> -----Original Message-----
> From: Maxime Coquelin <maxime.coque...@redhat.com>
> Sent: Tuesday, October 6, 2020 10:59 PM
> To: Liu, Yong <yong....@intel.com>; Xia, Chenbo <chenbo....@intel.com>;
> Wang, Zhihong <zhihong.w...@intel.com>
> Cc: dev@dpdk.org
> Subject: Re: [PATCH v2 4/5] vhost: add packed ring vectorized dequeue
>
>
>
> On 9/21/20 8:48 AM, Marvin Liu wrote:
> > Optimize vhost packed ring dequeue path with SIMD instructions. Four
> > descriptors status check and writeback are batched handled with AVX512
> > instructions. Address translation operations are also accelerated by
> > AVX512 instructions.
> >
> > If platform or compiler not support vectorization, will fallback to
> > default path.
> >
> > Signed-off-by: Marvin Liu <yong....@intel.com>
> >
> > diff --git a/lib/librte_vhost/meson.build b/lib/librte_vhost/meson.build
> > index cc9aa65c67..c1481802d7 100644
> > --- a/lib/librte_vhost/meson.build
> > +++ b/lib/librte_vhost/meson.build
> > @@ -8,6 +8,22 @@ endif
> > if has_libnuma == 1
> > dpdk_conf.set10('RTE_LIBRTE_VHOST_NUMA', true)
> > endif
> > +
> > +if arch_subdir == 'x86'
> > + if not machine_args.contains('-mno-avx512f')
> > + if cc.has_argument('-mavx512f') and cc.has_argument('-
> mavx512vl') and cc.has_argument('-mavx512bw')
> > + cflags += ['-DCC_AVX512_SUPPORT']
> > + vhost_avx512_lib =
> > static_library('vhost_avx512_lib',
> > + 'vhost_vec_avx.c',
> > + dependencies:
> > [static_rte_eal,
> static_rte_mempool,
> > + static_rte_mbuf,
> > static_rte_ethdev,
> static_rte_net],
> > + include_directories:
> > includes,
> > + c_args: [cflags,
> > '-mavx512f', '-mavx512bw', '-
> mavx512vl'])
> > + objs +=
> > vhost_avx512_lib.extract_objects('vhost_vec_avx.c')
> > + endif
> > + endif
> > +endif
>
> Not a Meson expert, but wonder how I can disable CC_AVX512_SUPPORT.
> I checked the DPDK doc, but I could not find how to pass -mno-avx512f to
> the machine_args.
Hi Maxime,
By now mno-avx512f flag will be set only if binutils check script found issue
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90028.
So avx512 code will be built-in if compiler support that. There's alternative
way is that introduce one new option in meson build.
Thanks,
Marvin
>
> > +
> > if (toolchain == 'gcc' and cc.version().version_compare('>=8.3.0'))
> > cflags += '-DVHOST_GCC_UNROLL_PRAGMA'
> > elif (toolchain == 'clang' and cc.version().version_compare('>=3.7.0'))
> > diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
> > index 4a81f18f01..fc7daf2145 100644
> > --- a/lib/librte_vhost/vhost.h
> > +++ b/lib/librte_vhost/vhost.h
> > @@ -1124,4 +1124,12 @@ virtio_dev_pktmbuf_alloc(struct virtio_net
> *dev, struct rte_mempool *mp,
> > return NULL;
> > }
> >
> > +int
> > +vhost_reserve_avail_batch_packed_avx(struct virtio_net *dev,
> > + struct vhost_virtqueue *vq,
> > + struct rte_mempool *mbuf_pool,
> > + struct rte_mbuf **pkts,
> > + uint16_t avail_idx,
> > + uintptr_t *desc_addrs,
> > + uint16_t *ids);
> > #endif /* _VHOST_NET_CDEV_H_ */
> > diff --git a/lib/librte_vhost/vhost_vec_avx.c
> b/lib/librte_vhost/vhost_vec_avx.c
> > new file mode 100644
> > index 0000000000..dc5322d002
> > --- /dev/null
> > +++ b/lib/librte_vhost/vhost_vec_avx.c
>
> For consistency it should be prefixed with virtio_net, not vhost.
