Re: [PATCH v4 0/3] virtio support cache indirect desc

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Nov 10, 2021 at 07:53:49AM -0500, Michael S. Tsirkin wrote:
> On Mon, Nov 08, 2021 at 10:47:40PM +0800, Xuan Zhuo wrote:
> > On Mon, 8 Nov 2021 08:49:27 -0500, Michael S. Tsirkin <mst@redhat.com> wrote:
> > >
> > > Hmm a bunch of comments got ignored. See e.g.
> > > https://lore.kernel.org/r/20211027043851-mutt-send-email-mst%40kernel.org
> > > if they aren't relevant add code comments or commit log text explaining the
> > > design choice please.
> > 
> > I should have responded to related questions, I am guessing whether some emails
> > have been lost.
> > 
> > I have sorted out the following 6 questions, if there are any missing questions,
> > please let me know.
> > 
> > 1. use list_head
> >   In the earliest version, I used pointers directly. You suggest that I use
> >   llist_head, but considering that llist_head has atomic operations. There is no
> >   competition problem here, so I used list_head.
> > 
> >   In fact, I did not increase the allocated space for list_head.
> > 
> >   use as desc array: | vring_desc | vring_desc | vring_desc | vring_desc |
> >   use as queue item: | list_head ........................................|
> 
> the concern is that you touch many cache lines when removing an entry.
> 
> I suggest something like:
> 
> llist: add a non-atomic list_del_first
> 
> One has to know what one's doing, but if one has locked the list
> preventing all accesses, then it's ok to just pop off an entry without
> atomics.
> 
> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> 
> ---
> 
> diff --git a/include/linux/llist.h b/include/linux/llist.h
> index 24f207b0190b..13a47dddb12b 100644
> --- a/include/linux/llist.h
> +++ b/include/linux/llist.h
> @@ -247,6 +247,17 @@ static inline struct llist_node *__llist_del_all(struct llist_head *head)
>  
>  extern struct llist_node *llist_del_first(struct llist_head *head);
>  
> +static inline struct llist_node *__llist_del_first(struct llist_head *head)
> +{
> +	struct llist_node *first = head->first;
> +
> +	if (!first)
> +		return NULL;
> +
> +	head->first = first->next;
> +	return first;
> +}
> +
>  struct llist_node *llist_reverse_order(struct llist_node *head);
>  
>  #endif /* LLIST_H */
> 
> 
> -----
> 
> 
> > 2.
> > > > +	if (vq->use_desc_cache && total_sg <= VIRT_QUEUE_CACHE_DESC_NUM) {
> > > > +		if (vq->desc_cache_chain) {
> > > > +			desc = vq->desc_cache_chain;
> > > > +			vq->desc_cache_chain = (void *)desc->addr;
> > > > +			goto got;
> > > > +		}
> > > > +		n = VIRT_QUEUE_CACHE_DESC_NUM;
> > >
> > > Hmm. This will allocate more entries than actually used. Why do it?
> > 
> > 
> > This is because the size of each cache item is fixed, and the logic has been
> > modified in the latest code. I think this problem no longer exists.
> > 
> > 
> > 3.
> > > What bothers me here is what happens if cache gets
> > > filled on one numa node, then used on another?
> > 
> > I'm thinking about another question, how did the cross-numa appear here, and
> > virtio desc queue also has the problem of cross-numa. So is it necessary for us
> > to deal with the cross-numa scene?
> 
> It's true that desc queue might be cross numa, and people are looking
> for ways to improve that. Not a reason to make things worse ...
> 

To add to that, given it's a concern, how about actually
testing performance for this config?

> > Indirect desc is used as virtio desc, so as long as it is in the same numa as
> > virito desc. So we can allocate indirect desc cache at the same time when
> > allocating virtio desc queue.
> 
> Using it from current node like we do now seems better.
> 
> > 4.
> > > So e.g. for rx, we are wasting memory since indirect isn't used.
> > 
> > In the current version, desc cache is set up based on pre-queue.
> > 
> > So if the desc cache is not used, we don't need to set the desc cache.
> > 
> > For example, virtio-net, as long as the tx queue and the rx queue in big packet
> > mode enable desc cache.
> 
> 
> I liked how in older versions adding indrect enabled it implicitly
> though without need to hack drivers.
> 
> > 5.
> > > Would a better API be a cache size in bytes? This controls how much
> > > memory is spent after all.
> > 
> > My design is to set a threshold. When total_sg is greater than this threshold,
> > it will fall back to kmalloc/kfree. When total_sg is less than or equal to
> > this threshold, use the allocated cache.
> > 
> 
> I know. My question is this, do devices know what a good threshold is?
> If yes how do they know?
> 
> > 6. kmem_cache_*
> > 
> > I have tested these, the performance is not as good as the method used in this
> > patch.
> 
> Do you mean kmem_cache_alloc_bulk/kmem_cache_free_bulk?
> You mentioned just kmem_cache_alloc previously.
> 
> > 
> > Thanks.