Re: [dpdk-users] DPDK 16.04 link changes cause PMD drivers to not be loaded

Re: [dpdk-users] DPDK 16.04 link changes cause PMD drivers to not be loaded

[Aurojit Panda]

[Cross-posting to dev]

Panu Matilainen wrote:
> On 04/20/2016 05:26 PM, Aurojit Panda wrote:
>> I am sorry that is a bit unintuitive considering:
>>
>> (a) This behavior differs between static and shared builds of DPDK.
>> - In fact this behavior was identical in 2.2, and even in mainline
>> before 948fd64befc3726 went in.
>> (b) You already know the EAL_PMD_PATH at build time, and it makes it
>> quite hard to ship scripts or anything to build DPDK, since now the
>> configuration file becomes dependent on location.
>
> The expectation is that shared libraries are installed to a shared, known location. Otherwise there's not that much
> point in using shared libraries.
>
> Note that you can also use the EAL option -d to load either single PMDs or entire directories at runtime (regardless of
> EAL_PMD_PATH) so you dont have to know the path at build time if that's what bothers you most.
>

I am actually bothered by needing to supply the path at all. You seem to assume that knowing an install path is 
essential for shared libraries, while you can in reality embed this in the rpath. Shared libraries are used for many 
reasons, including interop with some runtime, in which case they are not installed at a known place. I would prefer for 
those PMDs that are included with DPDK that either they be loaded from the "known" build path for them.

>>
>> I wonder if you would consider changing this, as it stands just building
>> DPDK after setting CONFIG_RTE_SHARED_LIB=y results in a testpmd that
>> cannot run.
>
> In shared library configuration, testpmd is not directly runnable regardless of the drivers since its missing all the
> other libraries too:
>
> [pmatilai@sopuli dpdk]$ build/app/testpmd
> build/app/testpmd: error while loading shared libraries: librte_distributor.so.1.1: cannot open shared object file: No
> such file or directory
> [pmatilai@sopuli dpdk]$
>
> You'll need to get those libraries into linkers path anyway, either by installing them to a common location or by
> extending LD_LIBRARY_PATH. In either case, you need to know where the libraries are anyway.
>
> All of which is not to say there might not be room for improvement, but the linking behavior is not going to change.
> Again, PMDs are *plugins* that are *meant* to be loaded at runtime. That allows for all sorts of flexibility especially
> for packaging and shipping, at some extra cost in setup complexity.

I am all for a plugin architecture, I was merely suggesting that you embed some path infromation at the beginning. Also 
please note:
(a) This behavior changed recently.
(b) This change is entirely undocumented, which is why I was reporting it in the first place.
(c) It is actually quite unintutive, because previously ensuring LD_LIBRARY_PATH was correct was all that was required 
to get any DPDK application to interact with ports.

>
> For your own purposes, you can of course tweak the linking settings as much as you like. Look for "plugins" in
> mk/rte.app.mk and change the shared lib condition on the line above to "y" and there you have it. But that's not the way
> plugins are meant to be used.

That is not a reasonable solution given that it makes it very hard to track future changes to DPDK without merges. My 
alternatives neither break people's abilities to use plugins, nor do they impact behavior.

>
> Oh and BTW, please don't top-post.
>
> - Panu -
>

Re: [dpdk-users] DPDK 16.04 link changes cause PMD drivers to not be loaded

[Thomas Monjalon]

2016-04-21 08:01, Aurojit Panda:
> Panu Matilainen wrote:
[...]
> > Again, PMDs are *plugins* that are *meant* to be loaded at runtime.
> > That allows for all sorts of flexibility especially
> > for packaging and shipping, at some extra cost in setup complexity.
> 
> I am all for a plugin architecture, I was merely suggesting that you
> embed some path infromation at the beginning. Also please note:
> (a) This behavior changed recently.

What changed recently?

> (b) This change is entirely undocumented, which is why I was reporting
> it in the first place.
> (c) It is actually quite unintutive, because previously ensuring
> LD_LIBRARY_PATH was correct was all that was required 
> to get any DPDK application to interact with ports.

?
Are you talking about combined library?

> > For your own purposes, you can of course tweak the linking settings
> > as much as you like. Look for "plugins" in mk/rte.app.mk and change
> > the shared lib condition on the line above to "y" and there you have it.
> > But that's not the way plugins are meant to be used.
> 
> That is not a reasonable solution given that it makes it very hard to
> track future changes to DPDK without merges.
> My alternatives neither break people's abilities to use plugins,
> nor do they impact behavior.

Please do not hesitate to send some patch to show your solution.
Thanks

Re: [dpdk-users] DPDK 16.04 link changes cause PMD drivers to not be loaded

[Aurojit Panda]

Thomas Monjalon wrote:
> 2016-04-21 08:01, Aurojit Panda:
>> Panu Matilainen wrote:
> [...]
>>> Again, PMDs are *plugins* that are *meant* to be loaded at runtime.
>>> That allows for all sorts of flexibility especially
>>> for packaging and shipping, at some extra cost in setup complexity.
>> I am all for a plugin architecture, I was merely suggesting that you
>> embed some path infromation at the beginning. Also please note:
>> (a) This behavior changed recently.
>
> What changed recently?
>
Change 948fd64befc3726 moved DPDK from building a shared library to using LD linker scripts.

