Add a section on changing a kfunc's prototype to the kfunc lifecycle expectations. libbpf now binds a ___-suffixed kfunc by prototype and resolve_btfids rejects kfuncs that libbpf can't tell apart, so kernel developers can keep BPF programs built against an old prototype loading, provided they follow the rules this section describes. Signed-off-by: Ihor Solodrai --- Documentation/bpf/kfuncs.rst | 96 ++++++++++++++++++++++++++++++++++++ 1 file changed, 96 insertions(+) diff --git a/Documentation/bpf/kfuncs.rst b/Documentation/bpf/kfuncs.rst index c7da603cab77..4d60a13ecbe3 100644 --- a/Documentation/bpf/kfuncs.rst +++ b/Documentation/bpf/kfuncs.rst @@ -800,6 +800,102 @@ procedure is followed for removal: 3. After the deprecation period the kfunc will be removed. At this point, BPF programs calling the kfunc will be rejected by the verifier. +3.2 Changing a kfunc's prototype +-------------------------------- + +libbpf drops the last ``___suffix`` from the name of a kfunc declared in a BPF +program and from the names of functions in kernel and module BTF, as it does +for CO-RE type names. Among the functions with the same remaining name, it +binds the one with a prototype that matches the declaration. Types are +compared kind by kind, and structs and unions by name only: + +.. code-block:: c + + typedef struct task_struct task_t; + + /* these match */ + int bpf_foo(struct task_struct *p, u32 flags); + int bpf_foo(const task_t *p, u64 flags); + int bpf_foo(struct task_struct___local *p, int flags); + + /* none of these match the ones above */ + int bpf_foo(struct file *p, u32 flags); + int bpf_foo(struct task_struct *p); + void bpf_foo(struct task_struct *p, u32 flags); + +If the old and new prototypes match, change the kfunc in place. Otherwise the +new prototype keeps the name, and the old one stays available as +``___v1`` (``___v2`` for the next change, and so on) for a deprecation +period, usually as a thin wrapper registered with the same flags: + +.. code-block:: c + + __bpf_kfunc int bpf_foo(struct task_struct *p, u64 flags) + { + ... + } + + /* bpf_foo() before flags was added */ + __bpf_kfunc int bpf_foo___v1(struct task_struct *p) + { + return bpf_foo(p, 0); + } + + BTF_ID_FLAGS(func, bpf_foo, KF_RCU) + BTF_ID_FLAGS(func, bpf_foo___v1, KF_RCU) + +BPF programs built against either prototype then load: + +.. code-block:: c + + /* built against the old prototype: binds bpf_foo___v1 */ + extern int bpf_foo(struct task_struct *p) __ksym; + + /* built against the new prototype: binds bpf_foo */ + extern int bpf_foo(struct task_struct *p, u64 flags) __ksym; + + /* the suffix doesn't select a kfunc: both bind bpf_foo___v1 */ + extern int bpf_foo___v1(struct task_struct *p) __ksym; + extern int bpf_foo___local(struct task_struct *p) __ksym; + + /* a stacked suffix binds nothing */ + extern int bpf_foo___v1___local(struct task_struct *p) __ksym; + +Only libbpf 1.9 and later look for ``bpf_foo___v1``. Older libbpf and light +skeletons see only ``bpf_foo``, as after an in-place change. + +If two kfunc names differ only in the ``___`` suffix, their prototypes must not +match, not even for aliases: libbpf can't tell them apart and fails to load a +program that declares either one, even as ``__weak``. For example, this widens +an offset to 64 bits and keeps the old prototype as ``bpf_bar___v1``: + +.. code-block:: c + + __bpf_kfunc int bpf_bar(const struct bpf_dynptr *p, u64 offset); + + /* bpf_bar() before offset was widened */ + __bpf_kfunc int bpf_bar___v1(const struct bpf_dynptr *p, u32 offset); + +``resolve_btfids`` rejects it:: + + ERROR: resolve_btfids: kfuncs bpf_bar and bpf_bar___v1 have matching prototypes + +Widening keeps the prototype matching, so it is done in place instead: drop +``bpf_bar___v1``, and programs built against the old prototype bind +``bpf_bar``. ``resolve_btfids`` checks registered kfuncs within vmlinux or one +module. libbpf catches other matches, such as between two modules, when loading +the program. + +Give the new prototype a name outside the family, not ``___*``, if: + +- programs loaded by older libbpf or light skeletons must keep working, or +- the prototypes match but the change breaks callers, e.g. two integer + parameters swap places. + +Don't rename a ``___vN`` kfunc while older libbpf or light skeletons +still load programs that call it by that name. libbpf 1.9 and later bind by +prototype, so a rename doesn't affect them. + 4. Core kfuncs ============== -- 2.56.1