Linking glibc in fully static mode is mostly unsupported. While such binaries
can easily be produced, conflicting symbols will often make them crash at
runtime. This happens because glibc will always (try to) load some dynamically
linked libraries, even when statically linked. This includes things like the
resolver, unicode/locale handling and others.
Internally at Google, this is not a concern due to the way glibc is being built
there. But in order to make all of our tests run in the open-source version of
this code, we need to change strategy a bit.
As a rule of thumb, glibc can safely be linked statically if a program is
resonably simple and does not use any networking of locale dependent
facilities. Calling syscalls directly instead of the corresponding libc
wrappers works as well, of course.
This change adjusts linker flags and sandbox policies to be more compatible
with regular Linux distributions.
Tested:
- `ctest -R '[A-Z].*'` (all SAPI/Sandbox2 tests)
PiperOrigin-RevId: 429025901
Change-Id: I46b677d9eb61080a8fe868002a34a77de287bf2d
- Link `zipe.c` statically (safe)
- Update policy to allow any use of `stat()`
PiperOrigin-RevId: 428971638
Change-Id: Ib0f5f496ea2389582986b41a8830592e6c1d4390
This fixes a couple of tests in the open source version of the code.
Internally, since we are using a different ELF loader, the page offset
will always be zero. Hence we never notices this was broken.
PiperOrigin-RevId: 427996428
Change-Id: I44c5b5610b074cf69b9f0c5eeb051be50923e351
Note that `//sandboxed_api/sandbox2:stack_trace_test` may still fail for
unrelated reasons, as we are linking libc statically, which is brittle. A
follow-up change will fix this.
PiperOrigin-RevId: 427175045
Change-Id: Ifb5ec2ac3d60f4bcc9708f26c834c83b75e769d7
Newer versions of libunwind use `PTRACE_GETREGSET` to obtain register data.
This change should make it easier to upgrade the libunwind dependency.
PiperOrigin-RevId: 420057842
Change-Id: Ib9abbeff574e457009709715f912ba5962033c5d
- Allow to specify multiple syscalls with `BlockSyscallsWithErrno()`
- Add functions to allow `unlink()` and `rename()` in all their spellings
PiperOrigin-RevId: 414987303
Change-Id: Ic0e680b785e8e3a3498f20e6a7403737e63fe876
__sanitizer_sandbox_on_notify is not tsan specific.
It's empty for other sanitizers now, but we are going to need it soon.
PiperOrigin-RevId: 414873197
Change-Id: I251ac38e5c886980b4baa7f05306643599a25090
Move VecStringToCharPtrArr before fork, so that it cannot deadlock when other thread holds allocation lock.
PiperOrigin-RevId: 414661912
Change-Id: Ie8aa5c36693e6f86c69d67a1da51b7e7ff1ec30b
Different versions of the `elf.h` header define their own integer types. For
example, even on LP64 systems, a 64-bit ELF integer types may decay into
`unsigned long long` instead of `unsigned long`.
This change replaces the various overloads with a single function template
that is well-defined for all integral types.
PiperOrigin-RevId: 410746713
Change-Id: I4b560f7541802372f01ae3d6f4a56554e51d70c8
Generate syscall jump table without using bpf_helper.
Check that any jump in the user provided policy is within the provided policy.
PiperOrigin-RevId: 409362089
Change-Id: I31493e52cf868e4b184ff79fcb26beeb75f49773
Note: This change allows `MAP_NORESERVE` generally, not just for MSAN. This follows
what we do for `AllowTcMalloc()/AllowSystemMalloc()`
PiperOrigin-RevId: 402231980
Change-Id: Ifa1c6b9f61f636dd6db231dde3765c3b4a40911b
These were previously dependent on an internal-only testing target.
For now, this only works with Bazel, but should enable us to have better test coverage in GitHub actions.
Eventually, all of these shell-based tests should be converted to `cc_test`s.
PiperOrigin-RevId: 400713615
Change-Id: I1cabb5b72977987ef4a1803480f699b58c4d56e9
For OSS, this change should be mostly a no-op. Visible edits are due to
changed order of code and/or includes.
