[dpdk-dev] [PATCHv7 0/6] Implement pmd hardware support exports

Neil Horman nhorman at tuxdriver.com
Thu Jun 9 19:46:55 CEST 2016


Hey all-
	So heres attempt number 2 at a method for exporting PMD hardware support
information.  As we discussed previously, the consensus seems to be that pmd
information should be:

1) Able to be interrogated on any ELF binary (application binary or individual
DSO)
2) Equally functional on statically linked applications or on DSO's
3) Resilient to symbol stripping
4) Script friendly
5) Show kernel dependencies
6) List driver options
7) Show driver name
8) Offer human readable output
9) Show DPDK version
10) Show driver version
11) Allow for expansion
12) Not place additional build environment dependencies on an application

I added item 12 myself, because I don't think its reasonable to require
applications to use a specific linker script to get hardware support information
(which precludes the use of a special .modinfo section like the kernel uses)

However, we still can use some tricks from the kernel to make this work.  In
this approach, what I've done is the following:

A) Modified the driver registration macro to also define a variable:
this_pmd_name<n>= "name"

Based on the unique name string pointed to by the above variable, we can
query for an arbitrary number of other symbols following the pattern:
<name>_<tag>

Where tag is some well known identifier for information we wish to export

B) Added a utility called pmdinfogen.  This utility is not meant for general use,
but rather is used by the dpdk build environment itself when compiling pmds.
for each object that uses the PMD_REGISTER_DRIVER macro, this utiity is run.  It
searches for the symbols in (A), and using those, extracts the hardware support
info, and module name from the object, using that to produce a new c file
containing a single variable in the following format:

static const char <name>[] __attribute__((used)) = "PMD_DRIVER_INFO=<json>";

The <name> is arbitrary, as its static and not referenced.  The relevant bit is
the string value assigned to it.  The <json> is a json encoded string of the
extracted hardware support information pmdinfo found for the corresponding
object.  This C file is suitable for compilation and relocatable linking back
into the parent object file.  The result of this operation is that the object
string table now contains a string that will not e removed by stripping, whos
leading text (PMD_DRIVER_INFO) can be easily searched for at any time weather
the symbol referring to it is stripped or not.

C) Added a utilty called pmdinfo.py.  This python script, searches the
string table of the .rodata section of any provided ELF file looking for the
PMD_DRIVER_INFO prefix on a string.  When found, it will interpret the remainder
of the string as json, and output the hardware support for that ELF file (if
any).


This approach ticks most of the above boxes:
1) Impervious to stripping
2) Works on static and dynamic binaries
3) Human and script friendly
4) Allows for expansion

Because of (4) the other items should be pretty easy to implement, as its just a
matter of modifying the macros to export the info, pmdinfo to encode it to json,
and pmd_hw_support.py to read it.  I'm not adding them now, as theres alot of
rote work to do to get some of them in place (e.g. DPDK has no current global
VERSION macro, drivers don't have a consistent version scheme, command line
strings have to be built for each driver, etc).  But once this is accepted,
those items can be done as time allows and should be fairly easy to implement.

Change Notes:
v2)
 * Made the export macros a bit easier to expand
 * Added a macro to optionally export command line strings
 * Renamed utilties to be more appropriate
   (pmdinfo -> pmdinfogen, pmd_hw_support.py -> pmdinfo.py)
 * Added search capabilities to pmdinfo.py so that we search for libraries
   linked using DT_NEEDINFO entries.  We search DT_RUNPATH, LD_LIBRARY_PATH,
   /usr/lib and /lib
 * Added serch abilities if the specified binary to pmdinfo isn't found, we
   search LD_LIBRARY_PATH, /usr/lib and /lib
 * Added an option to pmdinfo.py to pretty-print hardware support info using the
   pci.ids database

v3)
 * Made table mode printing the default mode
 * Added a default path to the pci.ids file
 * Modifed pmdinfo to use python rather than python3
 * Cleaned up some changelog entries
 * Added an export for RTE_EAL_PMD_PATH
 * Added a plugin option to pmdinfo to scan for autoloaded DSO's

v4)
 * Modified the operation of the -p option. As much as I don't like implying
that autoloaded pmds are guaranteed to be there at run time, I'm having a hard
time seeing how we can avoid specifying the application file to scan for the
autoload directory.  Without it we can't determine which library the user means
in a multiversion installation
 * Cleaned up the help text
 * Added a rule for an install target for pmdinfo
 * Guarded against some tracebacks in pmdinfo
 * Use DT_NEEDED entries to get versioned libraries in -p mode
 * Fixed traceback that occurs on lack of input arguments
 * Fixed some erroneous macro usage in drivers that aren't in the default build

v5)
 * Added a dpdk- prefix to pmdinfo symlink
 * Renamed build announcement to PMDINFOGEN/BUILD
 * Removed erroneous true statement from makefile
 * Removed duplicate rte.hostapp.mk makefile
 * Fixed some whitespace
 * Whitespace fixups
 * Fixed makefile if; then style
 * Renamed module string C file
 * Removed duplicate rte_pci_id definition
 * Clarified macro names
 * Removed PMD type attribute
 * Fixed tools usage for 32 bit arches
 * Removed some unused code
 * Added a few comments

v6)
 * Added some programming guide documentation
 * Reformatted python script with pep8

v7)
 * Fixed up copyright
 * Switched buildtool makefile to use hostapp.mk
 * Modified arch check to use RTE_ARCH_64
 * Modifed hostapp.mk to drop output in build/app
 * Additional testing on ppc64 to ensure big endian works
 * Modified TO_NATIVE macro to use rte_byteorder inlines 
   based on endianess of target ELF file
 * Ran checkpatch on commits
 * Fixed some typos in doc

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>













More information about the dev mailing list