17cba4f0f0
To improve performance of parking, keep the thread state in an atomic variable instead of protecting it with a mutex. However the mutex was used e.g. to force Unpark() to block while the safepoint operation was still running. Therefore the safepoint algorithm has to change as well. Park() and Unpark() use CAS operation to transition the state. Safepoint() uses a relaxed load for checking whether a safepoint was requested. Since Safepoint(), Park() and Unpark() all have a slow path, there is no need for busy-waiting on the main thread. We need two more ThreadStates: * SafepointRequested: This state is set by GlobalSafepoint to force Running threads into the slow path on Safepoint() and Park(). This state also replaces the separate atomic<bool> safepoint_requested_ field we used before. * ParkedSafepoint: This state is set by GlobalSafepoint as well to force parked threads into the slow path on Unpark(). When stopping all threads, GlobalSafepoint transitions states from Running --> SafepointRequested and Parked --> ParkedSafepoint to force the slow path for all three methods. After performing the transition for each thread we know the exact number of Running threads and wait until each of them either reached a safepoint or parked itself. Design doc: https://docs.google.com/document/d/1p9klWyqT_AScAnK_PdHZTcNhZGzoBiYWPkUciIh2C58/edit?usp=sharing Bug: chromium:1177144, v8:10315 Change-Id: I8697da915c7d18e2fb941f1bedf6181226408feb Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2704075 Reviewed-by: Ulan Degenbaev <ulan@chromium.org> Reviewed-by: Maya Lekova <mslekova@chromium.org> Commit-Queue: Dominik Inführ <dinfuehr@chromium.org> Cr-Commit-Position: refs/heads/master@{#73089} |
||
---|---|---|
.. | ||
bootstrap.sh | ||
BUILD.gn | ||
gcmole-test.cc | ||
gcmole-tools.tar.gz.sha1 | ||
gcmole.cc | ||
GCMOLE.gn | ||
gcmole.py | ||
ignored_files | ||
Makefile | ||
package.sh | ||
README | ||
run-gcmole.py | ||
suspects.whitelist | ||
test-expectations.txt |
DESCRIPTION ------------------------------------------------------------------- gcmole is a simple static analysis tool used to find possible evaluation order dependent GC-unsafe places in the V8 codebase and "stale" pointers to the heap (ones whose addresses got invalidated by the GC). For example the following code is GC-unsafe: Handle<Object> Foo(); // Assume Foo can trigger a GC. void Bar(Object, Object); Handle<Object> baz; baz->Qux(*Foo()); // (a) Bar(*Foo(), *baz); // (b) Both in cases (a) and (b) compiler is free to evaluate call arguments (that includes receiver) in any order. That means it can dereference baz before calling to Foo and save a raw pointer to a heap object in the register or on the stack. In terms of the AST analysis that gcmole does, it warns about places in the code which result in 2 subtrees, the order of execution of which is undefined by C++, one of which causes a GC and the other dereferences a Handle to a raw Object (or its subclasses). The following code triggers a stale variable warning (assuming that the Foo function was detected as potentially allocating, as in the previous example): JSObject raw_obj = ...; Foo(); raw_obj.Print(); Since Foo can trigger a GC, it might have moved the raw_obj. The solution is simply to store it as a Handle. PREREQUISITES ----------------------------------------------------------------- (1) Install Python $ sudo apt-get install python (2) Get LLVM 8.0 and Clang 8.0 sources and build them. Follow the instructions on http://clang.llvm.org/get_started.html. Make sure to pass -DCMAKE_BUILD_TYPE=Release to cmake to get Release build instead of a Debug one. (3) Build gcmole Clang plugin (libgcmole.so) In the tools/gcmole directory execute the following command: $ BUILD_ROOT=<path> LLVM_SRC_ROOT=<path> CLANG_SRC_ROOT=<path> make (*) Note that steps (2) and (3) can also be achieved by just using the included bootstrapping script in this directory: $ ./tools/gcmole/bootstrap.sh This will use "third_party/llvm+clang-build" as a build directory and checkout required sources in the "third_party" directory. USING GCMOLE ------------------------------------------------------------------ gcmole consists of driver script written in Python and Clang plugin that does C++ AST processing. Plugin (libgcmole.so) is expected to be in the same folder as driver (gcmole.py). To start analysis cd into the root of v8 checkout and execute the following command: CLANG_BIN=<path-to-clang-bin-folder> python tools/gcmole/gcmole.py [<arch>] where arch should be one of architectures supported by V8 (arm, ia32, x64). Analysis will be performed in 2 stages: - on the first stage driver will parse all files and build a global callgraph approximation to find all functions that might potentially cause GC, list of this functions will be written into gcsuspects file. - on the second stage driver will parse all files again and will locate all callsites that might be GC-unsafe based on the list of functions causing GC. Such places are marked with a "Possible problem with evaluation order." warning. Messages "Failed to resolve v8::internal::Object" are benign and can be ignored. If any errors were found driver exits with non-zero status. TESTING ----------------------------------------------------------------------- Tests are automatically run by the main python runner. Expectations are in test-expectations.txt and need to be updated whenever the sources of the tests in gcmole-test.cc are modified (line numbers also count). PACKAGING --------------------------------------------------------------------- gcmole is deployed on V8's buildbot infrastructure to run it as part of the continuous integration. A pre-built package of gcmole together with Clang is hosted on Google Cloud Storage for this purpose. To update this package to a newer version, use the provided packaging script: $ ./tools/gcmole/package.sh This will create a new "tools/gcmole/gcmole-tools.tar.gz" package with the corresponding SHA1 sum suitable to be used for this purpose. It assumes that Clang was built in "third_party/llvm+clang-build" (e.g. by the bootstrapping script "bootstrap.sh" mentioned above). TROUBLESHOOTING --------------------------------------------------------------- gcmole is tighly coupled with the AST structure that Clang produces. Therefore when upgrading to a newer Clang version, it might start producing bogus output or completely stop outputting warnings. In such occasion, one might start the debugging process by checking weather a new AST node type is introduced which is currently not supported by gcmole. Insert the following code at the end of the FunctionAnalyzer::VisitExpr method to see the unsupported AST class(es) and the source position which generates them: if (expr) { clang::Stmt::StmtClass stmtClass = expr->getStmtClass(); d_.Report(clang::FullSourceLoc(expr->getExprLoc(), sm_), d_.getCustomDiagID(clang::DiagnosticsEngine::Remark, "%0")) << stmtClass; } For instance, gcmole currently doesn't support AtomicExprClass statements introduced for atomic operations. A convenient way to observe the AST generated by Clang is to pass the following flags when invoking clang++ -Xclang -ast-dump -fsyntax-only