[ecmp] Deduplicate redundant multicast next hops via per-target ECMP selection

.github/buildomat/jobs/bench.sh

@@ -39,8 +39,19 @@ fi
 
 pfexec /usr/lib/brand/omicron1/baseline -w /var/run/brand/omicron1/baseline
 
+# Resolve the invoking user for ownership restoration. When this script is run
+# elevated locally, the effective `id -un` resolves to root and would chown the
+# tree to root. Prefer `$SUDO_USER` (set by sudo), then `logname`, which reports
+# the login user across pfexec. Fall back to `id` in CI, however.
+run_user=${SUDO_USER:-$(logname 2>/dev/null || id -un)}
+run_group=$(id -gn "$run_user" 2>/dev/null || id -gn)
+
 function cleanup {
-    pfexec chown -R `id -un`:`id -gn` .
+    # A restore-to-owner that resolves to root would defeat its own purpose.
+    # Skip rather than re-root the tree.
+    if [[ $run_user != root ]]; then
+        pfexec chown -R "$run_user":"$run_group" .
+    fi
     if [[ -z $BUILDOMAT_JOB_ID ]]; then
         pfexec rm -rf /input/xde
     fi
@@ -67,7 +78,14 @@ function get_artifact {
     return $curl_res
 }
 
-OUT_DIR=/work/bench-results
+# TGT_BASE allows one to run this more easily in their local
+# environment:
+#
+#   TGT_BASE=/var/tmp ./bench.sh
+#
+TGT_BASE=${TGT_BASE:=/work}
+
+OUT_DIR=$TGT_BASE/bench-results
 
 mkdir -p $OUT_DIR
 mkdir -p target/criterion
@@ -118,6 +136,6 @@ cargo ubench
 cp -r target/criterion $OUT_DIR
 cp -r target/xde-bench $OUT_DIR
 
-pushd /work
+pushd $TGT_BASE
 tar -caf bench-results.tgz bench-results
 popd

.github/buildomat/jobs/opteadm.sh

@@ -41,9 +41,16 @@ ptime -m cargo build --release
 
 popd
 
+# TGT_BASE allows one to run this more easily in their local
+# environment:
+#
+#   TGT_BASE=/var/tmp ./opteadm.sh
+#
+TGT_BASE=${TGT_BASE:=/work}
+
 for x in debug release
 do
-    mkdir -p /work/$x
-    cp target/$x/opteadm /work/$x/
-    sha256sum "target/$x/opteadm" > "/work/$x/opteadm.$x.sha256"
+    mkdir -p $TGT_BASE/$x
+    cp target/$x/opteadm $TGT_BASE/$x/
+    sha256sum "target/$x/opteadm" > "$TGT_BASE/$x/opteadm.$x.sha256"
 done

.github/buildomat/jobs/test.sh

@@ -22,17 +22,36 @@ set -o xtrace
 
 pfexec pkg install brand/omicron1 brand/omicron1/tools opte
 
+# TGT_BASE mirrors the artifact location used by xde.sh. Override it to match a
+# local xde.sh run, e.g. TGT_BASE=/var/tmp ./test.sh, so the test binaries are
+# found without forced writing to the root-owned /work.
+TGT_BASE=${TGT_BASE:=/work}
+
 if [[ -z $BUILDOMAT_JOB_ID ]]; then
     echo Note: if you are running this locally, you must run the xde.sh job first
     echo to have the artifacts at the expected spot.
     pfexec mkdir -p /input/xde
-    pfexec ln -s /work /input/xde/work
+    # Replace any stale symlink from an interrupted prior run so the link is
+    # idempotent across local re-runs.
+    pfexec rm -f /input/xde/work
+    pfexec ln -s $TGT_BASE /input/xde/work
 fi
 
 pfexec /usr/lib/brand/omicron1/baseline -w /var/run/brand/omicron1/baseline
 
+# Resolve the invoking user for ownership restoration. When this script is run
+# elevated locally, the effective `id -un` resolves to root and would chown the
+# tree to root. Prefer `$SUDO_USER` (set by sudo), then `logname`, which reports
+# the login user across pfexec. Fall back to `id` in CI, however.
+run_user=${SUDO_USER:-$(logname 2>/dev/null || id -un)}
+run_group=$(id -gn "$run_user" 2>/dev/null || id -gn)
+
 function cleanup {
-    pfexec chown -R `id -un`:`id -gn` .
+    # A restore-to-owner that resolves to root would defeat its own purpose.
+    # Skip rather than re-root the tree.
+    if [[ $run_user != root ]]; then
+        pfexec chown -R "$run_user":"$run_group" .
+    fi
     if [[ -z $BUILDOMAT_JOB_ID ]]; then
         pfexec rm -rf /input/xde
     fi
@@ -98,6 +117,9 @@ pfexec /input/xde/work/test/multicast_validation --nocapture --test-threads=1
 pfexec chmod +x /input/xde/work/test/multicast_source_filter
 pfexec /input/xde/work/test/multicast_source_filter --nocapture --test-threads=1
 
+pfexec chmod +x /input/xde/work/test/multicast_multi_nexthop
+pfexec /input/xde/work/test/multicast_multi_nexthop --nocapture --test-threads=1
+
 banner "teardown"
 # Ensure full driver teardown is exercised after tests complete
 pfexec rem_drv xde

