[dpdk-dev] app/testpmd: fix testpmd failure due to RSS offload check
Checks
Commit Message
After add RSS hash offload check, default rss_hf will fail on
devices that not support all bits, the patch take rss_hf as
a suggest value and only set bits that device supported base on
rte_eth_dev_get_info.
Fixes: 527624c663f8 ("ethdev: add supported hash function check")
Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
---
app/test-pmd/testpmd.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
Comments
On 4/24/2018 3:18 PM, Qi Zhang wrote:
> After add RSS hash offload check, default rss_hf will fail on
> devices that not support all bits, the patch take rss_hf as
> a suggest value and only set bits that device supported base on
> rte_eth_dev_get_info.
>
> Fixes: 527624c663f8 ("ethdev: add supported hash function check")
> Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
> ---
> app/test-pmd/testpmd.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
> index d6da41927..af3fc5388 100644
> --- a/app/test-pmd/testpmd.c
> +++ b/app/test-pmd/testpmd.c
> @@ -2265,13 +2265,16 @@ init_port_config(void)
> {
> portid_t pid;
> struct rte_port *port;
> + struct rte_eth_dev_info dev_info;
>
> RTE_ETH_FOREACH_DEV(pid) {
> port = &ports[pid];
> port->dev_conf.fdir_conf = fdir_conf;
> if (nb_rxq > 1) {
> + rte_eth_dev_info_get(pid, &dev_info);
> port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
> - port->dev_conf.rx_adv_conf.rss_conf.rss_hf = rss_hf;
> + port->dev_conf.rx_adv_conf.rss_conf.rss_hf =
> + rss_hf & dev_info.flow_type_rss_offloads;
"rss_hf" is a global variable, not per port. Adrien's patch [1] made that it has
been updated by "port config all rss ..." command, so that value can change.
Also Xueming's patch did it will set "0" if "default" parameter used [2], like
"port config all rss default"
So not sure if rss_hf is reliable at this point.
What do you think using dev_info.flow_type_rss_offloads directly?
[1] 75d5493eb302 ("app/testpmd: fix RSS flow action configuration")
[2] 8c1f4aff92a6 ("app/testpmd: new parameter for port config all RSS command")
> } else {
> port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
> port->dev_conf.rx_adv_conf.rss_conf.rss_hf = 0;
>
On 4/24/2018 3:39 PM, Ferruh Yigit wrote:
> On 4/24/2018 3:18 PM, Qi Zhang wrote:
>> After add RSS hash offload check, default rss_hf will fail on
>> devices that not support all bits, the patch take rss_hf as
>> a suggest value and only set bits that device supported base on
>> rte_eth_dev_get_info.
>>
>> Fixes: 527624c663f8 ("ethdev: add supported hash function check")
>> Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
>> ---
>> app/test-pmd/testpmd.c | 5 ++++-
>> 1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
>> index d6da41927..af3fc5388 100644
>> --- a/app/test-pmd/testpmd.c
>> +++ b/app/test-pmd/testpmd.c
>> @@ -2265,13 +2265,16 @@ init_port_config(void)
>> {
>> portid_t pid;
>> struct rte_port *port;
>> + struct rte_eth_dev_info dev_info;
>>
>> RTE_ETH_FOREACH_DEV(pid) {
>> port = &ports[pid];
>> port->dev_conf.fdir_conf = fdir_conf;
>> if (nb_rxq > 1) {
>> + rte_eth_dev_info_get(pid, &dev_info);
>> port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
>> - port->dev_conf.rx_adv_conf.rss_conf.rss_hf = rss_hf;
>> + port->dev_conf.rx_adv_conf.rss_conf.rss_hf =
>> + rss_hf & dev_info.flow_type_rss_offloads;
>
> "rss_hf" is a global variable, not per port. Adrien's patch [1] made that it has
> been updated by "port config all rss ..." command, so that value can change.
>
> Also Xueming's patch did it will set "0" if "default" parameter used [2], like
> "port config all rss default"
>
> So not sure if rss_hf is reliable at this point.
>
> What do you think using dev_info.flow_type_rss_offloads directly?
Prevent updating rss_hf in "port config all rss default" and this patch together
lgtm.
