[dpdk-dev] [PATCH] doc: update mlx4 usage and dependencies

[Olga Shern]

Hi Raghav, 

What OFED version did you install?

There are 2 ways to compile mlx4:
1. You can compile it dynamically with the libibverbs  that are coming with Mellanox_OFED 2.4
2. To get better performance: compile it statically with libibverbs and libmlx4 that can be downloaded from http://dpdk.org/download, the README inside librte_pmd_mlx4 explains how to do it
     If you compile the pmd statically then MLNX_OFED is needed only for kernel modules. 
 
Best Regards,
Olga


-----Original Message-----
From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Raghav Sethi
Sent: Friday, April 03, 2015 9:21 PM
To: Adrien Mazarguil; dev at dpdk.org; Xiaozhou Li
Subject: Re: [dpdk-dev] [PATCH] doc: update mlx4 usage and dependencies

Hi Adrien,

Quick update: So I just tried the setup again on a new machine with the exact same hardware - installed OFED and compiled DPDK 2.0.0. Works like a charm.

Still getting the issue on the old machine (also upgraded to 2.0.0 FWIW).
From the error message it appears that the libraries are incorrectly installed/configured. However, the OFED uninstall script removes the libraries in question (libmlx5-rdmav2.so, libibverbs/libmlx4-rdmav2.so,  libibverbs.so), I checked! In fact, I diffed the libibverbs.so from both machines, and they're exactly identical. So it seems either the OFED installation/uninstallation is somehow flawed, or the error message from testpmd is misleading.

Best,
Raghav

On Fri, Apr 3, 2015 at 12:52 PM Raghav Sethi <raghavs at cs.princeton.edu>
wrote:

