[dpdk-dev] [PATCH 4/4] pmd_hw_support.py: Add tool to query binaries for hw support information

[Neil Horman]

On Wed, May 18, 2016 at 03:48:12PM +0300, Panu Matilainen wrote:
> On 05/18/2016 03:03 PM, Neil Horman wrote:
> > On Wed, May 18, 2016 at 02:48:30PM +0300, Panu Matilainen wrote:
> > > On 05/16/2016 11:41 PM, Neil Horman wrote:
> > > > This tool searches for the primer sting PMD_DRIVER_INFO= in any ELF binary,
> > > > and, if found parses the remainder of the string as a json encoded string,
> > > > outputting the results in either a human readable or raw, script parseable
> > > > format
> > > > 
> > > > Signed-off-by: Neil Horman <nhorman at tuxdriver.com>
> > > > CC: Bruce Richardson <bruce.richardson at intel.com>
> > > > CC: Thomas Monjalon <thomas.monjalon at 6wind.com>
> > > > CC: Stephen Hemminger <stephen at networkplumber.org>
> > > > CC: Panu Matilainen <pmatilai at redhat.com>
> > > > ---
> > > >  tools/pmd_hw_support.py | 174 ++++++++++++++++++++++++++++++++++++++++++++++++
> > > >  1 file changed, 174 insertions(+)
> > > >  create mode 100755 tools/pmd_hw_support.py
> > > > 
> > > > diff --git a/tools/pmd_hw_support.py b/tools/pmd_hw_support.py
> > > > new file mode 100755
> > > > index 0000000..0669aca
> > > > --- /dev/null
> > > > +++ b/tools/pmd_hw_support.py
> > > > @@ -0,0 +1,174 @@
> > > > +#!/usr/bin/python3
> > > 
> > > I think this should use /usr/bin/python to be consistent with the other
> > > python scripts, and like the others work with python 2 and 3. I only tested
> > > it with python2 after changing this and it seemed to work fine so the
> > > compatibility side should be fine as-is.
> > > 
> > Sure, I can change the python executable, that makes sense.
> > 
> > > On the whole, AFAICT the patch series does what it promises, and works for
> > > both static and shared linkage. Using JSON formatted strings in an ELF
> > > section is a sound working technical solution for the storage of the data.
> > > But the difference between the two cases makes me wonder about this all...
> > You mean the difference between checking static binaries and dynamic binaries?
> > yes, there is some functional difference there
> > 
> > > 
> > > For static library build, you'd query the application executable, eg
> > Correct.
> > 
> > > testpmd, to get the data out. For a shared library build, that method gives
> > > absolutely nothing because the data is scattered around in individual
> > > libraries which might be just about wherever, and you need to somehow
> > Correct, I figured that users would be smart enough to realize that with
> > dynamically linked executables, they would need to look at DSO's, but I agree,
> > its a glaring diffrence.
> 
> Being able to look at DSOs is good, but expecting the user to figure out
> which DSOs might be loaded and not and where to look is going to be well
> above many users. At very least it's not what I would call user-friendly.
> 
I disagree, there is no linkage between an application and the dso's it opens
via dlopen that is exportable.  The only way to handle that is to have a
standard search path for the pmd_hw_info python script.  Thats just like modinfo
works (i.e. "modinfo bnx2" finds the bnx2 module for the running kernel).  We
can of course do something simmilar, but we have no existing implicit path
information to draw from to do that (because you can have multiple dpdk installs
side by side).  The only way around that is to explicitly call out the path on
the command line.

> > > discover the location + correct library files to be able to query that. For
> > > the shared case, perhaps the script could be taught to walk files in
> > > CONFIG_RTE_EAL_PMD_PATH to give in-the-ballpark correct/identical results
> > My initial thought would be to run ldd on the executable, and use a heuristic to
> > determine relevant pmd DSO's, and then feed each of those through the python
> > script.  I didn't want to go to that trouble unless there was consensus on it
> > though.
> 
> Problem is, ldd doesn't know about them either because the pmds are not
> linked to the executables at all anymore. They could be force-linked of
> course, but that means giving up the flexibility of plugins, which IMO is a
> no-go. Except maybe as an option, but then that would be a third case to
> support.
> 
Thats not true at all, or at least its a perfectly valid way to link the DSO's
in at link time via -lrte_pmd_<driver>.  Its really just the dlopen case we need
to worry about.  I would argue that, if they're not explicitly linked in like
that, then its correct to indicate that an application supports no hardware,
because it actually doesn't, it only supports the pmds that it chooses to list
on the command line.  And if a user is savy enough to specify a pmd on the
application command line, then they are perfectly capable of specifying the same
path to the hw_info script.

