A real-time (RT) system provides deterministic, bounded response times to
external events. A system qualifies as real-time if it is free of unbounded
latency, its worst-case response time can be calculated with precision, and all
tasks meet their scheduling deadlines within defined jitter limits.
Qualcomm® Linux ships an LTS RT kernel (6.18.x) built on top of the upstream
PREEMPT_RT patch set. The RT kernel is targeted at kernel-space processes
that require deterministic scheduling. User-space applications cannot achieve
RT behavior by using the RT kernel alone; they must also be written with RT
scheduling policies (SCHED_FIFO / SCHED_RR) and pinned to isolated CPU
cores.
How PREEMPT_RT changes the kernel
The standard Linux kernel (CONFIG_PREEMPT or CONFIG_PREEMPT_VOLUNTARY)
allows certain critical sections to run non-preemptibly, causing unbounded
scheduling latency spikes. PREEMPT_RT eliminates this by:
| Change | Effect |
|---|
| Converting spinlocks to sleeping mutexes (RT mutexes) | Critical sections become preemptible, reducing worst-case latency |
| Making interrupt handlers threaded by default | ISRs run as kernel threads with a configurable priority, removing them from the non-preemptible interrupt context |
| Serialising RCU (Read-Copy-Update) callbacks | RCU callbacks run in dedicated threads rather than softirq context, reducing jitter |
Converting local_lock_t to per-CPU sleeping locks | Per-CPU code paths become preemptible |
Required Kconfig options
The Qualcomm Linux meta-qcom layer enables CONFIG_PREEMPT_RT by default
through the rt.config configuration fragment included in the
linux-qcom-rt_6.18.bb recipe. The following table describes the key Kconfig
options and their role:
Table: RT kernel Kconfig options
| Option | Default in RT build | Description |
|---|
CONFIG_PREEMPT_RT | y | Enables full kernel preemption. Mandatory for RT. |
CONFIG_NO_HZ_COMMON | y | Infrastructure for tickless idle and full-tickless operation. Required by NO_HZ_FULL. |
CONFIG_NO_HZ_FULL | y | Suppresses scheduling-clock interrupts on CPUs with a single runnable task. Reduces jitter on isolated RT cores. |
CONFIG_CPUSETS | y | Allows grouping CPUs into sets for affinity control. Required for isolating RT tasks to dedicated cores. |
CONFIG_CPU_ISOLATION | y | Enables isolcpus= kernel parameter support to remove cores from the general scheduler. |
CONFIG_HZ_1000 | y | Sets the kernel timer frequency to 1000 Hz (1 ms tick period). Higher HZ values increase scheduling granularity at the cost of more timer overhead. |
zcat /proc/config.gz | grep -E "CONFIG_PREEMPT|CONFIG_NO_HZ|CONFIG_CPUSETS"
# CONFIG_PREEMPT_NONE is not set
# CONFIG_PREEMPT_VOLUNTARY is not set
# CONFIG_PREEMPT is not set
CONFIG_PREEMPT_RT=y
CONFIG_NO_HZ_COMMON=y
CONFIG_NO_HZ_FULL=y
CONFIG_CPUSETS=y
RT vs standard kernel trade-offs
Choosing the RT kernel involves accepting specific trade-offs:
| Dimension | Standard kernel | RT kernel |
|---|
| Worst-case scheduling latency | Unbounded (μs to ms) | Bounded (low μs on tuned hardware) |
| Throughput | Higher, fewer context switches | Slightly lower, more preemption points |
| Power consumption | Lower in idle | Higher, NO_HZ_FULL keeps RT cores active |
| Debugging complexity | Standard tools | RT-specific tools (cyclictest, ftrace with hwlat_detector) |
| Use case | General-purpose, multimedia | Industrial control, robotics, time-sensitive networking |
The RT kernel is supported on all Qualcomm Linux hardware platforms
(QCS6490, IQ-9075, IQ-8275, IQ-615). Latency numbers depend on the
specific SoC, CPU topology, and tuning applied. Always measure on target
hardware and do not rely solely on simulation results.
Verify RT is active
After booting the RT kernel, confirm full RT preemption is enabled:
# Kernel version string includes -rt suffix
uname -r
# Example: 6.18.0-rt1+
# Confirm PREEMPT_RT in the version banner
cat /proc/version
# Verify the Kconfig flag
zcat /proc/config.gz | grep CONFIG_PREEMPT_RT
# Expected: CONFIG_PREEMPT_RT=y
