When generating those files, it uses as input some symbols files provided by the maintainer. It looks for the following files (and uses the first that is found):
The main interest of those files is to provide the minimal version associated to each symbol provided by the libraries. Usually it corresponds to the first version of that package that provided the symbol, but it can be manually incremented by the maintainer if the ABI of the symbol is extended without breaking backwards compatibility. It's the responsibility of the maintainer to keep those files up-to-date and accurate, but dpkg-gensymbols helps with that.
When the generated symbols files differ from the maintainer supplied one, dpkg-gensymbols will print a diff between the two versions. Furthermore if the difference is too significant, it will even fail (you can customize how much difference you can tolerate, see the -c option).
Before applying any patch to the symbols file, the maintainer should double-check that it's sane. Public symbols are not supposed to disappear, so the patch should ideally only add new lines.
Note that you can put comments in symbols files: any line with '#' as the first character is a comment except if it starts with '#include' (see section Using includes). Lines starting with '#MISSING:' are special comments documenting symbols that have disappeared.
Do not forget to check if old symbol versions need to be increased. There is no way dpkg-gensymbols can warn about this. Blindly applying the diff or assuming there is nothing to change if there is no diff, without checking for such changes, can lead to packages with loose dependencies that claim they can work with older packages they cannot work with. This will introduce hard to find bugs with (partial) upgrades.
In some rare cases, the name of the library varies between architectures. To avoid hardcoding the name of the package in the symbols file, you can use the marker #PACKAGE#. It will be replaced by the real package name during installation of the symbols files. Contrary to the #MINVER# marker, #PACKAGE# will never appear in a symbols file inside a binary package.
Symbol tagging is useful for marking symbols that are special in some way. Any symbol can have an arbitrary number of tags associated with it. While all tags are parsed and stored, only some of them are understood by dpkg-gensymbols and trigger special handling of the symbols. See subsection Standard symbol tags for reference of these tags.
Tag specification comes right before the symbol name (no whitespace is allowed in between). It always starts with an opening bracket (, ends with a closing bracket ) and must contain at least one tag. Multiple tags are separated by the | character. Each tag can optionally have a value which is separated form the tag name by the = character. Tag names and values can be arbitrary strings except they cannot contain any of the special ) | = characters. Symbol names following a tag specification can optionally be quoted with either ' or " characters to allow whitespaces in them. However, if there are no tags specified for the symbol, quotes are treated as part of the symbol name which continues up until the first space.
(tag1=i am marked|tag name with space)"tagged quoted symbol"@Base 1.0
(optional)tagged_unquoted_symbol@Base 1.0 1
The first symbol in the example is named tagged quoted symbol and has two tags: tag1 with value i am marked and tag name with space that has no value. The second symbol named tagged_unquoted_symbol is only tagged with the tag named optional. The last symbol is an example of the normal untagged symbol.
Since symbol tags are an extension of the deb-symbols(5) format, they can only be part of the symbols files used in source packages (those files should then be seen as templates used to build the symbols files that are embedded in binary packages). When dpkg-gensymbols is called without the -t option, it will output symbols files compatible to the deb-symbols(5) format: it fully processes symbols according to the requirements of their standard tags and strips all tags from the output. On the contrary, in template mode (-t) all symbols and their tags (both standard and unknown ones) are kept in the output and are written in their original form as they were loaded.
This tag is useful for symbols which are private where their disappearance do not cause ABI breakage. For example, most of C++ template instantiations fall into this category. Like any other tag, this one may also have an arbitrary value: it could be used to indicate why the symbol is considered optional.
When operating in the default non-template mode, among arch-specific symbols only those that match the current host architecture are written to the symbols file. On the contrary, all arch-specific symbols (including those from foreign arches) are always written to the symbol file when operating in template mode.
The format of architecture-list is the same as the one used in the Build-Depends field of debian/control (except the enclosing square brackets ). For example, the first symbol from the list below will be considered only on alpha, any-amd64 and ia64 architectures, the second only on linux architectures, while the third one anywhere except on armel.
(arch=alpha any-amd64 ia64)a_64bit_specific_symbol@Base 1.0
The architecture-bits is either 32 or 64.
The architecture-endianness is either little or big.
Multiple restrictions can be chained.
Unlike a standard symbol specification, a pattern may cover multiple real symbols from the library. dpkg-gensymbols will attempt to match each pattern against each real symbol that does not have a specific symbol counterpart defined in the symbol file. Whenever the first matching pattern is found, all its tags and properties will be used as a basis specification of the symbol. If none of the patterns matches, the symbol will be considered as new.
A pattern is considered lost if it does not match any symbol in the library. By default this will trigger a dpkg-gensymbols failure under -c1 or higher level. However, if the failure is undesired, the pattern may be marked with the optional tag. Then if the pattern does not match anything, it will only appear in the diff as MISSING. Moreover, like any symbol, the pattern may be limited to the specific architectures with the arch tag. Please refer to Standard symbol tags subsection above for more information.
