Bionic comes with a set of 'clean' Linux kernel headers that can safely be included by userland applications and libraries without fear of hideous conflicts. for more information why this is needed, see the "RATIONALE" section at the end of this document. these clean headers are automatically generated by several scripts located in the 'bionic/kernel/tools' directory, which process a set of original and unmodified kernel headers in order to get rid of many annoying declarations and constructs that usually result in compilation failure. the 'clean headers' only contain type and macro definitions, with the exception of a couple static inline functions used for performance reason (e.g. optimized CPU-specific byte-swapping routines) they can be included from C++, or when compiling code in strict ANSI mode. they can be also included before or after any Bionic C library header. the generation process works as follows: * 'external/kernel-headers/original/' contains a set of kernel headers as normally found in the 'include' directory of a normal Linux kernel source tree. note that this should only contain the files that are really needed by Android (use 'find_headers.py' to find these automatically). * 'bionic/libc/kernel/common' contains the non-arch-specific clean headers and directories (e.g. linux, asm-generic and mtd) * 'bionic/libc/kernel/arch-arm/' contains the ARM-specific directory tree of clean headers. * 'bionic/libc/kernel/arch-arm/asm' contains the real ARM-specific headers * 'bionic/libc/kernel/arch-x86' 'bionic/libc/kernel/arch-x86/asm' similarly contains all headers and symlinks to be used on x86 * 'bionic/libc/kernel/tools' contains various Python and shell scripts used to manage and re-generate the headers the tools you can use are: * tools/find_users.py scans a list of source files or directories and prints which ones do include Linux headers. * tools/find_headers.py scans a list of source files or directories and recursively finds all the original kernel headers they need. * tools/clean_header.py prints the clean version of a given kernel header. with the -u option, this will also update the corresponding clean header file if its content has changed. you can also process more than one file with -u * tools/update_all.py automatically update all clean headers from the content of 'external/kernel-headers/original'. this is the script you're likely going to run whenever you update the original headers. HOW TO BUILD BIONIC AND OTHER PROGRAMS WITH THE CLEAN HEADERS: ============================================================== add bionic/kernel/common and bionic/kernel/arch- to your C include path. that should be enough. Note that Bionic will not compile properly if you don't. HOW TO SUPPORT ANOTHER ARCHITECTURE: ==================================== see the content of tools/defaults.py, you will need to make a few updates here: - add a new item to the 'kernel_archs' list of supported architectures - add a proper definition for 'kernel_known__statics' with relevant definitions. - update 'kernel_known_statics' to map "" to 'kernel_known__statics' then, add the new architecture-specific headers to original/asm-. (please ensure that these are really needed, e.g. with tools/find_headers.py) finally, run tools/update_all.py HOW TO UPDATE THE HEADERS WHEN NEEDED: ====================================== IMPORTANT IMPORTANT: WHEN UPDATING THE HEADERS, ALWAYS CHECK THAT THE NEW CLEAN HEADERS DO NOT BREAK THE KERNEL <-> USER ABI, FOR EXAMPLE BY CHANGING THE SIZE OF A GIVEN TYPE. THIS TASK CANNOT BE EASILY AUTOMATED AT THE MOMENT copy any updated kernel header into the corresponding location under 'bionic/kernel/original'. for any new kernel header you want to add, first run tools/find_headers.py to be sure that it is really needed by the Android sources. then add it to 'bionic/kernel/original' then, run tools/update_all.py to re-run the auto-cleaning HOW THE CLEANUP PROCESS WORKS: ============================== this section describes the action performed by the cleanup program(s) when they process the original kernel headers into clean ones: 1. Optimize well-known macros (e.g. __KERNEL__, __KERNEL_STRICT_NAMES) this pass gets rid of everything that is guarded by a well-known macro definition. this means that a block like #ifdef __KERNEL__ .... #endif will be totally omitted from the output. the optimizer is smart enough to handle all complex C-preprocessor conditional expression appropriately. this means that, for example: #if defined(__KERNEL__) || defined(FOO) ... #endif will be transformed into: #ifdef FOO ... #endif see tools/defaults.py for the list of well-known macros used in this pass, in case you need to update it in the future. note that this also remove any reference to a kernel-specific configuration macro like CONFIG_FOO from the clean headers. 