Hi Boqun,
> On Dec 18, 2025, at 9:07 PM, Boqun Feng <[email protected]> wrote:
>
> On Thu, Dec 18, 2025 at 06:36:00PM -0500, Mathieu Desnoyers wrote:
>> On 2025-12-18 15:22, Boqun Feng wrote:
>> [...]
>>>>> Could you utilize this[1] to see a
>>>>> comparison of the reader-side performance against RCU/SRCU?
>>>>
>>>> Good point ! Let's see.
>>>>
>>>> On a AMD 2x EPYC 9654 96-Core Processor with 192 cores,
>>>> hyperthreading disabled,
>>>> CONFIG_PREEMPT=y,
>>>> CONFIG_PREEMPT_RCU=y,
>>>> CONFIG_PREEMPT_HAZPTR=y.
>>>>
>>>> scale_type ns
>>>> -----------------------
>>>> hazptr-smp-mb 13.1 <- this implementation
>>>> hazptr-barrier 11.5 <- replace smp_mb() on acquire with
>>>> barrier(), requires IPIs on synchronize.
>>>> hazptr-smp-mb-hlist 12.7 <- replace per-task hp context and
>>>> per-cpu overflow lists by hlist.
>>>> rcu 17.0
>>>> srcu 20.0
>>>> srcu-fast 1.5
>>>> rcu-tasks 0.0
>>>> rcu-trace 1.7
>>>> refcnt 1148.0
>>>> rwlock 1190.0
>>>> rwsem 4199.3
>>>> lock 41070.6
>>>> lock-irq 46176.3
>>>> acqrel 1.1
>>>>
>>>> So only srcu-fast, rcu-tasks, rcu-trace and a plain acqrel
>>>> appear to beat hazptr read-side performance.
>>>>
>>>
>>> Could you also see the reader-side performance impact when the percpu
>>> hazard pointer slots are used up? I.e. the worst case.
>>
>> I've modified the code to populate "(void *)1UL" in the 7 first slots
>> at bootup, here is the result:
>>
>> hazptr-smp-mb-7-fail 16.3 ns
>>
>> So we go from 13.1 ns to 16.3 ns when all but one slots are used.
>>
>> And if we pre-populate the 8 slots for each cpu, and thus force
>> fallback to overflow list:
>>
>> hazptr-smp-mb-8-fail 67.1 ns
>>
>
> Thank you! So involving locking seems to hurt performance more than
> per-CPU/per-task operations. This may suggest that enabling
> PREEMPT_HAZPTR by default has an acceptable performance.
My impression is we do other locking on preemption anyway, such as the block
list for preempted RCU critical sections. So maybe that's okay. As you said, it
is an acceptable performance.
>>>
>>>> [...]
>>>>
>>>>>> +/*
>>>>>> + * Perform piecewise iteration on overflow list waiting until "addr" is
>>>>>> + * not present. Raw spinlock is released and taken between each list
>>>>>> + * item and busy loop iteration. The overflow list generation is checked
>>>>>> + * each time the lock is taken to validate that the list has not changed
>>>>>> + * before resuming iteration or busy wait. If the generation has
>>>>>> + * changed, retry the entire list traversal.
>>>>>> + */
>>>>>> +static
>>>>>> +void hazptr_synchronize_overflow_list(struct overflow_list
>>>>>> *overflow_list, void *addr)
>>>>>> +{
>>>>>> + struct hazptr_backup_slot *backup_slot;
>>>>>> + uint64_t snapshot_gen;
>>>>>> +
>>>>>> + raw_spin_lock(&overflow_list->lock);
>>>>>> +retry:
>>>>>> + snapshot_gen = overflow_list->gen;
>>>>>> + list_for_each_entry(backup_slot, &overflow_list->head, node) {
>>>>>> + /* Busy-wait if node is found. */
>>>>>> + while (smp_load_acquire(&backup_slot->slot.addr) == addr) { /*
>>>>>> Load B */
>>>>>> + raw_spin_unlock(&overflow_list->lock);
>>>>>> + cpu_relax();
>>>>>
>>>>> I think we should prioritize the scan thread solution [2] instead of
>>>>> busy waiting hazrd pointer updaters, because when we have multiple
>>>>> hazard pointer usages we would want to consolidate the scans from
>>>>> updater side.
Yeah the scan thread idea also fixes the scan cost issue with per-task slots if
we batch. If we implement a separate hazard pointer flavor along those lines,
then maybe we should definitely do a worker thread.
>>>> I agree that batching scans with a worker thread is a logical next step.
>>>>
>>>>> If so, the whole ->gen can be avoided.
>>>>
>>>> How would it allow removing the generation trick without causing long
>>>> raw spinlock latencies ?
>>>>
>>>
>>> Because we won't need to busy-wait for the readers to go away, we can
>>> check whether they are still there in the next scan.
>>>
>>> so:
>>>
>>> list_for_each_entry(backup_slot, &overflow_list->head, node) {
>>> /* Busy-wait if node is found. */
>>> if (smp_load_acquire(&backup_slot->slot.addr) == addr) { /* Load B */
>>> <mark addr as unable to free and move on>
>>
>> But then you still iterate on a possibly large list of overflow nodes,
>> with a raw spinlock held. That raw spinlock is taken by the scheduler
>> on context switch. This can cause very long scheduler latency.
>>
>
> That's fair.
What about combining both approaches?
Can we do the generation trick along with worker thread scanning? I feel that
should be doable.
>
>> So breaking up the iteration into pieces is not just to handle
>> busy-waiting, but also to make sure we don't increase the
>> system latency by holding a raw spinlock (taken with rq lock
>> held) for more than the little time needed to iterate to the next
>> node.
>>
>
> I agree that it helps reduce the latency, but I feel like with a scan
> thread in the picture (and we don't need to busy-wait), we should use
> a forward-progress-guaranteed way in the updater side scan, which means
> we may need to explore other solutions for the latency (e.g.
> fine-grained locking hashlist for the overflow list) than the generation
> counter.
Hmm.. That only works I guess if there is no interference between the finger
grained list being iterated and the list being overflowed into. Otherwise, I
think it might run into the same issue, but I could be missing something about
the idea.
thanks,
- Joel
>
> Regards,
> Boqun
>
>> Thanks,
>>
>> Mathieu
>>
>> --
>> Mathieu Desnoyers
>> EfficiOS Inc.
>> https://www.efficios.com
>
