Re: [PATCH v4 -rcu 1/4] rcu/segcblist: Do not depend on rcl->len to store the segcb len during merge

[Paul E. McKenney]

On Mon, Aug 24, 2020 at 10:48:39PM -0400, Joel Fernandes (Google) wrote:
> The donecbs's ->len field is used to store the total count of the segmented
> callback list's length. This ->len field is then added to the destination segcb
> list.
> 
> However, this presents a problem for per-segment length counting which is added
> in a future patch. This future patch sets the rcl->len field as we move
> segments of callbacks between source and destination lists, thus becoming
> incompatible with the donecb's ->len field.

OK, I will bite.  What is "rcl"?  A placeholder for donecbs and pendcbs?
If so, please just name them both.  If not, please explain.

> This commit therefore avoids depending on the ->len field in this way. IMHO,
> this is also less error-prone and is more accurate - the donecb's ->len field
> should be the length of the done segment and not just used as a temporarily
> variable.

Please also mention why ->len is handled specially at all, namely
interactions between rcu_barrier() and callback invocation.  This is
the answer to "why not just make all this work like normal lists?"
This might go well in the first paragraph.

> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> ---
>  kernel/rcu/rcu_segcblist.c | 8 ++++++--
>  1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/rcu/rcu_segcblist.c b/kernel/rcu/rcu_segcblist.c
> index 2d2a6b6b9dfb..b70d4154433c 100644
> --- a/kernel/rcu/rcu_segcblist.c
> +++ b/kernel/rcu/rcu_segcblist.c
> @@ -513,14 +513,18 @@ void rcu_segcblist_merge(struct rcu_segcblist *dst_rsclp,
>  {
>  	struct rcu_cblist donecbs;
>  	struct rcu_cblist pendcbs;
> +	long src_len;
>  
>  	rcu_cblist_init(&donecbs);
>  	rcu_cblist_init(&pendcbs);
> -	rcu_segcblist_extract_count(src_rsclp, &donecbs);
> +
> +	src_len = rcu_segcblist_xchg_len(src_rsclp, 0);

Given that both rcu_segcblist_xchg_len() and rcu_segcblist_extract_count()
have only one callsite each, why not get rid of one of them?

Or better yet, please see below, which should allow getting rid of both
of them.

>  	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
>  	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
> -	rcu_segcblist_insert_count(dst_rsclp, &donecbs);
> +
> +	rcu_segcblist_add_len(dst_rsclp, src_len);
>  	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
>  	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);

Rather than adding the blank lines, why not have the rcu_cblist structures
carry the lengths?  You are already adjusting one of the two call sites
that care (rcu_do_batch()), and the other is srcu_invoke_callbacks().
That should shorten this function a bit more.  And make callback handling
much more approachable, I suspect.

There would still be the callback-invocation need to be careful with
->cblist.len due to rcu_barrier() and srcu_barrier().  But both of
those should be excluded by this code.  (But don't take my word for it,
ask KCSAN.)

							Thanx, Paul

> +
>  	rcu_segcblist_init(src_rsclp);
>  }
> -- 
> 2.28.0.297.g1956fa8f8d-goog
>