cpuirq

#!/bin/sh

# Enumerate all interrupts used by either the default or a specified
# network interface, and map them to separate CPUs. Each CPU that gets
# assigned interrupts will also have its C-states deeper than C1
# disabled to reduce the probability of rx queue overflow.
#
# TODO: NUMA interrupt scheduling.
#

# Space-separated list of interfaces.
iface="$IFACE"

# Space-separated list of CPUs that should not be assigned interrupts.
cpu_blacklist="${CPU_BLACKLIST:-0}"

# Space-separated list of CPUs that should be assigned interrupts.
# Defaults to [0, 1024]. Blacklisted or offline CPUs will not be
# assigned any interrupts.
cpu_whitelist="${CPU_WHITELIST:-`seq -s ' ' 0 1024`}"

# Convenience vars.
sysfs_net="/sys/class/net"
sysfs_cpu="/sys/devices/system/cpu"

# Bail out if we're missing essential sysfs directories.
for dir in "$sysfs_net" "$sysfs_cpu"
do
        test -d "$dir" || {
                echo >&2 "ERROR: $dir not found."
                exit 1
        }
done

# Return a list of interrupts for a given network interface.
iface_irqs () {
        iface_irqs=""
        test -d "${sysfs_net}"/${1}/device/msi_irqs && {
                for f in "${sysfs_net}"/${1}/device/msi_irqs/*
                do
                        iface_irqs="$iface_irqs `basename $f`"
                done

                echo "$iface_irqs"
                return
        }

        # Fall back to naive parsing of /proc/interrupts.
        grep "\<${1}\>" /proc/interrupts | cut -d: -f1 | tr '\n' ' '
}

# Return a list of all interface interrupts.
avail_irqs () {
        avail_irqs=""

        # Enumerate bonding slaves, if any.
        test -f "${sysfs_net}"/${iface}/bonding/slaves &&
                iface=`cat "${sysfs_net}"/${iface}/bonding/slaves`

        for i in $iface
        do
                avail_irqs="$avail_irqs `iface_irqs $i`"
        done

        echo "$avail_irqs"
}

# Return 1 if CPU is blacklisted.
bl_cpu () {
        test -n "$cpu_blacklist" && {
                for bl_cpu in $cpu_blacklist
                do
                        test "$1" -eq "$bl_cpu" 2>/dev/null && return 1
                done
        }

        # Explicit zero return in case of test(1) error.
        return 0
}

# Return a list of available CPUs.
avail_cpus () {
        avail_cpus=""

        # Iterate over all CPUs.
        for cpu in $cpu_whitelist
        do
                # Break if no more CPUs.
                test -d "${sysfs_cpu}"/cpu${cpu} || break

                # Check if CPU > 0 is online.
                online=1
                test "$cpu" -gt 0 && {
                        online=`cat "${sysfs_cpu}"/cpu${cpu}/online`
                }
                test "$online" -gt 0 || {
                        echo >&2 "WARNING: Skipping offline CPU $cpu."
                        continue
                }

                # Skip CPUs in blacklist.
                bl_cpu "$cpu" || {
                        echo >&2 "WARNING: Skipping blacklisted CPU $cpu."
                        continue
                }

                # Add CPU to available list.
                avail_cpus="$avail_cpus $cpu"
        done

        echo "$avail_cpus"
}

# Summary: Write two files into the temporary directory: One
# containing IRQs, and one containing CPUs. Repeat the CPU list as
# many times as needed to ensure all IRQs are handled, trim the list
# to exactly match the number of IRQs, and then merge them together.
do_start () {
        tmpdir=`mktemp -d`

        # Populate the CPU list.
        ncpus="0"
        for cpu in `avail_cpus`
        do
                echo "$cpu" >> "$tmpdir"/cpus_init
                ncpus=$(( $ncpus + 1 ))
        done

        # Populate the IRQ list.
        nirqs="0"
        for irq in `avail_irqs`
        do
                echo "$irq" >> "$tmpdir"/irqs
                nirqs=$(( $nirqs + 1 ))
        done

        # Number of times to repeat the CPU list.
        cpu_reps=$(( $nirqs / $ncpus + 1 ))
        for i in `seq $cpu_reps`
        do
                cat "$tmpdir"/cpus_init >> "$tmpdir"/cpus_rep
        done

        # Trim the CPU list length to match the number of IRQs.
        head -n "$nirqs" "$tmpdir"/cpus_rep > "$tmpdir"/cpus

        # Merge IRQ and CPU lists.
        paste "$tmpdir"/irqs "$tmpdir"/cpus 2>/dev/null | while read irq cpu
        do
                # Map interrupt to CPU.
                echo TESTING $cpu \> /proc/irq/${irq}/smp_affinity_list

                # Disable high-latency CPU C-states.
                for f in "${sysfs_cpu}"/cpu${cpu}/cpuidle/state[2-9]/disable
                do
                        test -f "$f" && echo TESTING 1 \> "$f"
                done
        done
        rm -rf "$tmpdir"

        # Warn if no C-state toggles are available (since 3a53396b in
        # v3.4-rc1; per-cpu since dc7fd275 in v3.6-rc1).
        test -f "${sysfs_cpu}"/cpu0/cpuidle/state0/disable ||
                echo >&2 "WARNING: No C-state toggles available. Old kernel?"
}

# Simply re-enable all CPU C-states. We'll leave the interrupt mapping
# untouched, as it doesn't really have any ill effects on the system.
do_stop () {
        for cpu in `avail_cpus`
        do
                for f in "${sysfs_cpu}"/cpu${cpu}/cpuidle/state[0-9]/disable
                do
                        test -f "$f" && echo TESTING 0 \> "$f"
                done
        done

        # Explicit zero return in case of test(1) error.
        return 0
}

# Verify that we have a valid interface. Fall back to using
# the default route interface.
test -n "$iface" && test -d "${sysfs_net}"/${iface} ||
        iface=`route -n|grep "^0\.0\.0\.0"|awk '{print $NF}'`


case "$1" in
        start|"")
                do_start
                ;;
        stop)
                do_stop
                ;;
        *)
                echo >&2 "Usage: $0 [start|stop]"
                exit 1
                ;;
esac