PiperOrigin-RevId: 394177395
Change-Id: I1d32f9fd175579e8f05c051b1307953b249d139d
This mainly a debugging facility.
It makes diagnosing problems where sandboxed process just randomly exits whereas unsandboxed one runs to completion due to differences in the setup/environment much easier.
PiperOrigin-RevId: 391005548
Change-Id: Ia19fe6632748da93c1f4291bb55e895f50a4e2b0
Otherwise starting forkserver multiple times will result in zombie processes lingering around.
PiperOrigin-RevId: 388926497
Change-Id: Ia9947cce3d9e909edd709b0d3525e1ae8b8bbc51
Also really own `exec_fd_` as previously if the executor is destructed without calling `StartSubProcess` the file descriptor would leak.
PiperOrigin-RevId: 388901766
Change-Id: I6bbb15ced37a0a832ec5a5228452a3d54ef46ee9
Calling `Terminate()` issues additional syscalls that may clobber the `errno`
value. Reordering the log statements ensures we actually log the initial error
in `read()`/`write()`.
PiperOrigin-RevId: 387576942
Change-Id: I0f9c8c6001e6dc4ca098abe02cd251029f92a737
1. In many cases, sandboxes need to allow /proc/stat and /proc/cpuinfo so that
get_nprocs(3) will work; otherwise, per-CPU logic can't determine how many CPUs
there are. Unfortunately, some of those sandboxes also disable namespaces. The
solution is to provide two functions: AllowRestartableSequencesWithProcFiles(),
which allows syscalls and files; and AllowRestartableSequences(), which allows
syscalls only. Sandboxes should usually call the former; sandboxes that disable
namespaces should instead call the latter and are responsible for allowing the
files via the deprecated Fs mechanism.
2. Make the mmap(2) policy evaluate prot AND flags, not prot OR flags.
3. Order the code and the comments identically for better readability.
PiperOrigin-RevId: 386414028
Change-Id: I016b1854ed1da9c9bcff7b351c5e0041093b8193
The existing function signature took a `unique_ptr<>` owning a vector, and
took `nullptr` to mean an empty set of capabilities. This is more naturally
modeled by taking the vector directly and `std::move`-ing it.
PiperOrigin-RevId: 384214849
Change-Id: I177f04a06803ae00429b19a1f3f12e7be04d2908
- Assign to `*mutable_XXX()` instead of looping
- Use a const ref for capabilities
PiperOrigin-RevId: 384192675
Change-Id: I4db3d0c8ce0d7f6acc9fd486a2409962516b5fe7
This bug only manifests if a lot of fds are open when global forkserver is started.
If the allocated exec_fd number was equal Comms::kSandbox2ClientCommsFD then it would be replaced by the comms fd and result in EACCESS at execveat.
PiperOrigin-RevId: 380805414
Change-Id: I31427fa929abfc60890477b55790cc14c749f7f5
Recenly, Debian based distribution kernels started activating the Tomoyo Linux
Security Module by default. Even if it is not used, this changes the behavior
of `/dev/fd` (pointing to `/proc/self/fd` by default), which Sandbox2 needs during
`execveat()`.
As a result, Sandbox2 and Sandboxed API always fail without one of the following
conditions
- `/proc` mounted within the sandboxee
- `/dev` mounted
- `/dev/fd` symlinked to `/proc/self/fd` in the sandboxee's mount namespace
Some code pointers to upstream Linux 5.12.2:
- https://elixir.bootlin.com/linux/v5.12.2/source/fs/exec.c#L1775
- https://elixir.bootlin.com/linux/v5.12.2/source/security/tomoyo/tomoyo.c#L107
- https://elixir.bootlin.com/linux/v5.12.2/source/security/tomoyo/domain.c#L729
To find out whether your system has Tomoyo enabled, use this command, similar to
what this change does in code:
```
$ cat /sys/kernel/security/lsm | grep tomoyo && echo "Tomoyo active"
capability,yama,apparmor,tomoyo
Tomoyo active
```
The config setting `CONFIG_DEFAULT_SECURITY` controls which LSMs are built into
the kernel by default.
