An upcoming change to scudo will cause us to start calling
android_unsafe_frame_pointer_chase() from within the allocator. Since this
function uses ScopedDisableMTE, this would otherwise make it unsafe to use
the allocator from within ScopedDisableMTE. This seems like an unreasonable
restriction, so make ScopedDisableMTE save the PSTATE.TCO state in the
constructor and restore it in the destructor.
Bug: 135772972
Change-Id: I47e18d5fb2929efd5a58676488180cd85731007b
This patch enables GWP-ASan with process sampling.
**Note**: If you are visiting this patch because this broke a test or
otherwise is causing failures, please contact mitchp@ directly (or
respond to this patchset). GWP-ASan is designed to cause heap-based
memory safety bugs to manifest in SEGV on a sampled basis.
Bug: 135634846
Test: atest bionic-unit-tests gwp_asan_unittest
Change-Id: I58ca9373def105fdd718cf283482b3220b770698
GWP-ASan + heapprofd don't currently play nice together in some
circumstances. heapprofd thinks it's still an only child, and refuses to
accept the existence of its little brother, GWP-ASan.
If GWP-ASan is installed before heapprofd, then heapprofd is *required*
to respect that libc has a favourite child. If an allocation/free is passed
to heapprofd, then heapprofd *must* (eventually) pass that allocation/free to
GWP-ASan. If heapprofd doesn't do this, then a free() of a GWP-ASan
allocation can be passed to the system allocator.
This can happen in two places right now:
1. The heapprofd hooks simply clobber any trace of what was
previously in the default_dispatch_table when enabled through the
heapprofd signal.
2. Heapprofd can die when the system is under significant pressure.
Some pipes can timeout, which ends up in the client calling ShutdownLazy()
-> mallopt(M_RESET_HOOKS) -> DispatchReset(). This also clobbers any
trace of the previous default_dispatch_table.
To fix both these problems, we fix heapprofd to restore the previous
default_dispatch_table whenever either circumstance happens. We do some
tricky copying to avoid race conditions on the malloc_dispatch_table in
fixing #1.
Bug: 135634846
Test: Run HeapprofdEndToEnd.NativeProfilingActiveAtProcessExit/ForkMode
a significant number of times with large amounts of system pressure (I
just run bionic-unit-tests-scudo in parallel). You will see some test
failures where heapprofd died due to system pressure, but never a death
from the allocator. Tests should never fail when the system isn't under
immense pressure.
Change-Id: I20ab340d4bdc35d6d1012da5ee1a25634428d097
When enabled, GWP-ASan sets the current dispatch table. Then, when a
shim layer (malloc_debug, malloc_hooks, heapprofd) comes along, they
should (by design) overwrite the current dispatch table.
Currently, these shim layers check to see whether malloc_limit is
installed by checking the current dispatch table against nullptr.
Because GWP-ASan owns the current dispatch table, the shim thinks that
malloc_limit is installed and falls back to only use the default
dispatch, thinking that malloc_limit will call them. This is not the
case, and they should take over the current dispatch pointer.
Bug: 135634846
Test: atest bionic
Change-Id: Ifb6f8864a15af9ac7f20d9364c40f73c5dd9d870
The WriteProtected mutator for __libc_globals isn't reentrant.
Previously we were calling __libc_globals.mutate() inside of GWP-ASan's
libc initialisation, which is called inside the __libc_globals.mutate().
This causes problems with malloc_debug and other malloc shims, as they
fail to install when GWP-ASan is sampling their processes.
Bug: 135634846
Test: atest bionic
Change-Id: Iae51faa8d78677eeab6204b6ab4f3ae1b7517ba5
Scudo still isn't quite at the same RSS as jemalloc for the svelte config
so only enable this for normal config.
Bug: 137795072
Test: Built svelte config and verified it is still jemalloc.
Test: Ran performance tests on normal config (bionic benchmarks).
Test: Ran trace tests (system/extras/memory_replay).
Test: Ran scudo unit tests.
Test: Ran bionic unit tests.
Test: Ran libmemunreachable tests.
Test: Ran atest CtsRsBlasTestCases on cuttlefish.
Test: Ran atest AslrMallocTest.
Test: Ran atest CtsHiddenApiKillswitchWildcardTestCases and verified it has
Test: the same runtime as the jemalloc.
Change-Id: I241165feb8fe9ea814b7b166e3aaa6563d18524a
I had hoped that this would then let us remove more of the "introduced
in" annotations, but it looks like that's not really going to happen
until the NDK's minimum supported API is 21.
Also remove a .c file that wasn't referenced anywhere.
Test: treehugger
Change-Id: I8b4a61c082293f8445195a4fa5ee30595d70444e
This patch introduces GWP-ASan - a sampled allocator framework that
finds use-after-free and heap-buffer-overflow bugs in production
environments.
GWP-ASan is being introduced in an always-disabled mode. This means that
GWP-ASan will be permanently disabled until a further patch turns on
support. As such, there should be no visible functional change for the
time being.
GWP-ASan requires -fno-emulated-tls wherever it's linked from. We
intentionally link GWP-ASan into libc so that it's part of the initial
set of libraries, and thus has static TLS storage (so we can use
Initial-Exec TLS instead of Global-Dynamic). As a benefit, this reduces
overhead for a sampled process.
