On 2/4/2021 10:02 AM, Ido Goshen wrote:
-----Original Message-----
From: Ferruh Yigit <ferruh.yi...@intel.com>
Sent: Thursday, 4 February 2021 11:26
To: Ido Goshen <i...@cgstowernetworks.com>
Cc: dev@dpdk.org
Subject: Re: [PATCH v3 1/1] net/pcap: imissed stats support
On 2/4/2021 7:56 AM, Ido Goshen wrote:
-----Original Message-----
From: Ferruh Yigit <ferruh.yi...@intel.com>
Sent: Thursday, 4 February 2021 2:13
To: Ido Goshen <i...@cgstowernetworks.com>
Cc: dev@dpdk.org
Subject: Re: [PATCH v3 1/1] net/pcap: imissed stats support
On 2/3/2021 11:07 PM, Ido Goshen wrote:
get value from pcap_stats.ps_drop (see man pcap_stats) the value is
adjusted in this cases:
- port stop - pcap is closed and will lose count
- stats reset - pcap doesn't provide reset api
- rollover - pcap counter size is u_32 only
Signed-off-by: Ido Goshen <i...@cgstowernetworks.com>
---
v3:
* code cleanup by dedicated struct and functions extraction
* multi stop support by menmonic+= accumulation
* rollover fixup
v2:
* sum all queues (rx_missed_total += fix)
* null pcap protection
* inter stop/start persistancy (counter won't reset on stop)
drivers/net/pcap/rte_eth_pcap.c | 59
+++++++++++++++++++++++++++++++++
1 file changed, 59 insertions(+)
diff --git a/drivers/net/pcap/rte_eth_pcap.c
b/drivers/net/pcap/rte_eth_pcap.c index a32b1f3f3..16e8752f3 100644
--- a/drivers/net/pcap/rte_eth_pcap.c
+++ b/drivers/net/pcap/rte_eth_pcap.c
@@ -60,11 +60,21 @@ struct queue_stat {
volatile unsigned long err_pkts;
};
+struct queue_missed_stat {
+ /* last value retrieved from pcap */
+ volatile unsigned int pcap;
+ /* stores values lost by pcap stop or rollover */
+ volatile unsigned long mnemonic;
+ /* value on last reset */
+ volatile unsigned long reset;
+};
I am aware other stats has 'volatile' keyword, but as far as I can
see it is not needed, since these are new ones can you please drop the
'volatile'?
ok
+
struct pcap_rx_queue {
uint16_t port_id;
uint16_t queue_id;
struct rte_mempool *mb_pool;
struct queue_stat rx_stat;
+ struct queue_missed_stat missed_stat;
char name[PATH_MAX];
char type[ETH_PCAP_ARG_MAXLEN];
@@ -144,6 +154,49 @@ RTE_LOG_REGISTER(eth_pcap_logtype,
pmd.net.pcap, NOTICE);
rte_log(RTE_LOG_ ## level, eth_pcap_logtype, \
"%s(): " fmt "\n", __func__, ##args)
+static struct queue_missed_stat*
+queue_missed_stat_update(struct rte_eth_dev *dev, unsigned int qid) {
+ struct pmd_internals *internals = dev->data->dev_private;
+ struct queue_missed_stat *missed_stat =
+ &internals->rx_queue[qid].missed_stat;
+ const struct pmd_process_private *pp = dev->process_private;
+ pcap_t *pcap = pp->rx_pcap[qid];
+ struct pcap_stat stat;
Can you please put an empty line after variable declarations, and before
return.
ok
+ if (!pcap || (pcap_stats(pcap, &stat) != 0))
+ return missed_stat;
+ /* rollover check - best effort fixup assuming single rollover */
+ if (stat.ps_drop < missed_stat->pcap)
+ missed_stat->mnemonic += UINT_MAX;
+ missed_stat->pcap = stat.ps_drop;
here.
+ return missed_stat;
+}
+
+static void
+queue_missed_stat_on_stop_update(struct rte_eth_dev *dev, unsigned
+int qid) {
+ struct queue_missed_stat *missed_stat =
+ queue_missed_stat_update(dev, qid);
+ missed_stat->mnemonic += missed_stat->pcap;
Better to reset 'missed_stat->pcap' afterwards, in case stats
requested before port started again:
missed_stat->pcap = 0;
right, should be careful not to double count it but maybe better to
set it to 0 in queue_missed_stat_update in the stop return case
if (!pcap || (pcap_stats(pcap, &stat) != 0))
{
missed_stat->pcap = 0;
return missed_stat;
}
this way the missed_stat->pcap will always represent the current value
from pcap and not hold old value specifically in case port is stopped
it will be 0 and not re-added agree?