>
> > @@ -0,0 +1,181 @@
> > +/* SPDX-License-Identifier: BSD-3-Clause
> > + * Copyright(c) 2010-2016 Intel Corporation
> > + */
> > +#include <stdint.h>
> > +
> > +#include "vhost.h"
> > +
> > +#define BYTE_SIZE 8
> > +/* reference count offset in mbuf rearm data */
> > +#define REFCNT_BITS_OFFSET ((offsetof(struct rte_mbuf, refcnt) - \
> > + offsetof(struct rte_mbuf, rearm_data)) * BYTE_SIZE)
> > +/* segment number offset in mbuf rearm data */
> > +#define SEG_NUM_BITS_OFFSET ((offsetof(struct rte_mbuf, nb_segs) - \
> > + offsetof(struct rte_mbuf, rearm_data)) * BYTE_SIZE)
> > +
> > +/* default rearm data */
> > +#define DEFAULT_REARM_DATA (1ULL << SEG_NUM_BITS_OFFSET | \
> > + 1ULL << REFCNT_BITS_OFFSET)
> > +
> > +#define DESC_FLAGS_SHORT_OFFSET (offsetof(struct vring_packed_desc,
> flags) / \
> > + sizeof(uint16_t))
> > +
> > +#define DESC_FLAGS_SHORT_SIZE (sizeof(struct vring_packed_desc) / \
> > + sizeof(uint16_t))
> > +#define BATCH_FLAGS_MASK (1 << DESC_FLAGS_SHORT_OFFSET | \
> > + 1 << (DESC_FLAGS_SHORT_OFFSET + DESC_FLAGS_SHORT_SIZE) | \
> > + 1 << (DESC_FLAGS_SHORT_OFFSET + DESC_FLAGS_SHORT_SIZE * 2) |
> \
> > + 1 << (DESC_FLAGS_SHORT_OFFSET + DESC_FLAGS_SHORT_SIZE * 3))
> > +
> > +#define FLAGS_BITS_OFFSET ((offsetof(struct vring_packed_desc, flags) - \
> > + offsetof(struct vring_packed_desc, len)) * BYTE_SIZE)
> > +
> > +#define PACKED_FLAGS_MASK ((0ULL | VRING_DESC_F_AVAIL |
> VRING_DESC_F_USED) \
> > + << FLAGS_BITS_OFFSET)
> > +#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 DESC_FLAGS_POS 0xaa
> > +#define MBUF_LENS_POS 0x6666
> > +
> > +int
> > +vhost_reserve_avail_batch_packed_avx(struct virtio_net *dev,
> > + struct vhost_virtqueue *vq,
> > + struct rte_mempool *mbuf_pool,
> > + struct rte_mbuf **pkts,
> > + uint16_t avail_idx,
> > + uintptr_t *desc_addrs,
> > + uint16_t *ids)
> > +{
> > + struct vring_packed_desc *descs = vq->desc_packed;
> > + uint32_t descs_status;
> > + void *desc_addr;
> > + uint16_t i;
> > + uint8_t cmp_low, cmp_high, cmp_result;
> > + uint64_t lens[PACKED_BATCH_SIZE];
> > + struct virtio_net_hdr *hdr;
> > +
> > + if (unlikely(avail_idx & PACKED_BATCH_MASK))
> > + return -1;
> > +
> > + /* load 4 descs */
> > + desc_addr = &vq->desc_packed[avail_idx];
> > + __m512i desc_vec = _mm512_loadu_si512(desc_addr);
>
> Unlike split ring, packed ring specification does not mandate the ring
> size to be a power of two. So checking avail_idx is aligned on 64 bytes
> is not enough given a descriptor is 16 bytes.
>
> You need to also check against ring size to prevent out of bounds
> accesses.
>
> I see the non vectorized batch processing you introduced in v19.11 also
> do that wrong assumption. Please fix it.
>
> Also, I wonder whether it is assumed that &vq->desc_packed[avail_idx];
> is aligned on a cache-line. Meaning, does below intrinsics have such a
> requirement?
>
Got, packed ring size may arbitrary number. In v19.11 batch handling function
has already checked available index not oversized.
I forgot that in vectorized path, will fix it in next release.
In vectorized path, loading function mm512_loadu_si512 do not need
cache-aligned memory. So no special requirement is needed.
> > + /* burst check four status */
> > + __m512i avail_flag_vec;
> > + if (vq->avail_wrap_counter)
> > +#if defined(RTE_ARCH_I686)
> > + avail_flag_vec = _mm512_set4_epi64(PACKED_AVAIL_FLAG,
> 0x0,
> > + PACKED_FLAGS_MASK, 0x0);
> > +#else
> > + avail_flag_vec =
> _mm512_maskz_set1_epi64(DESC_FLAGS_POS,
> > + PACKED_AVAIL_FLAG);
> > +
> > +#endif
> > + else
> > +#if defined(RTE_ARCH_I686)
> > + avail_flag_vec =
> _mm512_set4_epi64(PACKED_AVAIL_FLAG_WRAP,
> > + 0x0, PACKED_AVAIL_FLAG_WRAP,
> 0x0);
> > +#else
> > + avail_flag_vec =
> _mm512_maskz_set1_epi64(DESC_FLAGS_POS,
> > + PACKED_AVAIL_FLAG_WRAP);
> > +#endif
> > +
> > + descs_status = _mm512_cmp_epu16_mask(desc_vec, avail_flag_vec,
> > + _MM_CMPINT_NE);
> > + if (descs_status & BATCH_FLAGS_MASK)
> > + 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_RO);
> > +
> > + if (!desc_addrs[i])
> > + goto free_buf;
> > + lens[i] = descs[avail_idx + i].len;
> > + rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
> > +
> > + pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool,
> > + lens[i]);
> > + if (!pkts[i])
> > + goto free_buf;
> > + }
> > + } 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))
> > + goto free_buf;
> > +
> > + desc_addrs[i] = addr_low +
> > + dev->mem->regions[index].host_user_addr -
> > + dev->mem->regions[index].guest_phys_addr;
> > + lens[i] = descs[avail_idx + i].len;
> > + rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
> > +
> > + pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool,
> > + lens[i]);
> > + if (!pkts[i])
> > + goto free_buf;
> > + }
> > + }
> > +
> > + if (virtio_net_with_host_offload(dev)) {
> > + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
> > + hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
> > + vhost_dequeue_offload(hdr, pkts[i]);
> > + }
> > + }
> > +
> > + if (unlikely(virtio_net_is_inorder(dev))) {
> > + ids[PACKED_BATCH_SIZE - 1] =
> > + descs[avail_idx + PACKED_BATCH_SIZE - 1].id;
>
> Isn't in-order a likely case? Maybe just remove the unlikely.
