On 11/07/2018 04:17 AM, Zhao1, Wei wrote:
Hi, Luca Boccassi
The purpose of this patch is to reduce the mailbox interrupt from vf to
pf, but there seem some point need for discussion in this patch.
First, I do not know why do you change code of function ixgbe_check_mac_link_vf(), because in rte_eth_link_get_nowait() and rte_eth_link_get(),
it will call ixgbe_dev_link_update()->ixgbe_dev_link_update_share()->
ixgbevf_check_link() for VF, NOT ixgbe_check_mac_link_vf() in your patch!
Second, in function ixgbevf_check_link(), there is mailbox message read
operation for vf,
" if (mbx->ops.read(hw, &in_msg, 1, 0))", that is ixgbe_read_mbx_vf() ,
This will cause interrupt from vf to pf, this is just the point of this patch,
it is also the problem that you want to solve.
So, you use autoneg_wait_to_complete flag to control this mailbox message read
operation, maybe you will use rte_eth_link_get_nowait(), Which set
autoneg_wait_to_complete = 0, then the interrupt from vf to pf can be reduced.
But I do not think this patch is necessary, because in ixgbevf_check_link(),
it,has
I think you are right here. This patch dates to before the addition
of the vf argument to ixgbe_dev_link_update_share() and the split of
.link_update between ixgbe and ixgbevf. At one point, this patch was
especially beneficial if you were running bonding (which tends to make
quite a few link status checks).
So this patch probably hasn't been helping at this point. I will try
to get some time to locally test this.
"
bool no_pflink_check = wait_to_complete == 0;
////////////////////////
if (no_pflink_check) {
if (*speed == IXGBE_LINK_SPEED_UNKNOWN)
mac->get_link_status = true;
else
mac->get_link_status = false;
goto out;
}
"
Comment of "for a quick link status checking, wait_to_compelet == 0, skip PF link
status checking " is clear.
That means in rte_eth_link_get_nowait(), code will skip this mailbox read
interrupt, only in
rte_eth_link_get() there will be this interrupt, so I think what you need to is
just replace
rte_eth_link_get() with rte_eth_link_get_nowait() in your APP,
that will reduce interrupt from vf to pf in mailbox read.
-----Original Message-----
From: dev [mailto:dev-boun...@dpdk.org] On Behalf Of Luca Boccassi
Sent: Wednesday, August 15, 2018 10:15 PM
To: dev@dpdk.org
Cc: Lu, Wenzhuo <wenzhuo...@intel.com>; Ananyev, Konstantin
<konstantin.anan...@intel.com>; Luca Boccassi <bl...@debian.org>;
sta...@dpdk.org
Subject: [dpdk-dev] [PATCH] net/ixgbe: reduce PF mailbox interrupt rate
We have observed high rate of NIC PF interrupts when VNF is using DPDK
APIs rte_eth_link_get_nowait() and rte_eth_link_get() functions, as they
are causing VF driver to send many MBOX ACK messages.
With these changes, the interrupt rates go down significantly. Here's some
testing results:
Without the patch:
$ egrep 'CPU|ens1f' /proc/interrupts ; sleep 10; egrep 'CPU|ens1f'
/proc/interrupts
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5
CPU6 CPU7
CPU8 CPU9 CPU10 CPU11 CPU12 CPU13 CPU14
CPU15
34: 88 0 0 0 0 41
30 509 0 0 350
24 88 114 461 562 PCI-MSI 1572864-edge
