[ovs-dev] [PATCH v5 06/12] dpif-avx512: Add ISA implementation of dpif

Harry van Haaren harry.van.haaren at intel.com
Thu Dec 3 19:37:41 UTC 2020


This commit adds the AVX512 implementation of DPIF functionality,
specifically the dp_netdev_input_outer_avx512 function. This function
only handles outer (no re-circulations), and is optimized to use the
AVX512 ISA for packet batching and other DPIF work.

Sparse is not able to handle the AVX512 intrinsics, causing compile
time failures, so it is disabled for this file.

Signed-off-by: Harry van Haaren <harry.van.haaren at intel.com>
Co-authored-by: Cian Ferriter <cian.ferriter at intel.com>
Signed-off-by: Cian Ferriter <cian.ferriter at intel.com>

---

v2:
- Add BMI flag for _blsr_u64() instruction
- Add sparse avoidance to fix compilation with --enable-sparse

v3:
- Fix minor grammar mistake
- Fix lines which are too long
- Simplify how *keys is used in AVX512 dp_netdev_input()
- Fix C++ style comments
- Fix spacing after in some for statements
- Remove TODO which gets taken care of in next commit
- Move dpif_netdev_packet_get_rss_hash_orig_pkt() in refactor commit
  rather than here.
- Remove memset changes in fast_path_processing() and dpcls_lookup()
- Fix count_1bits() of miniflow bits

v4:
- Bitmask based processing EMC (and SMC in a later patch)

v5:
- Document TODO for key->hash better, refactor key->len to share function
---
 lib/automake.mk                  |   5 +-
 lib/dpif-netdev-avx512.c         | 238 +++++++++++++++++++++++++++++++
 lib/dpif-netdev-private-dpif.h   |  32 +++++
 lib/dpif-netdev-private-thread.h |   9 +-
 lib/dpif-netdev-private.h        |  25 ++++
 lib/dpif-netdev.c                |  42 ++++--
 6 files changed, 335 insertions(+), 16 deletions(-)
 create mode 100644 lib/dpif-netdev-avx512.c
 create mode 100644 lib/dpif-netdev-private-dpif.h