PiperOrigin-RevId: 372919524
Change-Id: I2181819c04f15f57d96c44ea9977d0def4a1b623
This is needed for some compiler versions where `absl::string_view` == `std::string_view`.
PiperOrigin-RevId: 367392064
Change-Id: Id91d23510501df4745f386475ef9049d94062e1b
This changes the workflow definition so that we always try to install
compiler toolchains that we need.
See https://github.com/actions/virtual-environments/issues/2950 for more
context.
Drive-by:
- Mini fix to enable compilation under Clang 6.0
Signed-off-by: Christian Blichmann <cblichmann@google.com>
Now unwinding will properly handle binaries inside bind-mounted directories.
Drive-by:
- Get rid of n^2 path handling
- Get rid of namespace alias
PiperOrigin-RevId: 358353666
Change-Id: Ieec7690ec6a1ae6d358de375220566b69e8cb094
Using C++17 means we can get rid of many `#ifdef`s by using `if constexpr`.
This way, we ensure that both branches compile and still retain zero runtime
overhead.
Note that open source builds of Sandboxed API do not ship with sanitizer
configurations yet. This will be added in follow-up changes.
PiperOrigin-RevId: 354932160
Change-Id: I3678dffc47ea873919f0a8c01f3a7d999fc29a5b
Also accept `absl::string_view` and `absl::Span<const std::string>` arguments.
Drive-by:
- Move using declaration into namespace
PiperOrigin-RevId: 354271016
Change-Id: Iadd873377e51cac7fa3800aab1f9e85ff94bd4e9
Only externally visible changes should be a few changed includes as well as
some formatting changes.
PiperOrigin-RevId: 353226662
Change-Id: Iebf5be13774efcbd94c5d5a17b9b27e47275b229
This change should make it less confusing where utility code comes from.
Having it in two places made sense when we were debating whether to publish
Sandbox2 separately, but not any longer.
Follow-up changes will move `sandbox2/util.h` and rename the remaining
`sandbox2/util` folder.
PiperOrigin-RevId: 351601640
Change-Id: I6256845261f610e590c25e2c59851cc51da2d778
This change enables support for 32-bit ARM, as used by embedded controllers and older phones.
Note: This does not support 32-bit sandboxees on AArch64. Both sandboxee and host code must have the same bitness.
PiperOrigin-RevId: 347835193
Change-Id: I6395882677530f9862f118d2dc10230a61049836
Because any change that touches continuous integration needs a companion o.O
PiperOrigin-RevId: 347769780
Change-Id: I20525aaac2ce41c48f619b641baa31e880432e50
- Ubuntu 18.04 ships with GCC 7, which needs `std::move()` when returning an `absl::StatusOr<>`
- Ignore C++ AST nodes of type `cindex.TypeKind.UNEXPOSED` in Python generator
- Remove default values in `ubuntu-cmake.yml`
PiperOrigin-RevId: 347605109
Change-Id: Ibe167249ecf4ef1af1654d63c2e067fc02e5782d
This allows resource-constrained environments to benefit from the
space savings of dynamic linking. This is not meant to be used in
the general case.
PiperOrigin-RevId: 347398828
Change-Id: Ia634959148a31159878f48c44255dd733424a2b8
FS checks are an internal feature that has been deprecated for a while in
favor of user namespaces.
PiperOrigin-RevId: 347378761
Change-Id: I1d7956cecd6db47b2b96fdedaada0b2a36f9b112
The defined raw logging macros should be compatible with Abseil and
we can remove our version once Abseil releases theirs.
PiperOrigin-RevId: 347354273
Change-Id: I178a89cfd2e19bcd707a06fa9dfd7b767e2b654b
- Bazel: Use "incompatible" flag to fix fully static linking. The flag will
become the default in Bazel 4.0.
