3. DPDK ABI/API policy

3.1. Description

This document details some methods for handling ABI management in the DPDK.

3.2. General Guidelines

  1. Whenever possible, ABI should be preserved
  2. ABI/API may be changed with a deprecation process
  3. The modification of symbols can generally be managed with versioning
  4. Libraries or APIs marked in experimental state may change without constraint
  5. New APIs will be marked as experimental for at least one release to allow any issues found by users of the new API to be fixed quickly
  6. The addition of symbols is generally not problematic
  7. The removal of symbols generally is an ABI break and requires bumping of the LIBABIVER macro
  8. Updates to the minimum hardware requirements, which drop support for hardware which was previously supported, should be treated as an ABI change.

3.2.1. What is an ABI

An ABI (Application Binary Interface) is the set of runtime interfaces exposed by a library. It is similar to an API (Application Programming Interface) but is the result of compilation. It is also effectively cloned when applications link to dynamic libraries. That is to say when an application is compiled to link against dynamic libraries, it is assumed that the ABI remains constant between the time the application is compiled/linked, and the time that it runs. Therefore, in the case of dynamic linking, it is critical that an ABI is preserved, or (when modified), done in such a way that the application is unable to behave improperly or in an unexpected fashion.

3.3. ABI/API Deprecation

3.3.1. The DPDK ABI policy

ABI versions are set at the time of major release labeling, and the ABI may change multiple times, without warning, between the last release label and the HEAD label of the git tree.

ABI versions, once released, are available until such time as their deprecation has been noted in the Release Notes for at least one major release cycle. For example consider the case where the ABI for DPDK 2.0 has been shipped and then a decision is made to modify it during the development of DPDK 2.1. The decision will be recorded in the Release Notes for the DPDK 2.1 release and the modification will be made available in the DPDK 2.2 release.

ABI versions may be deprecated in whole or in part as needed by a given update.

Some ABI changes may be too significant to reasonably maintain multiple versions. In those cases ABI’s may be updated without backward compatibility being provided. The requirements for doing so are:

  1. At least 3 acknowledgments of the need to do so must be made on the dpdk.org mailing list.
    • The acknowledgment of the maintainer of the component is mandatory, or if no maintainer is available for the component, the tree/sub-tree maintainer for that component must acknowledge the ABI change instead.
    • It is also recommended that acknowledgments from different “areas of interest” be sought for each deprecation, for example: from NIC vendors, CPU vendors, end-users, etc.
  2. The changes (including an alternative map file) can be included with deprecation notice, in wrapped way by the RTE_NEXT_ABI option, to provide more details about oncoming changes. RTE_NEXT_ABI wrapper will be removed when it become the default ABI. More preferred way to provide this information is sending the feature as a separate patch and reference it in deprecation notice.
  3. A full deprecation cycle, as explained above, must be made to offer downstream consumers sufficient warning of the change.

Note that the above process for ABI deprecation should not be undertaken lightly. ABI stability is extremely important for downstream consumers of the DPDK, especially when distributed in shared object form. Every effort should be made to preserve the ABI whenever possible. The ABI should only be changed for significant reasons, such as performance enhancements. ABI breakage due to changes such as reorganizing public structure fields for aesthetic or readability purposes should be avoided.

Note

Updates to the minimum hardware requirements, which drop support for hardware which was previously supported, should be treated as an ABI change, and follow the relevant deprecation policy procedures as above: 3 acks and announcement at least one release in advance.

3.3.2. Examples of Deprecation Notices

The following are some examples of ABI deprecation notices which would be added to the Release Notes:

  • The Macro #RTE_FOO is deprecated and will be removed with version 2.0, to be replaced with the inline function rte_foo().
  • The function rte_mbuf_grok() has been updated to include a new parameter in version 2.0. Backwards compatibility will be maintained for this function until the release of version 2.1
  • The members of struct rte_foo have been reorganized in release 2.0 for performance reasons. Existing binary applications will have backwards compatibility in release 2.0, while newly built binaries will need to reference the new structure variant struct rte_foo2. Compatibility will be removed in release 2.2, and all applications will require updating and rebuilding to the new structure at that time, which will be renamed to the original struct rte_foo.
  • Significant ABI changes are planned for the librte_dostuff library. The upcoming release 2.0 will not contain these changes, but release 2.1 will, and no backwards compatibility is planned due to the extensive nature of these changes. Binaries using this library built prior to version 2.1 will require updating and recompilation.

3.3.3. New API replacing previous one

If a new API proposed functionally replaces an existing one, when the new API becomes non-experimental then the old one is marked with __rte_deprecated. Deprecated APIs are removed completely just after the next LTS.

