What is Thread Sanitizer?¶
Thread Sanitizer (TSan) is a fast data race detector for C/C++ and Rust programs. It uses a compile-time instrumentation to check all non-race-free memory access at runtime. Unlike other tools, it understands compiler-builtin atomics and synchronization and therefore provides very accurate results with no false positives (except if unsupported synchronization primitives like inline assembly or memory fences are used). More information on how TSan works can be found on the Thread Sanitizer wiki.
A meta bug called tsan is maintained to keep track of all the bugs found with TSan.
Note that unlike other sanitizers, TSan is currently only supported on Linux.
Downloading artifact builds¶
The easiest way to get Firefox builds with Thread Sanitizer is to download a continuous integration TSan build of mozilla-central (updated at least daily):
mozilla-central optimized builds: linux
The fuzzing team also offers a tool called
fuzzfetch to download this and many
other CI builds. It makes downloading and unpacking these builds much easier and
can be used not just for fuzzing but for all purposes that require a CI build download.
Afterwards, you can run
$ python -m fuzzfetch --tsan -n firefox-tsan
to get the build mentioned above unpacked into a directory called
Creating Try builds¶
If for some reason you can’t use the pre-built binaries mentioned in the previous section (e.g. you need to test a patch), you can either build Firefox yourself (see the following section) or use the try server to create the customized build for you. Pushing to try requires L1 commit access. If you don’t have this access yet you can request access (see Becoming A Mozilla Committer and Mozilla Commit Access Policy for the requirements).
mach try fuzzy --full you can select the
and related tests (if required).
Creating local builds on Linux¶
The TSan instrumentation is implemented as an LLVM pass and integrated
into Clang. We strongly recommend that you use the Clang version supplied
as part of the
mach bootstrap process, as we backported several required
fixes for TSan on Firefox.
Sanitizer support in Rust is genuinely experimental, so our build system only works with a limited range of rust nightlies. To install and use that specific version, run the following in the root of your mozilla-central checkout:
rustup toolchain install nightly-2020-11-14 --component rust-src rustup override set nightly-2020-11-14
Getting the source¶
Using that or any later revision, all you need to do is to get yourself a clone of mozilla-central.
Adjusting the build configuration¶
Create the build configuration file
mozconfig with the following
content in your mozilla-central directory:
# Combined .mozconfig file for TSan on Linux+Mac mk_add_options MOZ_OBJDIR=@TOPSRCDIR@/objdir-ff-tsan # Enable ASan specific code and build workarounds ac_add_options --enable-thread-sanitizer # This ensures that we also instrument Rust code. export RUSTFLAGS="-Zsanitizer=thread" # rustfmt is currently missing in Rust nightly unset RUSTFMT # Current Rust Nightly has warnings ac_add_options --disable-warnings-as-errors # These are required by TSan ac_add_options --disable-jemalloc ac_add_options --disable-crashreporter ac_add_options --disable-elf-hack ac_add_options --disable-profiling # The Thread Sanitizer is not compatible with sandboxing # (see bug 1182565) ac_add_options --disable-sandbox # Keep symbols to symbolize TSan traces later export MOZ_DEBUG_SYMBOLS=1 ac_add_options --enable-debug-symbols ac_add_options --disable-install-strip # Settings for an opt build (preferred) # The -gline-tables-only ensures that all the necessary debug information for ASan # is present, but the rest is stripped so the resulting binaries are smaller. ac_add_options --enable-optimize="-O2 -gline-tables-only" ac_add_options --disable-debug # Settings for a debug+opt build #ac_add_options --enable-optimize #ac_add_options --enable-debug
Starting the build process¶
Now you start the build process using the regular
After the build has completed,
./mach run with the usual options for
running in a debugger (
rr, etc.) work fine, as do
--disable-e10s and other options.
Firefox build, the build script below will probably help you to do so.
Execute this script in the
js/src/ subdirectory and pass a directory
name as the first parameter. The build will then be created in a new
subdirectory with that name.
