[dpdk-dev] [PATCH] bus/vdev: scan and probe vdev in secondary processes

Tan, Jianfeng jianfeng.tan at intel.com
Fri Dec 1 06:59:27 CET 2017


Hi Kyle,


On 12/1/2017 6:27 AM, Kyle Larose wrote:
> Hey Jianfeng,
>
>> -----Original Message-----
>> From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Jianfeng Tan
>> Sent: Thursday, November 30, 2017 1:59 PM
>> To: dev at dpdk.org
>> Cc: Jianfeng Tan
>> Subject: [dpdk-dev] [PATCH] bus/vdev: scan and probe vdev in
>> secondary processes
>>
>> Base on primary/secondary communication channel [1], we add vdev
>> action to scan virtual devices in secondary processes.
>>
>>    [1] http://dpdk.org/dev/patchwork/patch/31838/
>>
>> Signed-off-by: Jianfeng Tan <jianfeng.tan at intel.com>
>> ---
>>   drivers/bus/vdev/vdev.c | 104
>> ++++++++++++++++++++++++++++++++++++++++++++++--
>>   1 file changed, 100 insertions(+), 4 deletions(-)
>>
>
>
> We recently stumbled across a problem where running dpdk-pdump against
> a primary which had an rte_eth_bond virtual device would lead to a crash.
>
> We realized this was happening because the vdev bus wasn't being built
> properly on the secondary, leading to it corrupting the ethdev structures
> for the eth_bond device on the primary. This corruption occurred because pdump
> used what it thought was the first free dpdk port id for its pcap_pmd. That
> port id collided with the eth_bond pmd, causing the problem.
>
> My feeling is that this patch will fix that problem, since probing the virtual
> bus from the primary, rather than just from the eal cmdline arguments, will
> allow the secondary to choose an used port id. Am I correct in my understanding?

Yes, that's an issue we want to address in this patch set: those portid 
(either physical port or virtual port) will be synchronized to 
secondary, if the secondary wants to add port (private to the 
secondary), it will use unused portid.

But we still not address the problem that, both the primary and the 
secondary are allocating new devices which needs a centralized portid 
management mechanism.

Thanks,
Jianfeng

>
> Thanks,
>
> Kyle
>
>



More information about the dev mailing list