On Mon, 29 Sep 2025 12:12:59 +0200 Peter Zijlstra <[email protected]> wrote:
> On Mon, Sep 29, 2025 at 11:38:08AM +0206, John Ogness wrote: > > > >> Problem: > > >> 1. CPU0 executes (1) assigning tp_event->perf_events = list > > > > smp_wmb() > > > > >> 2. CPU0 executes (2) enabling kprobe functionality via class->reg() > > >> 3. CPU1 triggers and reaches kprobe_dispatcher > > >> 4. CPU1 checks TP_FLAG_PROFILE - condition passes (step 2 completed) > > > > smp_rmb() > > > > >> 5. CPU1 calls kprobe_perf_func() and crashes at (3) because > > >> call->perf_events is still NULL > > >> > > >> The issue: Assignment in step 1 may not be visible to CPU1 due to > > >> missing memory barriers before step 2 sets TP_FLAG_PROFILE flag. > > > > A better explanation of the issue would be: CPU1 sees that kprobe > > functionality is enabled but does not see that perf_events has been > > assigned. > > > > Add pairing read and write memory barriers to guarantee that if CPU1 > > sees that kprobe functionality is enabled, it must also see that > > perf_events has been assigned. > > > > Note that this could also be done more efficiently using a store_release > > when setting the flag (in step 2) and a load_acquire when loading the > > flag (in step 4). > > The RELEASE+ACQUIRE is a better pattern for these cases. > > And I'll argue the barrier should be in 2 not 1, since it is 2 that sets > the flag checked in 4. Any store before that flag might be affected, > not just the ->perf_events list. RELEASE+ACQUIRE ensures the memory ordering on the `same` CPU, so do we still need smp_rmb() and smp_wmb()? e.g. perf_trace_event_init() ----- tp_event->perf_events = list /*(1)*/ smp_store_release(&tp->event->flags, newflag); ----- kprobe_dispatcher() ----- smp_load_acquire(&tk->tp.flags); /*(2)*/ this_cpu_ptr(call->perf_events); ----- This ensures - the flags update is shown on other CPUs - perf_events update is done before the flags update on the same CPU - perf_events load is done after the flags load on the same CPU But we are still not sure the perf_events update is shown on other CPUs? To ensure that we still need smp_wmb() and smp_rmb() at (1) and (2). Or we don't need smp_*mb()? Thank you, -- Masami Hiramatsu (Google) <[email protected]>
