@@ -789,6 +789,7 @@ fib6_gc_test()
set -e
EXPIRE=5
+ GC_WAIT_TIME=$((EXPIRE * 2 + 2))
# Check expiration of routes every $EXPIRE seconds (GC)
$NS_EXEC sysctl -wq net.ipv6.route.gc_interval=$EXPIRE
@@ -805,7 +806,7 @@ fib6_gc_test()
$IP -6 route add 2001:20::$i \
via 2001:10::2 dev dummy_10 expires $EXPIRE
done
- sleep $(($EXPIRE * 2 + 1))
+ sleep $GC_WAIT_TIME
$NS_EXEC sysctl -wq net.ipv6.route.flush=1
check_rt_num 0 $($IP -6 route list |grep expires|wc -l)
log_test $ret 0 "ipv6 route garbage collection"
@@ -823,7 +824,8 @@ fib6_gc_test()
$IP -6 route add 2001:20::$i \
via 2001:10::2 dev dummy_10 expires $EXPIRE
done
- sleep $(($EXPIRE * 2 + 1))
+ # Wait for GC
+ sleep $GC_WAIT_TIME
check_rt_num 0 $($IP -6 route list |grep expires|wc -l)
log_test $ret 0 "ipv6 route garbage collection (with permanent routes)"
@@ -840,10 +842,8 @@ fib6_gc_test()
$IP -6 route replace 2001:20::$i \
via 2001:10::2 dev dummy_10 expires $EXPIRE
done
- check_rt_num_clean 5 $($IP -6 route list |grep expires|wc -l) || return
# Wait for GC
- sleep $(($EXPIRE * 2 + 1))
- $NS_EXEC sysctl -wq net.ipv6.route.flush=1
+ sleep $GC_WAIT_TIME
check_rt_num 0 $($IP -6 route list |grep expires|wc -l)
log_test $ret 0 "ipv6 route garbage collection (replace with expires)"
@@ -863,8 +863,7 @@ fib6_gc_test()
check_rt_num_clean 0 $($IP -6 route list |grep expires|wc -l) || return
# Wait for GC
- sleep $(($EXPIRE * 2 + 1))
-
+ sleep $GC_WAIT_TIME
check_rt_num 5 $($IP -6 route list |grep -v expires|grep 2001:20::|wc -l)
log_test $ret 0 "ipv6 route garbage collection (replace with permanent)"
@@ -901,9 +900,7 @@ fib6_gc_test()
check_rt_num_clean 1 $($IP -6 route list|grep expires|wc -l) || return
# Wait for GC
- sleep $(($EXPIRE * 2 + 1))
-
- $NS_EXEC sysctl -wq net.ipv6.route.flush=1
+ sleep $GC_WAIT_TIME
check_rt_num 0 $($IP -6 route list |grep expires|wc -l)
log_test $ret 0 "ipv6 route garbage collection (RA message)"
ipv6_gc fails occasionally. According to the study, fib6_run_gc() using jiffies_round() to round the GC interval could increase the waiting time up to 750ms (3/4 seconds). The timer has a granularity of 512ms at the range 4s to 32s. That means a route with an expiration time E seconds can wait for more than E * 2 + 1 seconds if the GC interval is also E seconds. E * 2 + 2 seconds should be enough for waiting for removing routes. Also remove a check immediately after replacing 5 routes since it is very likely to remove some of routes before completing the last route with a slow environment. Signed-off-by: Kui-Feng Lee <thinker.li@gmail.com> --- I made the follow change to simulate the case that a route misses the first GC, and the waiting time of the next GC is rounded up by round_jiffies()) adding 750ms extra waiting time. diff --git a/net/ipv6/ip6_fib.c b/net/ipv6/ip6_fib.c index 6540d877d369..88479daa6ff7 100644 --- a/net/ipv6/ip6_fib.c +++ b/net/ipv6/ip6_fib.c @@ -1487,8 +1487,10 @@ int fib6_add(struct fib6_node *root, struct fib6_info *rt, list_add(&rt->nh_list, &rt->nh->f6i_list); __fib6_update_sernum_upto_root(rt, fib6_new_sernum(info->nl_net)); - if (rt->fib6_flags & RTF_EXPIRES) + if (rt->fib6_flags & RTF_EXPIRES) { + printk(KERN_CRIT "fib6_add expires\n"); fib6_add_gc_list(rt); + } fib6_start_gc(info->nl_net, rt); } @@ -2296,6 +2298,7 @@ static void fib6_flush_trees(struct net *net) static int fib6_age(struct fib6_info *rt, struct fib6_gc_args *gc_args) { unsigned long now = jiffies; + static int expire_cnt = 0; /* * check addrconf expiration here. @@ -2303,8 +2306,9 @@ static int fib6_age(struct fib6_info *rt, struct fib6_gc_args *gc_args) */ if (rt->fib6_flags & RTF_EXPIRES && rt->expires) { - if (time_after(now, rt->expires)) { + if (time_after(now, rt->expires) && expire_cnt++) { pr_debug("expiring %p\n", rt); + printk(KERN_CRIT "fib6_age expiring\n"); return -1; } gc_args->more++; @@ -2376,8 +2380,7 @@ void fib6_run_gc(unsigned long expires, struct net *net, bool force) if (gc_args.more) mod_timer(&net->ipv6.ip6_fib_timer, - round_jiffies(now - + net->ipv6.sysctl.ip6_rt_gc_interval)); + now + net->ipv6.sysctl.ip6_rt_gc_interval + HZ * 3 / 4); else del_timer(&net->ipv6.ip6_fib_timer); spin_unlock_bh(&net->ipv6.fib6_gc_lock); The following is the test case. fib6_gc_test() { setup echo echo "Fib6 garbage collection test" set -e EXPIRE=5 # Check expiration of routes every $EXPIRE seconds (GC) $NS_EXEC sysctl -wq net.ipv6.route.gc_interval=$EXPIRE $IP link add dummy_10 type dummy $IP link set dev dummy_10 up $IP -6 address add 2001:10::1/64 dev dummy_10 $NS_EXEC sysctl -wq net.ipv6.route.flush=1 # Temporary routes $IP -6 route add 2001:20::1 \ via 2001:10::2 dev dummy_10 expires $EXPIRE sleep $(($EXPIRE * 4 + 1)) set +e cleanup &> /dev/null } According to what I found from the kernel messages, the route waited 11.258s before being removed. It is longer than $EXPIRE * 2 + 1 seconds where EXPIRE=5. [ 8.674004] fib6_add expires [ 19.932557] fib6_age expiring Waiting for EXPIRE * 2 + 2 seconds should be enough for non-debug build. Changes from v2: - GC_WAIT_TIME to replace $(($EXPIRE * 2 + 2)) v2: https://lore.kernel.org/all/20240305013734.872968-1-thinker.li@gmail.com/ --- tools/testing/selftests/net/fib_tests.sh | 17 +++++++---------- 1 file changed, 7 insertions(+), 10 deletions(-)