[dpdk-dev] [PATCH] net/ixgbevf: fix stats update after a PF reset

Olivier Matz olivier.matz at 6wind.com
Wed Mar 29 12:08:08 CEST 2017


Hi Wei,

On Wed, 29 Mar 2017 09:22:55 +0000, "Dai, Wei" <wei.dai at intel.com> wrote:
> First of all, I don't agree the method in this patch because the register can also rewind to 0 in normal mode except
> reset/PF down.
> 
> In the function void ixgbe_down(struct ixgbe_adapter *adapter) of ixgbe_main.c in the ixgbe kernel PF driver (version 5.0.4),
> the PF in kernel driver will ping all the active vfs to let them know PF is going down by mailbox messages.
> 
> On other side, ixgbe VF in DPDK PMD will handle the ping mailbox message in ixgbevf_dev_interrupt_handler( ) which is
> registered interrupt routine. 
> And ixgbevf_dev_interrupt_handler( ) will call ixgbevf_mbx_process( ) to process mailbox message from PF.
> 
> So it is in ixgbevf_mbx_process( ) where the SW stats registers of VF can be fixed.
> There is more than one line to send ping message to VF in kernel PF drvier codes.

Thank you for the explanation.

Just to be sure I correctly understand:
1- the ping message is properly sent by ixgbe pf linux driver
2- the ping message is not yet sent by ixgbe pf dpdk driver
3- the ping message is properly received by ixgbe vf dpdk driver, but
   it is not used yet to stop wrong stats accumulation

> I am studying it to make it clear how to identify the ping message due to PF down.

Are you saying that you are currently working on a fix for 2 and 3 (that
would be great!) or are you expecting that I work on it? In any case, I
can help, at least to test the patches.


> So can we defer this patch first.
> And another patch in above way should be submitted.


I'm of course ok to defer the initial patch, I agree it's more a workaround
than a proper solution. I was stuck because I did not know how the VF is notified
by the PF when it is reset.


Thanks
Olivier



More information about the dev mailing list