>
In_order option is depended on feature negotiation , will remove unlikely.
> > + } else {
> > + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
> > + ids[i] = descs[avail_idx + i].id;
> > + }
> > +
> > + uint64_t addrs[PACKED_BATCH_SIZE << 1];
> > + /* store mbuf data_len, pkt_len */
> > + vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
> > + addrs[i << 1] = (uint64_t)pkts[i]->rx_descriptor_fields1;
> > + addrs[(i << 1) + 1] = (uint64_t)pkts[i]->rx_descriptor_fields1
> > + + sizeof(uint64_t);
> > + }
> > +
> > + /* save pkt_len and data_len into mbufs */
> > + __m512i value_vec =
> _mm512_maskz_shuffle_epi32(MBUF_LENS_POS, desc_vec,
> > + 0xAA);
> > + __m512i offsets_vec = _mm512_maskz_set1_epi32(MBUF_LENS_POS,
> > + (uint32_t)-12);
> > + value_vec = _mm512_add_epi32(value_vec, offsets_vec);
> > + __m512i vindex = _mm512_loadu_si512((void *)addrs);
> > + _mm512_i64scatter_epi64(0, vindex, value_vec, 1);
> > +
> > + return 0;
> > +free_buf:
> > + for (i = 0; i < PACKED_BATCH_SIZE; i++)
> > + rte_pktmbuf_free(pkts[i]);
> > +
> > + return -1;
> > +}
> > diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> > index 6107662685..e4d2e2e7d6 100644
> > --- a/lib/librte_vhost/virtio_net.c
> > +++ b/lib/librte_vhost/virtio_net.c
> > @@ -2249,6 +2249,28 @@ vhost_reserve_avail_batch_packed(struct
> virtio_net *dev,
> > return -1;
> > }
> >
> > +static __rte_always_inline int
> > +vhost_handle_avail_batch_packed(struct virtio_net *dev,
> > + struct vhost_virtqueue *vq,
> > + struct rte_mempool *mbuf_pool,
> > + struct rte_mbuf **pkts,
> > + uint16_t avail_idx,
> > + uintptr_t *desc_addrs,
> > + uint16_t *ids)
> > +{
> > + if (unlikely(dev->vectorized))
> > +#ifdef CC_AVX512_SUPPORT
> > + return vhost_reserve_avail_batch_packed_avx(dev, vq,
> mbuf_pool,
> > + pkts, avail_idx, desc_addrs, ids);
> > +#else
> > + return vhost_reserve_avail_batch_packed(dev, vq,
> mbuf_pool,
> > + pkts, avail_idx, desc_addrs, ids);
> > +
> > +#endif
> > + return vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
> > + avail_idx, desc_addrs, ids);
> > +}
>
>
> 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 vhost_reserve_avail_batch_packed_avx(dev, vq,
> mbuf_pool,
> pkts, avail_idx, desc_addrs, ids);
> #else
> return vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
> avail_idx, desc_addrs, ids);
> #endif
Got, will change in next release.
> > +
> > static __rte_always_inline int
> > virtio_dev_tx_batch_packed(struct virtio_net *dev,
> > struct vhost_virtqueue *vq,
> > @@ -2261,8 +2283,9 @@ virtio_dev_tx_batch_packed(struct virtio_net
> *dev,
> > uint16_t ids[PACKED_BATCH_SIZE];
> > uint16_t i;
> >
> > - if (vhost_reserve_avail_batch_packed(dev, vq, mbuf_pool, pkts,
> > - avail_idx, desc_addrs, ids))
> > +
> > + if (vhost_handle_avail_batch_packed(dev, vq, mbuf_pool, pkts,
> > + avail_idx, desc_addrs, ids))
> > return -1;
> >
> > vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
> >
