Re: Re: [PATCH v2 net-next] xsk: Avoid starving xsk at the end of the list

[黄杰]

Magnus Karlsson <magnus.karlsson@xxxxxxxxx> 于2023年10月16日周一 14:41写道：
>
> On Mon, 16 Oct 2023 at 05:17, Albert Huang
> <huangjie.albert@xxxxxxxxxxxxx> wrote:
> >
> > In the previous implementation, when multiple xsk sockets were
> > associated with a single xsk_buff_pool, a situation could arise
> > where the xsk_tx_list maintained data at the front for one xsk
> > socket while starving the xsk sockets at the back of the list.
> > This could result in issues such as the inability to transmit packets,
> > increased latency, and jitter. To address this problem, we introduced
> > a new variable called tx_budget_cache, which limits each xsk to transmit
> > a maximum of MAX_XSK_TX_BUDGET tx descriptors. This allocation ensures
> > equitable opportunities for subsequent xsk sockets to send tx descriptors.
> > The value of MAX_XSK_TX_BUDGET is temporarily set to 16.
>
> Hi Albert. Yes you are correct that there is nothing hindering this to
> happen in the code at the moment, so let us fix it.
>
thanks.

> > Signed-off-by: Albert Huang <huangjie.albert@xxxxxxxxxxxxx>
> > ---
> >  include/net/xdp_sock.h |  6 ++++++
> >  net/xdp/xsk.c          | 18 ++++++++++++++++++
> >  2 files changed, 24 insertions(+)
> >
> > diff --git a/include/net/xdp_sock.h b/include/net/xdp_sock.h
> > index 69b472604b86..f617ff54e38c 100644
> > --- a/include/net/xdp_sock.h
> > +++ b/include/net/xdp_sock.h
> > @@ -44,6 +44,7 @@ struct xsk_map {
> >         struct xdp_sock __rcu *xsk_map[];
> >  };
> >
> > +#define MAX_XSK_TX_BUDGET 16
>
> I think something like MAX_PER_SOCKET_BUDGET would be clearer.
>

 OK, this will be considered  in the next patch.

> >  struct xdp_sock {
> >         /* struct sock must be the first member of struct xdp_sock */
> >         struct sock sk;
> > @@ -63,6 +64,11 @@ struct xdp_sock {
> >
> >         struct xsk_queue *tx ____cacheline_aligned_in_smp;
> >         struct list_head tx_list;
> > +       /* Record the actual number of times xsk has transmitted a tx
> > +        * descriptor, with a maximum limit not exceeding MAX_XSK_TX_BUDGET
> > +        */
> > +       u32 tx_budget_cache;
> > +
> >         /* Protects generic receive. */
> >         spinlock_t rx_lock;
> >
> > diff --git a/net/xdp/xsk.c b/net/xdp/xsk.c
> > index f5e96e0d6e01..087f2675333c 100644
> > --- a/net/xdp/xsk.c
> > +++ b/net/xdp/xsk.c
> > @@ -413,16 +413,25 @@ EXPORT_SYMBOL(xsk_tx_release);
> >
> >  bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
> >  {
> > +       u32 xsk_full_count = 0;
>
> Enough with a bool;
>
> >         struct xdp_sock *xs;
> >
> >         rcu_read_lock();
> > +again:
> >         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
> > +               if (xs->tx_budget_cache >= MAX_XSK_TX_BUDGET) {
> > +                       xsk_full_count++;
> > +                       continue;
> > +               }
>
> The problem here is that the fixed MAX_XSK_TX_BUDGET is only useful
> for the <= 2 socket case. If I have 3 sockets sharing a
> netdev/queue_id, the two first sockets can still starve the third one
> since the total budget per send is 32.

Why is there a limit of 32? I'm not quite clear on the implications of these,
Did I miss something?
BR
Albert

>You need to go through the list
> of sockets in the beginning to compute the MAX_XSK_TX_BUDGET to
> compute this dynamically before each call. Or cache this value
> somehow, in the pool for example. Actually, the refcount in the
> buf_pool will tell you how many sockets are sharing the same buf_pool.
> Try using that to form MAX_XSK_TX_BUDGET on the fly.
>
> Another simpler way of accomplishing this would be to just reorder the
> list every time. Put the first socket last in the list every time. The
> drawback of this is that you need to hold the xsk_tx_list_lock while
> doing this so might be slower. The per socket batch size would also be
> 32 and you would not receive "fairness" over a single call to
> sendto(). Would that be a problem for you?
>

Yes, I did consider this approach, but I abandoned it because it would lose
the performance advantages of lock-free operations(RCU read)
thanks
Albert


> > +
> >                 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
> >                         if (xskq_has_descs(xs->tx))
> >                                 xskq_cons_release(xs->tx);
> >                         continue;
> >                 }
> >
> > +               xs->tx_budget_cache++;
> > +
> >                 /* This is the backpressure mechanism for the Tx path.
> >                  * Reserve space in the completion queue and only proceed
> >                  * if there is space in it. This avoids having to implement
> > @@ -436,6 +445,14 @@ bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
> >                 return true;
> >         }
> >
> > +       if (unlikely(xsk_full_count > 0)) {
> > +               list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
> > +                       xs->tx_budget_cache = 0;
> > +               }
> > +               xsk_full_count = 0;
> > +               goto again;
> > +       }

this section of code only enters when it's unable to acquire any TX
descriptors and
xsk_full_count > 0.

> > +
> >  out:
> >         rcu_read_unlock();
> >         return false;
> > @@ -1230,6 +1247,7 @@ static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
> >         xs->zc = xs->umem->zc;
> >         xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG);
> >         xs->queue_id = qid;
> > +       xs->tx_budget_cache = 0;
> >         xp_add_xsk(xs->pool, xs);
> >
> >  out_unlock:
> > --
> > 2.20.1
> >
> >