Willem, this is only my 2nd patch, and my last one was a one liner. I’ll try to work through this, but let me know if I am doing a rookie mistake (learning curve and all).
On May 23, 2019, at 2:56 PM, Willem de Bruijn willemdebruijn.kernel@gmail.com wrote:
On Thu, May 23, 2019 at 5:11 PM Fred Klassen fklassen@appneta.com wrote:
This enhancement adds the '-a' option, which will count all CMSG messages on the error queue and print a summary report.
Fixes: 3a687bef148d ("selftests: udp gso benchmark")
Also not a fix, but an extension.
I’ll make a v2 patch and remove “Fixes:".
Example:
# ./udpgso_bench_tx -4uT -a -l5 -S 1472 -D 172.16.120.189 udp tx: 492 MB/s 8354 calls/s 8354 msg/s udp tx: 477 MB/s 8106 calls/s 8106 msg/s udp tx: 488 MB/s 8288 calls/s 8288 msg/s udp tx: 882 MB/s 14975 calls/s 14975 msg/s Summary over 5.000 seconds ... sum udp tx: 696 MB/s 57696 calls (11539/s) 57696 msgs (11539/s) Tx Timestamps: received: 57696 errors: 0
This can be useful in tracking loss of messages when under load. For example, adding the '-z' option results in loss of TX timestamp messages:
# ./udpgso_bench_tx -4ucT -a -l5 -S 1472 -D 172.16.120.189 -p 3239 -z udp tx: 490 MB/s 8325 calls/s 8325 msg/s udp tx: 500 MB/s 8492 calls/s 8492 msg/s udp tx: 883 MB/s 14985 calls/s 14985 msg/s udp tx: 756 MB/s 12823 calls/s 12823 msg/s Summary over 5.000 seconds ... sum udp tx: 657 MB/s 54429 calls (10885/s) 54429 msgs (10885/s) Tx Timestamps: received: 34046 errors: 0 Zerocopy acks: received: 54422 errors: 0
This would probably also be more useful as regression test if it is in the form of a pass/fail test: if timestamps are requested and total count is zero, then the feature is broken and the process should exit with an error.
I’ll add a hard failure for zero response for TX Timestamps or Zerocopy, or if any errors occur.
Fixes: 3a687bef148d ("selftests: udp gso benchmark")
Repeated
Will fix.
Signed-off-by: Fred Klassen fklassen@appneta.com
tools/testing/selftests/net/udpgso_bench_tx.c | 152 +++++++++++++++++++------- 1 file changed, 113 insertions(+), 39 deletions(-)
diff --git a/tools/testing/selftests/net/udpgso_bench_tx.c b/tools/testing/selftests/net/udpgso_bench_tx.c index 56e0d890b066..9924342a0b03 100644 --- a/tools/testing/selftests/net/udpgso_bench_tx.c +++ b/tools/testing/selftests/net/udpgso_bench_tx.c @@ -62,10 +62,19 @@ static bool cfg_tcp; static uint32_t cfg_tx_ts = SOF_TIMESTAMPING_TX_SOFTWARE; static bool cfg_tx_tstamp; static uint32_t cfg_tos; +static bool cfg_audit; static bool cfg_verbose; static bool cfg_zerocopy; static int cfg_msg_nr; static uint16_t cfg_gso_size; +static unsigned long total_num_msgs; +static unsigned long total_num_sends; +static unsigned long stat_tx_ts; +static unsigned long stat_tx_ts_errors; +static unsigned long tstart; +static unsigned long tend; +static unsigned long stat_zcopies; +static unsigned long stat_zcopy_errors;
static socklen_t cfg_alen; static struct sockaddr_storage cfg_dst_addr; @@ -137,8 +146,11 @@ static void flush_cmsg(struct cmsghdr *cmsg) struct my_scm_timestamping *tss;
tss = (struct my_scm_timestamping *)CMSG_DATA(cmsg);
fprintf(stderr, "tx timestamp = %lu.%09lu\n",tss->ts[i].tv_sec, tss->ts[i].tv_nsec);
if (tss->ts[i].tv_sec == 0)stat_tx_ts_errors++;if (cfg_verbose)fprintf(stderr, "tx timestamp = %lu.%09lu\n",tss->ts[i].tv_sec, tss->ts[i].tv_nsec);changes unrelated to this feature?
I’ll remove. Do you think that I should pull out any messages related to “cfg_verbose”?