Review of RCU-related patches in -rt
Hello, Sebastian,
Finally getting around to taking another look:
c7e07056a108 EXP rcu: skip the workqueue path on RT
This one makes sense given the later commit setting the rcu_normal_after_boot kernel parameter. Otherwise, it is slowing down expedited grace periods for no reason. But should the check also include rcu_normal_after_boot and rcu_normal? For example:
if ((IS_ENABLED(CONFIG_PREEMPT_RT_FULL) &&
(rcu_normal || rcu_normal_after_boot) ||
!READ_ONCE(rcu_par_gp_wq) || rcu_scheduler_active != RCU_SCHEDULER_RUNNING || rcu_is_last_leaf_node(rnp)) {
Alternatively, one approach would be to take the kernel parameters out in -rt:
static int rcu_normal_after_boot = IS_ENABLED(CONFIG_PREEMPT_RT_FULL); #ifndef CONFIG_PREEMPT_RT_FULL module_param(rcu_normal_after_boot, int, 0);
#endif
And similar for rcu_normal and rcu_expedited.
Or is there some reason to allow run-time expedited grace periods in CONFIG_PREEMPT_RT_FULL=y kernels?
d1f52391bd8a rcu: Disable RCU_FAST_NO_HZ on RT
Looks good. More complexity could be added if too many people get themselves in trouble via "select RCU_FAST_NO_HZ".
42b346870326 rcu: make RCU_BOOST default on RT
To avoid complaints about this showing up when people don't expected, could you please instead "select RCU_BOOST" in the Kconfig definition of PREEMPT_RT_FULL?
Or do people really want to be able to disable boosting?
457c1b0d9c0e sched: Do not account rcu_preempt_depth on RT in might_sleep()
The idea behind this one is to avoid false-positive complaints about -rt's sleeping spinlocks, correct?
7ee13e640b01 rbtree: don't include the rcu header c9b0c9b87081 rtmutex: annotate sleeping lock context
No specific comments.
7912d002ebf9 rcu: Eliminate softirq processing from rcutree
This hasn't caused any problems in -rcu from what I can see. I am therefore planning to submit the -rcu variant of this to mainline during the next merge window.
f06d34ebdbbb srcu: Remove srcu_queue_delayed_work_on()
Looks plausible. I will check more carefully for mainline.
aeb04e894cc9 srcu: replace local_irqsave() with a locallock e48989b033ad irqwork: push most work into softirq context
These look to still be -rt only.
Thanx, Paul
Quoted 1 line · Paul E. McKenney
Hello, Sebastian,
Hi Paul,
Quoted 4 lines
Finally getting around to taking another look:
c7e07056a108EXP rcu: skip the workqueue path on RT
Quoted 11 lines
This one makes sense given the later commit setting the
rcu_normal_after_boot kernel parameter. Otherwise, it is
slowing down expedited grace periods for no reason. But
should the check also include rcu_normal_after_boot and
rcu_normal? For example:
if ((IS_ENABLED(CONFIG_PREEMPT_RT_FULL) &&
(rcu_normal || rcu_normal_after_boot) ||
!READ_ONCE(rcu_par_gp_wq) ||
rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
rcu_is_last_leaf_node(rnp)) {I recently dropped that patch from the queue because the workqueue problem vanished.
Quoted 9 lines
Alternatively, one approach would be to take the kernel parameters out in -rt: static int rcu_normal_after_boot = IS_ENABLED(CONFIG_PREEMPT_RT_FULL); #ifndef CONFIG_PREEMPT_RT_FULL module_param(rcu_normal_after_boot, int, 0); #endif And similar for rcu_normal and rcu_expedited.
This makes sense.
Quoted 2 lines
Or is there some reason to allow run-time expedited grace periods in
CONFIG_PREEMPT_RT_FULL=ykernels?
No, I doubt there is any need to use the `expedited' version. The problem is that it increases latencies.
Quoted 4 lines
d1f52391bd8arcu: DisableRCU_FAST_NO_HZon RTLooks good. More complexity could be added if too many people get themselves in trouble via "select
RCU_FAST_NO_HZ".
That patch disables RCU_FAST_NO_HZ and claims that it has something to do with a timer_list timer and IRQ-off section. We couldn't schedule timers from IRQ-off regions but not anymore. Only del_timer_sync() can't be invoked from IRQ-off regions.
I just booted a box with this enabled together with NO_HZ/ NO_HZ_FULL and I not complains yet. So I might drop that…
Quoted 7 lines
42b346870326rcu: makeRCU_BOOSTdefault on RTTo avoid complaints about this showing up when people don't expected, could you please instead "select
RCU_BOOST" in theKconfigdefinition ofPREEMPT_RT_FULL?Or do people really want to be able to disable boosting?
I have no idea. I guess most people don't know what it does and stay with the default. It become default on RT once a few people complained that they run OOM during boot on some "memory contrained systems". That option avoided it.
So yes, will make it depend on RT.
Quoted 4 lines
457c1b0d9c0esched: Do not accountrcu_preempt_depthon RT inmight_sleep()The idea behind this one is to avoid false-positive complaints about -rt's sleeping spinlocks, correct?