ens1f0-TxRx-0
35: 49 24 0 0 65 130
64 29 67 0 10
0 0 46 38 764 PCI-MSI 1572865-edge
ens1f0-TxRx-1
36: 53 0 0 64 15 85
132 71 108 0
30 0 165 215 303 104 PCI-MSI
1572866-edge ens1f0-
TxRx-2
37: 46 196 0 0 10 48
62 68 51 0 0
0 103 82 54 192 PCI-MSI 1572867-edge
ens1f0-TxRx-3
38: 226 0 0 0 159 145
749 265 0 0
202 0 69229 166 450 0 PCI-MSI
1572868-edge ens1f0
52: 95 896 0 0 0 18
53 0 494 0 0
0 0 265 79 124 PCI-MSI 1574912-edge
ens1f1-TxRx-0
53: 50 0 18 0 72 33
0 168 330 0 0
0 141 22 12 65 PCI-MSI 1574913-edge
ens1f1-TxRx-1
54: 65 0 0 0 239 104
166 49 442 0
0 0 126 26 307 0 PCI-MSI 1574914-edge
ens1f1-TxRx-2
55: 57 0 0 0 123 35
83 54 157 106
0 0 26 29 312 97 PCI-MSI 1574915-edge
ens1f1-TxRx-3
56: 232 0 13910 0 16 21
0 54422 0 0
0 24 25 0 78 0 PCI-MSI 1574916-edge
ens1f1
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5
CPU6 CPU7
CPU8 CPU9 CPU10 CPU11 CPU12 CPU13 CPU14
CPU15
34: 88 0 0 0 0 41
30 509 0 0 350
24 88 119 461 562 PCI-MSI 1572864-edge
ens1f0-TxRx-0
35: 49 24 0 0 65 130
64 29 67 0 10
0 0 46 38 771 PCI-MSI 1572865-edge
ens1f0-TxRx-1
36: 53 0 0 64 15 85
132 71 108 0
30 0 165 215 303 113 PCI-MSI
1572866-edge ens1f0-
TxRx-2
37: 46 196 0 0 10 48
62 68 56 0 0
0 103 82 54 192 PCI-MSI 1572867-edge
ens1f0-TxRx-3
38: 226 0 0 0 159 145
749 265 0 0
202 0 71281 166 450 0 PCI-MSI
1572868-edge ens1f0
52: 95 896 0 0 0 18
53 0 494 0 0
0 0 265 79 133 PCI-MSI 1574912-edge
ens1f1-TxRx-0
53: 50 0 18 0 72 33
0 173 330 0 0
0 141 22 12 65 PCI-MSI 1574913-edge
ens1f1-TxRx-1
54: 65 0 0 0 239 104
166 49 442 0
0 0 126 26 312 0 PCI-MSI 1574914-edge
ens1f1-TxRx-2
55: 57 0 0 0 123 35
83 59 157 106
0 0 26 29 312 97 PCI-MSI 1574915-edge
ens1f1-TxRx-3
56: 232 0 15910 0 16 21
0 54422 0 0
0 24 25 0 78 0 PCI-MSI 1574916-edge
ens1f1
During the 10s interval, CPU2 jumped by 2000 interrupts, CPU12 by 2051
interrupts, for about 200 interrupts/second. That's on the order of what we
expect. I would have guessed 100/s but perhaps there are two mailbox
messages.
With the patch:
$ egrep 'CPU|ens1f' /proc/interrupts ; sleep 10; egrep 'CPU|ens1f'
/proc/interrupts
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5
CPU6 CPU7
CPU8 CPU9 CPU10 CPU11 CPU12 CPU13 CPU14
CPU15
34: 88 0 0 0 0 25
19 177 0 0 350
24 88 100 362 559 PCI-MSI 1572864-edge
ens1f0-TxRx-0
35: 49 19 0 0 65 130
64 29 67 0 10
0 0 46 38 543 PCI-MSI 1572865-edge
ens1f0-TxRx-1
36: 53 0 0 64 15 53
85 71 108 0 24
0 85 215 292 31 PCI-MSI 1572866-edge
ens1f0-TxRx-2
37: 46 196 0 0 10 43
57 39 19 0 0
0 78 69 49 149 PCI-MSI 1572867-edge
ens1f0-TxRx-3
38: 226 0 0 0 159 145
749 247 0 0
202 0 58250 0 450 0 PCI-MSI
1572868-edge ens1f0
52: 95 896 0 0 0 18
53 0 189 0 0
0 0 265 79 25 PCI-MSI 1574912-edge
ens1f1-TxRx-0
53: 50 0 18 0 72 33
0 90 330 0 0
0 136 5 12 0 PCI-MSI 1574913-edge
ens1f1-TxRx-1
54: 65 0 0 0 10 104
166 49 442 0 0
0 126 26 226 0 PCI-MSI 1574914-edge
ens1f1-TxRx-2
55: 57 0 0 0 61 35
83 30 157 101 0
0 26 15 312 0 PCI-MSI 1574915-edge
ens1f1-TxRx-3
56: 232 0 2062 0 16 21
0 54422 0 0
0 24 25 0 78 0 PCI-MSI 1574916-edge
ens1f1
CPU0 CPU1 CPU2 CPU3 CPU4 CPU5
CPU6 CPU7
CPU8 CPU9 CPU10 CPU11 CPU12 CPU13 CPU14
CPU15