.github/buildomat/jobs/xde.sh

@@ -18,6 +18,7 @@
 #:   "=/work/test/multicast_multi_sub",
 #:   "=/work/test/multicast_validation",
 #:   "=/work/test/multicast_source_filter",
+#:   "=/work/test/multicast_multi_nexthop",
 #:   "=/work/xde.conf",
 #: ]
 #:
@@ -62,7 +63,7 @@ install_pkg jq
 
 pushd xde
 
-cp xde.conf /work/xde.conf
+cp xde.conf $TGT_BASE/xde.conf
 
 header "check style"
 ptime -m cargo +$NIGHTLY fmt -p xde -p xde-link -- --check
@@ -140,9 +141,15 @@ multicast_source_filter_test=$(
     cargo build -q --test multicast_source_filter --message-format=json |\
     jq -r "select(.profile.test == true) | .filenames[]"
 )
-mkdir -p /work/test
-cp $loopback_test /work/test/loopback
-cp $multicast_rx_test /work/test/multicast_rx
-cp $multicast_multi_sub_test /work/test/multicast_multi_sub
-cp $multicast_validation_test /work/test/multicast_validation
-cp $multicast_source_filter_test /work/test/multicast_source_filter
+cargo build --test multicast_multi_nexthop
+multicast_multi_nexthop_test=$(
+    cargo build -q --test multicast_multi_nexthop --message-format=json |\
+    jq -r "select(.profile.test == true) | .filenames[]"
+)
+mkdir -p $TGT_BASE/test
+cp $loopback_test $TGT_BASE/test/loopback
+cp $multicast_rx_test $TGT_BASE/test/multicast_rx
+cp $multicast_multi_sub_test $TGT_BASE/test/multicast_multi_sub
+cp $multicast_validation_test $TGT_BASE/test/multicast_validation
+cp $multicast_source_filter_test $TGT_BASE/test/multicast_source_filter
+cp $multicast_multi_nexthop_test $TGT_BASE/test/multicast_multi_nexthop

Cargo.toml

@@ -48,6 +48,7 @@ crc32fast = { version = "1", default-features = false }
 criterion = "0.8"
 ctor = "0.10"
 darling = "0.23"
+dlpi = { git = "https://github.com/oxidecomputer/dlpi-sys", default-features = false }
 dyn-clone = "1.0"
 heapless = "0.8"
 ingot = "0.1.1"

dtrace/opte-mcast-delivery.d

@@ -137,6 +137,9 @@ mcast-local-delivery {
 	@by_vni["DELIVER", this->vni] = count();
 	@by_port[this->port] = count();
 	@by_group["DELIVER", this->group_str] = count();
+
+	/* Per-(port,group,vni) delivery matrix for end-to-end fan-out verification. */
+	@deliver_by_port_group[this->port, this->group_str, this->vni] = count();
 }
 
 mcast-local-delivery
@@ -165,6 +168,9 @@ mcast-underlay-fwd {
 	@by_vni["UNDERLAY", this->vni] = count();
 	@by_underlay["UNDERLAY", this->underlay_str] = count();
 	@by_nexthop_unicast[this->next_hop_str] = count();
+
+	/* Per-(group,vni,next-hop) fan-out; cross-checks ddm-peers rear-port count. */
+	@fwd_by_group_nh[this->underlay_str, this->vni, this->next_hop_str] = count();
 }
 
 mcast-underlay-fwd
@@ -421,6 +427,11 @@ mcast-source-filtered {
 	@by_vni["FILTERED", this->vni] = count();
 	@by_port[this->port] = count();
 	@filtered_by_mode[this->mode_str] = count();
+
+	/* Per-(event,scope,group,vni) drops for end-to-end loss attribution.
+	 * The scope names the address space of the group column: overlay for the
+	 * inner multicast group, underlay for the outer delivery address. */
+	@drops["FILTERED", "overlay", this->dst_str, this->vni] = count();
 }
 
 mcast-source-filtered
@@ -454,6 +465,9 @@ mcast-fwd-source-filtered {
 	@by_vni["FWD_FILT", this->vni] = count();
 	@by_nexthop_unicast[this->next_hop_str] = count();
 	@filtered_by_mode[this->mode_str] = count();
+
+	/* Per-(event,scope,group,vni) drops; see mcast-source-filtered. */
+	@drops["FWD_FILT", "overlay", this->dst_str, this->vni] = count();
 }
 
 mcast-fwd-source-filtered
@@ -473,9 +487,13 @@ mcast-no-fwd-entry {
 	/* arg0=underlay_ptr, arg1=vni */
 	this->underlay = (in6_addr_t *)arg0;
 	this->vni = arg1;
+	this->underlay_str = inet_ntoa6(this->underlay);
 
 	/* Always track aggregations */
 	@by_event["NOFWD"] = count();
+
+	/* Per-(event,scope,group,vni) drops; see mcast-source-filtered. */
+	@drops["NOFWD", "underlay", this->underlay_str, this->vni] = count();
 }
 
 mcast-no-fwd-entry
@@ -499,10 +517,16 @@ END
 	printa(@by_underlay);
 	printf("\nLocal delivery by port:\n");
 	printa(@by_port);
+	printf("\nDelivery matrix (port, group, vni):\n");
+	printa(@deliver_by_port_group);
 	printf("\nForwarding by unicast next hop (routing address):\n");
 	printa(@by_nexthop_unicast);
+	printf("\nForwarding fan-out (underlay group, vni, next hop):\n");
+	printa(@fwd_by_group_nh);
 	printf("\nSource filtering by mode:\n");
 	printa(@filtered_by_mode);
+	printf("\nDrops (event, scope, group, vni):\n");
+	printa(@drops);
 	printf("\nConfig ops:\n");
 	printa(@cfg_counts);
 }

xde-tests/Cargo.toml

@@ -12,6 +12,7 @@ opte-test-utils.workspace = true
 oxide-vpc.workspace = true
 
 anyhow.workspace = true
+dlpi.workspace = true
 libnet.workspace = true
 rand.workspace = true
 slog.workspace = true