Reminder that old API should follow deprecation process to be removed.

3.4. Experimental APIs

APIs marked as experimental are not considered part of the ABI and may change without warning at any time. Since changes to APIs are most likely immediately after their introduction, as users begin to take advantage of those new APIs and start finding issues with them, new DPDK APIs will be automatically marked as experimental to allow for a period of stabilization before they become part of a tracked ABI.

Note that marking an API as experimental is a multi step process. To mark an API as experimental, the symbols which are desired to be exported must be placed in an EXPERIMENTAL version block in the corresponding libraries’ version map script. Secondly, the corresponding prototypes of those exported functions (in the development header files), must be marked with the __rte_experimental tag (see rte_compat.h). The DPDK build makefiles perform a check to ensure that the map file and the C code reflect the same list of symbols. This check can be circumvented by defining ALLOW_EXPERIMENTAL_API during compilation in the corresponding library Makefile.

In addition to tagging the code with __rte_experimental, the doxygen markup must also contain the EXPERIMENTAL string, and the MAINTAINERS file should note the EXPERIMENTAL libraries.

For removing the experimental tag associated with an API, deprecation notice is not required. Though, an API should remain in experimental state for at least one release. Thereafter, normal process of posting patch for review to mailing list can be followed.

3.5. Library versioning

Downstreams might want to provide different DPDK releases at the same time to support multiple consumers of DPDK linked against older and newer sonames.

Also due to the interdependencies that DPDK libraries can have applications might end up with an executable space in which multiple versions of a library are mapped by ld.so.

Think of LibA that got an ABI bump and LibB that did not get an ABI bump but is depending on LibA.

Note

Application -> LibA.old -> LibB.new -> LibA.new

That is a conflict which can be avoided by setting CONFIG_RTE_MAJOR_ABI. If set, the value of CONFIG_RTE_MAJOR_ABI overwrites all - otherwise per library - versions defined in the libraries LIBABIVER. An example might be CONFIG_RTE_MAJOR_ABI=16.11 which will make all libraries librte<?>.so.16.11 instead of librte<?>.so.<LIBABIVER>.

3.6. ABI versioning

3.6.1. Versioning Macros

When a symbol is exported from a library to provide an API, it also provides a calling convention (ABI) that is embodied in its name, return type and arguments. Occasionally that function may need to change to accommodate new functionality or behavior. When that occurs, it is desirable to allow for backward compatibility for a time with older binaries that are dynamically linked to the DPDK.

To support backward compatibility the rte_compat.h header file provides macros to use when updating exported functions. These macros are used in conjunction with the rte_<library>_version.map file for a given library to allow multiple versions of a symbol to exist in a shared library so that older binaries need not be immediately recompiled.

The macros exported are:

  • VERSION_SYMBOL(b, e, n): Creates a symbol version table entry binding versioned symbol b@DPDK_n to the internal function b_e.
  • BIND_DEFAULT_SYMBOL(b, e, n): Creates a symbol version entry instructing the linker to bind references to symbol b to the internal symbol b_e.
  • MAP_STATIC_SYMBOL(f, p): Declare the prototype f, and map it to the fully qualified function p, so that if a symbol becomes versioned, it can still be mapped back to the public symbol name.

3.6.1.1. Examples of ABI Macro use

3.6.1.1.1. Updating a public API

Assume we have a function as follows

/*
 * Create an acl context object for apps to
 * manipulate
 */
struct rte_acl_ctx *
rte_acl_create(const struct rte_acl_param *param)
{
       ...
}

Assume that struct rte_acl_ctx is a private structure, and that a developer wishes to enhance the acl api so that a debugging flag can be enabled on a per-context basis. This requires an addition to the structure (which, being private, is safe), but it also requires modifying the code as follows

/*
 * Create an acl context object for apps to
 * manipulate
 */
struct rte_acl_ctx *
rte_acl_create(const struct rte_acl_param *param, int debug)
{
       ...
}

Note also that, being a public function, the header file prototype must also be changed, as must all the call sites, to reflect the new ABI footprint. We will maintain previous ABI versions that are accessible only to previously compiled binaries

The addition of a parameter to the function is ABI breaking as the function is public, and existing application may use it in its current form. However, the compatibility macros in DPDK allow a developer to use symbol versioning so that multiple functions can be mapped to the same public symbol based on when an application was linked to it. To see how this is done, we start with the requisite libraries version map file. Initially the version map file for the acl library looks like this

