On Thu, Jul 15, 2021 at 4:52 AM Nguyen, Anthony L <anthony.l.nguyen@xxxxxxxxx> wrote: > > On Fri, 2021-07-09 at 15:13 +0800, Jason Xing wrote: > > Oh, one more thing I missed in the last email is that all the > > failures > > are happening on the combination of X722 10GbE and 1GbE. So the value > > of @num_tx_qp the driver fetches is 384 while the value is 768 > > without x722 1GbE. > > > > I get that information back here: > > $ lspci | grep -i ether > > 5a:00.0 Ethernet controller: Intel Corporation Ethernet Connection > > X722 for 10GbE SFP+ (rev 09) > > 5a:00.1 Ethernet controller: Intel Corporation Ethernet Connection > > X722 for 10GbE SFP+ (rev 09) > > 5a:00.2 Ethernet controller: Intel Corporation Ethernet Connection > > X722 for 1GbE (rev 09) > > 5a:00.3 Ethernet controller: Intel Corporation Ethernet Connection > > X722 for 1GbE (rev 09) > > > > I know it's really stupid to control the number of online cpus, but > > finding a good way only to limit the @alloc_queue_pairs is not easy > > to > > go. So could someone point out a better way to fix this issue and > > take > > care of some relatively old nics with the number of cpus increasing? > > Hi Jason, > > Sorry for the slow response; I was trying to talk to the i40e team > about this. Thanks for your kind help really. It indeed has a big impact on thousands of machines. > > I agree, the limiting of number of online CPUs doesn't seem like a > solution we want to pursue. The team is working on a patch that deals As I said above, if the machine is equipped with only 10GbE nic, the maximum online cpus would be 256 and so on. For now, it depends on the num of cpus. > with the same, or similiar, issue; it is reworking the allocations of > the queue pile. I'll make sure that they add you on the patch when it It's not easy to cover all kinds of cases. But I still believe it's the only proper way to fix the issue. Looking forward to your patch :) > is sent so that you can try this and see if it resolves your issue. > Yeah, sure, I will double-check and then see if it's really fixed. Thanks, Jason > Thanks, > Tony >