diff --git a/lib/automake.mk b/lib/automake.mk
index eccfaf3e3..650207940 100644
--- a/lib/automake.mk
+++ b/lib/automake.mk
@@ -33,11 +33,13 @@ lib_libopenvswitchavx512_la_CFLAGS = \
 	-mavx512f \
 	-mavx512bw \
 	-mavx512dq \
+	-mbmi \
 	-mbmi2 \
 	-fPIC \
 	$(AM_CFLAGS)
 lib_libopenvswitchavx512_la_SOURCES = \
-	lib/dpif-netdev-lookup-avx512-gather.c
+	lib/dpif-netdev-lookup-avx512-gather.c \
+	lib/dpif-netdev-avx512.c
 lib_libopenvswitchavx512_la_LDFLAGS = \
 	-static
 endif
@@ -113,6 +115,7 @@ lib_libopenvswitch_la_SOURCES = \
 	lib/dpif-netdev.h \
 	lib/dpif-netdev-private-dfc.h \
 	lib/dpif-netdev-private-dpcls.h \
+	lib/dpif-netdev-private-dpif.h \
 	lib/dpif-netdev-private-flow.h \
 	lib/dpif-netdev-private-hwol.h \
 	lib/dpif-netdev-private-thread.h \
diff --git a/lib/dpif-netdev-avx512.c b/lib/dpif-netdev-avx512.c
new file mode 100644
index 000000000..5bf6cf073
--- /dev/null
+++ b/lib/dpif-netdev-avx512.c
@@ -0,0 +1,238 @@
+/*
+ * Copyright (c) 2020 Intel.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifdef __x86_64__
+/* Sparse cannot handle the AVX512 instructions */
+#if !defined(__CHECKER__)
+
+#include <config.h>
+
+#include "dpif-netdev.h"
+#include "dpif-netdev-perf.h"
+
+#include "dpif-netdev-private.h"
+#include "dpif-netdev-private-dpcls.h"
+#include "dpif-netdev-private-flow.h"
+#include "dpif-netdev-private-thread.h"
+
+#include "dp-packet.h"
+#include "netdev.h"
+
+#include "immintrin.h"
+
+
+/* Structure to contain per-packet metadata that must be attributed to the
+ * dp netdev flow. This is unfortunate to have to track per packet, however
+ * it's a bit awkward to maintain them in a performant way. This structure
+ * helps to keep two variables on a single cache line per packet.
+ */
+struct pkt_flow_meta {
+    uint16_t bytes;
+    uint16_t tcp_flags;
+};
+
+int32_t
+dp_netdev_input_outer_avx512(struct dp_netdev_pmd_thread *pmd,
+                             struct dp_packet_batch *packets,
+                             odp_port_t in_port)
+{
+    OVS_ALIGNED_VAR(CACHE_LINE_SIZE)
+        struct netdev_flow_key keys[NETDEV_MAX_BURST];
+    OVS_ALIGNED_VAR(CACHE_LINE_SIZE)
+        struct netdev_flow_key *key_ptrs[NETDEV_MAX_BURST];
+    OVS_ALIGNED_VAR(CACHE_LINE_SIZE)
+        struct pkt_flow_meta pkt_meta[NETDEV_MAX_BURST];
+
+    /* Stores the computed output: a rule pointer for each packet */
+    /* The AVX512 DPIF implementation handles rules in a way that is optimized
+     * for reducing data-movement between HWOL/EMC/SMC and DPCLS. This is
+     * achieved by separating the rule arrays. Bitmasks are kept for each
+     * packet, indicating if it matched in the HWOL/EMC/SMC array or DPCLS
+     * array. Later the two arrays are merged by AVX-512 expand instructions.
+     */
+    struct dpcls_rule *rules[NETDEV_MAX_BURST];
+    struct dpcls_rule *dpcls_rules[NETDEV_MAX_BURST];
+    uint32_t dpcls_key_idx = 0;
+
+    for (uint32_t i = 0; i < NETDEV_MAX_BURST; i += 8) {
+        _mm512_storeu_si512(&rules[i], _mm512_setzero_si512());
+        _mm512_storeu_si512(&dpcls_rules[i], _mm512_setzero_si512());
+    }
+
+    /* Prefetch each packet's metadata */
+    const size_t batch_size = dp_packet_batch_size(packets);
+    for (int i = 0; i < batch_size; i++) {
+        struct dp_packet *packet = packets->packets[i];
+        OVS_PREFETCH(dp_packet_data(packet));
+        pkt_metadata_prefetch_init(&packet->md);
+    }
+
+    /* Check if EMC or SMC are enabled */
+    struct dfc_cache *cache = &pmd->flow_cache;
+    const uint32_t emc_enabled = pmd->ctx.emc_insert_min != 0;
+    uint32_t emc_hits = 0;
+
+    /* a 1 bit in this mask indidcates a hit, so no DPCLS lookup on the pkt. */
+    uint32_t hwol_emc_smc_hitmask = 0;
+
+    /* Perform first packet interation */
+    uint32_t lookup_pkts_bitmask = (1ULL << batch_size) - 1;
+    uint32_t iter = lookup_pkts_bitmask;
+    while (iter) {
+        uint32_t i = __builtin_ctz(iter);
+        iter = _blsr_u64(iter);
+
+        /* Initialize packet md and do miniflow extract */
+        struct dp_packet *packet = packets->packets[i];
+        pkt_metadata_init(&packet->md, in_port);
+        struct netdev_flow_key *key = &keys[i];
+        miniflow_extract(packet, &key->mf);
+
+        /* Cache TCP and byte values for all packets */
+        pkt_meta[i].bytes = dp_packet_size(packet);
+        pkt_meta[i].tcp_flags = miniflow_get_tcp_flags(&key->mf);
+
+        key->len = netdev_flow_key_size(miniflow_n_values(&key->mf));
+
+        /* TODO: This hash is only required for EMC/SMC usage, DPCLS does not
+         * require it. Deferring its calculation may help performance in some
+         * configurations.
+         */
+        key->hash = dpif_netdev_packet_get_rss_hash_orig_pkt(packet, &key->mf);
+
+        if (emc_enabled) {
+            struct dp_netdev_flow *f = emc_lookup(&cache->emc_cache, key);
+            if (f) {
+                rules[i] = &f->cr;
+                emc_hits++;
+                hwol_emc_smc_hitmask |= (1 << i);
+                continue;
+            }
+        };
+
+        /* The flow pointer was not found in HWOL/EMC/SMC, so add it to the
+         * dpcls input keys array for batch lookup later.
+         */
+        key_ptrs[dpcls_key_idx] = &keys[i];
+        dpcls_key_idx++;
+    }
+
+
+    /* DPCLS handles any packets missed by HWOL/EMC/SMC. It operates on the
+     * key_ptrs[] for input miniflows to match, storing results in the
+     * dpcls_rules[] array.
+     */
+    if (dpcls_key_idx > 0) {
+        struct dpcls *cls = dp_netdev_pmd_lookup_dpcls(pmd, in_port);