> 
> > 
> > > when querying the executable as with static builds. If identical operation
> > > between static and shared versions is a requirement (without running the app
> > > in question) then query through the executable itself is practically the
> > > only option. Unless some kind of (auto-generated) external config file
> > > system ala kernel depmod / modules.dep etc is brought into the picture.
> > Yeah, I'm really trying to avoid that, as I think its really not a typical part
> > of how user space libraries are interacted with.
> > 
> > > 
> > > For shared library configurations, having the data in the individual pmds is
> > > valuable as one could for example have rpm autogenerate provides from the
> > > data to ease/automate installation (in case of split packaging and/or 3rd
> > > party drivers). And no doubt other interesting possibilities. With static
> > > builds that kind of thing is not possible.
> > Right.
> > 
> > Note, this also leaves out PMD's that are loaded dynamically (i.e. via dlopen).
> > For those situations I don't think we have any way of 'knowing' that the
> > application intends to use them.
> 
> Hence my comment about CONFIG_RTE_EAL_PMD_PATH above, it at least provides a
> reasonable heuristic of what would be loaded by the app when run. But
> ultimately the only way to know what hardware is supported at a given time
> is to run an app which calls rte_eal_init() to load all the drivers that are
> present and work from there, because besides CONFIG_RTE_EAL_PMD_PATH this
> can be affected by runtime commandline switches and applies to both shared
> and static builds.
> 
I'm not sure I agree with that.  Its clearly tempting to use, but its not
at all guaranteed to be accurate (the default is just set to "", and there is no
promise anyone will set it properly). And it also requires that the binary will
be tied to a specific release.  I really think that, given the fact that
distributions generally try to package dpdk in such a way that multiple dpdk
versions might be available, the right solution is to just require a full path
specification if you want to get hw info for a DSO that is dynamically loaded
via dlopen from the command line.  Otherwise you're going to fall into this trap
where you might be looking implicitly at an older version of the PMD while your
application may use a newer version.

> > > 
> > > Calling up on the list of requirements from
> > > http://dpdk.org/ml/archives/dev/2016-May/038324.html, I see a pile of
> > > technical requirements but perhaps we should stop for a moment to think
> > > about the use-cases first?
> > 
> > To ennumerate the list:
> > 
> > - query all drivers in static binary or shared library (works)
> > - stripping resiliency (works)
> > - human friendly (works)
> > - script friendly (works)
> > - show driver name (works)
> > - list supported device id / name (works)
> > - list driver options (not yet, but possible)
> > - show driver version if available (nope, but possible)
> > - show dpdk version (nope, but possible)
> > - show kernel dependencies (vfio/uio_pci_generic/etc) (nope)
> > - room for extra information? (works)
> > 
> > Of the items that are missing, I've already got a V2 started that can do driver
> > options, and is easier to expand.  Adding in the the DPDK and PMD version should
> > be easy (though I think they can be left out, as theres currently no globaly
> > defined DPDK release version, its all just implicit, and driver versions aren't
> > really there either).  I'm also hesitant to include kernel dependencies without
> > defining exactly what they mean (just module dependencies, or feature
> > enablement, or something else?).  Once we define it though, adding it can be
> > easy.
> 
> Yup. I just think the shared/static difference needs to be sorted out
> somehow, eg requiring user to know about DSOs is not human-friendly at all.
> That's why I called for the higher level use-cases in my previous email.
> 

I disagree with that.  While its reasonable to give users the convienience of
scanning the DT_NEEDED entries of a binary and scanning those DSO's.  If a user
has to specify the PMD to load in an application (either on the command line or
via a configuration file), then its reasonable assume that they (a) know where
to find that pmd and (b) are savy enough to pass that same path to a hardware
info tool.  Thats the exact same way that modinfo works (save for the fact that
modinfo can implicitly check the running kernel version to find the appropriate
path for a modular driver).

The only other thing that seems reasonable to me would be to scan
LD_LIBRARY_PATH.  I would assume that, if an application is linked dynamically,
the individual DSO's (librte_sched.so, etc), need to be in LD_LIBRARY_PATH.  If
thats the case, then we can assume that the appropriate PMD DSO's are there too,
and we can search there.  We can also check the standard /usr/lib and /lib paths
with that.  I think that would make fairly good sense.


> > 
> > I'll have a v2 posted soon, with the consensus corrections you have above, as
> > well as some other cleanups
> > 
> > Best
> > Neil
> > 
> > > 
> > > To name some from the top of my head:
> > > - user wants to know whether the hardware on the system is supported
> > > - user wants to know which package(s) need to be installed to support the
> > > system hardware
> > > - user wants to list all supported hardware before going shopping
> > > - [what else?]
> > > 
> > > ...and then think how these things would look like from the user
> > > perspective, in the light of the two quite dramatically differing cases of
> > > static vs shared linkage.
> 
> 
> 	- Panu -
>