2. remove variable and function declarations: this pass scans non-directive text and only keeps things that look like a typedef/struct/union/enum declaration. this allows to get rid of any variable or function declaration that should only be used within the kernel anyway (and which normally *should* be guarded in a #ifdef __KERNEL__ ... #endif block, if the kernel writers were not so messy) there are however a few exceptions: it is seldom useful to keep the definition of some static inline functions performing very simple operations. a good example is the optimized 32-bit byte-swap function found in arch-arm/asm/byteorder.h the list of exceptions is in tools/defaults.py in case you need to update it in the future. note that we do *not* remove macro definitions, including these macro that perform a call to one of these kernel-header functions, or even define other functions. we consider it safe since userland applications have no business using them anyway. 3. whitespace cleanup: the final pass remove any comments and empty lines from the final headers. 4. add a standard disclaimer: prepended to each generated header, contains a message like "do not edit directly - file was auto-generated by ...." RATIONALE: ========== OVERVIEW OF THE CURRENT KERNEL HEADER MESS: ------------------------------------------- The original kernel headers are not easily usable from userland applications. they contain many declarations and construct that will result in a compilation failure or even worse, incorrect behaviour. for example: - some headers try to define Posix types (e.g. size_t, ssize_t) that can conflict with the corresponding definitions provided by your C library. - some headers use constructs that cannot be compiled in ANSI C mode. - some headers use constructs do not compile with C++ at all. - some headers contain invalid "legacy" definitions for the benefit of old C libraries (e.g. glibc5) but result in incorrect behaviour if used directly. e.g. gid_t being defined in as a 16-bit type while the kernel uses 32-bit ids. this results in problems when getgroups() or setgroups() are called, since they operate on gid_t arrays. unfortunately, these headers are also the only source of some really extensive constant and type definitions that are required by userland applications. think any library/program that need to access ALSA, or Video4Linux, or anything related to a specific device or Linux-specific system interface (e.g. IOCTLS, etc...) As a consequence, every Linux distribution provides a set of patched kernel headers to be used by userland applications (which installs in /usr/include/linux/, /usr/include/asm/, etc...). these are manually maintained by distribution packagers, and generated either manually or with various scripts. these headers are also tailored to GNU LibC and cannot be reused easily by Bionic. for a really long period, the kernel authors have stated that they don't want to fix the problem, even when someone proposed a patch to start cleaning the official headers. from their point of view this is purely a library author problem. fortunately, enlightnment happened, and the kernel now provides a way to install a set of "user-friendly" headers that are generated from the official ones by stripping the __KERNEL__ protected declarations. unfortunately, this is not enough for Bionic because the result still contains a few broken declarations that are difficult to route around. (see below for a little bit of details). we plan to be able to support these kernel-generated user-land headers in the future, but the priority on this issue is very low. WHAT WE DO: ----------- so we're doomed to repeat the same effort than anyone else. the big difference here is that we want to automate as much as possible the generation of the clean headers to easily support additional architectures in the future, and keep current with upstream changes in the header definitions with the least possible hassle. of course, this is only a race to the bottom. the kernel maintainers still feel free to randomly break the structure of their headers (e.g. moving the location of some files) occasionally, so we'll need to keep up with that by updating our build script/original headers as these cases happen. what we do is keep a set of "original" kernel headers, and process them automatically to generate a set of "clean" headers that can be used from userland and the C library. note that the "original" headers can be tweaked a little to avoid some subtle issues. for example: - when the location of various USB-related headers changes in the kernel source tree, we want to keep them at the same location in our generated headers (there is no reason to break the userland API for something like that). - sometimes, we prefer to take certain things out of blocks guarded by a #ifdef __KERNEL__ .. #endif. for example, on recent kernels only includes when in kernel mode. we make it available to userland as well since some code out there assumes that this is the case. - sometimes, the header is simply incorrect (e.g. it uses a type without including the header that defines it before-hand)