GWP-ASan is always initialised via. a call to
mallopt(M_INITIALIZE_GWP_ASAN, which must be done before a process is
multithreaded).
More information about GWP-ASan can be found in the upstream
documentation: http://llvm.org/docs/GwpAsan.html
Bug: 135634846
Test: atest bionic
Change-Id: Ib9bd33337d17dab39ac32f4536bff71bd23498b0
tracefs will be mounted at /sys/kernel/tracing when debugfs
is not mounted.
Bug: 134669095
Test: atest bionic-unit-tests-static
Change-Id: Ic224cf13500efc570da8b6a27ce925bbcf068fdd
For consistency, linker namespace for apex modules use its apex name
instead of hard-coded short name.
Bug: 148826508
Test: m / boot
Change-Id: I4bf565cd528d744fc42841fd2d9f8bf652d4d346
We've optimized the ctype functions to the point where they're pretty
much all down to one instruction. This change takes the obvious next
step of just inlining them.
On Android these function have only ever been for ASCII. You need the
<wctype.h> functions for non-ASCII.
libc++ currently has its own inlines for the _l variants, so if we want
to just inline them in bionic directly, we'll need to coordinate that.
Bug: http://b/144165498
Test: treehugger plus benchmarks
Change-Id: I4cc8aa96f7994ae710a562cfc9d4f220ab7babd6
Over the last year, LLVM apparently learned how to optimize many
FORTIFY'ed functions. I went through the list of functions it optimizes,
and simplified their implementations here.
This is more than a code health thing; __bos_trivially_ge expands to a
branch that's not eliminated until after inlining, so it can actually
cause some functions (like one of std::string's ctors) to become
uninlineable.
Bug: 148189733
Test: hand-checked the IR we get for each of the changed functions. Many
get optimized to their non-_chk variant when appropriate. Others
will get optimized to non-_chk versions when bos == -1. Bug repro
also now shows all 'inline's.
Change-Id: Ic360818ad9daaeda3958e1282af41087f85122a3
These just cause confusion because they often have different
values/layouts, but they're never actually used.
Test: treehugger
Change-Id: I424034088e017c919f62fcefa7d6d3f903f31cfb
This function will be used by Scudo and GWP-ASan to efficiently collect
stack traces for frames built with frame pointers.
Bug: 135634846
Bug: 135772972
Change-Id: Ic63efdbafe11dfbb1226b5b4b403d53c4dbf28f3
Merged-In: Ic63efdbafe11dfbb1226b5b4b403d53c4dbf28f3
If we remove the mips uapi headers, versioner fails because it assumes
they're available. We'll need a new versioner prebuilt beforre we can
remove the libc/versioner-dependencies/mips* symlinks.
Test: treehugger
Change-Id: Ife6df0cb57938f806a31ec334d648df9694c3d17
Add a hook that's called upon file descriptor creation to libc, and a
library that uses it to capture backtraces for file descriptor creation,
to make it easier to hunt down file descriptor leaks.
Currently, this doesn't capture all of the ways of creating a file
descriptor, but completeness isn't required for this to be useful as
long as leaked file descriptors are created with a function that is
tracked. The primary unhandled case is binder, which receives file
descriptors as a payload in a not-trivially-parsable byte blob, but
there's a chance that the leak we're currently trying to track down
isn't of a file descriptor received over binder, so leave that for
later.
Bug: http://b/140703823
Test: manual
Change-Id: I308a14c2e234cdba4207157b634ab6b8bc539dd9
(cherry picked from commit b7eccd4b15)
Many of our header files are very sensitive to the order in which
their #includes appear...
Bug: N/A
Test: N/A
Change-Id: I2c21cac5e9bd49b7e80620d14971af8fefa17e91
mte_supported() lets code efficiently detect the presence of MTE, and
ScopedDisableMTE lets code disable MTE RAII-style in a particular region
of code.
Bug: 135772972
Change-Id: I628a054b50d79f67f39f35d44232b7a2ae166afb
Even with formatting off, clang still tries to rearrange the include
files or the using statements, so disable that too.
Test: Verified that the include directories are not rearranged.
Change-Id: I991a1b2bfa94a8202c5a486664658d654f1c7811
On Android, fcntl is always implemented by fcntl64(2). This means that
an LP32 binary can `fcntl(F_SETLK, struct flock)` (because fcntl64(2)
passes through to the 32-bit fcntl(2) to handle F_SETLK), and it can
also `fcntl(F_SETLK64, struct flock64)`. What it can't do before this
patch is set _FILE_OFFSET_BITS=64 and then `fcntl(F_SETLK, struct
flock)` where that `struct flock` is actually implicitly `struct
flock64`.
Move the kernel uapi structs out of the way, define them ourselves based
on __LP64__ and _FILE_OFFSET_BITS, and fix up the relevant F_ constants.
(Also add a .clang-format to turn off clang-format in libc/include/.)
Bug: N/A
Test: treehugger (and strace!)
Change-Id: Iccd6c83d9133e1efcf93a7b49a6ae0f1bbd3d58b
Use the .clang-format-2 found in system/core instead of this which is
not actually being used.
Also, enable clang-format running by default.
All upstream directories are marked as ignoring formatting so that
their source files are not modified.
Test: NA
Change-Id: Icee6030f373fa5f072df162f97e6f34320e3d89a
