On Sun, Oct 14, 2018 at 07:33:28PM -0700, Paul E. McKenney wrote: > On Sun, Oct 14, 2018 at 07:13:49PM -0700, Joel Fernandes wrote: > > On Sun, Oct 14, 2018 at 07:08:27PM -0700, Joel Fernandes wrote: > > > On Sun, Oct 14, 2018 at 04:17:31PM -0700, Paul E. McKenney wrote: > > > > On Sun, Oct 14, 2018 at 02:29:55PM -0700, Joel Fernandes (Google) wrote: > > > > > The Requirements.html document says "Disabling Preemption Does Not > > > > > Block > > > > > Grace Periods". However this is no longer true with the RCU > > > > > consolidation. Lets remove the obsolete (non-)requirement entirely. > > > > > > > > > > Signed-off-by: Joel Fernandes (Google) <j...@joelfernandes.org> > > > > > > > > Good catch, queued, thank you! > > > > > > Thanks! By the way after I sent the patch, I also tried Oleg's experiment > > > to > > > confirm that this is indeed obsolete. :) > > > > > > One thing interesting came up when I tried synchronize_rcu_expedited() > > > instead of synchronize_rcu() in Oleg's experiment, I still saw a multiple > > > millisecond delay between when the rcu read section completely and the > > > synchronize_rcu_expedited returns: > > > > > > For example, with synchronize_rcu_expedited, the 'SPIN done' and the 'SYNC > > > done' are about 3 millisecond apart: > > > [ 77.599142] SPIN start > > > [ 77.601595] SYNC start > > > [ 82.604950] SPIN done! > > > [ 82.607836] SYNC done! > > > I saw anywhere from 2-6 milliseconds. > > > > > > The reason I bring this up is according to Requirements.html: In some > > > cases, > > > the multi-millisecond synchronize_rcu() latencies are unacceptable. In > > > these > > > cases, synchronize_rcu_expedited() may be used instead,.. so either I > > > messed > > > something up in the experiment, or I need to update this part of the > > > document ;-) > > In normal testing, 2-6 milliseconds is indeed excessive. Could you please > point me at Oleg's experiment? Also, what CONFIG_PREEMPT setting were > you using? (My guess is CONFIG_PREEMPT=y.)
The CONFIG_PREEMPT config I am using is CONFIG_PREEMPT=y. > > So I realized I'm running in Qemu so it could also be a scheduling delay of > > the vcpu thread. So apologies about the noise if the experiment works fine > > for you. > > I used rcuperf, which might not be doing the same thing as Oleg's > experiment. The experiment is mentioned at: https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg912055.html If you apply the below diff, it applies cleanly on rcu/dev. And then run: taskset 2 perl -e 'syscall 157, 666, 5000' & taskset 1 perl -e 'syscall 157, 777' diff --git a/kernel/sys.c b/kernel/sys.c index cf5c67533ff1..b654b7566ca3 100644 --- a/kernel/sys.c +++ b/kernel/sys.c @@ -2261,6 +2261,9 @@ int __weak arch_prctl_spec_ctrl_set(struct task_struct *t, unsigned long which, return -EINVAL; } +#include <linux/delay.h> + + SYSCALL_DEFINE5(prctl, int, option, unsigned long, arg2, unsigned long, arg3, unsigned long, arg4, unsigned long, arg5) { @@ -2274,6 +2277,19 @@ SYSCALL_DEFINE5(prctl, int, option, unsigned long, arg2, unsigned long, arg3, error = 0; switch (option) { + case 666: + preempt_disable(); + pr_crit("SPIN start\n"); + while (arg2--) + mdelay(1); + pr_crit("SPIN done!\n"); + preempt_enable(); + break; + case 777: + pr_crit("SYNC start\n"); + synchronize_rcu(); + pr_crit("SYNC done!\n"); + break; case PR_SET_PDEATHSIG: if (!valid_signal(arg2)) { error = -EINVAL;
