Re: [dpdk-dev] [PATCH] app/testpmd: fix testpmd failure due to RSS offload check

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

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

-- 
Adrien Mazarguil
6WIND