Home | History | Annotate | Download | only in bionic
      1 Working on bionic
      2 =================
      3 
      4 What are the big pieces of bionic?
      5 ----------------------------------
      6 
      7 #### libc/ --- libc.so, libc.a
      8 
      9 The C library. Stuff like `fopen(3)` and `kill(2)`.
     10 
     11 #### libm/ --- libm.so, libm.a
     12 
     13 The math library. Traditionally Unix systems kept stuff like `sin(3)` and
     14 `cos(3)` in a separate library to save space in the days before shared
     15 libraries.
     16 
     17 #### libdl/ --- libdl.so
     18 
     19 The dynamic linker interface library. This is actually just a bunch of stubs
     20 that the dynamic linker replaces with pointers to its own implementation at
     21 runtime. This is where stuff like `dlopen(3)` lives.
     22 
     23 #### libstdc++/ --- libstdc++.so
     24 
     25 The C++ ABI support functions. The C++ compiler doesn't know how to implement
     26 thread-safe static initialization and the like, so it just calls functions that
     27 are supplied by the system. Stuff like `__cxa_guard_acquire` and
     28 `__cxa_pure_virtual` live here.
     29 
     30 #### linker/ --- /system/bin/linker and /system/bin/linker64
     31 
     32 The dynamic linker. When you run a dynamically-linked executable, its ELF file
     33 has a `DT_INTERP` entry that says "use the following program to start me".  On
     34 Android, that's either `linker` or `linker64` (depending on whether it's a
     35 32-bit or 64-bit executable). It's responsible for loading the ELF executable
     36 into memory and resolving references to symbols (so that when your code tries to
     37 jump to `fopen(3)`, say, it lands in the right place).
     38 
     39 #### tests/ --- unit tests
     40 
     41 The `tests/` directory contains unit tests. Roughly arranged as one file per
     42 publicly-exported header file.
     43 
     44 #### benchmarks/ --- benchmarks
     45 
     46 The `benchmarks/` directory contains benchmarks.
     47 
     48 
     49 What's in libc/?
     50 ----------------
     51 
     52 <pre>
     53 libc/
     54   arch-arm/
     55   arch-arm64/
     56   arch-common/
     57   arch-mips/
     58   arch-mips64/
     59   arch-x86/
     60   arch-x86_64/
     61     # Each architecture has its own subdirectory for stuff that isn't shared
     62     # because it's architecture-specific. There will be a .mk file in here that
     63     # drags in all the architecture-specific files.
     64     bionic/
     65       # Every architecture needs a handful of machine-specific assembler files.
     66       # They live here.
     67     include/
     68       machine/
     69         # The majority of header files are actually in libc/include/, but many
     70         # of them pull in a <machine/something.h> for things like limits,
     71         # endianness, and how floating point numbers are represented. Those
     72         # headers live here.
     73     string/
     74       # Most architectures have a handful of optional assembler files
     75       # implementing optimized versions of various routines. The <string.h>
     76       # functions are particular favorites.
     77     syscalls/
     78       # The syscalls directories contain script-generated assembler files.
     79       # See 'Adding system calls' later.
     80 
     81   include/
     82     # The public header files on everyone's include path. These are a mixture of
     83     # files written by us and files taken from BSD.
     84 
     85   kernel/
     86     # The kernel uapi header files. These are scrubbed copies of the originals
     87     # in external/kernel-headers/. These files must not be edited directly. The
     88     # generate_uapi_headers.sh script should be used to go from a kernel tree to
     89     # external/kernel-headers/ --- this takes care of the architecture-specific
     90     # details. The update_all.py script should be used to regenerate bionic's
     91     # scrubbed headers from external/kernel-headers/.
     92 
     93   private/
     94     # These are private header files meant for use within bionic itself.
     95 
     96   dns/
     97     # Contains the DNS resolver (originates from NetBSD code).
     98 
     99   upstream-freebsd/
    100   upstream-netbsd/
    101   upstream-openbsd/
    102     # These directories contain unmolested upstream source. Any time we can
    103     # just use a BSD implementation of something unmodified, we should.