This also works, but there is other condition in that if block, I don't know
when
'pcap_stats()' can fail, it has a risk of unexpected side affect to clear the
stats
randomly, I think safer to do it in 'queue_missed_stat_on_stop_update()' and I
believe logic is more clear that way.
ok
+}
+
+static void
+queue_missed_stat_reset(struct rte_eth_dev *dev, unsigned int qid) {
+ struct queue_missed_stat *missed_stat =
+ queue_missed_stat_update(dev, qid);
+ missed_stat->reset = missed_stat->pcap;
I guess this should be:
"missed_stat->reset = missed_stat->pcap + missed_stat->mnemonic;"
I don't think so
reset should only remember where pcap was at the reset point and not store
old values
trying it immediately results in
testpmd> show port stats 0
######################## NIC statistics for port 0
########################
RX-packets: 0 RX-missed: 1940 RX-bytes: 0
RX-errors: 0
RX-nombuf: 0
TX-packets: 0 TX-errors: 0 TX-bytes: 0
Throughput (since last show)
Rx-pps: 0 Rx-bps: 0
Tx-pps: 0 Tx-bps: 0
#################################################################
###########
testpmd> clear port stats 0
NIC statistics for port 0 cleared
testpmd> show port stats 0
######################## NIC statistics for port 0
########################
RX-packets: 0 RX-missed: 18446744073709550646 RX-bytes: 0
RX-errors: 0
RX-nombuf: 0
TX-packets: 0 TX-errors: 0 TX-bytes: 0
Throughput (since last show)
Rx-pps: 0 Rx-bps: 0
Tx-pps: 0 Tx-bps: 0
#################################################################
###########
What above shows, is it output after suggested change?
Yes it is **with** the suggested change (missed_stat->reset = missed_stat->pcap +
missed_stat->mnemonic;)
Assume this sequence
port stop -> mnemonic = N, pcap = 0
reset stats -> reset = 0 + N, mnemonic = 0
show -> imissed = 0 + 0 - N
it works good w/o the suggested change (missed_stat->reset = missed_stat->pcap;)
testpmd> show port stats 0
######################## NIC statistics for port 0 ########################
RX-packets: 0 RX-missed: 3932 RX-bytes: 0
RX-errors: 0
RX-nombuf: 0
TX-packets: 0 TX-errors: 0 TX-bytes: 0
Throughput (since last show)
Rx-pps: 0 Rx-bps: 0
Tx-pps: 0 Tx-bps: 0
############################################################################
testpmd> clear port stats 0
NIC statistics for port 0 cleared
testpmd> show port stats 0
######################## NIC statistics for port 0 ########################
RX-packets: 0 RX-missed: 0 RX-bytes: 0
RX-errors: 0
RX-nombuf: 0
TX-packets: 0 TX-errors: 0 TX-bytes: 0
Throughput (since last show)
Rx-pps: 0 Rx-bps: 0
Tx-pps: 0 Tx-bps: 0
############################################################################
I missed 'mnemonic' get zeroed out in the "clear stats", existing implementation
looks good, thanks for clarification.
missed = "(pcap + mnemonic) - reset"
Lets assume overflow limit is 16, with original code:
reset: 0, pcap: 2, mnemonic: 16, missed: 18
clear stats:
reset: 2, pcap: 2, mnemonic: 16, missed: 16 (wrong, it should be 0)
OR
reset: 0, pcap: 2, mnemonic: 0, missed: 2
port stop:
reset: 0, pcap: 0, mnemonic: 2, missed: 2
port start, some traffic:
reset: 0, pcap: 3, mnemonic: 2, missed: 5
clear stats:
reset: 3, pcap: 3, mnemonic: 2, missed: 2 (wrong, it should be 0)
'mnemonic' becomes part of the stats after overflow or "port stop", so I think
it should be stored in the reset.
As far as I can see suggested code change is required if I am not missing
anything.
+ missed_stat->mnemonic = 0;
+}
+
+static unsigned long
+queue_missed_stat_get(struct rte_eth_dev *dev, unsigned int qid) {
+ const struct queue_missed_stat *missed_stat =
+ queue_missed_stat_update(dev, qid);
+ return missed_stat->pcap + missed_stat->mnemonic -
+missed_stat->reset; }
+
<...>
