This issue was discovered under the case of software vm2vm
fowarding. When pkts are received from virtio device 0 and
tx_route to virtio device 1, tx of device 0 is not updated.
To fix this problem, we check each rx/tx branch to update
stats. Besides, the stats are printed at a separated thread,
so we design a mechanism to make sure unreasonable data will
not show.
Signed-off-by: Jianfeng Tan <jianfeng.tan at intel.com>
Tested-by: Qian Xu <qian.q.xu at intel.com>
---
examples/vhost/main.c | 71 ++++++++++++++++++++++++++++++++++++---------------
1 file changed, 51 insertions(+), 20 deletions(-)
diff --git a/examples/vhost/main.c b/examples/vhost/main.c
index 9bfda6d..1863dc3 100644
--- a/examples/vhost/main.c
+++ b/examples/vhost/main.c
@@ -312,12 +312,25 @@ struct ipv4_hdr {
/* Per-device statistics struct */
struct device_statistics {
- uint64_t tx_total;
+ /* rx: from vhost to virtio; tx: from virtio to vhost */
+
+ /* write-only by datapath threads */
+ /* for non zero-copy case, pkts may be enqueued by any lcore */
rte_atomic64_t rx_total_atomic;
- uint64_t rx_total;
- uint64_t tx;
rte_atomic64_t rx_atomic;
+ /* for zero-copy case */
+ uint64_t rx_total;
uint64_t rx;
+ /* write-only by corresponding datapath thread */
+ uint64_t tx_total;
+ uint64_t tx;
+
+ /* write-only by print-stats threads */
+ uint64_t rx_total_p;
+ uint64_t rx_p;
+ uint64_t tx_total_p;
+ uint64_t tx_p;
+
} __rte_cache_aligned;
struct device_statistics dev_statistics[MAX_DEVICES];
@@ -2787,7 +2800,6 @@ static void
print_stats(void)
{
struct virtio_net_data_ll *dev_ll;
- uint64_t tx_dropped, rx_dropped;
uint64_t tx, tx_total, rx, rx_total;
uint32_t device_fh;
const char clr[] = { 27, '[', '2', 'J', '\0' };
@@ -2804,9 +2816,18 @@ print_stats(void)
dev_ll = ll_root_used;
while (dev_ll != NULL) {
device_fh = (uint32_t)dev_ll->vdev->dev->device_fh;
- tx_total = dev_statistics[device_fh].tx_total;
- tx = dev_statistics[device_fh].tx;
- tx_dropped = tx_total - tx;
+
+ tx_total = dev_statistics[device_fh].tx_total -
+ dev_statistics[device_fh].tx_total_p;
+ tx = dev_statistics[device_fh].tx -
+ dev_statistics[device_fh].tx_p;
+ /* Because we do not use a lock to control the access
of tx_total
+ * and tx in dev_statistics, tx may be greater than
tx_total. If
+ * this happens, we'll count those redundant tx next
time.
+ */
+ if (unlikely(tx_total < tx))
+ tx = tx_total;
+
if (zero_copy == 0) {
rx_total = rte_atomic64_read(
&dev_statistics[device_fh].rx_total_atomic);
@@ -2816,22 +2837,32 @@ print_stats(void)
rx_total = dev_statistics[device_fh].rx_total;
rx = dev_statistics[device_fh].rx;
}
- rx_dropped = rx_total - rx;
+ rx_total -= dev_statistics[device_fh].rx_total_p;
+ rx -= dev_statistics[device_fh].rx_p;
+ if (unlikely(rx_total < rx))
+ rx = rx_total;
+
+ dev_statistics[device_fh].rx_total_p += rx_total;
+ dev_statistics[device_fh].rx_p += rx;
+ dev_statistics[device_fh].tx_total_p += tx_total;
+ dev_statistics[device_fh].tx_p += tx;
printf("\nStatistics for device %"PRIu32"
------------------------------"
- "\nTX total: %"PRIu64""
- "\nTX dropped: %"PRIu64""
- "\nTX successful:
%"PRIu64""
- "\nRX total: %"PRIu64""
- "\nRX dropped: %"PRIu64""
- "\nRX successful:
%"PRIu64"",
+ "\n\tTX total:\t\t\t%"PRIu64""
+ "\n\tTX dropped:\t\t\t%"PRIu64""
+ "\n\tTX successful:\t\t\t%"PRIu64""
+ "\n\tRX total:\t\t\t%"PRIu64""
+ "\n\tRX dropped:\t\t\t%"PRIu64""
+ "\n\tRX successful:\t\t\t%"PRIu64"",
device_fh,
- tx_total,
- tx_dropped,
- tx,
- rx_total,
- rx_dropped,
- rx);
+ dev_statistics[device_fh].tx_total_p,
+ (dev_statistics[device_fh].tx_total_p
+ -
dev_statistics[device_fh].tx_p),
+ dev_statistics[device_fh].tx_p,
+ dev_statistics[device_fh].rx_total_p,
+ (dev_statistics[device_fh].rx_total_p
+ -
dev_statistics[device_fh].rx_p),
+ dev_statistics[device_fh].rx_p);
dev_ll = dev_ll->next;
}
--
2.1.4
