Re: [PATCH v4 3/3] rcu: Use _full() API to debug synchronize_rcu()

[Joel Fernandes <joelagnelf@nvidia.com>]

On 3/4/2025 9:54 PM, Boqun Feng wrote:
> On Tue, Mar 04, 2025 at 11:56:18AM +0100, Uladzislau Rezki wrote:
>> On Tue, Mar 04, 2025 at 11:52:26AM +0100, Uladzislau Rezki wrote:
>>>>> Did I get that right?
>>>>>
>>>>
>>>> Other than I'm unable to follow what do you mean "WH has not been
>>>> injected, so nothing to wait on", maybe because I am missing some
>>>> terminology from you ;-) I think it's a good analysis, thank you!
>>>>
>>>>> I think this is a real bug AFAICS, hoping all the memory barriers are in
>>>>> place to make sure the code reordering also correctly orders the accesses.
>>>>> I'll double check that.
>>>>>
>>>>> I also feel its 'theoretical', because as long as rcu_gp_init() and
>>>>> rcu_gp_cleanup() are properly ordered WRT pre-existing readers, then
>>>>> synchronize_rcu_normal() still waits for pre-existing readers even though its
>>>>> a bit confused about the value of the cookies.
>>>>>
>>>>> For the fix,
>>>>> Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
>>>>>
>>>>> (If possible, include a Link: to my (this) post so that the sequence of
>>>>> events is further clarified.)
>>>>>
>>>>
>>>> Will add the tag (with the email you really want ;-)) and a link to this
>>>> email to the patch. Thanks!
>>>>
>>>
>>> CPU_1:                                | CPU_2:
>>>   # Increase a seq-number             |
>>>   rcu_seq_start(&rcu_state.gp_seq);   |
>>>                                       | add_client() {
>>> 				      |   # Record a gp-sec state 
>>> 				      |   get_state_synchronize_rcu_full(&rs.oldstate);
>>>                                       | }
>>>                                       |
>>>                                       | rcu_sr_normal_gp_init() {
>>> 				      |   add a dummy-wait-head;
>>> 				      | }
>>>
>>>
>>> A client has been added with already updated gp-sec number, i.e.
>>> "oldstate" would refer to this GP, not to previous. A poll_state_synchronize_rcu_full()
>>> will complain because this GP is not passed, it will on a next iteration.
>>>
>>> This is how i see this.
>>>
>> Updated the plain-text, removed tabs:
>>
>> CPU_1:                                 | CPU_2:
>>    # Increase a seq-number             |
>>    rcu_seq_start(&rcu_state.gp_seq);   |
>>                                        | add_client() {
>>                                        |   # Record a gp-sec state 
>>                                        |   get_state_synchronize_rcu_full(&rs.oldstate);
>>                                        | }
>>                                        |
>>                                        | rcu_sr_normal_gp_init() {
>>                                        |   add a dummy-wait-head;
>>                                        | }
>>
> 
> Thank you. I added links from you and Joel as the detailed explanation
> to the commit log, and the comment I proposed[1].
> 
> [1]: https://lore.kernel.org/rcu/Z8SnhS_LnzN_wvxr@tardis/
> 

Yep, I am in line with Vlad's explanation as well, and add links to both
explanations sounds perfect, thanks!


 - Joel