> Hi Adrien,
>
> I've copied dev at dpdk.
>
> I re-installed Mellanox OFED, and then re-compiled DPDK without the 
> flags for the static library. I'm running Ubuntu 14.04. However, I 
> still get the same error message (below).
>
> I then tried re-installing the Mellanox libraries on dpdk.org, and 
> compiled another copy of DPDK with them statically linked. This 
> appears to work perfectly. I no longer get the permission issues I was 
> getting before, possibly that was unrelated/fixed by OFED. Now both 
> OFED, and OFED+ dpdk.org libs statically linked have the same issue:
>
> EAL: PCI device 0000:06:00.0 on NUMA socket 0
> EAL:   probe driver: 15b3:1003 librte_pmd_mlx4
> libibverbs: Warning: couldn't load driver '/usr/lib/libibverbs/libmlx5':
> /usr/lib/libibverbs/libmlx5-rdmav2.so: symbol ibv_exp_cmd_getenv, 
> version
> IBVERBS_1.1 not defined in file libibverbs.so.1 with link time 
> reference
> libibverbs: Warning: couldn't load driver '/usr/lib/libibverbs/libmlx4':
> /usr/lib/libibverbs/libmlx4-rdmav2.so: symbol 
> ibv_exp_cmd_destroy_flow, version IBVERBS_1.1 not defined in file 
> libibverbs.so.1 with link time reference
> libibverbs: Warning: no userspace device-specific driver found for
> /sys/class/infiniband_verbs/uverbs0
> PMD: librte_pmd_mlx4: cannot access device, is mlx4_ib loaded?
> EAL: No probed ethernet devices
>
> The OFED install was successful (and it apparently removes/overwrites 
> old versions of the libraries). The installer also checks the firmware 
> versions, and it says I'm on the latest. Loading the kernel modules is 
> also successful, so I can't for the life of me figure out what this issue is now.
>
> Again, just to be clear, the full list of steps was:
> 1. Install OFED, run /etc/init.d/openibd restart 2. Compile DPDK with 
> the CONFIG_RTE_LIBRTE_MLX4_PMD=y 3. Set up hugepages 4. Load kernel 
> modules 5. Run testpmd
>
> Best,
> Raghav
>
> On Fri, Apr 3, 2015 at 3:34 AM Adrien Mazarguil < 
> adrien.mazarguil at 6wind.com> wrote:
>
>> Hi Raghav,
>>
>> On Thu, Apr 02, 2015 at 10:05:57PM +0000, Raghav Sethi wrote:
>> > Hi Adrien,
>> >
>> > Thanks for your reply - I hope it's OK that I'm emailing you off-list.
>> I'm
>> > a grad student at Princeton working on a research project.
>>
>> It's fine, but let's discuss on the ML as much as possible next time 
>> so the community can benefit from it. Since very few people use this 
>> driver at the moment (DPDK 2.0 is not even out yet), I have my own 
>> questions for you.
>>
>> > Given your patch (which was very helpful, thanks), here is the 
>> > series of all steps that I'm taking to run testpmd on a ConnectX-3 
>> > EN 40G
>> >
>> > 1. Compile and install the mellanox-specific libraries listed here:
>> > http://dpdk.org/download/mlx4/2015-03-26-DPDK-v2.0.0/
>>
>> This is fine, but did you use the provided Makefile to compile these 
>> libraries statically? If so, no issues?
>>
>> Also you did not install Mellanox OFED? You should, because updated 
>> kernel modules come from that distribution. It won't work properly 
>> (or at all) without them.
>>
>> The libraries from DPDK.org are actually optional once you've 
>> installed Mellanox OFED. They only improve performance. Remember to 
>> update firmwares also.
>>
>> > 2. export EXTRA_CFLAGS and EXTRA_LDFLAGS 3. Create a new 
>> > configuration which includes the CONFIG_RTE_LIBRTE_MLX4_PMD=y flag.
>> > 4. Compile DPDK 2.0 with the new configuration using setup.sh 5. 
>> > Set up hugepages using setup.sh 6. Load the kernel modules 
>> > ib_uverbs mlx4_en mlx4_core mlx4_ib 7. Run testpmd with the right 
>> > PCI bus address
>> >
>> > However, I consistently get:
>> > sudo: unable to execute ./testpmd: Permission denied
>>
>> Now that doesn't make sense, isn't that file missing permissions or 
>> something? Are you running RHEL with a restrictive SELinux policy?
>> Is sudo configured to allow you to run testpmd?
>>
>> In any case, I don't think it's related to librte_pmd_mlx4.
>>
>> > I tried taking the interface down with ifconfig because 
>> > (./tools/dpdk_nic_bind.py --status shows the interface as active), 
>> > but
>> that
>> > didn't change anything.
>>
>> No, you should leave the interfaces up, it's not supposed to 
>> disappear, the PMD will bring them back up anyway. Actually the Linux 
>> kernel interface remains fully functional even while DPDK 
>> applications are running. You can send packets out of it but as long 
>> as both the kernel and DPDK applications share the same MAC address, 
>> only DPDK applications will receive replies.
>> If
>> you change the MAC address on the kernel interface however, you'll be 
>> able to use it normally.
>>
>> > I tried installing the mellanox libraries globally, but then I see 
>> > these error messages:
>> >
>> > -------
>> > EAL: PCI device 0000:06:00.0 on NUMA socket 0
>> > EAL:   probe driver: 15b3:1003 librte_pmd_mlx4
>> > libibverbs: Warning: couldn't load driver '/usr/lib/libibverbs/libmlx5':
>> > /usr/lib/libibverbs/libmlx5-rdmav2.so: symbol ibv_exp_cmd_getenv,
>> version
>> > IBVERBS_1.1 not defined in file libibverbs.so.1 with link time 
>> > reference
>> > libibverbs: Warning: couldn't load driver '/usr/lib/libibverbs/libmlx4':
>> > /usr/lib/libibverbs/libmlx4-rdmav2.so: symbol 
>> > ibv_exp_cmd_destroy_flow, version IBVERBS_1.1 not defined in file 
>> > libibverbs.so.1 with link time reference
>> > libibverbs: Warning: no userspace device-specific driver found for
>> > /sys/class/infiniband_verbs/uverbs0
>> > PMD: librte_pmd_mlx4: cannot access device, is mlx4_ib loaded?
>> > EAL: No probed ethernet devices
>> > Interactive-mode selected
>> > EAL: Error - exiting with code: 1
>> >   Cause: Creation of mbuf pool for socket 0 failed
>> > --------
>>
>> That's because the libibverbs version you're using is not compatible 
>> with the userspace drivers installed on your system (libibverbs uses 
>> dlopen() to load liblmx4). Even though you've installed both from 
>> DPDK.org, seems you have another version installed by your 
>> distribution packages.
>>
>> If you intend to use dynamic libraries, you should uninstall 
>> libibverbs,
>> libmlx4 and related packages first.
>>
>> Regards,
>>
>> --
>> Adrien Mazarguil
>> 6WIND
>>
>