Hi Maxime,
> -----Original Message-----
> From: Fu, Patrick
> Sent: Wednesday, July 29, 2020 9:40 AM
> To: 'Maxime Coquelin' <maxime.coque...@redhat.com>; dev@dpdk.org; Xia,
> Chenbo <chenbo....@intel.com>
> Subject: RE: [PATCH v4] vhost: fix async copy fail on multi-page buffers
>
> Hi Maxime,
>
> > -----Original Message-----
> > From: Maxime Coquelin <maxime.coque...@redhat.com>
> > Sent: Tuesday, July 28, 2020 9:55 PM
> > To: Fu, Patrick <patrick...@intel.com>; dev@dpdk.org; Xia, Chenbo
> > <chenbo....@intel.com>
> > Subject: Re: [PATCH v4] vhost: fix async copy fail on multi-page
> > buffers
> >
> >
> >
> > On 7/28/20 5:28 AM, patrick...@intel.com wrote:
> > > From: Patrick Fu <patrick...@intel.com>
> > >
> > > Async copy fails when single ring buffer vector is splited on
> > > multiple physical pages. This happens because current hpa address
> > > translation function doesn't handle multi-page buffers. A new gpa to
> > > hpa address conversion function, which returns the hpa on the first
> > > hitting host pages, is implemented in this patch. Async data path
> > > recursively calls this new function to construct a multi-segments
> > > async copy descriptor for ring buffers crossing physical page boundaries.
> > >
> > > Fixes: cd6760da1076 ("vhost: introduce async enqueue for split
> > > ring")
> > >
> > > Signed-off-by: Patrick Fu <patrick...@intel.com>
> > > ---
> > > v2:
> > > - change commit message and title
> > > - v1 patch used CPU to copy multi-page buffers; v2 patch split the
> > > copy into multiple async copy segments whenever possible
> > >
> > > v3:
> > > - added fixline
> > >
> > > v4:
> > > - fix miss translation of the gpa which is the same length with host
> > > page size
> > >
> > > lib/librte_vhost/vhost.h | 50
> +++++++++++++++++++++++++++++++++++
> > > lib/librte_vhost/virtio_net.c | 40 +++++++++++++++++-----------
> > > 2 files changed, 75 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/lib/librte_vhost/virtio_net.c
> > > b/lib/librte_vhost/virtio_net.c index 95a0bc19f..124a33a10 100644
> > > --- a/lib/librte_vhost/virtio_net.c
> > > +++ b/lib/librte_vhost/virtio_net.c
> > > @@ -980,6 +980,7 @@ async_mbuf_to_desc(struct virtio_net *dev,
> > > struct
> > vhost_virtqueue *vq,
> > > struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
> > > struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
> > > int error = 0;
> > > + uint64_t mapped_len;
> > >
> > > uint32_t tlen = 0;
> > > int tvec_idx = 0;
> > > @@ -1072,24 +1073,31 @@ async_mbuf_to_desc(struct virtio_net *dev,
> > > struct vhost_virtqueue *vq,
> > >
> > > cpy_len = RTE_MIN(buf_avail, mbuf_avail);
> > >
> > > - if (unlikely(cpy_len >= cpy_threshold)) {
> > > - hpa = (void *)(uintptr_t)gpa_to_hpa(dev,
> > > - buf_iova + buf_offset, cpy_len);
> > > + while (unlikely(cpy_len && cpy_len >= cpy_threshold)) {
> > > + hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
> > > + buf_iova + buf_offset,
> > > + cpy_len, &mapped_len);
> > >
> > > - if (unlikely(!hpa)) {
> > > - error = -1;
> > > - goto out;
> > > - }
> > > + if (unlikely(!hpa || mapped_len < cpy_threshold))
> > > + break;
> > >
> > > async_fill_vec(src_iovec + tvec_idx,
> > > (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
> > > - mbuf_offset), cpy_len);
> > > + mbuf_offset), (size_t)mapped_len);
> > >
> > > - async_fill_vec(dst_iovec + tvec_idx, hpa, cpy_len);
> > > + async_fill_vec(dst_iovec + tvec_idx,
> > > + hpa, (size_t)mapped_len);
> > >
> > > - tlen += cpy_len;
> > > + tlen += (uint32_t)mapped_len;
> > > + cpy_len -= (uint32_t)mapped_len;
> > > + mbuf_avail -= (uint32_t)mapped_len;
> > > + mbuf_offset += (uint32_t)mapped_len;
> > > + buf_avail -= (uint32_t)mapped_len;
> > > + buf_offset += (uint32_t)mapped_len;
> > > tvec_idx++;
> > > - } else {
> > > + }
> > > +
> > > + if (likely(cpy_len)) {
> > > if (unlikely(vq->batch_copy_nb_elems >= vq->size)) {
> > > rte_memcpy(
> > > (void *)((uintptr_t)(buf_addr + buf_offset)),
> > @@ -1112,10
> > > +1120,12 @@ async_mbuf_to_desc(struct virtio_net *dev, struct
> > vhost_virtqueue *vq,
> > > }
> > > }
> > >
> > > - mbuf_avail -= cpy_len;
> > > - mbuf_offset += cpy_len;
> > > - buf_avail -= cpy_len;
> > > - buf_offset += cpy_len;
> > > + if (cpy_len) {
> > > + mbuf_avail -= cpy_len;
> > > + mbuf_offset += cpy_len;
> > > + buf_avail -= cpy_len;
> > > + buf_offset += cpy_len;
> > > + }
> >
> > Is that really necessary to check if copy length is not 0?
> >
> The intension is to optimize for the case that ring buffers are NOT split
> (which
> should be the most common case). In that case, cpy_len will be zero and by
> this "if" statement we can save couple of cycles. With that said, the actual
> difference is minor. I'm open with either adding an "unlikely" to the "if", or
> removing this the "if". Would like to hear your option and submit modified
> patch.
>
I have a better way to handle the case (combine this "if" logic with the
previous one).
Please review my v5 patch for the code change.
Thanks,
Patrick