DPDK_2.0 {
     global:

     rte_acl_add_rules;
     rte_acl_build;
     rte_acl_classify;
     rte_acl_classify_alg;
     rte_acl_classify_scalar;
     rte_acl_create;
     rte_acl_dump;
     rte_acl_find_existing;
     rte_acl_free;
     rte_acl_ipv4vlan_add_rules;
     rte_acl_ipv4vlan_build;
     rte_acl_list_dump;
     rte_acl_reset;
     rte_acl_reset_rules;
     rte_acl_set_ctx_classify;

     local: *;
};

This file needs to be modified as follows

DPDK_2.0 {
     global:

     rte_acl_add_rules;
     rte_acl_build;
     rte_acl_classify;
     rte_acl_classify_alg;
     rte_acl_classify_scalar;
     rte_acl_create;
     rte_acl_dump;
     rte_acl_find_existing;
     rte_acl_free;
     rte_acl_ipv4vlan_add_rules;
     rte_acl_ipv4vlan_build;
     rte_acl_list_dump;
     rte_acl_reset;
     rte_acl_reset_rules;
     rte_acl_set_ctx_classify;

     local: *;
};

DPDK_2.1 {
     global:
     rte_acl_create;

} DPDK_2.0;

The addition of the new block tells the linker that a new version node is available (DPDK_2.1), which contains the symbol rte_acl_create, and inherits the symbols from the DPDK_2.0 node. This list is directly translated into a list of exported symbols when DPDK is compiled as a shared library

Next, we need to specify in the code which function map to the rte_acl_create symbol at which versions. First, at the site of the initial symbol definition, we need to update the function so that it is uniquely named, and not in conflict with the public symbol name

 struct rte_acl_ctx *
-rte_acl_create(const struct rte_acl_param *param)
+rte_acl_create_v20(const struct rte_acl_param *param)
{
       size_t sz;
       struct rte_acl_ctx *ctx;
       ...

Note that the base name of the symbol was kept intact, as this is conducive to the macros used for versioning symbols. That is our next step, mapping this new symbol name to the initial symbol name at version node 2.0. Immediately after the function, we add this line of code

VERSION_SYMBOL(rte_acl_create, _v20, 2.0);

Remembering to also add the rte_compat.h header to the requisite c file where these changes are being made. The above macro instructs the linker to create a new symbol rte_acl_create@DPDK_2.0, which matches the symbol created in older builds, but now points to the above newly named function. We have now mapped the original rte_acl_create symbol to the original function (but with a new name)

Next, we need to create the 2.1 version of the symbol. We create a new function name, with a different suffix, and implement it appropriately

struct rte_acl_ctx *
rte_acl_create_v21(const struct rte_acl_param *param, int debug);
{
     struct rte_acl_ctx *ctx = rte_acl_create_v20(param);

     ctx->debug = debug;

     return ctx;
}

This code serves as our new API call. Its the same as our old call, but adds the new parameter in place. Next we need to map this function to the symbol rte_acl_create@DPDK_2.1. To do this, we modify the public prototype of the call in the header file, adding the macro there to inform all including applications, that on re-link, the default rte_acl_create symbol should point to this function. Note that we could do this by simply naming the function above rte_acl_create, and the linker would chose the most recent version tag to apply in the version script, but we can also do this in the header file

struct rte_acl_ctx *
-rte_acl_create(const struct rte_acl_param *param);
+rte_acl_create(const struct rte_acl_param *param, int debug);
+BIND_DEFAULT_SYMBOL(rte_acl_create, _v21, 2.1);

The BIND_DEFAULT_SYMBOL macro explicitly tells applications that include this header, to link to the rte_acl_create_v21 function and apply the DPDK_2.1 version node to it. This method is more explicit and flexible than just re-implementing the exact symbol name, and allows for other features (such as linking to the old symbol version by default, when the new ABI is to be opt-in for a period.

One last thing we need to do. Note that we’ve taken what was a public symbol, and duplicated it into two uniquely and differently named symbols. We’ve then mapped each of those back to the public symbol rte_acl_create with different version tags. This only applies to dynamic linking, as static linking has no notion of versioning. That leaves this code in a position of no longer having a symbol simply named rte_acl_create and a static build will fail on that missing symbol.

