chore: release 1.3.1

docs/GUEST-SETUP.md

@@ -75,12 +75,6 @@ Enable-NetFirewallRule -DisplayGroup "Remote Desktop"
 Put your public key in `C:\ProgramData\ssh\administrators_authorized_keys` for an Administrator
 account. `vm-native-verify` uses that key.
 
-SSH into the guest fails with `Corrupted MAC on input` until the host has the e1000e offload rule
-from the `vfio-native` package. The emulated NIC's TX offloads corrupt integrity-checked traffic on
-the host side of the tap; SMB tolerates it, SSH does not. The package installs a udev rule that
-turns the offloads off on every libvirt tap as it appears, and `vm-native-setup` says so if it is
-missing.
-
 ## 3. Make the NVMe driver boot-critical
 
 The disk is about to move from virtio to emulated NVMe, and Windows only loads boot-start drivers
@@ -127,9 +121,14 @@ Two things it asks or warns about:
 
 Before the first boot, if the host has less free memory than the guest's RAM, free and compact
 it so the guest lands on transparent hugepages; `vm-native-setup` prints the two commands when it
-applies. The NIC stays `e1000e`, so the network survives the driver removal in the next step. Do not use
+applies. The NIC stays `igb`, so the network survives the driver removal in the next step. Do not use
 virtiofs for host files: it is a virtio device the scanner names, and its shared memory backing
-blocks transparent hugepages for the whole guest. Share over SMB on the e1000e link instead.
+blocks transparent hugepages for the whole guest. Share over SMB on the `igb` link instead.
+
+The interface asks for MTU 9000, and the package's hook turns GSO and GRO on for the tap. Together
+they take the inbound link from 2.9 to 14 Gbit/s; neither does anything alone. The libvirt network
+needs `<mtu size='9000'/>` too, and the guest needs *Jumbo Packet* 9014 with its interface MTU at
+9000 - setting the adapter property alone leaves the IP MTU at 1500 and gains nothing.
 
 ## 5. Remove the virtio drivers and the agents
 

packaging/vfio-native/PKGBUILD

@@ -8,7 +8,7 @@
 #   vfio-native-qemu       QEMU 11.1.1 with the platform-identity patches, in /opt
 
 pkgname=vfio-native
-pkgver=1.1.1
+pkgver=1.3.1
 pkgrel=1
 pkgdesc="Present a libvirt guest as a self-consistent physical machine, and tune it"
 arch=('any')
@@ -48,7 +48,6 @@ package() {
     # this coexists with whatever hook the host already has.
     install -Dm755 scripts/libvirt-hook-cpu-isolation.sh \
         "${pkgdir}/etc/libvirt/hooks/qemu.d/10-cpu-isolation.sh"
-    # e1000e offloads corrupt integrity-checked traffic on libvirt taps; host-wide by nature
+    install -Dm755 scripts/libvirt-hook-vnet-offload.sh \
-    install -Dm644 scripts/99-vfio-native-vnet-offload.rules \
+        "${pkgdir}/etc/libvirt/hooks/qemu.d/20-vnet-offload.sh"
-        "${pkgdir}/usr/lib/udev/rules.d/99-vfio-native-vnet-offload.rules"
 }

packaging/vfio-native/vfio-native.install

@@ -49,11 +49,11 @@ post_install() {
 
   'native' is a good default if you would rather not maintain anything.
 
-  A udev rule (99-vfio-native-vnet-offload.rules) turns TX offloads off on every
+  A hook at /etc/libvirt/hooks/qemu.d/20-vnet-offload.sh turns GSO and GRO on
-  libvirt tap as it appears. The emulated e1000e NIC corrupts integrity-checked
+  for the guest's tap. With the interface's MTU 9000 that takes the inbound link
-  traffic with them on; SSH to the guest fails with "Corrupted MAC on input".
+  from 2.9 to 14 Gbit/s; neither does anything alone. The libvirt network needs
-  It applies to every VM on the host; the throughput cost on a host<->guest
+  <mtu size='9000'/> too, and the guest needs Jumbo Packet 9014 with its
-  link is not measurable.
+  interface MTU at 9000.
 