Correct. Maybe we could invoke a different schedule() primitiv so RCU is aware that this is a sleeping spinlock and not a regular sleeping lock.
Quoted 10 lines
7ee13e640b01rbtree: don't include the rcu headerc9b0c9b87081rtmutex: annotate sleeping lock contextNo specific comments.
7912d002ebf9rcu: Eliminate softirq processing from rcutreeThis hasn't caused any problems in -rcu from what I can see. I am therefore planning to submit the -rcu variant of this to mainline during the next merge window.
wonderful.
Quoted 3 lines
f06d34ebdbbbsrcu: Removesrcu_queue_delayed_work_on()Looks plausible. I will check more carefully for mainline.
Hmmm. I though this was already upstream.
That said, we can now schedule work from a preempt_disable() section but I still like the negative diffstat here :)
Quoted 4 lines
aeb04e894cc9srcu: replacelocal_irqsave()with a locallocke48989b033adirqwork: push most work into softirq contextThese look to still be -rt only.
I might get rid of the local_lock in srcu. Will have to check.
Thank you Paul.
Quoted 1 line
Thanx, Paul
Sebastian
Quoted 88 lines · Sebastian Andrzej Siewior
Quoted 1 line · Paul E. McKenney
Hello, Sebastian,
Hi Paul,
Quoted 4 lines
Finally getting around to taking another look:
c7e07056a108EXP rcu: skip the workqueue path on RTQuoted 11 lines
This one makes sense given the later commit setting the rcu_normal_after_boot kernel parameter. Otherwise, it is slowing down expedited grace periods for no reason. But should the check also include rcu_normal_after_boot and rcu_normal? For example: if ((IS_ENABLED(CONFIG_PREEMPT_RT_FULL) && (rcu_normal || rcu_normal_after_boot) || !READ_ONCE(rcu_par_gp_wq) || rcu_scheduler_active != RCU_SCHEDULER_RUNNING || rcu_is_last_leaf_node(rnp)) {I recently dropped that patch from the queue because the workqueue problem vanished.
Quoted 9 lines
Alternatively, one approach would be to take the kernel parameters out in -rt: static int rcu_normal_after_boot = IS_ENABLED(CONFIG_PREEMPT_RT_FULL); #ifndef CONFIG_PREEMPT_RT_FULL module_param(rcu_normal_after_boot, int, 0); #endif And similar for rcu_normal and rcu_expedited.This makes sense.
Quoted 2 lines
Or is there some reason to allow run-time expedited grace periods in
CONFIG_PREEMPT_RT_FULL=ykernels?No, I doubt there is any need to use the `expedited' version. The problem is that it increases latencies.
Quoted 4 lines
d1f52391bd8arcu: DisableRCU_FAST_NO_HZon RTLooks good. More complexity could be added if too many people get themselves in trouble via "select
RCU_FAST_NO_HZ".That patch disables
RCU_FAST_NO_HZand claims that it has something to do with atimer_listtimer and IRQ-off section. We couldn't schedule timers from IRQ-off regions but not anymore. Onlydel_timer_sync()can't be invoked from IRQ-off regions.
I just booted a box with this enabled together with NO_HZ/NO_HZ_FULLand I not complains yet. So I might drop that…Quoted 7 lines
42b346870326rcu: makeRCU_BOOSTdefault on RTTo avoid complaints about this showing up when people don't expected, could you please instead "select
RCU_BOOST" in theKconfigdefinition ofPREEMPT_RT_FULL?Or do people really want to be able to disable boosting?
I have no idea. I guess most people don't know what it does and stay with the default. It become default on RT once a few people complained that they run OOM during boot on some "memory contrained systems". That option avoided it.
So yes, will make it depend on RT.Quoted 4 lines
457c1b0d9c0esched: Do not accountrcu_preempt_depthon RT inmight_sleep()The idea behind this one is to avoid false-positive complaints about -rt's sleeping spinlocks, correct?
Correct. Maybe we could invoke a different
schedule()primitiv so RCU is aware that this is a sleeping spinlock and not a regular sleeping lock.Quoted 10 lines
7ee13e640b01rbtree: don't include the rcu headerc9b0c9b87081rtmutex: annotate sleeping lock contextNo specific comments.
7912d002ebf9rcu: Eliminate softirq processing from rcutreeThis hasn't caused any problems in -rcu from what I can see. I am therefore planning to submit the -rcu variant of this to mainline during the next merge window.
wonderful.
Quoted 7 lines
Quoted 3 lines
f06d34ebdbbbsrcu: Removesrcu_queue_delayed_work_on()Looks plausible. I will check more carefully for mainline.
Hmmm. I though this was already upstream.
That said, we can now schedule work from apreempt_disable()section but I still like the negative diffstat here :)
Right you are! e81baf4cb19a ("srcu: Remove srcu_queue_delayed_work_on()") is in v5.1.
Quoted 8 lines
Quoted 4 lines
aeb04e894cc9srcu: replacelocal_irqsave()with a locallocke48989b033adirqwork: push most work into softirq contextThese look to still be -rt only.
I might get rid of the
local_lockin srcu. Will have to check.Thank you Paul.
And you! I will check again in a few months, for some definition of "a few".
Thanx, Paul