[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <Z91yk90LZy9yJexG@mini-arch>
Date: Fri, 21 Mar 2025 07:07:15 -0700
From: Stanislav Fomichev <stfomichev@...il.com>
To: Kuniyuki Iwashima <kuniyu@...zon.com>
Cc: "David S. Miller" <davem@...emloft.net>,
David Ahern <dsahern@...nel.org>,
Eric Dumazet <edumazet@...gle.com>,
Jakub Kicinski <kuba@...nel.org>, Paolo Abeni <pabeni@...hat.com>,
Simon Horman <horms@...nel.org>,
Kuniyuki Iwashima <kuni1840@...il.com>, netdev@...r.kernel.org
Subject: Re: [PATCH v1 net-next 00/13] ipv6: No RTNL for IPv6 routing table.
On 03/20, Kuniyuki Iwashima wrote:
> IPv6 routing tables are protected by each table's lock and work in
> the interrupt context, which means we basically don't need RTNL to
> modify an IPv6 routing table itself.
>
> Currently, the control paths require RTNL because we may need to
> perform device and nexthop lookups; we must prevent dev/nexthop from
> going away from the netns.
>
> This, however, can be achieved by RCU as well.
>
> If we are in the RCU critical section while adding an IPv6 route,
> synchronize_net() in netif_change_net_namespace() and
> unregister_netdevice_many_notify() guarantee that the dev will not be
> moved to another netns or removed.
>
> Also, nexthop is guaranteed not to be freed during the RCU grace period.
>
> If we care about a race between nexthop removal and IPv6 route addition,
> we can get rid of RTNL from the control paths.
>
> Patch 1 moves a validation for RTA_MULTIPATH earlier.
> Patch 2 removes RTNL for SIOCDELRT and RTM_DELROUTE.
> Patch 3 ~ 10 move validation and memory allocation earlier.
> Patch 11 prevents a race between two requests for the same table.
> Patch 12 prevents the race mentioned above.
> Patch 13 removes RTNL for SIOCADDRT and RTM_NEWROUTE.
>
>
> Test:
>
> The script [0] lets each CPU-X create 100000 routes on table-X in a
> batch.
>
> On c7a.metal-48xl EC2 instance with 192 CPUs,
>
> With this series:
>
> $ sudo ./route_test.sh
> start adding routes
> added 19200000 routes (100000 routes * 192 tables).
> Time elapsed: 189154 milliseconds.
>
> Without series:
>
> $ sudo ./route_test.sh
> start adding routes
> added 19200000 routes (100000 routes * 192 tables).
> Time elapsed: 62531 milliseconds.
>
> I changed the number of routes (1000 ~ 100000 per CPU/table) and
> constantly saw it complete 3x faster with this series.
>
>
> [0]
> #!/bin/bash
>
> mkdir tmp
>
> NS="test"
> ip netns add $NS
> ip -n $NS link add veth0 type veth peer veth1
> ip -n $NS link set veth0 up
> ip -n $NS link set veth1 up
>
> TABLES=()
> for i in $(seq $(nproc)); do
> TABLES+=("$i")
> done
>
> ROUTES=()
> for i in {1..100}; do
> for j in {1..1000}; do
> ROUTES+=("2001:$i:$j::/64")
> done
> done
>
> for TABLE in "${TABLES[@]}"; do
> FILE="./tmp/batch-table-$TABLE.txt"
> > $FILE
> for ROUTE in "${ROUTES[@]}"; do
> echo "route add $ROUTE dev veth0 table $TABLE" >> $FILE
> done
> done
>
> echo "start adding routes"
>
> START_TIME=$(date +%s%3N)
> for TABLE in "${TABLES[@]}"; do
> ip -n $NS -6 -batch "./tmp/batch-table-$TABLE.txt" &
> done
>
> wait
> END_TIME=$(date +%s%3N)
> ELAPSED_TIME=$((END_TIME - START_TIME))
>
> echo "added $((${#ROUTES[@]} * ${#TABLES[@]})) routes (${#ROUTES[@]} routes * ${#TABLES[@]} tables)."
> echo "Time elapsed: ${ELAPSED_TIME} milliseconds."
> echo $(ip -n $NS -6 route show table all | wc -l) # Just for debug
>
> ip netns del $NS
> rm -fr ./tmp/
Lockdep is not supper happy about some patch:
https://netdev-3.bots.linux.dev/vmksft-forwarding-dbg/results/42463/38-gre-multipath-nh-res-sh/stderr
---
pw-bot: cr
Powered by blists - more mailing lists