   A libvirt hook is installed at /etc/libvirt/hooks/qemu.d/10-cpu-isolation.sh.
   It keeps host processes off the cores the guest is pinned to, automatically,

scripts/99-vfio-native-vnet-offload.rules

@@ -1,8 +0,0 @@
-# vfio-native: the emulated e1000e NIC's TX checksum and segmentation offloads
-# corrupt packets on the host side of a libvirt tap. SMB tolerates it; SSH fails
-# with "Corrupted MAC on input" and any integrity-checked protocol breaks.
-# Measured on a Zen 4 host with QEMU 11.1.1. Disabling the offloads on every
-# libvirt tap as it appears fixes it; on a host<->guest link the throughput
-# cost is not measurable. libvirt's <driver><host .../> attributes are ignored
-# for e1000e, and a libvirt hook must not call virsh, hence udev.
-ACTION=="add", SUBSYSTEM=="net", KERNEL=="vnet*", RUN+="/usr/bin/ethtool -K %k tx off gso off gro off tso off"

scripts/gpu-passthrough.sh

@@ -57,6 +57,24 @@ group_members() {
     done
 }
 
+# The functions to hand to the guest with GPU $1: every device in its IOMMU group
+# (mandatory for vfio), plus any sibling function of the same PCI device that is
+# itself a display or HDMI/DP audio controller. An APU parks its PSP and USB
+# controllers on the same PCI device in separate IOMMU groups - those are the
+# host's, so siblings are filtered by class and never pulled in blindly.
+passthrough_devs() {
+    local d="$1" m cls
+    {
+        group_members "$d"
+        for m in "/sys/bus/pci/devices/${d%.*}".*; do
+            [ -e "$m" ] || continue
+            m=$(basename "$m")
+            cls=$(cat "/sys/bus/pci/devices/$m/class" 2>/dev/null)
+            case "$cls" in 0x03*|0x0403*) echo "$m";; esac
+        done
+    } | sort -u
+}
+
 driver_of() {
     local l
     l=$(readlink -f "/sys/bus/pci/devices/$1/driver" 2>/dev/null)
@@ -79,7 +97,7 @@ drives_display() {
 
 ids_of() {   # vendor:device for vfio-pci binding
     local d
-    for d in $(group_members "$1"); do
+    for d in $(passthrough_devs "$1"); do
         local cls; cls=$(cat "/sys/bus/pci/devices/$d/class" 2>/dev/null)
         # only real functions of the card: skip bridges (class 0x0604xx)
         case "$cls" in 0x0604*) continue;; esac
@@ -219,7 +237,7 @@ prepare)
 
     # 1. stop whatever is holding the DRM device
     dm=$(systemctl list-units --type=service --state=running --no-legend 2>/dev/null |
-         awk '{print $1}' | grep -xE '(gdm|sddm|lightdm|lxdm|greetd|display-manager)\.service' | head -1)
+         awk '{print $1}' | grep -xE '((gdm|sddm|lightdm|lxdm|greetd|ly|emptty|lemurs)(@[a-z0-9-]+)?|display-manager)\.service' | head -1)
     if [ -n "$dm" ]; then
         echo "$dm" > "$STATE/dm"
         log "stopping $dm"
@@ -303,7 +321,7 @@ apply() {
     "$v" -c qemu:///system dumpxml --inactive "$dom" > "$backup/$dom.gpu-current.xml"
     grep -q 'ua-vfionative-gpu' "$backup/$dom.gpu-current.xml" || cp "$backup/$dom.gpu-current.xml" "$backup/$dom.before-gpu.xml"
     local devs="" m
-    for m in $( { group_members "$d"; ls -d "/sys/bus/pci/devices/${d%.*}".* | xargs -n1 basename; } | sort -u); do
+    for m in $(passthrough_devs "$d"); do
         case "$(cat "/sys/bus/pci/devices/$m/class" 2>/dev/null)" in 0x0604*) continue;; esac
         IFS=':.' read -r dm bs sl fn <<< "$m"
         devs+="    <hostdev mode='subsystem' type='pci' managed='yes'>\n      <source>\n"
@@ -330,7 +348,7 @@ xml() {
     echo "Add this to the domain, inside <devices>. Every device in the GPU's"
     echo "IOMMU group has to go together:"
     echo
-    for m in $(group_members "$d"); do
+    for m in $(passthrough_devs "$d"); do
         local cls; cls=$(cat "/sys/bus/pci/devices/$m/class" 2>/dev/null)
         case "$cls" in 0x0604*) continue;; esac
         IFS=':. ' read -r dom bus slot fn <<< "$(echo "$m" | tr ':.' '  ')"

scripts/libvirt-hook-vnet-offload.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+# libvirt qemu hook: turn GSO and GRO on for the guest's tap.
+#
+# QEMU leaves them off and sets the tap up after udev has run, so it has to
+# happen here. Paired with MTU 9000 they are the difference between 2.5 and
+# 14 Gbit/s into the guest; neither helps alone. Tap names come from the domain
+# XML on stdin, so the hook never calls virsh, which would deadlock libvirtd.
+
+[ "$2" = started ] || exit 0
+
+for tap in $(grep -oE "<target dev='(vnet|tap|macvtap)[^']*'" | sed "s/.*dev='//; s/'$//"); do
+    ethtool -K "$tap" gso on gro on 2>/dev/null
+done
+
+exit 0

