Thread (10 messages) 10 messages, 5 authors, 2021-01-12

Re: [dpdk-dev] [PATCH] net/iavf: fix vector id assignment

From: Yu, DapengX <hidden>
Date: 2021-01-12 06:46:21

Hi Ting,

The following test case is for i40e. if you want to test the case on ice, just ignore step 1. 

    Step1: switch i40evf to iavf
    # sed -i '/{ RTE_PCI_DEVICE(IAVF_INTEL_VENDOR_ID, IAVF_DEV_ID_ADAPTIVE_VF) },/a { RTE_PCI_DEVICE(IAVF_INTEL_VENDOR_ID, IAVF_DEV_ID_VF) },' drivers/net/iavf/iavf_ethdev.c
    # sed -i -e '/I40E_DEV_ID_VF/s/0x154C/0x164C/g'  drivers/net/i40e/base/i40e_devids.h
    # sed -i -e '/DEV_RX_OFFLOAD_CHECKSUM,/d' ./examples/l3fwd-power/main.c
    
    Step2. build dpdk and l3fwd-power
    
    Step3. create 1 vf
    # echo 1 > /sys/bus/pci/devices/0000\:81\:00.0/sriov_numvfs
    
    Step4. bind vf to vfio-pci
    # usertools/dpdk-devbind.py --force --bind=vfio-pci 81:02.0
    
    Step5. launch l3fwd-power
    # ./x86_64-native-linuxapp-gcc/examples/dpdk-l3fwd-power -l 0-1 -n 4 -- -P -p 0x1 --config='(0,0,0),(0,1,1)'

Expected result:
    start  l3fwd-power successfully
Actual result:
    fail to start l3fwd-power

-----Original Message-----
From: Xu, Ting 
Sent: Tuesday, January 12, 2021 2:27 PM
To: Yu, DapengX <redacted>; Zhang, Qi Z <redacted>; Wu, Jingjing <redacted>; Xing, Beilei <redacted>
Cc: dev@dpdk.org; Yu, DapengX <redacted>; stable@dpdk.org
Subject: RE: [PATCH] net/iavf: fix vector id assignment
-----Original Message-----
From: dapengx.yu@intel.com <redacted>
Sent: Friday, January 8, 2021 6:21 PM
To: Zhang, Qi Z <redacted>; Wu, Jingjing 
[off-list ref]; Xing, Beilei [off-list ref]; Xu, 
Ting [off-list ref]
Cc: dev@dpdk.org; Yu, DapengX <redacted>; stable@dpdk.org
Subject: [PATCH] net/iavf: fix vector id assignment

From: YU DAPENG <redacted>

The number of MSI-X interrupts on Rx shall be the minimal value of the 
number of available MSI-X interrupts per VF - 1 (the 1 is for 
miscellaneous
interrupt) and the number of configured Rx queues.
The current code break the rule because the number of available MSI-X 
interrupts is used as the first value, but code does not subtract 1 from it.

In normal situation, the first value is larger than the second value. 
So each queue can be assigned a unique vector_id.

For example: 17 available MSI-X interrupts, and 16 available Rx queues 
per VF; but only 4 Rx queues are configured when device is started.
vector_id:0 is for misc interrupt, vector_id:1 for Rx queue0,
vector_id:2 for Rx queue1, vector_id:3 for Rx queue2, vector_id:4 for 
Rx queue3.

Current code breaks the rule in this normal situation, because when 
assign vector_ids to interrupt handle, for example, it does not assign
vector_id:4 to the queue3, but assign vector_id:1 to it, because the 
condition used causes vector_id wrap around too early.
Hi, Dapeng,

Could you please further explain in which condition will this error happen? Seems it requires vf->nb_msix = 3 to make it happen, but I do not notice such situation.
I know it may be an example, is there any more specific case?

Thanks.
quoted hunk ↗ jump to hunk
In iavf_config_irq_map(), the current code does not write data into 
the last element of vecmap[], because of the previous code break. 
Which cause wrong data is sent to PF with opcode 
VIRTCHNL_OP_CONFIG_IRQ_MAP and cause
error: VIRTCHNL_STATUS_ERR_PARAM(-5).

If kernel driver supports large VFs (up to 256 queues), different 
queues can be assigned same vector_id.

In order to adapt to large VFs and avoid wrapping early, the condition 
is replaced from vec >= vf->nb_msix to vec >= vf->vf_res->max_vectors.

Fixes: d6bde6b5eae9 ("net/avf: enable Rx interrupt")
Cc: stable@dpdk.org

Signed-off-by: YU DAPENG <redacted>
---
 drivers/net/iavf/iavf_ethdev.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/drivers/net/iavf/iavf_ethdev.c 
b/drivers/net/iavf/iavf_ethdev.c index 7e3c26a94..d730bb156 100644
--- a/drivers/net/iavf/iavf_ethdev.c
+++ b/drivers/net/iavf/iavf_ethdev.c
@@ -483,6 +483,7 @@ static int iavf_config_rx_queues_irqs(struct 
rte_eth_dev *dev,  struct iavf_qv_map *qv_map;  uint16_t interval, i;  
int vec;
+uint16_t max_vectors;

 if (rte_intr_cap_multiple(intr_handle) &&
     dev->data->dev_conf.intr_conf.rxq) { @@ -570,15 +571,16 @@ static 
int iavf_config_rx_queues_irqs(struct rte_eth_dev *dev,
 /* If Rx interrupt is reuquired, and we can use
  * multi interrupts, then the vec is from 1
  */
-vf->nb_msix = RTE_MIN(vf->vf_res->max_vectors,
-      intr_handle->nb_efd);
+max_vectors =
+vf->vf_res->max_vectors -
IAVF_RX_VEC_START;
+vf->nb_msix = RTE_MIN(max_vectors, intr_handle-
quoted
nb_efd);
 vf->msix_base = IAVF_RX_VEC_START;
 vec = IAVF_RX_VEC_START;
 for (i = 0; i < dev->data->nb_rx_queues; i++) {  qv_map[i].queue_id = 
i;  qv_map[i].vector_id = vec;  intr_handle->intr_vec[i] = vec++; -if 
(vec >= vf->nb_msix)
+if (vec >= vf->vf_res->max_vectors)
 vec = IAVF_RX_VEC_START;
 }
 vf->qv_map = qv_map;
--
2.27.0
  
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help