>
>
> [1] 75d5493eb302 ("app/testpmd: fix RSS flow action configuration")
>
> [2] 8c1f4aff92a6 ("app/testpmd: new parameter for port config all RSS command")
>
>> } else {
>> port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
>> port->dev_conf.rx_adv_conf.rss_conf.rss_hf = 0;
>>
>
On Tue, Apr 24, 2018 at 05:13:46PM +0100, Ferruh Yigit wrote:
> On 4/24/2018 3:39 PM, Ferruh Yigit wrote:
> > On 4/24/2018 3:18 PM, Qi Zhang wrote:
> >> After add RSS hash offload check, default rss_hf will fail on
> >> devices that not support all bits, the patch take rss_hf as
> >> a suggest value and only set bits that device supported base on
> >> rte_eth_dev_get_info.
> >>
> >> Fixes: 527624c663f8 ("ethdev: add supported hash function check")
> >> Signed-off-by: Qi Zhang <qi.z.zhang@intel.com>
> >> ---
> >> app/test-pmd/testpmd.c | 5 ++++-
> >> 1 file changed, 4 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/app/test-pmd/testpmd.c b/app/test-pmd/testpmd.c
> >> index d6da41927..af3fc5388 100644
> >> --- a/app/test-pmd/testpmd.c
> >> +++ b/app/test-pmd/testpmd.c
> >> @@ -2265,13 +2265,16 @@ init_port_config(void)
> >> {
> >> portid_t pid;
> >> struct rte_port *port;
> >> + struct rte_eth_dev_info dev_info;
> >>
> >> RTE_ETH_FOREACH_DEV(pid) {
> >> port = &ports[pid];
> >> port->dev_conf.fdir_conf = fdir_conf;
> >> if (nb_rxq > 1) {
> >> + rte_eth_dev_info_get(pid, &dev_info);
> >> port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
> >> - port->dev_conf.rx_adv_conf.rss_conf.rss_hf = rss_hf;
> >> + port->dev_conf.rx_adv_conf.rss_conf.rss_hf =
> >> + rss_hf & dev_info.flow_type_rss_offloads;
> >
> > "rss_hf" is a global variable, not per port. Adrien's patch [1] made that it has
> > been updated by "port config all rss ..." command, so that value can change.
> >
> > Also Xueming's patch did it will set "0" if "default" parameter used [2], like
> > "port config all rss default"
> >
> > So not sure if rss_hf is reliable at this point.
> >
> > What do you think using dev_info.flow_type_rss_offloads directly?
>
> Prevent updating rss_hf in "port config all rss default" and this patch together
> lgtm.
I have to take back my previous reply [3] regarding the original series.
After testing it I confirm mlx4 is broken with testpmd as well. The fact
mlx4 doesn't implement the legacy RSS API means it won't ever satisfy the
global rss_hf settings and rte_eth_dev_configure() will always fail.
As an immediate fix, Qi's patch above should deal with this problem.
Now what's needed in testpmd is the ability to distinguish a *configured*
rss_hf value (--rss-ip, --rss-udp command-line parameters or interactively
through "port config all rss $anything_other_than_default") from a default
one (no command-line parameters or "port config all rss default").
I think a second global is needed:
int rss_hf_enable = 0; /* Take rss_hf into account. */
If unset, rss_hf is never taken into account and PMDs are fed their default
settings (whatever they report as supported), otherwise rss_hf is
enforced as requested by user configuration.
> > [1] 75d5493eb302 ("app/testpmd: fix RSS flow action configuration")
> >
> > [2] 8c1f4aff92a6 ("app/testpmd: new parameter for port config all RSS command")
> >
> >> } else {
> >> port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
> >> port->dev_conf.rx_adv_conf.rss_conf.rss_hf = 0;
> >>
> >
>
[3] http://dpdk.org/ml/archives/dev/2018-April/097904.html
@@ -2265,13 +2265,16 @@ init_port_config(void)
{
portid_t pid;
struct rte_port *port;
+ struct rte_eth_dev_info dev_info;
RTE_ETH_FOREACH_DEV(pid) {
port = &ports[pid];
port->dev_conf.fdir_conf = fdir_conf;
if (nb_rxq > 1) {
+ rte_eth_dev_info_get(pid, &dev_info);
port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
- port->dev_conf.rx_adv_conf.rss_conf.rss_hf = rss_hf;
+ port->dev_conf.rx_adv_conf.rss_conf.rss_hf =
+ rss_hf & dev_info.flow_type_rss_offloads;
} else {
port->dev_conf.rx_adv_conf.rss_conf.rss_key = NULL;
port->dev_conf.rx_adv_conf.rss_conf.rss_hf = 0;