- Bazel: Deduplicate features into `FULLY_STATIC_FEATURES` variable
- CMake: Remove the testcase properties. `sapi::base` already sets
`POSITION_INDEPENDENT_CODE`. Note that `-pie` is incompatible with `-static`
and `-static-pie` requires GCC 8 and GLIBC 2.27.
PiperOrigin-RevId: 346952478
Change-Id: I7a317c90a3bec9691b13df1a00e3fddf4481df4d
The default policy causes immediate termination of a sandboxee that
calls `bpf`(2).
This does not allow for try-call use of `bpf()` to test for optional
features.
To support such try-call use cases, sandboxes would like to say:
```
sandbox2::PolicyBuilder builder;
builder.BlockSyscallWithErrno(__NR_bpf, EPERM);
```
but this doesn't work because the default policy unconditionally treats
`bpf()` as a sandbox violation.
Remove the bpf violation check from the policy if `bpf()` is explicitly
blocked with an errno.
PiperOrigin-RevId: 345239389
Change-Id: I7fcfd3a938c610c8679edf8e1fa0238b32cc9db4
The lookup is not on the hot path and this removes the SYSCALLS_UNUSED macros.
PiperOrigin-RevId: 344240762
Change-Id: I324bd798945851ac0b92e257206525eab4ec36e5
Updates syscall arguments mostly according to this list and more recent kernel sources:
https://chromium.googlesource.com/chromiumos/docs/+/master/constants/syscalls.md#x86_64-64_bit
The list includes some more syscalls that were recently added.
Follow-up changes will do the same for x86-32, POWER and AArch64.
PiperOrigin-RevId: 341016698
Change-Id: If1771fd37a47b227ca8f572704a64190e4621a38
No need for the smart pointer indirection when an `std::vector` can also hold
the BPF policy.
PiperOrigin-RevId: 340809220
Change-Id: I8a63567e8042d9ff875cba739e8552db87b6901a
- Use a `constexpr inline` string constant for the forkserver env var
- Add annotation for the comms channel mutex
PiperOrigin-RevId: 335395005
Change-Id: Ic058c19c3704f182aa7ed7b8e8964b2fc5082800
This is a work in progress:
- Syscall tables need work
- Only tested on real hardware using one of our test hosts
As a drive-by, this change also enables the open source version to function on
POWER.
Another side-effect of this change is that the default policies no longer
check for different host architectures at runtime. On x86_64, we do not need
to check for PPC or AArch64 specifice and vice versa.
PiperOrigin-RevId: 331137472
Change-Id: Ic6d6be5cbe61d83dbe13d5a0be036871754b2eb8
This allows us to remove some uses of macros.
Related changes:
- Make it clear that we support hosting sandboxed binaries from 64-bit
processes only. CPU architectures are x86-64 and POWER64 (little endian).
- Introduced CPU architecture macros, abstracting away compiler specifics
PiperOrigin-RevId: 330918134
Change-Id: Ife7ad5f14723eec9f68055127b0583b8aecd38dd
This removes our own fork of `absl::StatusOr<>`. Sandboxed API still includes
a custom matcher for Googletest, as that is not open source yet. For
compatibility, the `statusor.h` header is still retained and now aliases
`sapi::StatusOr<>` to `absl::StatusOr<>`.
PiperOrigin-RevId: 329916309
Change-Id: I0544b73a9e312dce499bc4128c28457e04ab9929
- Drop `delim` argument from the `GetStackTrace()` family of functions.
We only ever used plain spaces.
- Use an `std::vector<std::string>` for the symbolized stack frames and
adjust the unwind proto accordingly.