Patterns are an extension of the deb-symbols(5) format hence they are only valid in symbol file templates. Pattern specification syntax is not any different from the one of a specific symbol. However, symbol name part of the specification serves as an expression to be matched against name@version of the real symbol. In order to distinguish among different pattern types, a pattern will typically be tagged with a special tag.
At the moment, dpkg-gensymbols supports three basic pattern types:
libdummy.so.1 libdummy1 #MINVER#
(c++)"non-virtual thunk to NSB::ClassD::~ClassD()@Base" 1.0
The demangled name above can be obtained by executing the following command:
$ echo '_ZThn8_N3NSB6ClassDD1Ev@Base' | c++filt
Please note that while mangled name is unique in the library by definition, this is not necessarily true for demangled names. A couple of distinct real symbols may have the same demangled name. For example, that's the case with non-virtual thunk symbols in complex inheritance configurations or with most constructors and destructors (since g++ typically generates two real symbols for them). However, as these collisions happen on the ABI level, they should not degrade quality of the symbol file.
libc.so.6 libc6 #MINVER#
All symbols associated with versions GLIBC_2.0 and GLIBC_2.7 will lead to minimal version of 2.0 and 2.7 respectively with the exception of the symbol access@GLIBC_2.0. The latter will lead to a minimal dependency on libc6 version 2.2 despite being in the scope of the "(symver)GLIBC_2.0" pattern because specific symbols take precedence over patterns.
Please note that while old style wildcard patterns (denoted by "*@version" in the symbol name field) are still supported, they have been deprecated by new style syntax "(symver|optional)version". For example, "*@GLIBC_2.0 2.0" should be written as "(symver|optional)GLIBC_2.0 2.0" if the same behaviour is needed.
libdummy.so.1 libdummy1 #MINVER#
Symbols like "mystack_new@Base", "mystack_push@Base", "mystack_pop@Base" etc. will be matched by the first pattern while e.g. "ng_mystack_new@Base" won't. The second pattern will match all symbols having the string "private" in their names and matches will inherit optional tag from the pattern.
Basic patterns listed above can be combined where it makes sense. In that case, they are processed in the order in which the tags are specified. For example, both
will match symbols "_ZN3NSA6ClassA7Private11privmethod1Ei@Base" and "_ZN3NSA6ClassA7Private11privmethod2Ei@Base". When matching the first pattern, the raw symbol is first demangled as C++ symbol, then the demangled name is matched against the regular expression. On the other hand, when matching the second pattern, regular expression is matched against the raw symbol name, then the symbol is tested if it is C++ one by attempting to demangle it. A failure of any basic pattern will result in the failure of the whole pattern. Therefore, for example, "__N3NSA6ClassA7Private11privmethod\dEi@Base" will not match either of the patterns because it is not a valid C++ symbol.
In general, all patterns are divided into two groups: aliases (basic c++ and symver) and generic patterns (regex, all combinations of multiple basic patterns). Matching of basic alias-based patterns is fast (O(1)) while generic patterns are O(N) (N - generic pattern count) for each symbol. Therefore, it is recommended not to overuse generic patterns.
When multiple patterns match the same real symbol, aliases (first c++, then symver) are preferred over generic patterns. Generic patterns are matched in the order they are found in the symbol file template until the first success. Please note, however, that manual reordering of template file entries is not recommended because dpkg-gensymbols generates diffs based on the alphanumerical order of their names.
When the set of exported symbols differ between architectures, it may become inefficient to use a single symbol file. In those cases, an include directive may prove to be useful in a couple of ways:
As a result, all symbols included from file-to-include will be considered to be tagged with tag ... tagN by default. You can use this feature to create a common package.symbols file which includes architecture specific symbol files:
(arch=amd64 ia64 alpha)#include "package.symbols.64bit"
(arch=!amd64 !ia64 !alpha)#include "package.symbols.32bit"
The symbols files are read line by line, and include directives are processed as soon as they are encountered. This means that the content of the included file can override any content that appeared before the include directive and that any content after the directive can override anything contained in the included file. Any symbol (or even another #include directive) in the included file can specify additional tags or override values of the inherited tags in its tag specification. However, there is no way for the symbol to remove any of the inherited tags.
An included file can repeat the header line containing the SONAME of the library. In that case, it overrides any header line previously read. However, in general it's best to avoid duplicating header lines. One way to do it is the following:
A well-maintained library has the following features:
While maintaining the symbols file, it's easy to notice appearance and disappearance of symbols. But it's more difficult to catch incompatible API and ABI change. Thus the maintainer should read thoroughly the upstream changelog looking for cases where the rules of good library management have been broken. If potential problems are discovered, the upstream author should be notified as an upstream fix is always better than a Debian specific work-around.
This value can be overridden by the environment variable DPKG_GENSYMBOLS_CHECK_LEVEL.