Re: [PATCH 4/4] Docs/RCU/rculist_nulls: Drop unnecessary '_release' in insert function

[Alan Huang]

Hi SJ,

> 2023年6月10日 08:20，SeongJae Park <sj@xxxxxxxxxx> 写道：
> 
> On Fri, 9 Jun 2023 16:42:59 -0700 "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:
> 
>> On Fri, Jun 09, 2023 at 07:12:06PM +0000, SeongJae Park wrote:
>>> On Fri, 19 May 2023 14:52:50 -0400 Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>>> 
>>>> On Thu, May 18, 2023 at 6:40 PM SeongJae Park <sj@xxxxxxxxxx> wrote:
>>>>> 
>>>>> The document says we can avoid extra smp_rmb() in lockless_lookup() and
>>>>> extra _release() in insert function when hlist_nulls is used.  However,
>>>>> the example code snippet for the insert function is still using the
>>>>> extra _release().  Drop it.
>>>>> 
>>>>> Signed-off-by: SeongJae Park <sj@xxxxxxxxxx>
>>>>> ---
>>>>> Documentation/RCU/rculist_nulls.rst | 2 +-
>>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>> 
>>>>> diff --git a/Documentation/RCU/rculist_nulls.rst b/Documentation/RCU/rculist_nulls.rst
>>>>> index 5cd6f3f8810f..463270273d89 100644
>>>>> --- a/Documentation/RCU/rculist_nulls.rst
>>>>> +++ b/Documentation/RCU/rculist_nulls.rst
>>>>> @@ -191,7 +191,7 @@ scan the list again without harm.
>>>>>   obj = kmem_cache_alloc(cachep);
>>>>>   lock_chain(); // typically a spin_lock()
>>>>>   obj->key = key;
>>>>> -  atomic_set_release(&obj->refcnt, 1); // key before refcnt
>>>>> +  atomic_set(&obj->refcnt, 1);
>>>>>   /*
>>>>>    * insert obj in RCU way (readers might be traversing chain)
>>>>>    */
>>>> 
>>>> If write to ->refcnt of 1 is reordered with setting of ->key, what
>>>> prevents the 'lookup algorithm' from doing a key match (obj->key ==
>>>> key) before the refcount has been initialized?
>>>> 
>>>> Are we sure the reordering mentioned in the document is the same as
>>>> the reordering prevented by the atomic_set_release()?
>>> 
>>> Paul, may I ask your opinion?
>> 
>> The next line of code is this:
>> 
>> hlist_nulls_add_head_rcu(&obj->obj_node, list);
>> 
>> If I understand the code correctly, obj (and thus *obj) are not
>> visible to readers before the hlist_nulls_add_head_rcu().  And
>> hlist_nulls_add_head_rcu() uses rcu_assign_pointer() to ensure that
>> initialization (including both ->key and ->refcnt) is ordered before
>> list insertion.
>> 
>> Except that this memory is being allocated from a slab cache that was
>> created with SLAB_TYPESAFE_BY_RCU.  This means that there can be readers
>> who gained a reference before this object was freed, and who still hold
>> their references.
>> 
>> Unfortunately, the implementation of try_get_ref() is not shown.  However,
>> if ->refcnt is non-zero, this can succeed, and if it succeeds, we need
>> the subsequent check of obj->key with key in the lookup algorithm to
>> be stable.  For this check to be stable, try_get_ref() needs to use an
>> atomic operation with at least acquire semantics (kref_get_unless_zero()
>> would work), and this must pair with something in the initialization.
>> 
>> So I don't see how it is safe to weaken that atomic_set_release() to
>> atomic_set(), even on x86.
> 
> Thank you for the nice explanation, and I agree.
> 
>> 
>> Or am I missing something subtle here?
> 
> I found the text is saying extra _release() in insert function is not
> needed[1], and I thought it means the atomic_set_release().  Am I misreading
> it?  If not, would it be better to fix the text, for example, like below?

The original text is:

	“With hlist_nulls we can avoid extra smp_rmb() in lockless_lookup()
	 and extra smp_wmb() in insert function.”

We can avoid the extra smp_wmb(), but the _release is required,

As Paul said,

>> Except that this memory is being allocated from a slab cache that was
>> created with SLAB_TYPESAFE_BY_RCU.  This means that there can be readers
>> who gained a reference before this object was freed, and who still hold
>> their references.

Without the _release, we can get the old ‘key’ after the invocation of
try_get_ref (although try_get_ref noticed the effect of atomic_set).

Thanks,
Alan

> 
> ```
> --- a/Documentation/RCU/rculist_nulls.rst
> +++ b/Documentation/RCU/rculist_nulls.rst
> @@ -129,8 +129,7 @@ very very fast (before the end of RCU grace period)
> Avoiding extra smp_rmb()
> ========================
> 
> -With hlist_nulls we can avoid extra smp_rmb() in lockless_lookup()
> -and extra _release() in insert function.
> +With hlist_nulls we can avoid extra smp_rmb() in lockless_lookup().
> 
> For example, if we choose to store the slot number as the 'nulls'
> end-of-list marker for each slot of the hash table, we can detect
> @@ -182,6 +181,9 @@ scan the list again without harm.
> 2) Insert algorithm
> -------------------
> 
> +Same to the above one, but uses hlist_nulls_add_head_rcu() instead of
> +hlist_add_head_rcu().
> +
> ::
> 
>   /*
> @@ -191,7 +193,7 @@ scan the list again without harm.
>   obj = kmem_cache_alloc(cachep);
>   lock_chain(); // typically a spin_lock()
>   obj->key = key;
> -  atomic_set_release(&obj->refcnt, 1); // key before refcnt
> +  atomic_set(&obj->refcnt, 1);
>   /*
>    * insert obj in RCU way (readers might be traversing chain)
>    */
> ```
> 
> [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/RCU/rculist_nulls.rst#n133
> 
> 
> Thanks,
> SJ
> 
>> 
>> Thanx, Paul
>> 
>>> Thanks,
>>> SJ
>>> 
>>>> 
>>>> For the other 3 patches, feel free to add:
>>>> Reviewed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
>>>> 
>>>> thanks,
>>>> 
>>>> - Joel