diff --git a/hosts/skydick/datapool.nix b/hosts/skydick/datapool.nix
index d0cd745..cdc06c9 100644
--- a/hosts/skydick/datapool.nix
+++ b/hosts/skydick/datapool.nix
@@ -393,7 +393,7 @@
   services.rpcbind.enable = true;
 
   # NFS-over-RDMA listener (additive to TCP — clients pick one transport
-  # per mount). skydick's ConnectX-5 / mlx5_bond_0 RoCEv2 fabric is the
+  # per mount). skydick's ConnectX-3 Pro / mlx4_0 (bond40g, enp130s0) RoCEv2 fabric is the
   # transport; nfsd binds to UDP/TCP 2049 by default plus RDMA 20049
   # when this oneshot writes to /proc/fs/nfsd/portlist. Clients use
   # the RDMA path with `proto=rdma,port=20049` mount options. Saves a
@@ -401,6 +401,43 @@
   # equivalent needed — RDMA QPs are intrinsically parallel).
   boot.kernelModules = [ "rpcrdma" ];
 
+  # RoCEv2 QoS — DSCP-marking half (decision "C", 2026-06-25). Stamp ToS 96 =
+  # DSCP 24 on RDMA-CM connections so the switch + the mlx5 RDMA clients classify
+  # skydick's RoCE egress into priority 3 (the PFC/lossless queue). Only affects
+  # NEW rdma_cm connections, and must be set before nfsd accepts client QPs.
+  # MARKING ONLY: end-to-end lossless also needs PFC on prio 3 at every hop;
+  # skydick's ConnectX-3 (mlx4) does PCP-based PFC with no upstream trust-dscp, so
+  # the PFC half is deferred (move RoCE to the idle ConnectX-4-LX/mlx5 for a clean
+  # DSCP/PFC fabric, or mlx4 pfctx/pfcrx — open decision A/B).
+  systemd.services.roce-rdma-cm-tos = {
+    description = "RoCEv2 RDMA-CM default ToS = 96 (DSCP 24) for prio-3 QoS";
+    requires = [ "sys-kernel-config.mount" ];
+    after = [ "sys-kernel-config.mount" "systemd-modules-load.service" ];
+    before = [ "nfsd-rdma-listener.service" "nfs-server.service" ];
+    wantedBy = [ "multi-user.target" ];
+    path = [ pkgs.coreutils ];
+    serviceConfig = {
+      Type = "oneshot";
+      RemainAfterExit = true;
+      ExecStartPre = "${pkgs.kmod}/bin/modprobe rdma_cm";
+    };
+    # Best-effort over all present RDMA devices (mlx4_0 now; covers mlx5 if option
+    # B ever brings it up). Always exits 0 so a transient device state can't trip
+    # deploy-rs magic-rollback.
+    script = ''
+      for d in /sys/class/infiniband/*; do
+        [ -e "$d" ] || continue
+        dev=$(basename "$d")
+        mkdir -p /sys/kernel/config/rdma_cm/"$dev" 2>/dev/null || continue
+        for tos in /sys/kernel/config/rdma_cm/"$dev"/ports/*/default_roce_tos; do
+          [ -e "$tos" ] || continue
+          echo 96 > "$tos" 2>/dev/null && echo "set $tos = 96"
+        done
+      done
+      exit 0
+    '';
+  };
+
   systemd.services.nfsd-rdma-listener = {
     description = "Enable NFS-over-RDMA listener on port 20049";
     # `after` orders the unit at boot.