On 7/25/2018 8:02 AM, Yonghong Song wrote:
On 7/24/18 3:40 PM, John Fastabend wrote:On 07/24/2018 08:45 AM, Yonghong Song wrote:In one of our production machines, tools/testing/selftests/bpf test_sockmap failed randomly like below: ... [TEST 78]: (512, 1, 1, sendmsg, pass,apply 1,): rx thread exited with err 1. FAILED ... ... [TEST 80]: (2, 1024, 256, sendmsg, pass,apply 1,): rx thread exited with err 1. FAILED ... ... [TEST 83]: (100, 1, 5, sendpage, pass,apply 1,): rx thread exited with err 1. FAILED ... ... [TEST 79]: (512, 1, 1, sendpage, pass,apply 1,): rx thread exited with err 1. FAILED ... The command line is just `test_sockmap`. The machine has 80 cpus, 256G memory. The kernel is based on 4.16 but backported with latest bpf-next bpf changes. The failed test number (78, 79, 80, or 83) is random. But they all share similar characteristics: . the option rate is greater than one, i.e., more than one sendmsg/sendpage in the sender forked process. . The txmsg_apply is not 0 I debugged a little bit. It happens in msg_loop() function below "unexpected timeout" path. ... slct = select(max_fd + 1, &w, NULL, NULL, &timeout); if (slct == -1) { perror("select()");clock_gettime(CLOCK_MONOTONIC, &s->end);goto out_errno; } else if (!slct) { if (opt->verbose) fprintf(stderr, "unexpected timeout\n"); errno = -EIO;clock_gettime(CLOCK_MONOTONIC, &s->end);goto out_errno; } ... It appears that when the error happens, the receive process does not receive all bytes sent from the send process and eventually times out. Has anybody seen this issue as well? John, any comments on this failure?Can you run the test with verbose enabled so we can determine if the tx side is even sending the message? Sample patch below. This will allow us to see the tx bytes and rx bytes, although it will be a bit noisy. I notice that the test program is not smart enough to (re)send bytes if the sendmsg call doesn't consume all bytes. This is a valid error if we get a enomem or other normal error on the tx side. With apply this is more likely because every byte (in apply = 1 case) is being sent through BPF prog.The following are some of logs for the failed tests: ...[TEST 78]: (512, 1, 1, sendmsg, pass,apply 1,): connected sockets: c1 <-> p1, c2 <-> p2cgroups binding: c1(24) <-> s1(22) - - - c2(25) <-> s2(23)tx_sendmsg: TX: 512B 0.000000B/s 0.000000 GB/s RX: 0B 0.000000B/s 0.000000GB/sunexpected timeout msg_loop_rx: iov_count 1 iov_buf 1 cnt 512 err -5rx_sendmsg: TX: 0B 0.000000B/s 0.000000GB/s RX: 147B 147.000000B/s 0.000000GB/srx thread exited with err 1. FAILED ...[TEST 82]: (100, 1, 5, sendmsg, pass,apply 1,): connected sockets: c1 <-> p1, c2 <-> p2cgroups binding: c1(24) <-> s1(22) - - - c2(25) <-> s2(23)tx_sendmsg: TX: 500B 0.000000B/s 0.000000 GB/s RX: 0B 0.000000B/s 0.000000GB/sunexpected timeout msg_loop_rx: iov_count 1 iov_buf 5 cnt 100 err -5rx_sendmsg: TX: 0B 0.000000B/s 0.000000GB/s RX: 366B 366.000000B/s 0.000000GB/srx thread exited with err 1. FAILED[TEST 402]: (512, 1, 1, sendmsg, pass,apply 1,): connected sockets: c1 <-> p1, c2 <-> p2cgroups binding: c1(42) <-> s1(40) - - - c2(43) <-> s2(41)tx_sendmsg: TX: 512B 0.000000B/s 0.000000 GB/s RX: 0B 0.000000B/s 0.000000GB/sunexpected timeout msg_loop_rx: iov_count 1 iov_buf 1 cnt 512 err -5rx_sendmsg: TX: 0B 0.000000B/s 0.000000GB/s RX: 147B 147.000000B/s 0.000000GB/srx thread exited with err 1. FAILED[TEST 80]: (2, 1024, 256, sendmsg, pass,apply 1,): connected sockets: c1 <-> p1, c2 <-> p2cgroups binding: c1(24) <-> s1(22) - - - c2(25) <-> s2(23)tx_sendmsg: TX: 524288B 0.000000B/s 0.000000 GB/s RX: 0B 0.000000B/s 0.000000GB/sunexpected timeout msg_loop_rx: iov_count 1024 iov_buf 256 cnt 2 err -5rx_sendmsg: TX: 0B 0.000000B/s 0.000000GB/s RX: 458805B 458805.000000B/s 0.000459GB/srx thread exited with err 1. FAILEDIf this is not the case I can do some more digging but I've not seen this before.I cannot reproduce it in my FC27 VM (with latest bpf-next) either.The bug only shows up on our production service which is a lot busier and has more cpus/memorys and some sysctl configurations are different from my VM setup.
In the past I have noticed that on busy machines sometimes data is delayed so much that select() timeout is triggered. Although it isn't a solution but you can try to increase the timeout value to check if that is the case.
-Prashant