    104     # The structure under these directories mimics the upstream tree,
    105     # but there's also...
    106     android/
    107       include/
    108         # This is where we keep the hacks necessary to build BSD source
    109         # in our world. The *-compat.h files are automatically included
    110         # using -include, but we also provide equivalents for missing
    111         # header/source files needed by the BSD implementation.
    112 
    113   bionic/
    114     # This is the biggest mess. The C++ files are files we own, typically
    115     # because the Linux kernel interface is sufficiently different that we
    116     # can't use any of the BSD implementations. The C files are usually
    117     # legacy mess that needs to be sorted out, either by replacing it with
    118     # current upstream source in one of the upstream directories or by
    119     # switching the file to C++ and cleaning it up.
    120 
    121   malloc_debug/
    122     # The code that implements the functionality to enable debugging of
    123     # native allocation problems.
    124 
    125   stdio/
    126     # These are legacy files of dubious provenance. We're working to clean
    127     # this mess up, and this directory should disappear.
    128 
    129   tools/
    130     # Various tools used to maintain bionic.
    131 
    132   tzcode/
    133     # A modified superset of the IANA tzcode. Most of the modifications relate
    134     # to Android's use of a single file (with corresponding index) to contain
    135     # time zone data.
    136   zoneinfo/
    137     # Android-format time zone data.
    138     # See 'Updating tzdata' later.
    139 </pre>
    140 
    141 
    142 Adding system calls
    143 -------------------
    144 
    145 Adding a system call usually involves:
    146 
    147   1. Add entries to SYSCALLS.TXT.
    148      See SYSCALLS.TXT itself for documentation on the format.
    149   2. Run the gensyscalls.py script.
    150   3. Add constants (and perhaps types) to the appropriate header file.
    151      Note that you should check to see whether the constants are already in
    152      kernel uapi header files, in which case you just need to make sure that
    153      the appropriate POSIX header file in libc/include/ includes the
    154      relevant file or files.
    155   4. Add function declarations to the appropriate header file.
    156   5. Add the function name to the correct section in libc/libc.map.txt and
    157      run `./libc/tools/genversion-scripts.py`.
    158   6. Add at least basic tests. Even a test that deliberately supplies
    159      an invalid argument helps check that we're generating the right symbol
    160      and have the right declaration in the header file, and that you correctly
    161      updated the maps in step 5. (You can use strace(1) to confirm that the
    162      correct system call is being made.)
    163 
    164 
    165 Updating kernel header files
    166 ----------------------------
    167 
    168 As mentioned above, this is currently a two-step process:
    169 
    170   1. Use generate_uapi_headers.sh to go from a Linux source tree to appropriate
    171      contents for external/kernel-headers/.
    172   2. Run update_all.py to scrub those headers and import them into bionic.
    173 
    174 
    175 Updating tzdata
    176 ---------------
    177 
    178 This is fully automated (and these days handled by the libcore team, because
    179 they own icu, and that needs to be updated in sync with bionic):
    180 
    181   1. Run update-tzdata.py in external/icu/tools/.
    182 
    183 
    184 Verifying changes
    185 -----------------
    186 
    187 If you make a change that is likely to have a wide effect on the tree (such as a
    188 libc header change), you should run `make checkbuild`. A regular `make` will
    189 _not_ build the entire tree; just the minimum number of projects that are
    190 required for the device. Tests, additional developer tools, and various other
    191 modules will not be built. Note that `make checkbuild` will not be complete
    192 either, as `make tests` covers a few additional modules, but generally speaking
    193 `make checkbuild` is enough.
    194 
    195 
    196 Running the tests
    197 -----------------
    198 
    199 The tests are all built from the tests/ directory.
    200 
    201 ### Device tests
    202 
    203     $ mma # In $ANDROID_ROOT/bionic.