>> (b) This change is entirely undocumented, which is why I was reporting
>> it in the first place.
>> (c) It is actually quite unintutive, because previously ensuring
>> LD_LIBRARY_PATH was correct was all that was required
>> to get any DPDK application to interact with ports.
>
> ?
> Are you talking about combined library?
Yes, if you build a shared combined library, PMD drivers are not automatically loaded anymore. This implies that 
programs do not enumerate the set of ports on the machine. (Unless CONFIG_EAL_PMD_PATH is correctly set to the absolute 
directory where the built DPDK will live).
>
>>> For your own purposes, you can of course tweak the linking settings
>>> as much as you like. Look for "plugins" in mk/rte.app.mk and change
>>> the shared lib condition on the line above to "y" and there you have it.
>>> But that's not the way plugins are meant to be used.
>> That is not a reasonable solution given that it makes it very hard to
>> track future changes to DPDK without merges.
>> My alternatives neither break people's abilities to use plugins,
>> nor do they impact behavior.
>
> Please do not hesitate to send some patch to show your solution.

My initial e-mail was sent because I don't entirely understand LD linker scripts (the PMD libraries are listed within 
the built libdpdk.so file), so the only patch I can suggest is undoing 948fd64befc3726 in which case the behavior as I 
describe.

> Thanks


Thanks

Panda

Re: [dpdk-users] DPDK 16.04 link changes cause PMD drivers to not be loaded

[Panu Matilainen]

On 04/21/2016 06:24 PM, Aurojit Panda wrote:
>
>
> Thomas Monjalon wrote:
>> 2016-04-21 08:01, Aurojit Panda:
>>> Panu Matilainen wrote:
>> [...]
>>>> Again, PMDs are *plugins* that are *meant* to be loaded at runtime.
>>>> That allows for all sorts of flexibility especially
>>>> for packaging and shipping, at some extra cost in setup complexity.
>>> I am all for a plugin architecture, I was merely suggesting that you
>>> embed some path infromation at the beginning. Also please note:
>>> (a) This behavior changed recently.
>>
>> What changed recently?
>>
> Change 948fd64befc3726 moved DPDK from building a shared library to
> using LD linker scripts.
>
>>> (b) This change is entirely undocumented, which is why I was reporting
>>> it in the first place.
>>> (c) It is actually quite unintutive, because previously ensuring
>>> LD_LIBRARY_PATH was correct was all that was required
>>> to get any DPDK application to interact with ports.
>>
>> ?
>> Are you talking about combined library?
> Yes, if you build a shared combined library, PMD drivers are not
> automatically loaded anymore. This implies that programs do not
> enumerate the set of ports on the machine. (Unless CONFIG_EAL_PMD_PATH
> is correctly set to the absolute directory where the built DPDK will live).
>>
>>>> For your own purposes, you can of course tweak the linking settings
>>>> as much as you like. Look for "plugins" in mk/rte.app.mk and change
>>>> the shared lib condition on the line above to "y" and there you have
>>>> it.
>>>> But that's not the way plugins are meant to be used.
>>> That is not a reasonable solution given that it makes it very hard to
>>> track future changes to DPDK without merges.
>>> My alternatives neither break people's abilities to use plugins,
>>> nor do they impact behavior.
>>
>> Please do not hesitate to send some patch to show your solution.
>
> My initial e-mail was sent because I don't entirely understand LD linker
> scripts (the PMD libraries are listed within the built libdpdk.so file),
> so the only patch I can suggest is undoing 948fd64befc3726 in which case
> the behavior as I describe.

Reverting would be counter-productive, the change was introduced for a 
reason. Many of them in fact.

Any application wanting to force linkage to plugins always needed to 
link them inside --whole-archive section with static or individual 
shared libraries, now that is true for the combined shared library as well.

Within DPDK itself it would be easy enough to add an option to force 
link-in of the plugins (see the comment about mk/rte.app.mk change), but 
not sure we want to have yet another build option. And anything else 
still would need to use --whole-archive if they want that behavior.

So yes it's a behavior change for that particular corner which nobody 
thought of or brought up during review, and so ended up being 
undocumented too.

Like said there are always ways to improve things, once the use-case is 
understood, actual examples of how you are deploying DPDK would perhaps 
help. One possibility could be attempting to load known plugins from 
LD_LIBRARY_PATH, but that requires inserting that information somewhere 
in the binaries where it currently is not needed.

