On 4/30/25 23:40, Sami Tolvanen wrote: > In rare situations where distributions must make significant > changes to otherwise opaque data structures that have > inadvertently been included in the published ABI, keeping > symbol versions stable using the existing kABI macros can > become tedious. > > For example, Android decided to switch to a newer io_uring > implementation in the 5.10 GKI kernel "to resolve a huge number > of potential, and known, problems with the codebase," requiring > "horrible hacks" with genksyms: > > "A number of the io_uring structures get used in other core > kernel structures, only as "opaque" pointers, so there is > not any real ABI breakage. But, due to the visibility of > the structures going away, the CRC values of many scheduler > variables and functions were changed." > -- https://r.android.com/2425293 > > While these specific changes probably could have been hidden > from gendwarfksyms using the existing kABI macros, this may not > always be the case. > > Add a last resort kABI rule that allows distribution > maintainers to fully override a type string for a symbol or a > type. Also add a more informative error message in case we find > a non-existent type references when calculating versions. > > Suggested-by: Giuliano Procida <gproc...@google.com> > Signed-off-by: Sami Tolvanen <samitolva...@google.com> > [...] > @@ -372,9 +405,81 @@ static void type_expand(struct die *cache, struct > type_expansion *type, > cache_free(&expansion_cache); > } > > +static void type_parse(const char *name, const char *str, > + struct type_expansion *type) > +{ > + char *fragment; > + size_t start = 0; > + size_t end; > + size_t pos; > + > + if (!*str) > + error("empty type string override for '%s'", name); > + > + type_expansion_init(type); > + > + for (pos = 1; str[pos]; ++pos) { > + bool empty; > + char marker = ' '; > + > + if (!is_type_prefix(&str[pos - 1])) > + continue; > + > + end = pos + 1; > + > + /* > + * Find the end of the type reference. If the type name contains > + * spaces, it must be in single quotes. > + */ > + if (str[end] == '\'') { > + marker = '\''; > + ++end; > + } > + while (str[end] && str[end] != marker) > + ++end; > + > + /* Check that we have a non-empty type name */ > + if (marker == '\'') { > + if (str[end] != marker) > + error("incomplete %c# type reference for '%s' > (string : '%s')", > + str[pos - 1], name, str); > + empty = end == pos + 2; > + ++end; > + } else { > + empty = end == pos + 1; > + } > + if (empty) > + error("empty %c# type name for '%s' (string: '%s')", > + str[pos - 1], name, str); > + > + /* Append the part of the string before the type reference */ > + if (pos > start + 1) { > + fragment = xstrndup(&str[start], pos - start - 1); > + type_expansion_append(type, fragment, fragment); > + } > + > + /* > + * Append the type reference -- note that if the reference > + * is invalid, i.e. points to a non-existent type, we will > + * print out an error when calculating versions. > + */ > + fragment = xstrndup(&str[pos - 1], end - pos + 1); > + type_expansion_append(type, fragment, fragment); > + > + pos = start = end; > + if (!str[pos]) > + break; > + } > + > + /* Append the rest of the type string, if there's any left */ > + if (str[start]) > + type_expansion_append(type, &str[start], NULL); > +} > +
I'd find this mini-parser more straightforward, if its main loop started at pos=0 and looked ahead, instead of starting at pos=1 and looking behind. The patch otherwise makes sense to me. It looks similar to the override keyword supported by genksyms. -- Thanks, Petr