    204     $ adb root && adb remount && adb sync
    205     $ adb shell /data/nativetest/bionic-unit-tests/bionic-unit-tests32
    206     $ adb shell \
    207         /data/nativetest/bionic-unit-tests-static/bionic-unit-tests-static32
    208     # Only for 64-bit targets
    209     $ adb shell /data/nativetest64/bionic-unit-tests/bionic-unit-tests64
    210     $ adb shell \
    211         /data/nativetest64/bionic-unit-tests-static/bionic-unit-tests-static64
    212 
    213 Note that we use our own custom gtest runner that offers a superset of the
    214 options documented at
    215 <https://github.com/google/googletest/blob/master/googletest/docs/AdvancedGuide.md#running-test-programs-advanced-options>,
    216 in particular for test isolation and parallelism (both on by default).
    217 
    218 ### Device tests via CTS
    219 
    220 Most of the unit tests are executed by CTS. By default, CTS runs as
    221 a non-root user, so the unit tests must also pass when not run as root.
    222 Some tests cannot do any useful work unless run as root. In this case,
    223 the test should check `getuid() == 0` and do nothing otherwise (typically
    224 we log in this case to prevent accidents!). Obviously, if the test can be
    225 rewritten to not require root, that's an even better solution.
    226 
    227 Currently, the list of bionic CTS tests is generated at build time by
    228 running a host version of the test executable and dumping the list of
    229 all tests. In order for this to continue to work, all architectures must
    230 have the same number of tests, and the host version of the executable
    231 must also have the same number of tests.
    232 
    233 Running the gtests directly is orders of magnitude faster than using CTS,
    234 but in cases where you really have to run CTS:
    235 
    236     $ make cts # In $ANDROID_ROOT.
    237     $ adb unroot # Because real CTS doesn't run as root.
    238     # This will sync any *test* changes, but not *code* changes:
    239     $ cts-tradefed \
    240         run singleCommand cts --skip-preconditions -m CtsBionicTestCases
    241 
    242 ### Host tests
    243 
    244 The host tests require that you have `lunch`ed either an x86 or x86_64 target.
    245 Note that due to ABI limitations (specifically, the size of pthread_mutex_t),
    246 32-bit bionic requires PIDs less than 65536. To enforce this, set /proc/sys/kernel/pid_max
    247 to 65536.
    248 
    249     $ ./tests/run-on-host.sh 32
    250     $ ./tests/run-on-host.sh 64   # For x86_64-bit *targets* only.
    251 
    252 You can supply gtest flags as extra arguments to this script.
    253 
    254 ### Against glibc
    255 
    256 As a way to check that our tests do in fact test the correct behavior (and not
    257 just the behavior we think is correct), it is possible to run the tests against
    258 the host's glibc.
    259 
    260     $ ./tests/run-on-host.sh glibc
    261 
    262 
    263 Gathering test coverage
    264 -----------------------
    265 
    266 For either host or target coverage, you must first:
    267 
    268  * `$ export NATIVE_COVERAGE=true`
    269      * Note that the build system is ignorant to this flag being toggled, i.e. if
    270        you change this flag, you will have to manually rebuild bionic.
    271  * Set `bionic_coverage=true` in `libc/Android.mk` and `libm/Android.mk`.
    272 
    273 ### Coverage from device tests
    274 
    275     $ mma
    276     $ adb sync
    277     $ adb shell \
    278         GCOV_PREFIX=/data/local/tmp/gcov \
    279         GCOV_PREFIX_STRIP=`echo $ANDROID_BUILD_TOP | grep -o / | wc -l` \
    280         /data/nativetest/bionic-unit-tests/bionic-unit-tests32
    281     $ acov
    282 
    283 `acov` will pull all coverage information from the device, push it to the right
    284 directories, run `lcov`, and open the coverage report in your browser.
