VF is still resetting

Nobin Mathew nobin.mathew at gmail.com
Wed Jul 20 06:56:09 CEST 2022


Any pointers?

i40e driver forum asked to submit question  here.

iavf in-kernel VF driver has a larger timeout,
 #define IAVF_RESET_WAIT_COMPLETE_COUNT  2000
and also reset detection loop timeout
#define IAVF_RESET_WAIT_DETECTED_COUNT  500


We tried even
+#define IAVF_RESET_WAIT_CNT       3000

but no luck....

-Nobin

On Tue, Jul 19, 2022 at 9:48 PM Nobin Mathew <nobin.mathew at gmail.com> wrote:
>
> Hi,
>
> We are running a dpdk app inside a pod, and orchestrating the app very
> frequently(test app).
>
> 1/100 or so we are getting an error:
>
> 2022-07-17T22:34:24.620291289+03:00 iavf_check_vf_reset_done(): reset
> VFR value: 3
> 2022-07-17T22:34:24.620310455+03:00 iavf_init_vf(): VF is still resetting
> 2022-07-17T22:34:24.620339697+03:00 iavf_dev_init(): Init vf failed
> 2022-07-17T22:34:24.620390802+03:00 EAL: Releasing PCI mapped resource
> for 0000:3b:0f.5
> 2022-07-17T22:34:24.620397381+03:00 EAL: Calling pci_unmap_resource
> for 0000:3b:0f.5 at 0x2101000000
> 2022-07-17T22:34:24.620442514+03:00 EAL: Calling pci_unmap_resource
> for 0000:3b:0f.5 at 0x2101010000
> 2022-07-17T22:34:24.729012277+03:00 EAL: Requested device 0000:3b:0f.5
> cannot be used
> 2022-07-17T22:34:24.729028758+03:00 EAL: Bus (pci) probe failed.
>
> we added one log in dpdk lib to print the VFGEN_RSTAT register of the
> VF. In problematic cases, we are seeing the value 3 which maps to
> 0xDEADBEEF
>
> / VF reset states - these are written into the RSTAT register:
> * VFGEN_RSTAT on the VF
> * When the PF initiates a reset, it writes 0
> * When the reset is complete, it writes 1
> * When the PF detects that the VF has recovered, it writes 2
> * VF checks this register periodically to determine if a reset has occurred,
> * then polls it to know when the reset is complete.
> * If either the PF or VF reads the register while the hardware
> * is in a reset state, it will return DEADBEEF, which, when masked
> * will result in 3.
> /
> enum virtchnl_vfr_states {
> VIRTCHNL_VFR_INPROGRESS = 0,
> VIRTCHNL_VFR_COMPLETED,
> VIRTCHNL_VFR_VFACTIVE,
> };
>
> We tried this patch also, increasing the poll time, no help.
> https://github.com/DPDK/dpdk/commit/be7226980c9ad4963b92b489c8afb17f08899953
>
> Details of the setup:
>
> DPDK library version
> 21.11
> VF Driver:-
> intel-iavf version 4.0.1-3.2
> PF driver:-
> sudo ethtool -i enp94s0f1
> driver: i40e
> version: 2.14.13
> firmware-version: 8.15 0x800096ca 20.0.17
>
> Since we are seeing 0xDEADBEEF, I am assuming VF-PF reset mailbox msg
> is received by PF, and PF initiated the RESET sequence by writing
> VFSWR to VPGEN_VFRTRIG register.
>
> I am not seeing
> " dev_err(&pf->pdev->dev, "VF reset check timeout on VF %d\n", "
> anywhere in syslog.
>
> Any pointers?, why does this happen(why VF reset is not complete)?...
>
> One more question, what is the sequence of calls in the reset path?
> i40e_vc_process_vf_msg() -> VIRTCHNL_OP_RESET_VF i40e_vc_reset_vf() ->
> i40e_reset_vf() -> i40e_trigger_vf_reset() & i40e_cleanup_reset_vf()
>
> this one?
>
> -Nobin


More information about the users mailing list