Re: [PATCH memory-model 2/4] tools/memory-model: Add example for heuristic lockless reads

[Manfred Spraul <manfred@colorfullife.com>]

Hi Alan,

On 7/23/21 3:05 PM, Alan Stern wrote:
> On Fri, Jul 23, 2021 at 08:52:50AM +0200, Manfred Spraul wrote:
>> Hi Alan,
> Hi.
>
>> On 7/23/21 4:08 AM, Alan Stern wrote:
>>> On Wed, Jul 21, 2021 at 02:10:01PM -0700, Paul E. McKenney wrote:
>>>> This commit adds example code for heuristic lockless reads, based loosely
>>>> on the sem_lock() and sem_unlock() functions.
>>>>
>>>> Reported-by: Manfred Spraul <manfred@colorfullife.com>
>>>> [ paulmck: Update per Manfred Spraul and Hillf Danton feedback. ]
>>>> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
>>>> ---
>>>>    .../Documentation/access-marking.txt          | 94 +++++++++++++++++++
>>>>    1 file changed, 94 insertions(+)
>>>>
>>>> diff --git a/tools/memory-model/Documentation/access-marking.txt b/tools/memory-model/Documentation/access-marking.txt
>>>> index 58bff26198767..be7d507997cf8 100644
>>>> --- a/tools/memory-model/Documentation/access-marking.txt
>>>> +++ b/tools/memory-model/Documentation/access-marking.txt
>>>> @@ -319,6 +319,100 @@ of the ASSERT_EXCLUSIVE_WRITER() is to allow KCSAN to check for a buggy
>>>>    concurrent lockless write.
>>>> +Lock-Protected Writes With Heuristic Lockless Reads
>>>> +---------------------------------------------------
>>>> +
>>>> +For another example, suppose that the code can normally make use of
>>>> +a per-data-structure lock, but there are times when a global lock
>>>> +is required.  These times are indicated via a global flag.  The code
>>>> +might look as follows, and is based loosely on nf_conntrack_lock(),
>>>> +nf_conntrack_all_lock(), and nf_conntrack_all_unlock():
>>>> +
>>>> +	bool global_flag;
>>>> +	DEFINE_SPINLOCK(global_lock);
>>>> +	struct foo {
>>>> +		spinlock_t f_lock;
>>>> +		int f_data;
>>>> +	};
>>>> +
>>>> +	/* All foo structures are in the following array. */
>>>> +	int nfoo;
>>>> +	struct foo *foo_array;
>>>> +
>>>> +	void do_something_locked(struct foo *fp)
>>>> +	{
>>>> +		bool gf = true;
>>>> +
>>>> +		/* IMPORTANT: Heuristic plus spin_lock()! */
>>>> +		if (!data_race(global_flag)) {
>>>> +			spin_lock(&fp->f_lock);
>>>> +			if (!smp_load_acquire(&global_flag)) {
>>>> +				do_something(fp);
>>>> +				spin_unlock(&fp->f_lock);
>>>> +				return;
>>>> +			}
>>>> +			spin_unlock(&fp->f_lock);
>>>> +		}
>>>> +		spin_lock(&global_lock);
>>>> +		/* Lock held, thus global flag cannot change. */
>>>> +		if (!global_flag) {
>>> How can global_flag ever be true at this point?  The only line of code
>>> that sets it is in begin_global() below, it only runs while global_lock
>>> is held, and global_flag is set back to false before the lock is
>>> released.
>> It can't be true. The code is a simplified version of the algorithm in
>> ipc/sem.c.
>>
>> For the ipc/sem.c, global_flag can remain true even after dropping
>> global_lock.
>>
>> When transferring the approach to nf_conntrack_core, I didn't notice that
>> nf_conntrack doesn't need a persistent global_flag.
>>
>> Thus the recheck after spin_lock(&global_lock) is not needed.
> In fact, since global_flag is true if and only if global_lock is locked,
> perhaps it can be removed entirely and replaced with
> spin_is_locked(&global_lock).

I try to avoid spin_is_locked():

- spin_is_locked() is no memory barrier

- spin_lock() is an acquire memory barrier - for the read part. There is 
no barrier at all related to the write part.

With an explicit variable, the memory barriers can be controlled much 
better - and it is guaranteed to work in the same way on all architectures.


--

     Manfred