34: 88 0 0 0 0 25
19 177 0 0 350
24 88 102 362 562 PCI-MSI 1572864-edge
ens1f0-TxRx-0
35: 49 19 0 0 65 130
64 29 67 0 10
0 0 46 38 548 PCI-MSI 1572865-edge
ens1f0-TxRx-1
36: 53 0 0 64 15 53
85 71 108 0 24
0 85 215 292 36 PCI-MSI 1572866-edge
ens1f0-TxRx-2
37: 46 196 0 0 10 45
57 39 19 0 0
0 78 69 49 152 PCI-MSI 1572867-edge
ens1f0-TxRx-3
38: 226 0 0 0 159 145
749 247 0 0
202 0 58259 0 450 0 PCI-MSI
1572868-edge ens1f0
52: 95 896 0 0 0 18
53 0 194 0 0
0 0 265 79 25 PCI-MSI 1574912-edge
ens1f1-TxRx-0
53: 50 0 18 0 72 33
0 95 330 0 0
0 136 5 12 0 PCI-MSI 1574913-edge
ens1f1-TxRx-1
54: 65 0 0 0 10 104
166 49 442 0 0
0 126 26 231 0 PCI-MSI 1574914-edge
ens1f1-TxRx-2
55: 57 0 0 0 66 35
83 30 157 101 0
0 26 15 312 0 PCI-MSI 1574915-edge
ens1f1-TxRx-3
56: 232 0 2071 0 16 21
0 54422 0 0
0 24 25 0 78 0 PCI-MSI 1574916-edge
ens1f1
Note the interrupt rate has gone way down. During the 10s interval, we only
saw a handful of interrupts.
Note that this patch was originally provided by Intel directly to AT&T and
Vyatta, but unfortunately I am unable to find records of the exact author.
We have been using this in production for more than a year.
Fixes: af75078fece3 ("first public release")
Cc: sta...@dpdk.org
Signed-off-by: Luca Boccassi <bl...@debian.org>
---
drivers/net/ixgbe/base/ixgbe_vf.c | 33 ++++++++++++++++---------------
1 file changed, 17 insertions(+), 16 deletions(-)
diff --git a/drivers/net/ixgbe/base/ixgbe_vf.c
b/drivers/net/ixgbe/base/ixgbe_vf.c
index 5b25a6b4d4..16086670b1 100644
--- a/drivers/net/ixgbe/base/ixgbe_vf.c
+++ b/drivers/net/ixgbe/base/ixgbe_vf.c
@@ -586,7 +586,6 @@ s32 ixgbe_check_mac_link_vf(struct ixgbe_hw *hw,
ixgbe_link_speed *speed,
s32 ret_val = IXGBE_SUCCESS;
u32 links_reg;
u32 in_msg = 0;
- UNREFERENCED_1PARAMETER(autoneg_wait_to_complete);
/* If we were hit with a reset drop the link */
if (!mbx->ops.check_for_rst(hw, 0) || !mbx->timeout) @@ -643,23
+642,25 @@ s32 ixgbe_check_mac_link_vf(struct ixgbe_hw *hw,
ixgbe_link_speed *speed,
*speed = IXGBE_LINK_SPEED_UNKNOWN;
}
- /* if the read failed it could just be a mailbox collision, best wait
- * until we are called again and don't report an error
- */
- if (mbx->ops.read(hw, &in_msg, 1, 0))
- goto out;
+ if (autoneg_wait_to_complete) {
+ /* if the read failed it could just be a mailbox collision, best
wait
+ * until we are called again and don't report an error
+ */
+ if (mbx->ops.read(hw, &in_msg, 1, 0))
+ goto out;
- if (!(in_msg & IXGBE_VT_MSGTYPE_CTS)) {
- /* msg is not CTS and is NACK we must have lost CTS status
*/
- if (in_msg & IXGBE_VT_MSGTYPE_NACK)
+ if (!(in_msg & IXGBE_VT_MSGTYPE_CTS)) {
+ /* msg is not CTS and is NACK we must have lost CTS
status */
+ if (in_msg & IXGBE_VT_MSGTYPE_NACK)
+ ret_val = -1;
+ goto out;
+ }
+
+ /* the pf is talking, if we timed out in the past we reinit */
+ if (!mbx->timeout) {
ret_val = -1;
- goto out;
- }
-
- /* the pf is talking, if we timed out in the past we reinit */
- if (!mbx->timeout) {
- ret_val = -1;
- goto out;
+ goto out;
+ }
}
/* if we passed all the tests above then the link is up and we no
--
2.18.0
