On 9/21/20 8:48 AM, Marvin Liu wrote:
> Optimize vhost packed ring enqueue path with SIMD instructions. Four
> descriptors status and length are batched handled with AVX512
> instructions. Address translation operations are also accelerated
> by AVX512 instructions.
>
> Signed-off-by: Marvin Liu <yong....@intel.com>
>
> diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
> index fc7daf2145..b78b2c5c1b 100644
> --- a/lib/librte_vhost/vhost.h
> +++ b/lib/librte_vhost/vhost.h
> @@ -1132,4 +1132,10 @@ vhost_reserve_avail_batch_packed_avx(struct virtio_net
> *dev,
> uint16_t avail_idx,
> uintptr_t *desc_addrs,
> uint16_t *ids);
> +
> +int
> +virtio_dev_rx_batch_packed_avx(struct virtio_net *dev,
> + struct vhost_virtqueue *vq,
> + struct rte_mbuf **pkts);
> +
> #endif /* _VHOST_NET_CDEV_H_ */
> diff --git a/lib/librte_vhost/vhost_vec_avx.c
> b/lib/librte_vhost/vhost_vec_avx.c
> index dc5322d002..7d2250ed86 100644
> --- a/lib/librte_vhost/vhost_vec_avx.c
> +++ b/lib/librte_vhost/vhost_vec_avx.c
> @@ -35,9 +35,15 @@
> #define PACKED_AVAIL_FLAG ((0ULL | VRING_DESC_F_AVAIL) << FLAGS_BITS_OFFSET)
> #define PACKED_AVAIL_FLAG_WRAP ((0ULL | VRING_DESC_F_USED) << \
> FLAGS_BITS_OFFSET)
> +#define PACKED_WRITE_AVAIL_FLAG (PACKED_AVAIL_FLAG | \
> + ((0ULL | VRING_DESC_F_WRITE) << FLAGS_BITS_OFFSET))
> +#define PACKED_WRITE_AVAIL_FLAG_WRAP (PACKED_AVAIL_FLAG_WRAP | \
> + ((0ULL | VRING_DESC_F_WRITE) << FLAGS_BITS_OFFSET))
>
> #define DESC_FLAGS_POS 0xaa
> #define MBUF_LENS_POS 0x6666
> +#define DESC_LENS_POS 0x4444
> +#define DESC_LENS_FLAGS_POS 0xB0B0B0B0
>
> int
> vhost_reserve_avail_batch_packed_avx(struct virtio_net *dev,
> @@ -179,3 +185,154 @@ vhost_reserve_avail_batch_packed_avx(struct virtio_net
> *dev,
>
> return -1;
> }
> +
> +int
> +virtio_dev_rx_batch_packed_avx(struct virtio_net *dev,
> + struct vhost_virtqueue *vq,
> + struct rte_mbuf **pkts)
> +{
> + struct vring_packed_desc *descs = vq->desc_packed;
> + uint16_t avail_idx = vq->last_avail_idx;
> + uint64_t desc_addrs[PACKED_BATCH_SIZE];
> + uint32_t buf_offset = dev->vhost_hlen;
> + uint32_t desc_status;
> + uint64_t lens[PACKED_BATCH_SIZE];
> + uint16_t i;
> + void *desc_addr;
> + uint8_t cmp_low, cmp_high, cmp_result;
> +
> + if (unlikely(avail_idx & PACKED_BATCH_MASK))
> + return -1;
Same comment as for patch 4. Packed ring size may not be a pow2.
> + /* check refcnt and nb_segs */
> + __m256i mbuf_ref = _mm256_set1_epi64x(DEFAULT_REARM_DATA);
> +
> + /* load four mbufs rearm data */
> + __m256i mbufs = _mm256_set_epi64x(
> + *pkts[3]->rearm_data,
> + *pkts[2]->rearm_data,
> + *pkts[1]->rearm_data,
> + *pkts[0]->rearm_data);
> +
> + uint16_t cmp = _mm256_cmpneq_epu16_mask(mbufs, mbuf_ref);
> + if (cmp & MBUF_LENS_POS)
> + return -1;
> +
> + /* check desc status */
> + desc_addr = &vq->desc_packed[avail_idx];
> + __m512i desc_vec = _mm512_loadu_si512(desc_addr);
> +
> + __m512i avail_flag_vec;
> + __m512i used_flag_vec;
> + if (vq->avail_wrap_counter) {
> +#if defined(RTE_ARCH_I686)
Is supporting AVX512 on i686 really useful/necessary?
