Re: [PATCH 1/3] rculist: Fix list_entry_rcu to read ptr with rcu_dereference_raw

[Patrick Marlier <patrick.marlier@gmail.com>]

On 03/25/2015 03:30 PM, Paul E. McKenney wrote:
> On Tue, Mar 24, 2015 at 11:31:38AM +0100, Patrick Marlier wrote:
>> Change to read effectively ptr with rcu_dereference_raw and not the
>> __ptr variable on the stack.
>>
>> Signed-off-by: Patrick Marlier <patrick.marlier@gmail.com>
> Avoiding an extra load could be worthwhile in a number of situations,
> agreed.
Not only a load. It adds a store and a load on the stack and I think 
this creates a dependency in the processor pipeline.

> However, won't this change cause sparse to complain if invoked on a
> non-RCU-protected pointer?  The ability to use list-RCU API
> members on both RCU and non-RCU pointers was one of the points
> of the previous commit, right?
Probably we can put back the cast but I am not familiar enough with the 
RCU API.

Also, the problem here is that you probably want ACCESS_ONCE to happen 
on the content of 'ptr' and not on the stack variable '__ptr'.

(you have to follow this chain: rcu_dereference_raw -> 
rcu_dereference_check -> __rcu_dereference_check -> lockless_dereference 
-> ACCESS_ONCE)

#define lockless_dereference(p) \
({ \
     typeof(p) _________p1 = ACCESS_ONCE(p); \
     smp_read_barrier_depends(); /* Dependency order vs. p above. */ \
     (_________p1); \
})

#define __ACCESS_ONCE(x) ({ \
      __maybe_unused typeof(x) __var = (__force typeof(x)) 0; \
     (volatile typeof(x) *)&(x); })
#define ACCESS_ONCE(x) (*__ACCESS_ONCE(x))

Note that ACCESS_ONCE is doing "&" on x.

IMHO, I would prefer saving some useless instructions for better 
performance rather than giving too much flexibility on the API (also 
pretty sure the cast can be still done).
--
Patrick Marlier