    285 
    286 ### Coverage from host tests
    287 
    288 First, build and run the host tests as usual (see above).
    289 
    290     $ croot
    291     $ lcov -c -d $ANDROID_PRODUCT_OUT -o coverage.info
    292     $ genhtml -o covreport coverage.info # or lcov --list coverage.info
    293 
    294 The coverage report is now available at `covreport/index.html`.
    295 
    296 
    297 Running the benchmarks
    298 ----------------------
    299 
    300 ### Device benchmarks
    301 
    302     $ mma
    303     $ adb remount
    304     $ adb sync
    305     $ adb shell /data/nativetest/bionic-benchmarks/bionic-benchmarks
    306     $ adb shell /data/nativetest64/bionic-benchmarks/bionic-benchmarks
    307 
    308 You can use `--benchmark_filter=getpid` to just run benchmarks with "getpid"
    309 in their name.
    310 
    311 ### Host benchmarks
    312 
    313 See the "Host tests" section of "Running the tests" above.
    314 
    315 
    316 Attaching GDB to the tests
    317 --------------------------
    318 
    319 Bionic's test runner will run each test in its own process by default to prevent
    320 tests failures from impacting other tests. This also has the added benefit of
    321 running them in parallel, so they are much faster.
    322 
    323 However, this also makes it difficult to run the tests under GDB. To prevent
    324 each test from being forked, run the tests with the flag `--no-isolate`.
    325 
    326 
    327 32-bit ABI bugs
    328 ---------------
    329 
    330 ### `off_t` is 32-bit.
    331 
    332 On 32-bit Android, `off_t` is a signed 32-bit integer. This limits functions
    333 that use `off_t` to working on files no larger than 2GiB.
    334 
    335 Android does not require the `_LARGEFILE_SOURCE` macro to be used to make
    336 `fseeko` and `ftello` available. Instead they're always available from API
    337 level 24 where they were introduced, and never available before then.
    338 
    339 Android also does not require the `_LARGEFILE64_SOURCE` macro to be used
    340 to make `off64_t` and corresponding functions such as `ftruncate64` available.
    341 Instead, whatever subset of those functions was available at your target API
    342 level will be visible.
    343 
    344 There are a couple of exceptions to note. Firstly, `off64_t` and the single
    345 function `lseek64` were available right from the beginning in API 3. Secondly,
    346 Android has always silently inserted `O_LARGEFILE` into any open call, so if
    347 all you need are functions like `read` that don't take/return `off_t`, large
    348 files have always worked.
    349 
    350 Android support for `_FILE_OFFSET_BITS=64` (which turns `off_t` into `off64_t`
    351 and replaces each `off_t` function with its `off64_t` counterpart, such as
    352 `lseek` in the source becoming `lseek64` at runtime) was added late. Even when
    353 it became available for the platform, it wasn't available from the NDK until
    354 r15. Before NDK r15, `_FILE_OFFSET_BITS=64` silently did nothing: all code
    355 compiled with that was actually using a 32-bit `off_t`. With a new enough NDK,
    356 the situation becomes complicated. If you're targeting an API before 21, almost
    357 all functions that take an `off_t` become unavailable. You've asked for their
    358 64-bit equivalents, and none of them (except `lseek`/`lseek64`) exist. As you
    359 increase your target API level, you'll have more and more of the functions
    360 available. API 12 adds some of the `<unistd.h>` functions, API 21 adds `mmap`,
    361 and by API 24 you have everything including `<stdio.h>`. See the
    362 [linker map](libc/libc.map.txt) for full details.
    363 
    364 In the 64-bit ABI, `off_t` is always 64-bit.
    365 
    366 ### `sigset_t` is too small for real-time signals.
    367 
    368 On 32-bit Android, `sigset_t` is too small for ARM and x86 (but correct for
    369 MIPS). This means that there is no support for real-time signals in 32-bit
    370 code.
    371 
    372 In the 64-bit ABI, `sigset_t` is the correct size for every architecture.
    373 
    374 ### `time_t` is 32-bit.
    375 
    376 On 32-bit Android, `time_t` is 32-bit. The header `<time64.h>` and type
    377 `time64_t` exist as a workaround, but the kernel interfaces exposed on 32-bit
    378 Android all use the 32-bit `time_t`.
    379 
    380 In the 64-bit ABI, `time_t` is 64-bit.
    381