To correct this, we can simply map a function of our choosing back to the public symbol in the static build with the MAP_STATIC_SYMBOL macro. Generally the assumption is that the most recent version of the symbol is the one you want to map. So, back in the C file where, immediately after rte_acl_create_v21 is defined, we add this

struct rte_acl_ctx *
rte_acl_create_v21(const struct rte_acl_param *param, int debug)
{
     ...
}
MAP_STATIC_SYMBOL(struct rte_acl_ctx *rte_acl_create(const struct rte_acl_param *param, int debug), rte_acl_create_v21);

That tells the compiler that, when building a static library, any calls to the symbol rte_acl_create should be linked to rte_acl_create_v21

That’s it, on the next shared library rebuild, there will be two versions of rte_acl_create, an old DPDK_2.0 version, used by previously built applications, and a new DPDK_2.1 version, used by future built applications.

3.6.1.1.2. Deprecating part of a public API

Lets assume that you’ve done the above update, and after a few releases have passed you decide you would like to retire the old version of the function. After having gone through the ABI deprecation announcement process, removal is easy. Start by removing the symbol from the requisite version map file:

  DPDK_2.0 {
       global:

       rte_acl_add_rules;
       rte_acl_build;
       rte_acl_classify;
       rte_acl_classify_alg;
       rte_acl_classify_scalar;
       rte_acl_dump;
-      rte_acl_create
       rte_acl_find_existing;
       rte_acl_free;
       rte_acl_ipv4vlan_add_rules;
       rte_acl_ipv4vlan_build;
       rte_acl_list_dump;
       rte_acl_reset;
       rte_acl_reset_rules;
       rte_acl_set_ctx_classify;

       local: *;
  };

  DPDK_2.1 {
       global:
       rte_acl_create;
  } DPDK_2.0;

Next remove the corresponding versioned export.

-VERSION_SYMBOL(rte_acl_create, _v20, 2.0);

Note that the internal function definition could also be removed, but its used in our example by the newer version _v21, so we leave it in place. This is a coding style choice.

Lastly, we need to bump the LIBABIVER number for this library in the Makefile to indicate to applications doing dynamic linking that this is a later, and possibly incompatible library version:

-LIBABIVER := 1
+LIBABIVER := 2
3.6.1.1.3. Deprecating an entire ABI version

While removing a symbol from and ABI may be useful, it is often more practical to remove an entire version node at once. If a version node completely specifies an API, then removing part of it, typically makes it incomplete. In those cases it is better to remove the entire node

To do this, start by modifying the version map file, such that all symbols from the node to be removed are merged into the next node in the map

In the case of our map above, it would transform to look as follows

  DPDK_2.1 {
       global:

       rte_acl_add_rules;
       rte_acl_build;
       rte_acl_classify;
       rte_acl_classify_alg;
       rte_acl_classify_scalar;
       rte_acl_dump;
       rte_acl_create
       rte_acl_find_existing;
       rte_acl_free;
       rte_acl_ipv4vlan_add_rules;
       rte_acl_ipv4vlan_build;
       rte_acl_list_dump;
       rte_acl_reset;
       rte_acl_reset_rules;
       rte_acl_set_ctx_classify;

       local: *;
};

Then any uses of BIND_DEFAULT_SYMBOL that pointed to the old node should be updated to point to the new version node in any header files for all affected symbols.

-BIND_DEFAULT_SYMBOL(rte_acl_create, _v20, 2.0);
+BIND_DEFAULT_SYMBOL(rte_acl_create, _v21, 2.1);

Lastly, any VERSION_SYMBOL macros that point to the old version node should be removed, taking care to keep, where need old code in place to support newer versions of the symbol.

3.7. Running the ABI Validator

The devtools directory in the DPDK source tree contains a utility program, validate-abi.sh, for validating the DPDK ABI based on the Linux ABI Compliance Checker.

This has a dependency on the abi-compliance-checker and and abi-dumper utilities which can be installed via a package manager. For example:

sudo yum install abi-compliance-checker
sudo yum install abi-dumper

The syntax of the validate-abi.sh utility is:

./devtools/validate-abi.sh <REV1> <REV2>

Where REV1 and REV2 are valid gitrevisions(7) https://www.kernel.org/pub/software/scm/git/docs/gitrevisions.html on the local repo.

For example:

# Check between the previous and latest commit:
./devtools/validate-abi.sh HEAD~1 HEAD

# Check on a specific compilation target:
./devtools/validate-abi.sh -t x86_64-native-linux-gcc HEAD~1 HEAD

# Check between two tags:
./devtools/validate-abi.sh v2.0.0 v2.1.0

# Check between git master and local topic-branch "vhost-hacking":
./devtools/validate-abi.sh master vhost-hacking

After the validation script completes (it can take a while since it need to compile both tags) it will create compatibility reports in the ./abi-check/compat_report directory. Listed incompatibilities can be found as follows:

grep -lr Incompatible abi-check/compat_reports/