Re: [RFC PATCH] mm: thp: grab the lock before manipulation defer list

[Wei Yang <richardw.yang@linux.intel.com>]

On Wed, Jan 08, 2020 at 10:40:41AM +0100, Michal Hocko wrote:
>On Wed 08-01-20 08:35:43, Wei Yang wrote:
>> On Tue, Jan 07, 2020 at 09:38:08AM +0100, Michal Hocko wrote:
>> >On Tue 07-01-20 09:22:41, Wei Yang wrote:
>> >> On Mon, Jan 06, 2020 at 11:23:45AM +0100, Michal Hocko wrote:
>> >> >On Fri 03-01-20 22:34:07, Wei Yang wrote:
>> >> >> As all the other places, we grab the lock before manipulate the defer list.
>> >> >> Current implementation may face a race condition.
>> >> >
>> >> >Please always make sure to describe the effect of the change. Why a racy
>> >> >list_empty check matters?
>> >> >
>> >> 
>> >> Hmm... access the list without proper lock leads to many bad behaviors.
>> >
>> >My point is that the changelog should describe that bad behavior.
>> >
>> >> For example, if we grab the lock after checking list_empty, the page may
>> >> already be removed from list in split_huge_page_list. And then list_del_init
>> >> would trigger bug.
>> >
>> >And how does list_empty check under the lock guarantee that the page is
>> >on the deferred list?
>> 
>> Just one confusion, is this kind of description basic concept of concurrent
>> programming? How detail level we need to describe the effect?
>
>When I write changelogs for patches like this I usually describe, what
>is the potential race - e.g.
>	CPU1			CPU2
>	path1			path2
>	  check			  lock
>	  			    operation2
>				  unlock
>	    lock
>	    # check might not hold anymore
>	    operation1
>	    unlock
>

Nice, I would prepare a changelog like this.

>and what is the effect of the race - e.g. a crash, data corruption,
>pointless attempt for operation1 which fails with user visible effect
>etc.
>This helps reviewers and everybody reading the code in the future to
>understand the locking scheme.
>
>> To me, grab the lock before accessing the critical section is obvious.
>
>It might be obvious but in many cases it is useful to minimize the
>locking and do a potentially race check before the lock is taken if the
>resulting operation can handle that.
>
>> list_empty and list_del should be the critical section. And the
>> lock should protect the whole critical section instead of part of it.
>
>I am not disputing that. What I am trying to say is that the changelog
>should described the problem in the first place.
>
>Moreover, look at the code you are trying to fix. Sure extending the
>locking seem straightforward but does it result in a correct code
>though? See my question in the previous email. How do we know that the
>page is actually enqued in a non-empty list?

I may not get your point for the last sentence.

The list_empty() doesn't check the queue is empty but check the list, here is
the page, is not enqueued into any list. Is this your concern?

>-- 
>Michal Hocko
>SUSE Labs

-- 
Wei Yang
Help you, Help me