bpf: Prevent nesting overflow in bpf_try_get_buffers

bpf_try_get_buffers() returns one of multiple per-CPU buffers based on a
per-CPU nesting counter. This mechanism expects that buffers are not
endlessly acquired before being returned. migrate_disable() ensures that a
task remains on the same CPU, but it does not prevent the task from being
preempted by another task on that CPU.

Without disabled preemption, a task may be preempted while holding a
buffer, allowing another task to run on same CPU and acquire an
additional buffer. Several such preemptions can cause the per-CPU
nest counter to exceed MAX_BPRINTF_NEST_LEVEL and trigger the warning in
bpf_try_get_buffers(). Adding preempt_disable()/preempt_enable() around
buffer acquisition and release prevents this task preemption and
preserves the intended bounded nesting behavior.

Reported-by: [e-mail]
Closes: https://lore.kernel.org/all/[e-mail]/
Fixes: 4223bf833c849 ("bpf: Remove preempt_disable in bpf_try_get_buffers")
Suggested-by: Yonghong Song <[e-mail]>
Reviewed-by: Sebastian Andrzej Siewior <[e-mail]>
Signed-off-by: Sahil Chandna <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Alexei Starovoitov <[e-mail]>

kernel/bpf/helpers.c

···
 {
 	int nest_level;
 
+	preempt_disable();
 	nest_level = this_cpu_inc_return(bpf_bprintf_nest_level);
 	if (WARN_ON_ONCE(nest_level > MAX_BPRINTF_NEST_LEVEL)) {
 		this_cpu_dec(bpf_bprintf_nest_level);
+		preempt_enable();
 		return -EBUSY;
 	}
 	*bufs = this_cpu_ptr(&bpf_bprintf_bufs[nest_level - 1]);
···
 	if (WARN_ON_ONCE(this_cpu_read(bpf_bprintf_nest_level) == 0))
 		return;
 	this_cpu_dec(bpf_bprintf_nest_level);
+	preempt_enable();
 }
 
 void bpf_bprintf_cleanup(struct bpf_bprintf_data *data)