Re: [PATCH v3 1/1] powerpc: Enable dynamic preemption

On 1/30/25 20:24, Sebastian Andrzej Siewior wrote:

On 2025-01-06 10:49:19 [+0530], Shrikanth Hegde wrote:

quoted

--- a/arch/powerpc/kernel/interrupt.c

Thanks for taking a look.

quoted

+
  #ifdef CONFIG_PPC_BOOK3S_64
  DEFINE_STATIC_KEY_FALSE(interrupt_exit_not_reentrant);
  static inline bool exit_must_hard_disable(void)

@@ -396,7 +400,7 @@ notrace unsigned long interrupt_exit_kernel_prepare(struct pt_regs *regs)
  		/* Returning to a kernel context with local irqs enabled. */
  		WARN_ON_ONCE(!(regs->msr & MSR_EE));
  again:
-		if (IS_ENABLED(CONFIG_PREEMPTION)) {
+		if (preempt_model_preemptible()) {

CONFIG_HAVE_PREEMPT_DYNAMIC_KEY is the only option, right? Wouldn't

| #DEFINE need_irq_preemption() \
|          (static_branch_unlikely(&sk_dynamic_irqentry_exit_cond_resched))
|
| 	if (need_irq_preemption()) {

be a bit smaller/ quicker? This could be a fast path ;)

I am okay with either way. I did try both[1], there wasn't any significant difference,
hence chose a simpler one. May be system size, workload pattern might matter.

Let me do some more testing to see which one wins.
Is there any specific benchmark which might help here?

[1]: https://lore.kernel.org/all/b98b7795-070a-4d9c-9599-445c2ff55fd7@linux.ibm.com/ (local)

quoted

  			/* Return to preemptible kernel context */
  			if (unlikely(read_thread_flags() & _TIF_NEED_RESCHED)) {
  				if (preempt_count() == 0)

diff --git a/arch/powerpc/kernel/traps.c b/arch/powerpc/kernel/traps.c
index edf5cabe5dfd..2556fa8ec019 100644
--- a/arch/powerpc/kernel/traps.c
+++ b/arch/powerpc/kernel/traps.c

@@ -266,7 +266,11 @@ static int __die(const char *str, struct pt_regs *regs, long err)
  	printk("%s PAGE_SIZE=%luK%s%s%s%s%s%s %s\n",
  	       IS_ENABLED(CONFIG_CPU_LITTLE_ENDIAN) ? "LE" : "BE",
  	       PAGE_SIZE / 1024, get_mmu_str(),
-	       IS_ENABLED(CONFIG_PREEMPT) ? " PREEMPT" : "",
+	       preempt_model_none()      ? "none" :
+	       preempt_model_voluntary() ? "voluntary" :
+	       preempt_model_full()      ? "full" :
+	       preempt_model_lazy()      ? "lazy" :
+	       "",

So intend to rework this part. I have patches stashed at
	https://git.kernel.org/pub/scm/linux/kernel/git/bigeasy/staging.git/log/?h=preemption_string

which I didn't sent yet due to the merge window. Just a heads up ;)

Makes sense. I had seen at-least two places where this code was there, ftrace/powerpc.
There were way more places..

You want me to remove this part?

quoted

  	       IS_ENABLED(CONFIG_SMP) ? " SMP" : "",
  	       IS_ENABLED(CONFIG_SMP) ? (" NR_CPUS=" __stringify(NR_CPUS)) : "",
  	       debug_pagealloc_enabled() ? " DEBUG_PAGEALLOC" : "",

Sebastian