[dpdk-dev] Issue with patch "app/testpmd: detect numa socket count"

Qiu, Michael michael.qiu at intel.com
Sat Dec 19 02:33:17 CET 2015


Sorry, I forgot to do that, add CC to mailing list 

Thanks,
Michael

> 在 2015年12月19日,上午7:32,Stephen Hurd <shurd at broadcom.com> 写道:
> 
> Thanks, I can reproduce it now.
> 
> The intent was to set max_socket when detecting lcores, not when selecting active ones, so there's a bug in my patch.  I'll actually be working on DPDK next week, so I'll have fix on Monday or Tuesday.
> 
> As for parsing /sys/, that's not likely to work on FreeBSD, and since we already detect all lcores and their socket, it makes sense to set max_socket there.
> 
> My request for the additional logging was so I could see the lines that correspond to these:
> 
> EAL: Detected lcore 0 as core 0 on socket 0
> EAL: Detected lcore 1 as core 0 on socket 1
> EAL: Detected lcore 2 as core 1 on socket 0
> EAL: Detected lcore 3 as core 1 on socket 1
> 
> Which would indicate that EAL was able to detect a higher max socket even though it's not enabled.
> 
> Is there a reason we're not CCing the mailing list?
> 
> 
> -- Stephen Hurd
> 
> 
> -----Original Message-----
> From: Qiu, Michael [mailto:michael.qiu at intel.com] 
> Sent: Thursday, December 17, 2015 10:31 PM
> To: Stephen Hurd; De Lara Guarch, Pablo
> Cc: Tan, Jianfeng; Gonzalez Monroy, Sergio
> Subject: Re: Issue with patch "app/testpmd: detect numa socket count"
> 
> That's the bug, I have two socket in my system, also from log you could see:
> 
> EAL: Requesting 512 pages of size 2MB from socket 1
> 
> command is : ./testpmd -c 0x3 -n 4 -- -i -socket-num=1
> 
> the root cause is that you only check the enabled lcore which is on
> socket 0, and then mark max socket as 1,  also your could has issue when
> in --numa
> 
> Because when numa enabled, still we run with coremask 0x03, the max
> socket will never be 2, which will lead lots of issues.
> 
> Thanks,
> Michael
> 
> 
>> On 2015/12/18 9:57, Stephen Hurd wrote:
>> Can you provide more log info (the socket to core mapping especially) and the entire command line (or at least the entire EAL set)?  The error suggests that you only have one numa socket (socket 0) on that system and are attempting to allocate memory on socket 1.
>> 
>> My patch redefined the max socket to be the last socket the system has whereas the old code allowed specifying sockets that aren't physically present in the system.
>> 
>> 
>> -- Stephen Hurd
>> 
>> 
>> -----Original Message-----
>> From: Qiu, Michael [mailto:michael.qiu at intel.com] 
>> Sent: Thursday, December 17, 2015 5:42 PM
>> To: Stephen Hurd; De Lara Guarch, Pablo
>> Subject: Issue with patch "app/testpmd: detect numa socket count"
>> 
>> Hi, Stephen
>> 
>> I just see this patch and found some issue with it.
>> 
>> When I start testpmd with -c 0x3 but with socket-num 1, that means run
>> lcore in socket 0, but want hugepage allocated in socket 1, in previous,
>> it works, I don't know why you force it in the socket lcore locates.I
>> would like to give a warning instead of failure. just like before:
>> 
>> EAL: Requesting 512 pages of size 2MB from socket 1
>> EAL: TSC frequency is ~2294689 KHz
>> EAL: WARNING: Master core has no memory on local socket!
>> 
>> After your patch:
>> 
>> EAL: No probed ethernet devices
>> Interactive-mode selected
>> EAL: Error - exiting with code: 1
>>  Cause: The socket number should be < 1
>> 
>> 
>> Thanks,
>> Michael
> 
> 


More information about the dev mailing list