+        if (OVS_UNLIKELY(!cls)) {
+            return -1;
+        }
+        int any_miss = !dpcls_lookup(cls,
+                                    (const struct netdev_flow_key **) key_ptrs,
+                                    dpcls_rules, dpcls_key_idx, NULL);
+        if (OVS_UNLIKELY(any_miss)) {
+            return -1;
+        }
+
+        /* Merge DPCLS rules and HWOL/EMC/SMC rules. */
+        uint32_t dpcls_idx = 0;
+        for (int i = 0; i < NETDEV_MAX_BURST; i += 8) {
+            /* Indexing here is somewhat complicated due to DPCLS output rule
+             * load index depending on the hitmask of HWOL/EMC/SMC. More
+             * packets from HWOL/EMC/SMC bitmask means less DPCLS rules are
+             * used.
+             */
+            __m512i v_cache_rules = _mm512_loadu_si512(&rules[i]);
+            __m512i v_merged_rules =
+                        _mm512_mask_expandloadu_epi64(v_cache_rules,
+                                                      ~hwol_emc_smc_hitmask,
+                                                      &dpcls_rules[dpcls_idx]);
+            _mm512_storeu_si512(&rules[i], v_merged_rules);
+
+            /* Update DPCLS load index and bitmask for HWOL/EMC/SMC hits.
+             * There are 8 output pointer per register, subtract the
+             * HWOL/EMC/SMC lanes equals the number of DPCLS rules consumed.
+             */
+            uint32_t hitmask_FF = (hwol_emc_smc_hitmask & 0xFF);
+            dpcls_idx += 8 - __builtin_popcountll(hitmask_FF);
+            hwol_emc_smc_hitmask = (hwol_emc_smc_hitmask >> 8);
+        }
+    }
+
+    /* At this point we don't return error anymore, so commit stats here. */
+    pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_RECV, batch_size);
+    pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_EXACT_HIT, emc_hits);
+    pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MASKED_HIT,
+                            dpcls_key_idx);
+    pmd_perf_update_counter(&pmd->perf_stats, PMD_STAT_MASKED_LOOKUP,
+                            dpcls_key_idx);
+
+    /* Initialize the "Action Batch" for each flow handled below. */
+    struct dp_packet_batch action_batch;
+    action_batch.trunc = 0;
+    action_batch.do_not_steal = false;
+
+    while (lookup_pkts_bitmask) {
+        uint32_t rule_pkt_idx = __builtin_ctz(lookup_pkts_bitmask);
+        uint64_t needle = (uintptr_t) rules[rule_pkt_idx];
+
+        /* Parallel compare 8 flow* 's to the needle, create a bitmask. */
+        __mmask32 batch_bitmask = 0;
+        for (uint32_t j = 0; j < NETDEV_MAX_BURST; j += 8) {
+            /* Pre-calculate store addr */
+            uint32_t num_pkts_in_batch = __builtin_popcountll(batch_bitmask);
+            void *store_addr = &action_batch.packets[num_pkts_in_batch];
+
+            /* Search for identical flow* in burst, update bitmask. */
+            __m512i v_needle = _mm512_maskz_set1_epi64(-1, needle);
+            __m512i v_hay = _mm512_loadu_si512(&rules[j]);
+            uint16_t cmp_bits = _mm512_cmpeq_epi64_mask(v_needle, v_hay);
+            batch_bitmask |= cmp_bits << j;
+
+            /* Compress & Store the batched packets. */
+            struct dp_packet **packets_ptrs = &packets->packets[j];
+            __m512i v_pkt_ptrs = _mm512_loadu_si512(packets_ptrs);
+            _mm512_mask_compressstoreu_epi64(store_addr, cmp_bits, v_pkt_ptrs);
+        }
+
+        /* Strip all packets in this batch from the lookup_pkts_bitmask. */
+        lookup_pkts_bitmask &= (~batch_bitmask);
+        action_batch.count = __builtin_popcountll(batch_bitmask);
+
+        /* Loop over all packets in this batch, to gather the byte and tcp_flag
+         * values, and pass them to the execute function. It would be nice to
+         * optimize this away, however it is not easy to refactor in dpif.
+         */
+        uint32_t bytes = 0;
+        uint16_t tcp_flags = 0;
+        uint32_t bitmask_iter = batch_bitmask;
+        for (int i = 0; i < action_batch.count; i++) {
+            uint32_t idx = __builtin_ctzll(bitmask_iter);
+            bitmask_iter = _blsr_u64(bitmask_iter);
+
+            bytes += pkt_meta[idx].bytes;
+            tcp_flags |= pkt_meta[idx].tcp_flags;
+        }
+
+        dp_netdev_batch_execute(pmd, &action_batch, rules[rule_pkt_idx],
+                                bytes, tcp_flags);
+    }
+
+    return 0;
+}
+
+#endif
+#endif
diff --git a/lib/dpif-netdev-private-dpif.h b/lib/dpif-netdev-private-dpif.h
new file mode 100644
index 000000000..ae9068458
--- /dev/null
+++ b/lib/dpif-netdev-private-dpif.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2020 Intel Corperation.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef DPIF_NETDEV_PRIVATE_DPIF_H
+#define DPIF_NETDEV_PRIVATE_DPIF_H 1
+
+#include "openvswitch/types.h"
+
+/* Forward declarations to avoid including files */
+struct dp_netdev_pmd_thread;
+struct dp_packet_batch;
+
+/* Available implementations for dpif work */
+int32_t
+dp_netdev_input_outer_avx512(struct dp_netdev_pmd_thread *pmd,
+                             struct dp_packet_batch *packets,
+                             odp_port_t in_port);
+
+#endif /* netdev-private.h */
diff --git a/lib/dpif-netdev-private-thread.h b/lib/dpif-netdev-private-thread.h
index ecd6aa977..18387b81d 100644
--- a/lib/dpif-netdev-private-thread.h
+++ b/lib/dpif-netdev-private-thread.h
@@ -50,9 +50,12 @@ struct dp_netdev_pmd_thread_ctx {
 /* Forward declaration for typedef */
 struct dp_netdev_pmd_thread;
 
