callbacks
3	Hardirq (top half)	in_irq()	no	device IRQ handler, timer tick, IPI callback
4	NMI	in_nmi()	no	pseudo-NMI (arm64), perf, hard-lockup watchdog

Contexts 2–4 are collectively interrupt context (in_interrupt()), and none of them may sleep — no msleep, no mutex_lock, no smp_call_function_single. Only process context can block.

The nuance: process context has two sub-states

The question that actually matters — "can I sleep here?" — has a fifth answer that isn't one of the four contexts:

• preemptible (normal process context): can sleep, take mutexes, call the blocking primitives.
• atomic (in_atomic() true, but in_interrupt() false): still a task, but preemption is disabled or a spin_lock is held — must not sleep, even though you never left process context.

This is why locking_lab sleep_in_atomic=1 splats: the code runs in a kthread (process context) but under a spinlock, so preempt_count is nonzero → might_sleep() / CONFIG_DEBUG_ATOMIC_SLEEP catches the msleep() with "BUG: sleeping function called from invalid context".

So sleeping depends on preempt_count as a whole, not just on which of the four contexts you are in.

The meter itself: preempt_count

One word, split into bitfields (offsets from include/linux/preempt.h):

 bits  0– 7   preempt-disable count   (spin_lock, preempt_disable)
 bits  8–15   softirq count           SOFTIRQ_OFFSET  = 0x0000_0100
 bits 16–19   hardirq count           HARDIRQ_OFFSET  = 0x0001_0000
 bit  20      NMI                      NMI_OFFSET      = 0x0010_0000

in_atomic() is simply "any of these nonzero". in_irq=65536 you saw in the ipi lab is 0x10000 — the hardirq bit set — versus in_irq=0 when the callback happened to run directly in process context (sender == target).

Seen live — context_lab

context_lab runs one reporting function from each reachable context and prints the meter, so the bitfields are visible per context:

context_lab: process   pc=00000000  in_task=1 in_softirq=0 in_irq=0 in_nmi=0  in_atomic=0  cpu=6
context_lab: proc+spin pc=00000001  in_task=1 in_softirq=0 in_irq=0 in_nmi=0  in_atomic=1  cpu=6
context_lab: softirq   pc=00000101  in_task=0 in_softirq=1 in_irq=0 in_nmi=0  in_atomic=1  cpu=6
context_lab: hardirq   pc=00010001  in_task=0 in_softirq=0 in_irq=1 in_nmi=0  in_atomic=1  cpu=6

Read pc= by nibble:

• 00000000 — plain process context, fully preemptible.
• 00000001 — process context + spinlock: low byte = 1 (one level of preempt-disable). in_task still 1, but in_atomic flips to 1 — the "can't sleep here" trap.
• 00000101 — tasklet: +SOFTIRQ_OFFSET.
• 00010001 — hrtimer callback: +HARDIRQ_OFFSET (the low 01 is the hardirq path's own preempt-disable).

Why no NMI row

NMI is the fourth context but is not reachable from a module on this 4.19 arm64 kernel: arm64 only has NMIs with pseudo-NMI support, and the API to run code there (request_nmi()) arrived in v5.1. report_context() would print in_nmi=1 if it ever ran there.

Rule of thumb

Sleepable primitives (msleep, mutex_lock, smp_call_function_single, get_online_cpus) require process context with preemption on. In any atomic or interrupt context, reach for the non-sleeping variants instead (udelay, spin_lock, smp_call_function_single_async, irq_work).

Related labs

• context_lab — this one; reports the contexts.
• locking_lab (sleep_in_atomic=1) — sleeping under a spinlock; the DEBUG_ATOMIC_SLEEP splat.
• ipi_lab — in_irq=65536 (hardirq) vs in_irq=0 (self-call in process context).
• hrtimer_lab — timer callback in hardirq context.
• workqueue_lab — deferred work back in process context (a kworker).