#! /bin/sh if [ -z $1 ] ; then echo "usage: $0 <dirname>" elif [ -d $1 ] ; then echo "directory $1 already exists" else autoconf2.13 mkdir $1 cd $1 CC="/path/to/mozbuild/clang" \ CXX="/path/to/mozbuild/clang++" \ ../configure --disable-debug --enable-optimize="-O2 -gline-tables-only" --enable-thread-sanitizer --disable-jemalloc fi
Thread Sanitizer and Symbols¶
Unlike Address Sanitizer, TSan requires in-process symbolizing to work properly in the first place, as any kind of runtime suppressions will otherwise not work.
Hence, it is required that you have a copy of
PATH or pointed to by the
variable. This binary is included in your local mozbuild directory, obtained
TSan has the ability to suppress race reports at runtime. This can be used to silence a race while a fix is developed as well as to permanently silence a (benign) race that cannot be fixed.
Warning: Many races look benign but are indeed not. Please read the FAQ section carefully and think twice before attempting to suppress a race.
The runtime Suppression list is directly baked into Firefox at compile-time and located at mozglue/build/TsanOptions.cpp.
Important: When adding a suppression, always make sure to include
the bug number. If the suppression is supposed to be permanent, please
add the string
permanent in the same line as the bug number.
Important: When adding a suppression for a data race, always make sure to include a stack frame from each of the two race stacks. Adding only one suppression for one stack can cause intermittent failures that are later on hard to track. One exception to this rule is when suppressing races on global variables. In that case, a single race entry with the name of the variable is sufficient.
Troubleshooting / Known Problems¶
Known Sources of False Positives¶
TSan has a number of things that can cause false positives, namely:
The use of memory fences (e.g. Rust Arc)
The use of inline assembly for synchronization
Uninstrumented code (e.g. external libraries) using compiler-builtins for synchronization
A lock order inversion involving only a single thread can cause a false positive deadlock report (see also https://github.com/google/sanitizers/issues/488).
If none of these four items are involved, you should never assume that TSan is reporting a false positive to you without consulting TSan peers. It is very easy to misjudge a race to be a false positive because races can be highly complex and totally non-obvious due to compiler optimizations and the nature of parallel code.
Intermittent Broken Stacks¶
If you intermittently see race reports where one stack is missing with a
failed to restore the stack
message, this can indicate that a suppression is partially covering the race you are seeing.
Any race where only one of the two stacks is matched by a runtime suppression will show up if that particular stack fails to symbolize for some reason. The usual solution is to search the suppressions for potential candidates and disable them temporarily to check if your race report now becomes mostly consistent.
However, there are other reasons for broken TSan stacks, in particular if they are not intermittent.
See also the
history_size parameter in the TSan flags.
Intermittent Race Reports¶
Unfortunately, the TSan algorithm does not guarantee, that a race is detected 100% of the time. Intermittent failures with TSan are (to a certain degree) to be expected and the races involved should be filed and fixed to solve the problem.
Frequently Asked Questions about TSan¶
Why fix data races?¶
Data races are undefined behavior and can cause crashes as well as correctness issues. Compiler optimizations can cause racy code to have unpredictable and hard-to-reproduce behavior.
At Mozilla, we have already seen several dangerous races, causing random use-after-free crashes, intermittent test failures, hangs, performance issues and intermittent asserts. Such problems do not only decrease the quality of our code and user experience, but they also waste countless hours of developer time.
Since it is very hard to judge if a particular race could cause such a situation, we have decided to fix all data races wherever possible, since doing so is often cheaper than analyzing a race.
My race is benign, can we ignore it?¶
While it is possible to add a runtime suppression to ignore the race, we strongly encourage you to not do so, for two reasons:
Each suppressed race decreases the overall performance of the TSan build, as the race has to be symbolized each time when it occurs. Since TSan is already in itself a slow build, we need to keep the amount of suppressed races as low as possible.
Deciding if a race is truly benign is surprisingly hard. We recommend to read this blog post and this paper <https://www.usenix.org/legacy/events/hotpar11/tech/final_files/Boehm.pdf> on the effects of seemingly benign races.
Valid reasons to suppress a confirmed benign race include performance problems arising from fixing the race or cases where fixing the race would require an unreasonable amount of work.
Note that the use of atomics usually does not have the bad performance impact that developers tend to associate with it. If you assume that e.g. using atomics for synchronization will cause performance regressions, we suggest to perform a benchmark to confirm this. In many cases, the difference is not measurable.