-typedef void (*dp_netdev_input_func)(struct dp_netdev_pmd_thread *pmd,
-                                     struct dp_packet_batch *packets,
-                                     odp_port_t port_no);
+/* Typedef for DPIF functions.
+ * Returns a bitmask of packets to handle, possibly including upcall/misses.
+ */
+typedef int32_t (*dp_netdev_input_func)(struct dp_netdev_pmd_thread *pmd,
+                                        struct dp_packet_batch *packets,
+                                        odp_port_t port_no);
 
 /* PMD: Poll modes drivers.  PMD accesses devices via polling to eliminate
  * the performance overhead of interrupt processing.  Therefore netdev can
diff --git a/lib/dpif-netdev-private.h b/lib/dpif-netdev-private.h
index d7b6fd7ec..9c2237fac 100644
--- a/lib/dpif-netdev-private.h
+++ b/lib/dpif-netdev-private.h
@@ -31,4 +31,29 @@
 #include "dpif-netdev-private-dfc.h"
 #include "dpif-netdev-private-thread.h"
 
+/* Allow other implementations to lookup the DPCLS instances */
+struct dpcls *
+dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread *pmd,
+                           odp_port_t in_port);
+
+/* Allow other implementations to call dpcls_lookup() for subtable search */
+bool
+dpcls_lookup(struct dpcls *cls, const struct netdev_flow_key *keys[],
+             struct dpcls_rule **rules, const size_t cnt,
+             int *num_lookups_p);
+
+/* Allow other implementations to execute actions on a batch */
+void
+dp_netdev_batch_execute(struct dp_netdev_pmd_thread *pmd,
+                        struct dp_packet_batch *packets,
+                        struct dpcls_rule *rule,
+                        uint32_t bytes,
+                        uint16_t tcp_flags);
+
+/* Available implementations for dpif work */
+int32_t
+dp_netdev_input_outer_avx512(struct dp_netdev_pmd_thread *pmd,
+                             struct dp_packet_batch *packets,
+                             odp_port_t in_port);
+
 #endif /* netdev-private.h */
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 6c573a188..e8d47e539 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -185,10 +185,6 @@ static uint32_t dpcls_subtable_lookup_reprobe(struct dpcls *cls);
 static void dpcls_insert(struct dpcls *, struct dpcls_rule *,
                          const struct netdev_flow_key *mask);
 static void dpcls_remove(struct dpcls *, struct dpcls_rule *);
-static bool dpcls_lookup(struct dpcls *cls,
-                         const struct netdev_flow_key *keys[],
-                         struct dpcls_rule **rules, size_t cnt,
-                         int *num_lookups_p);
 