This change now prints each stack frame on its own line while skipping
duplicate ones:
```
I20200717 11:47:16.811381 3636246 monitor.cc:326] Stack trace: [
I20200717 11:47:16.811415 3636246 monitor.cc:337] map:/lib/x86_64-linux-gnu/libc-2.30.so+0xceee7(0x7fb871602ee7)
I20200717 11:47:16.811420 3636246 monitor.cc:337] Rot13File+0x130(0x55ed24615995)
I20200717 11:47:16.811424 3636246 monitor.cc:337] ffi_call_unix64+0x55(0x55ed2461f2dd)
I20200717 11:47:16.811429 3636246 monitor.cc:337] map:[stack]+0x1ec80(0x7ffee4257c80)
I20200717 11:47:16.811455 3636246 monitor.cc:339] (last frame repeated 196 times)
I20200717 11:47:16.811460 3636246 monitor.cc:347] ]
```
PiperOrigin-RevId: 322089140
Change-Id: I05b0de2f4118fed90fe920c06bbd70ea0d1119e2
Initializing `absl::Span`s like by assigning them from a temporary
array leaves them pointing to invalid data. Due to the way the linker
initializes these constant tables, _most_ of them will still be valid
_most_ of the time, leading to crashes when running sandboxees with the
`--sandbox2_danger_danger_permit_all_and_log` option.
PiperOrigin-RevId: 321112099
Change-Id: I891118da08cbb6000b3e2e275618bc4edaa1d020
Otherwise ExecuteProcess is implicitly `[[noreturn]]` and this
might cause policy violations in `__asan_handle_no_return`
for ASAN builds.
PiperOrigin-RevId: 319203128
Change-Id: I5c8ba71ce88261f803aa3f16730eccea0d803dd1
- Extract dependent types directly from the Clang AST and re-serialize
back into compilable code
- Collect types and emit diagnostics
- Format generated code
Signed-off-by: Christian Blichmann <mail@blichmann.eu>
Similar to what the Bazel build does, this change adds zlib as an additional
dependency when `SAPI_ENABLE_EXAMPLES` is set to `ON`.
PiperOrigin-RevId: 309203959
Change-Id: I201a9e6415789afb1e058bc48cebbc0fc0004fe9
BoringSSL (which is the crypto library used by most Google products) is starting to use madvise(_, _, MADV_WIPEONFORK) to protect random-number state from being duplicated by fork(). This causes extra madvise calls that sandboxes need to permit in order to continue functioning.
PiperOrigin-RevId: 309173849
Change-Id: I007dacc1ff1fd0ccc138caaa08735cfe5bc78234
When embedding SAPI in an external CMake project, the version of
`protobuf_generate_cpp` that we lifted from upstream protobuf produces
the wrong generated file paths.
For example, given this project structure:
```
/parent/
+-- myproject/
+-- myproject_build/ <- CMake build directory
+-- sandboxed-api/ <- Checkout from GitHub
```
And a CMake file in `myproject/CMakeLists.txt` that embeds SAPI like
this:
```
cmake_minimum_required(VERSION 3.12)
project(SandboxedTest LANGUAGES CXX)
set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
add_subdirectory(
${PROJECT_SOURCE_DIR}/../sandboxed-api
${PROJECT_BINARY_DIR}/sandboxed-api
)
```
Then `protobuf_generate_cpp` correctly invokes the protoc compiler to
generate
`/parent/myproject_build/sandboxed-api/sandboxed_api/proto_arg.proto.pb.cc'.
However, the path of the generated source file that is passed to the C++
compiler will be
`/parent/myproject_build/sandboxed-api/sandboxed_api/../../myproject_build/sandboxed-api/sandboxed_api/proto_arg.pb.cc`.
Note the duplicated project build directory component in the
canonicalized version:
`/parent/myproject_build/myproject_build/sandboxed-api/sandboxed_api/proto_arg.pb.cc`.
This change simple omits the computation of any relative file paths and
simply uses `_pb_PROTOC_OUT_DIR` which defauls to
`CMAKE_CURRENT_BINARY_DIR`, which should always contain the correct
path.
Signed-off-by: Christian Blichmann <mail@blichmann.eu>
Created documentation for network proxy. fixed 2 things in documentation (namespaces are enabled by default for a while).
PiperOrigin-RevId: 300321016
Change-Id: Id9c54b29551e8d3b70e814e2fdbfee594126aa90
Sending -1 as fd will fail and take forkserver down.
This should not happen normally so turned it into a check.
PiperOrigin-RevId: 285391908
Change-Id: Idbb05004c36cb0be57be1bd26df1c57cecfb0019