> + avail_flag_vec = _mm512_set4_epi64(PACKED_WRITE_AVAIL_FLAG,
> + 0x0, PACKED_WRITE_AVAIL_FLAG, 0x0);
> + used_flag_vec = _mm512_set4_epi64(PACKED_FLAGS_MASK, 0x0,
> + PACKED_FLAGS_MASK, 0x0);
> +#else
> + avail_flag_vec = _mm512_maskz_set1_epi64(DESC_FLAGS_POS,
> + PACKED_WRITE_AVAIL_FLAG);
> + used_flag_vec = _mm512_maskz_set1_epi64(DESC_FLAGS_POS,
> + PACKED_FLAGS_MASK);
> +#endif
> + } else {
> +#if defined(RTE_ARCH_I686)
> + avail_flag_vec = _mm512_set4_epi64(
> + PACKED_WRITE_AVAIL_FLAG_WRAP, 0x0,
> + PACKED_WRITE_AVAIL_FLAG, 0x0);
> + used_flag_vec = _mm512_set4_epi64(0x0, 0x0, 0x0, 0x0);
> +#else
> + avail_flag_vec = _mm512_maskz_set1_epi64(DESC_FLAGS_POS,
> + PACKED_WRITE_AVAIL_FLAG_WRAP);
> + used_flag_vec = _mm512_setzero_epi32();
> +#endif
> + }
> +
> + desc_status = _mm512_mask_cmp_epu16_mask(BATCH_FLAGS_MASK, desc_vec,
> + avail_flag_vec, _MM_CMPINT_NE);
> + if (desc_status)
> + return -1;
> +
> + if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)) {
> + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
> + uint64_t size = (uint64_t)descs[avail_idx + i].len;
> + desc_addrs[i] = __vhost_iova_to_vva(dev, vq,
> + descs[avail_idx + i].addr, &size,
> + VHOST_ACCESS_RW);
> +
> + if (!desc_addrs[i])
> + return -1;
> +
> + rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[i], void *,
> + 0));
> + }
> + } else {
> + /* check buffer fit into one region & translate address */
> + __m512i regions_low_addrs =
> + _mm512_loadu_si512((void *)&dev->regions_low_addrs);
> + __m512i regions_high_addrs =
> + _mm512_loadu_si512((void *)&dev->regions_high_addrs);
> + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
> + uint64_t addr_low = descs[avail_idx + i].addr;
> + uint64_t addr_high = addr_low +
> + descs[avail_idx + i].len;
> + __m512i low_addr_vec = _mm512_set1_epi64(addr_low);
> + __m512i high_addr_vec = _mm512_set1_epi64(addr_high);
> +
> + cmp_low = _mm512_cmp_epi64_mask(low_addr_vec,
> + regions_low_addrs, _MM_CMPINT_NLT);
> + cmp_high = _mm512_cmp_epi64_mask(high_addr_vec,
> + regions_high_addrs, _MM_CMPINT_LT);
> + cmp_result = cmp_low & cmp_high;
> + int index = __builtin_ctz(cmp_result);
> + if (unlikely((uint32_t)index >= dev->mem->nregions))
> + return -1;
> +
> + desc_addrs[i] = addr_low +
> + dev->mem->regions[index].host_user_addr -
> + dev->mem->regions[index].guest_phys_addr;
> + rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[i], void *,
> + 0));
> + }
> + }
> +
> + /* check length is enough */
> + __m512i pkt_lens = _mm512_set_epi32(
> + 0, pkts[3]->pkt_len, 0, 0,
> + 0, pkts[2]->pkt_len, 0, 0,
> + 0, pkts[1]->pkt_len, 0, 0,
> + 0, pkts[0]->pkt_len, 0, 0);
> +
> + __m512i mbuf_len_offset = _mm512_maskz_set1_epi32(DESC_LENS_POS,
> + dev->vhost_hlen);
> + __m512i buf_len_vec = _mm512_add_epi32(pkt_lens, mbuf_len_offset);
> + uint16_t lens_cmp = _mm512_mask_cmp_epu32_mask(DESC_LENS_POS,
> + desc_vec, buf_len_vec, _MM_CMPINT_LT);
> + if (lens_cmp)
> + return -1;
> +
> + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
> + rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
> + rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
> + pkts[i]->pkt_len);
> + }
> +
> + if (unlikely((dev->features & (1ULL << VHOST_F_LOG_ALL)))) {
> + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
> + lens[i] = descs[avail_idx + i].len;
> + vhost_log_cache_write_iova(dev, vq,
> + descs[avail_idx + i].addr, lens[i]);
> + }
> + }
> +
> + vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
> + vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
> + /* save len and flags, skip addr and id */
> + __m512i desc_updated = _mm512_mask_add_epi16(desc_vec,
> + DESC_LENS_FLAGS_POS, buf_len_vec,
> + used_flag_vec);
> + _mm512_storeu_si512(desc_addr, desc_updated);
> +
> + return 0;
> +}
> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> index e4d2e2e7d6..5c56a8d6ff 100644
> --- a/lib/librte_vhost/virtio_net.c
> +++ b/lib/librte_vhost/virtio_net.c
> @@ -1354,6 +1354,21 @@ virtio_dev_rx_single_packed(struct virtio_net *dev,
> return 0;
> }
>
> +static __rte_always_inline int
> +virtio_dev_rx_handle_batch_packed(struct virtio_net *dev,
> + struct vhost_virtqueue *vq,
> + struct rte_mbuf **pkts)
> +
> +{
> + if (unlikely(dev->vectorized))
> +#ifdef CC_AVX512_SUPPORT
> + return virtio_dev_rx_batch_packed_avx(dev, vq, pkts);
> +#else
> + return virtio_dev_rx_batch_packed(dev, vq, pkts);
> +#endif
> + return virtio_dev_rx_batch_packed(dev, vq, pkts);
It should be as below to not have any performance impact when
CC_AVX512_SUPPORT is not set:
#ifdef CC_AVX512_SUPPORT
if (unlikely(dev->vectorized))
return virtio_dev_rx_batch_packed_avx(dev, vq, pkts);
#else
return virtio_dev_rx_batch_packed(dev, vq, pkts);
#endif
> +}
> +
> static __rte_noinline uint32_t
> virtio_dev_rx_packed(struct virtio_net *dev,
> struct vhost_virtqueue *__rte_restrict vq,
> @@ -1367,8 +1382,8 @@ virtio_dev_rx_packed(struct virtio_net *dev,
> rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
>
> if (remained >= PACKED_BATCH_SIZE) {
> - if (!virtio_dev_rx_batch_packed(dev, vq,
> - &pkts[pkt_idx])) {
> + if (!virtio_dev_rx_handle_batch_packed(dev, vq,
> + &pkts[pkt_idx])) {
> pkt_idx += PACKED_BATCH_SIZE;
> remained -= PACKED_BATCH_SIZE;
> continue;
>