	- Panu -

Re: [dpdk-users] DPDK 16.04 link changes cause PMD drivers to not be loaded

[Aurojit Panda]

[The original report is included below for your convenience]

Thomas Monjalon wrote:
> 2016-04-21 08:01, Aurojit Panda:
>> Panu Matilainen wrote:
> [...]
>>> Again, PMDs are *plugins* that are *meant* to be loaded at runtime.
>>> That allows for all sorts of flexibility especially
>>> for packaging and shipping, at some extra cost in setup complexity.
>> I am all for a plugin architecture, I was merely suggesting that you
>> embed some path infromation at the beginning. Also please note:
>> (a) This behavior changed recently.
>
> What changed recently?
>
>> (b) This change is entirely undocumented, which is why I was reporting
>> it in the first place.
>> (c) It is actually quite unintutive, because previously ensuring
>> LD_LIBRARY_PATH was correct was all that was required
>> to get any DPDK application to interact with ports.
>
> ?
> Are you talking about combined library?
>
>>> For your own purposes, you can of course tweak the linking settings
>>> as much as you like. Look for "plugins" in mk/rte.app.mk and change
>>> the shared lib condition on the line above to "y" and there you have it.
>>> But that's not the way plugins are meant to be used.
>> That is not a reasonable solution given that it makes it very hard to
>> track future changes to DPDK without merges.
>> My alternatives neither break people's abilities to use plugins,
>> nor do they impact behavior.
>
> Please do not hesitate to send some patch to show your solution.
> Thanks

I was trying to run testpmd from DPDK 16.04 on Linux with kernel version 4.4.0-1 (ld version 2.26). My machine has two
XL710QDA2 NICs, and I built DPDK as a shared, combined library (i.e., CONFIG_RTE_BUILD_SHARED_LIB=y and
CONFIG_RTE_BUILD_COMBINE_LIBS=y in config/common_linuxapp). I found that the issue is due to ld not linking against all
libraries with the new linker script (introduced in 948fd64befc3726) but am not sure how to fix this. As evidence for
this being caused by this change:

$ LD_TRACE_LOADED_OBJECTS=y ./testpmd -c 0x1c00 -n 4 -w 82:00.0 -w 82:00.1 --file-prefix "send" -- -i
           linux-vdso.so.1 (0x00007ffe11b3e000)
           librte_distributor.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_distributor.so.1.1
(0x00007fa55eb85000)
           librte_reorder.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_reorder.so.1.1 (0x00007fa55e982000)
           librte_pipeline.so.3.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_pipeline.so.3.1 (0x00007fa55e77d000)
           librte_table.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_table.so.2.1 (0x00007fa55e55e000)
           librte_port.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_port.so.2.1 (0x00007fa55e34e000)
           librte_timer.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_timer.so.1.1 (0x00007fa55e145000)
           librte_hash.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_hash.so.2.1 (0x00007fa55df37000)
           librte_jobstats.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_jobstats.so.1.1 (0x00007fa55dd35000)
           librte_lpm.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_lpm.so.2.1 (0x00007fa55db2e000)
           librte_power.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_power.so.1.1 (0x00007fa55d91f000)
           librte_acl.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_acl.so.2.1 (0x00007fa55d705000)
           librte_meter.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_meter.so.1.1 (0x00007fa55d504000)
           librte_sched.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_sched.so.1.1 (0x00007fa55d2fd000)
           librte_vhost.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_vhost.so.2.1 (0x00007fa55d0e6000)
           librte_kvargs.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_kvargs.so.1.1 (0x00007fa55cee4000)
           librte_mbuf.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_mbuf.so.2.1 (0x00007fa55cce2000)
           librte_ip_frag.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_ip_frag.so.1.1 (0x00007fa55cada000)
           libethdev.so.3.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/libethdev.so.3.1 (0x00007fa55c84f000)
           librte_cryptodev.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_cryptodev.so.1.1 (0x00007fa55c647000)
           librte_mempool.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_mempool.so.1.1 (0x00007fa55c444000)
           librte_ring.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_ring.so.1.1 (0x00007fa55c242000)
           librte_eal.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_eal.so.2.1 (0x00007fa55bfe1000)
           librte_cmdline.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_cmdline.so.2.1 (0x00007fa55bdd8000)
           librte_cfgfile.so.2.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_cfgfile.so.2.1 (0x00007fa55bbd6000)
           librte_pmd_bond.so.1.1 => /home/apanda/e2d2/3rdparty/dpdk/build/lib/librte_pmd_bond.so.1.1 (0x00007fa55b9c4000)
           libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007fa55b7ae000)
           libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007fa55b5aa000)
           libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007fa55b38d000)
           libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007fa55afe9000)
           /lib64/ld-linux-x86-64.so.2 (0x00007fa55ed87000)
           libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007fa55aceb000)
           librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007fa55aae3000)


As you can see librte_pmd_i40e.so is missing from the list of files. Furthermore, using

$ LD_PRELOAD="../lib/librte_pmd_i40e.so" ./testpmd -c 0x1c00 -n 4 -w 82:00.0 -w 82:00.1 --file-prefix "send" -- -i
correctly enumerates the NICs.

I unfortunately do not how to fix the linker script as checked in, but wanted to report this bug in case someone has a fix.