On Thu, May 09, 2024 at 05:13:12PM +0200, Oleg Nesterov wrote: > On 05/07, Paul E. McKenney wrote: > > > > On Tue, May 07, 2024 at 10:54:41AM -0400, Oleg Nesterov wrote: > > > Hello, > > > > > > I feel I don't really like this patch but I am travelling without my working > > > laptop, can't read the source code ;) Quite possibly I am wrong, I'll return > > > to this when I get back on May 10. > > > > By the stricter data-race rules used in RCU code [1], this is a bug that > > needs to be fixed. > > Now that I can read the code... Sorry, still can't understand. > > > which is read locklessly, > > Where??? > > OK, OK, we have > > // rcu_sync_exit() > WARN_ON_ONCE(READ_ONCE(rsp->gp_count) == 0) > > and > > // rcu_sync_dtor() > WARN_ON_ONCE(READ_ONCE(rsp->gp_count)); > > other than that ->gp_count is always accessed under ->rss_lock. > > And yes, at least WARN_ON_ONCE() in rcu_sync_exit() can obviously race with > rcu_sync_enter/exit which update gp_count. I think this is fine correctness-wise. > > But OK, we need to please KCSAN (or is there another problem I missed ???) > > We can move these WARN_ON()'s into the ->rss_lock protected section. > > Or perhaps we can use data_race(rsp->gp_count) ? To be honest I thought > that READ_ONCE() should be enough... > > Or we can simply kill these WARN_ON_ONCE()'s. Or we could move those WARN_ON_ONCE() under the lock. If this would be a lock-contention issue, we could condition them with something like IS_ENABLED(CONFIG_PROVE_RCU). Then all accesses to those variables would always be protected by the lock, and the WRITE_ONCE() and READ_ONCE() calls could be dropped. (Or am I missing another lockless access?) Which would have the further advantage that if anyone accessed these without holding the lock, KCSAN would complain. > I don't understand why should we add more WRITE_ONCE()'s into the critical > section protected by ->rss_lock. There are indeed several ways to fix this. Which would you prefer? > Help! ;) ;-) ;-) ;-) Thanx, Paul > Oleg. > > > which in turn results in a data race. The fix is to mark > > the updates (as below) with WRITE_ONCE(). > > > > Or is there something in one or the other of these updates to ->gp_count > > that excludes lockless readers? (I am not seeing it, but you know this > > code way better than I do!) > > > > Thanx, Paul > > > > [1] https://docs.google.com/document/d/1FwZaXSg3A55ivVoWffA9iMuhJ3_Gmj_E494dLYjjyLQ/edit?usp=sharing > > > > > Oleg. > > > > > > On 05/07, Uladzislau Rezki (Sony) wrote: > > > > > > > > From: "Paul E. McKenney" <paulmck@xxxxxxxxxx> > > > > > > > > The rcu_sync structure's ->gp_count field is updated under the protection > > > > of ->rss_lock, but read locklessly, and KCSAN noted the data race. > > > > This commit therefore uses WRITE_ONCE() to do this update to clearly > > > > document its racy nature. > > > > > > > > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx> > > > > Cc: Oleg Nesterov <oleg@xxxxxxxxxx> > > > > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx> > > > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx> > > > > --- > > > > kernel/rcu/sync.c | 8 ++++++-- > > > > 1 file changed, 6 insertions(+), 2 deletions(-) > > > > > > > > diff --git a/kernel/rcu/sync.c b/kernel/rcu/sync.c > > > > index 86df878a2fee..6c2bd9001adc 100644 > > > > --- a/kernel/rcu/sync.c > > > > +++ b/kernel/rcu/sync.c > > > > @@ -122,7 +122,7 @@ void rcu_sync_enter(struct rcu_sync *rsp) > > > > * we are called at early boot time but this shouldn't happen. > > > > */ > > > > } > > > > - rsp->gp_count++; > > > > + WRITE_ONCE(rsp->gp_count, rsp->gp_count + 1); > > > > spin_unlock_irq(&rsp->rss_lock); > > > > > > > > if (gp_state == GP_IDLE) { > > > > @@ -151,11 +151,15 @@ void rcu_sync_enter(struct rcu_sync *rsp) > > > > */ > > > > void rcu_sync_exit(struct rcu_sync *rsp) > > > > { > > > > + int gpc; > > > > + > > > > WARN_ON_ONCE(READ_ONCE(rsp->gp_state) == GP_IDLE); > > > > WARN_ON_ONCE(READ_ONCE(rsp->gp_count) == 0); > > > > > > > > spin_lock_irq(&rsp->rss_lock); > > > > - if (!--rsp->gp_count) { > > > > + gpc = rsp->gp_count - 1; > > > > + WRITE_ONCE(rsp->gp_count, gpc); > > > > + if (!gpc) { > > > > if (rsp->gp_state == GP_PASSED) { > > > > WRITE_ONCE(rsp->gp_state, GP_EXIT); > > > > rcu_sync_call(rsp); > > > > -- > > > > 2.39.2 > > > > > > > > > >
