Re: [PATCH RFC v3 06/13] bpf: Take a Tasks Trace reader in the trampoline glue

[Josef Bacik <josef@toxicpanda.com>]

On Wed, 16 Sep 2026 03:45:16 +0000, Alexei Starovoitov wrote:
> On Tue Sep 15, 2026 at 1:17 PM UTC, Josef Bacik wrote:
> >  	__acquires(RCU)
> >  {
> > +	bpf_tramp_read_lock_trace();
> >  	rcu_read_lock_dont_migrate();
>
> This is double increment. rcu_read_lock_dont_migrate() includes
> rcu_read_lock_trace().

Unless I'm looking at the wrong tree it doesn't, on Linus' master and on
bpf-next it is

	static __always_inline void rcu_read_lock_dont_migrate(void)
	{
		if (IS_ENABLED(CONFIG_PREEMPT_RCU))
			migrate_disable();
		rcu_read_lock();
	}

so plain RCU plus migrate_disable(), no Tasks Trace reader. That is why
the non-sleepable glue needs one added here: on these architectures the
trampoline image the glue returns into is only kept alive by Tasks RCU
while the task is a rcu_read_lock_trace() reader, and rcu_read_lock()
does not give us that.

It is two counters for a non-sleepable prog on x86-64/arm64 though,
rcu_read_lock()'s and trc_reader_nesting plus the SRCU-fast percpu one,
if that is what you meant. I don't see a way around it short of not
using Tasks Trace as the trampoline reader: the prog still needs plain
RCU for everything it dereferences, and the image needs something that
survives preemption. It is compiled out on every other configuration and
nothing changes in the JITed image. If you would rather the reader be
taken once around the whole image in the JIT instead of per prog in the
glue (which would also let the fentry-only teardown stay a single grace
period), I can do that for x86 and arm64, it is what v2 did with the
private counter.

Separately, Junseo's "bpf: keep trampoline progs alive until image
release" also adds bpf_tramp_image::nr_progs; if that lands first I will
just use it here.

Thanks,

Josef