Hi Alexei,
Thank you for the feedback.
On 1/24/26 1:37 AM, Alexei Starovoitov wrote:
On Fri, Jan 23, 2026 at 12:43 AM Changwoo Min <[email protected]> wrote:
Introduce bpf_in_nmi(), bpf_in_hardirq(), bpf_in_serving_softirq(), and
bpf_in_task() kfuncs to allow BPF programs to query the current execution
context.
While BPF programs can sometimes infer context based on the attach point,
certain programs (such as those in sched_ext) may be called from multiple
contexts. These kfuncs provide a reliable way for logic to branch based on
whether the CPU is currently handling an interrupt or executing in task
context.
For example, this is particularly useful for sched_ext schedulers that need
to differentiate between task-to-task wake-ups and interrupt-to-task
wake-ups.
As the names imply, these helpers wrap the kernel's internal in_nmi(),
in_hardirq(), in_serving_softirq(), and in_task() macros.
Signed-off-by: Changwoo Min <[email protected]>
---
kernel/bpf/helpers.c | 44 ++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 44 insertions(+)
diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
index 637677815365..cb36bc7a80c6 100644
--- a/kernel/bpf/helpers.c
+++ b/kernel/bpf/helpers.c
@@ -4365,6 +4365,46 @@ __bpf_kfunc int bpf_dynptr_file_discard(struct
bpf_dynptr *dynptr)
return 0;
}
+/**
+ * bpf_in_nmi - Test if the current execution context is in NMI context.
+ *
+ * Return: true if we are in NMI context, false otherwise.
+ */
+__bpf_kfunc bool bpf_in_nmi(void)
+{
+ return in_nmi();
+}
+
+/**
+ * bpf_in_hardirq - Test if the current execution context is in hard IRQ
context.
+ *
+ * Return: true if we are in hard IRQ context, false otherwise.
+ */
+__bpf_kfunc bool bpf_in_hardirq(void)
+{
+ return in_hardirq();
+}
+
+/**
+ * bpf_in_serving_softirq - Test if the current execution context is in
softirq context.
+ *
+ * Return: true if we are in softirq context, false otherwise.
+ */
+__bpf_kfunc bool bpf_in_serving_softirq(void)
+{
+ return in_serving_softirq();
+}
+
+/**
+ * bpf_in_task - Test if the current execution context is in task context.
+ *
+ * Return: true if we are in task context, false otherwise.
+ */
+__bpf_kfunc bool bpf_in_task(void)
+{
+ return in_task();
+}
+
This functionality is already available.
See bpf_in_interrupt() and get_preempt_count() in bpf_experimental.h.
No need to replicate that as kfuncs.
I see. While bpf_in_interrupt() provides a general check, certain use
cases in sched_ext require more granular differentiation -- specifically
between hardirq and softirq contexts to handle different wake-up paths
correctly.
If something can be implemented as a bpf program it should stay as
pure bpf code, since it's more flexible and faster this way.
llvm together with JITs completely inline get_preempt_count().
I agree. Implementing these checks as pure BPF code is more efficient
due to JIT inlining. I was initially hesitant to duplicate kernel-side
macros in BPF, but since the precedent is already established in
bpf_experimental.h, I will follow that pattern.
I will drop the kfunc implementations in Patch 1 and instead add
bpf_in_nmi(), bpf_in_hardirq(), bpf_in_serving_softirq(), and
bpf_in_task() to bpf_experimental.h by wrapping get_preempt_count().
I will also update the selftests in Patch 2 to use these new headers and
submit a v2 shortly.
Regards,
Changwoo Min