 /* Set of supported meter flags */
 #define DP_SUPPORTED_METER_FLAGS_MASK \
@@ -482,7 +478,7 @@ static void dp_netdev_execute_actions(struct dp_netdev_pmd_thread *pmd,
                                       const struct flow *flow,
                                       const struct nlattr *actions,
                                       size_t actions_len);
-static void dp_netdev_input(struct dp_netdev_pmd_thread *,
+static int32_t dp_netdev_input(struct dp_netdev_pmd_thread *,
                             struct dp_packet_batch *, odp_port_t port_no);
 static void dp_netdev_recirculate(struct dp_netdev_pmd_thread *,
                                   struct dp_packet_batch *);
@@ -554,7 +550,7 @@ dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd,
                                bool purge);
 static int dpif_netdev_xps_get_tx_qid(const struct dp_netdev_pmd_thread *pmd,
                                       struct tx_port *tx);
-static inline struct dpcls *
+inline struct dpcls *
 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread *pmd,
                            odp_port_t in_port);
 
@@ -1917,7 +1913,7 @@ void dp_netdev_flow_unref(struct dp_netdev_flow *flow)
     }
 }
 
-static inline struct dpcls *
+inline struct dpcls *
 dp_netdev_pmd_lookup_dpcls(struct dp_netdev_pmd_thread *pmd,
                            odp_port_t in_port)
 {
@@ -2717,7 +2713,7 @@ dp_netdev_pmd_lookup_flow(struct dp_netdev_pmd_thread *pmd,
                           int *lookup_num_p)
 {
     struct dpcls *cls;
-    struct dpcls_rule *rule;
+    struct dpcls_rule *rule = NULL;
     odp_port_t in_port = u32_to_odp(MINIFLOW_GET_U32(&key->mf,
                                                      in_port.odp_port));
     struct dp_netdev_flow *netdev_flow = NULL;
@@ -4179,7 +4175,10 @@ dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
         }
 
         /* Process packet batch. */
-        pmd->netdev_input_func(pmd, &batch, port_no);
+        int32_t ret = pmd->netdev_input_func(pmd, &batch, port_no);
+        if (ret) {
+            dp_netdev_input(pmd, &batch, port_no);
+        }
 
         /* Assign processing cycles to rx queue. */
         cycles = cycle_timer_stop(&pmd->perf_stats, &timer);
@@ -6421,6 +6420,24 @@ packet_batch_per_flow_execute(struct packet_batch_per_flow *batch,
                               actions->actions, actions->size);
 }
 
+void
+dp_netdev_batch_execute(struct dp_netdev_pmd_thread *pmd,
+                        struct dp_packet_batch *packets,
+                        struct dpcls_rule *rule,
+                        uint32_t bytes,
+                        uint16_t tcp_flags)
+{
+    /* Gets action* from the rule */
+    struct dp_netdev_flow *flow = dp_netdev_flow_cast(rule);
+    struct dp_netdev_actions *actions = dp_netdev_flow_get_actions(flow);
+
+    dp_netdev_flow_used(flow, dp_packet_batch_size(packets), bytes,
+                        tcp_flags, pmd->ctx.now / 1000);
+    const uint32_t steal = 1;
+    dp_netdev_execute_actions(pmd, packets, steal, &flow->flow,
+                              actions->actions, actions->size);
+}
+
 static inline void
 dp_netdev_queue_batches(struct dp_packet *pkt,
                         struct dp_netdev_flow *flow, uint16_t tcp_flags,
@@ -6930,12 +6947,13 @@ dp_netdev_input__(struct dp_netdev_pmd_thread *pmd,
     }
 }
 
-static void
+static int32_t
 dp_netdev_input(struct dp_netdev_pmd_thread *pmd,
                 struct dp_packet_batch *packets,
                 odp_port_t port_no)
 {
     dp_netdev_input__(pmd, packets, false, port_no);
+    return 0;
 }
 
 static void
@@ -8300,7 +8318,7 @@ netdev_flow_key_gen_masks(const struct netdev_flow_key *tbl,
 
 /* Returns true if 'target' satisfies 'key' in 'mask', that is, if each 1-bit
  * in 'mask' the values in 'key' and 'target' are the same. */
-bool
+inline bool ALWAYS_INLINE
 dpcls_rule_matches_key(const struct dpcls_rule *rule,
                        const struct netdev_flow_key *target)
 {
@@ -8326,7 +8344,7 @@ dpcls_rule_matches_key(const struct dpcls_rule *rule,
  * priorities, instead returning any rule which matches the flow.
  *
  * Returns true if all miniflows found a corresponding rule. */
-static bool
+bool
 dpcls_lookup(struct dpcls *cls, const struct netdev_flow_key *keys[],
              struct dpcls_rule **rules, const size_t cnt,
              int *num_lookups_p)
-- 
2.25.1



More information about the dev mailing list