Hi Ferruh,
> -----Original Message-----
> From: Ferruh Yigit <ferruh.yi...@xilinx.com>
> Sent: Thursday, May 19, 2022 12:53 AM
> To: Joyce Kong <joyce.k...@arm.com>; Jakub Grajciar <jgraj...@cisco.com>
> Cc: Ruifeng Wang <ruifeng.w...@arm.com>; dev@dpdk.org; nd
> <n...@arm.com>
> Subject: Re: [PATCH v1 1/2] net/memif: add a Rx fast path
>
> On 5/17/2022 11:51 AM, Joyce Kong wrote:
> > For memif non-zero-copy mode, there is a branch to compare the mbuf
> > and memif buffer size during memory copying. Add a fast memory copy
> > path by removing this branch with mbuf and memif buffer size defined
> > at compile time. The removal of the branch leads to considerable
> > performance uplift.
> >
> > When memif <= buffer size, Rx chooses the fast memcpy path, otherwise
> > it would choose the original path.
> >
> > Test with 1p1q on Ampere Altra AArch64 server,
> > --------------------------------------------
> > buf size | memif <= mbuf | memif > mbuf |
> > --------------------------------------------
> > non-zc gain | 4.30% | -0.52% |
> > --------------------------------------------
> > zc gain | 2.46% | 0.70% |
> > --------------------------------------------
> >
> > Test with 1p1q on Cascade Lake Xeon X86server,
> > -------------------------------------------
> > buf size | memif <= mbuf | memif > mbuf |
> > -------------------------------------------
> > non-zc gain | 2.13% | -1.40% |
> > -------------------------------------------
> > zc gain | 0.18% | 0.48% |
> > -------------------------------------------
> >
>
>
> Hi Joyce,
>
> I have multiple questions,
>
> 1) The patch updates only non-zero-copy mode Rx path ('eth_memif_rx'), why
> zero-copy path performance also impacted?
>
For memif zero-copy mode, only client runs 'eth_memif_rx_zc', and server still
runs
'eth_memif_rx', so the patch would impacts zero-copy mode.
> 2) As far as I can see there is a behavior change, more details below
>
> 3) patch talking about memif buffer size being defined in compile time, is the
> big "memif <= mbuf" if block optimized out?
> Since 'pkt_buffer_size' is a devarg, so it can change from run to run and it
> is not
> known in compile time, I doubt that it is optimized out.
> Is having 'pkt_buffer_size' as devarg breaks your logic?
>
From memif run to run, run.pkt_buffer_size would change, and cfg.pkt_buffer_size
which is the reserved max buffer size would not change. For patch details, I use
cfg.pkt_buffer_size to implement the logic.
> 4) One option gains performance and other loose performance, do you think
> gain performance case is more common use case? Is there any data around it?
>
Yes, I think the gain performance case is more common case, as the default memif
buffer size equals to mbuf size. In theory, when memif buf size >= mbuf size,
the Rx
runs the original path, it would not lead to obvious impact.
>
> Jakub,
>
> Do you want to test this patch first before progressing with it?
>
> > Signed-off-by: Joyce Kong <joyce.k...@arm.com>
> > ---
> > drivers/net/memif/rte_eth_memif.c | 124 ++++++++++++++++++++----------
> > 1 file changed, 84 insertions(+), 40 deletions(-)
> >
> > diff --git a/drivers/net/memif/rte_eth_memif.c
> > b/drivers/net/memif/rte_eth_memif.c
> > index 587ad45576..f55776ca46 100644
> > --- a/drivers/net/memif/rte_eth_memif.c
> > +++ b/drivers/net/memif/rte_eth_memif.c
> > @@ -342,66 +342,111 @@ eth_memif_rx(void *queue, struct rte_mbuf
> **bufs, uint16_t nb_pkts)
> > goto refill;
> > n_slots = last_slot - cur_slot;
> >
> > - while (n_slots && n_rx_pkts < nb_pkts) {
> > - mbuf_head = rte_pktmbuf_alloc(mq->mempool);
> > - if (unlikely(mbuf_head == NULL))
> > - goto no_free_bufs;
> > - mbuf = mbuf_head;
> > - mbuf->port = mq->in_port;
> > + if (likely(mbuf_size >= pmd->cfg.pkt_buffer_size)) {
> > + while (n_slots && n_rx_pkts < nb_pkts) {
> > + mbuf_head = rte_pktmbuf_alloc(mq->mempool);
> > + if (unlikely(mbuf_head == NULL))
> > + goto no_free_bufs;
> > + mbuf = mbuf_head;
> > + mbuf->port = mq->in_port;
> > +
> > +next_slot1:
> > + s0 = cur_slot & mask;
> > + d0 = &ring->desc[s0];
> >
> > -next_slot:
> > - s0 = cur_slot & mask;
> > - d0 = &ring->desc[s0];
> > + cp_len = d0->length;
> >
> > - src_len = d0->length;
> > - dst_off = 0;
> > - src_off = 0;
> > + rte_pktmbuf_data_len(mbuf) = cp_len;
> > + rte_pktmbuf_pkt_len(mbuf) = cp_len;
> > + if (mbuf != mbuf_head)
> > + rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
> >
> > - do {
> > - dst_len = mbuf_size - dst_off;
> > - if (dst_len == 0) {
> > - dst_off = 0;
> > - dst_len = mbuf_size;
> > + rte_memcpy(rte_pktmbuf_mtod(mbuf, void *),
> > + (uint8_t *)memif_get_buffer(proc_private, d0),
> cp_len);
> > +
> > + cur_slot++;
> > + n_slots--;
> >
> > - /* store pointer to tail */
> > + if (d0->flags & MEMIF_DESC_FLAG_NEXT) {
> > mbuf_tail = mbuf;
> > mbuf = rte_pktmbuf_alloc(mq->mempool);
> > if (unlikely(mbuf == NULL))
> > goto no_free_bufs;
> > - mbuf->port = mq->in_port;
> > ret = memif_pktmbuf_chain(mbuf_head,
> mbuf_tail, mbuf);
> > if (unlikely(ret < 0)) {
> > MIF_LOG(ERR, "number-of-segments-
> overflow");
> > rte_pktmbuf_free(mbuf);
> > goto no_free_bufs;
> > }
> > + goto next_slot1;
> > }
>
> It is very hard to comment on the correct part of the patch, since it is
> mixed a
> lot, but
> - previously when memif buffer is segmented, and its size is less than mbuf;
> mbuf is filled with as much memif data as possible and later switched to next
> mbuf, like:
>
> memif buffer
> +-+ +-+ +-+ +-+
> |a|->|b|->|c|->|d|
> +-+ +-+ +-+ +-+
>
> +---+ +---+
> |abc|->|d |
> +---+ +---+
> mbuf
>
>
> - Now each memif segment is a mbuf,
>
> memif buffer
> +-+ +-+ +-+ +-+
> |a|->|b|->|c|->|d|
> +-+ +-+ +-+ +-+
>
> +---+ +---+ +---+ +---+
> |a |->|b |->|c |->|d |
> +---+ +---+ +---+ +---+
> mbuf
>
> Can you please confirm this behavior change? If so can you please highlight is
> more in the commit log?
> And is this tradeoff something preferred?