scripts/setup-vm.sh

@@ -637,8 +637,11 @@ if conformant and E["CONVERT"] == "1":
         if "device='disk'" not in d or ("bus='nvme'" in d and "<serial>" in d):
             return d
         d = re.sub(r"<target dev='([^']*)' bus='(virtio|sata|scsi)'/>", r"<target dev='\1' bus='nvme'/>", d)
+        # cache='none' is O_DIRECT: on btrfs the guest can change a page while the
+        # write is in flight, so the stored checksum never matches and later reads
+        # fail with EIO. Buffered writes hand the filesystem a stable page.
         d = re.sub(r"<driver name='qemu' type='([^']*)'[^/]*/>",
-                   r"<driver name='qemu' type='\1' cache='none' io='native' discard='unmap'/>", d)
+                   r"<driver name='qemu' type='\1' cache='writeback' io='threads' discard='unmap'/>", d)
         d = re.sub(r"\s*<address type='(pci|drive)'[^/]*/>", "", d)
         if "<serial>" not in d:
             serial = E["NVME_SERIAL"] if n[0] == 0 else E["NVME_SERIAL"][:-1] + "0123456789ABCDEF"[n[0] % 16]
@@ -658,7 +661,11 @@ if conformant and E["CONVERT"] == "1":
     s = re.sub(r"\s*<input type='[^']*' bus='virtio'/>", "", s)
     s = re.sub(r"<memballoon model='virtio'>.*?</memballoon>", "<memballoon model='none'/>", s, flags=re.S)
     s = re.sub(r"<memballoon model='virtio'/>", "<memballoon model='none'/>", s)
-    s = re.sub(r"<model type='virtio'/>(\s*<driver [^/]*/>)?", "<model type='e1000e'/>", s)
+    # jumbo frames are the single biggest win on the host<->guest link: the emulated
+    # NIC is packet-rate bound, so 9000-byte frames cut the per-packet cost the guest
+    # pays on receive. Measured 2922 -> 14232 Mbit/s inbound on igb, byte-exact clean.
+    s = re.sub(r"<model type='virtio'/>(\s*<driver [^/]*/>)?",
+               "<model type='igb'/>\n      <mtu size='9000'/>", s)
     if prof == "full":
         s = re.sub(r"<video>.*?</video>", "<video>\n      <model type='none'/>\n    </video>", s, flags=re.S)
 
@@ -824,10 +831,19 @@ if [ "$PROFILE" = full ]; then
 fi
 
 if [ ! -e /usr/lib/udev/rules.d/99-vfio-native-vnet-offload.rules ] && [ ! -e /etc/udev/rules.d/99-vfio-native-vnet-offload.rules ]; then
-    echo "NOTE: the e1000e offload udev rule is not installed. SSH into the guest will fail with"
+    echo "NOTE: the NIC offload udev rule is not installed. SSH into the guest will fail with"
     echo "      'Corrupted MAC on input' until it is:"
     echo "      sudo install -Dm644 $SELF/scripts/99-vfio-native-vnet-offload.rules /etc/udev/rules.d/ && sudo udevadm control --reload-rules"
 fi
+
+NET=$("${C[@]}" dumpxml "$DOM" 2>/dev/null | sed -n "s/.*<source network='\([^']*\)'.*/\1/p" | head -1)
+NETMTU=$([ -n "$NET" ] && "${C[@]}" net-dumpxml --inactive "$NET" 2>/dev/null | sed -n "s/.*<mtu size='\([0-9]*\)'.*/\1/p")
+if [ -n "$NET" ] && [ "${NETMTU:-1500}" -lt 9000 ]; then
+    echo "NOTE: the guest interface asks for MTU 9000 but libvirt network '$NET' is at ${NETMTU:-1500}."
+    echo "      Jumbo needs both ends; inbound throughput is ~5x with it. Add <mtu size='9000'/> to"
+    echo "      the network and restart it:  virsh net-edit $NET && virsh net-destroy $NET && virsh net-start $NET"
+    echo "      Then in the guest: set the NIC's Jumbo Packet to 9014 and the interface MTU to 9000."
+fi
 gov=$(cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor 2>/dev/null || echo unknown)
 [ "$gov" = performance ] || echo "host governor is '$gov' - run: sudo cpupower frequency-set -g performance"