On 12/6/19 4:17 AM, Thadeu Lima de Souza Cascardo wrote: > On Wed, Dec 04, 2019 at 12:03:57PM -0800, Eric Dumazet wrote: >> >> >> On 12/4/19 11:53 AM, Thadeu Lima de Souza Cascardo wrote: >>> When using fragments with size 8 and payload larger than 8000, the backlog >>> might fill up and packets will be dropped, causing the test to fail. This >>> happens often enough when conntrack is on during the IPv6 test. >>> >>> As the larger payload in the test is 10000, using a backlog of 1250 allow >>> the test to run repeatedly without failure. At least a 1000 runs were >>> possible with no failures, when usually less than 50 runs were good enough >>> for showing a failure. >>> >>> As netdev_max_backlog is not a pernet setting, this sets the backlog to >>> 1000 during exit to prevent disturbing following tests. >>> >> >> Hmmm... I would prefer not changing a global setting like that. >> This is going to be flaky since we often run tests in parallel (using different netns) >> >> What about adding a small delay after each sent packet ? >> >> diff --git a/tools/testing/selftests/net/ip_defrag.c b/tools/testing/selftests/net/ip_defrag.c >> index c0c9ecb891e1d78585e0db95fd8783be31bc563a..24d0723d2e7e9b94c3e365ee2ee30e9445deafa8 100644 >> --- a/tools/testing/selftests/net/ip_defrag.c >> +++ b/tools/testing/selftests/net/ip_defrag.c >> @@ -198,6 +198,7 @@ static void send_fragment(int fd_raw, struct sockaddr *addr, socklen_t alen, >> error(1, 0, "send_fragment: %d vs %d", res, frag_len); >> >> frag_counter++; >> + usleep(1000); >> } >> >> static void send_udp_frags(int fd_raw, struct sockaddr *addr, >> > > That won't work because the issue only shows when we using conntrack, as the > packet will be reassembled on output, then fragmented again. When this happens, > the fragmentation code is transmitting the fragments in a tight loop, which > floods the backlog. Interesting ! So it looks like the test is correct, and exposed a long standing problem in this code. We should not adjust the test to some kernel-of-the-day-constraints, and instead fix the kernel bug ;) Where is this tight loop exactly ? If this is feeding/bursting ~1000 skbs via netif_rx() in a BH context, maybe we need to call a variant that allows immediate processing instead of (ab)using the softnet backlog. Thanks !
