Message-ID: <20211110074326-mutt-send-email-mst@kernel.org>
Date:   Wed, 10 Nov 2021 07:53:44 -0500
From:   "Michael S. Tsirkin" <mst@...hat.com>
To:     Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
Cc:     virtualization@...ts.linux-foundation.org, netdev@...r.kernel.org,
        Jason Wang <jasowang@...hat.com>,
        "David S. Miller" <davem@...emloft.net>,
        Jakub Kicinski <kuba@...nel.org>
Subject: Re: [PATCH v4 0/3] virtio support cache indirect desc

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@...hat.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@...hat.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 ...


> 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.

Powered by blists - more mailing lists