Correct TSAN tasking integration (JuliaLang#36929)

In order for tsan to work, any setjmp/longjmp must be executed while
that task's tsan state is active. As a result, we cannot switch the
tsan state until said setjmp is completed, but need to do it
(as basically the only thing happening) between the setjmp and the
subsequent longjmp. To facilitate this without too much disruption,
move the tsan state into the jl_ucontext_t, which seems appropriate
since it's additional state that needs to be restored on context
switch.

Also forbid using TSAN from Clang < 11, where the runtime library
has bugs that cause us to exhaust the maximum number of allowed
mappings.

Make.inc

@@ -431,6 +431,10 @@ CXX_DISABLE_ASSERTION := -DJL_NDEBUG
 DISABLE_ASSERTIONS := -DNDEBUG -DJL_NDEBUG
 endif
 
+ifeq ($(LLVM_ASSERTIONS),0)
+CXX_DISABLE_ASSERTION += -DNDEBUG
+endif
+
 # Compiler specific stuff
 
 ifeq ($(USEMSVC), 1)

src/aotcompile.cpp

@@ -20,12 +20,8 @@
 #include <llvm/Transforms/IPO.h>
 #include <llvm/Transforms/Scalar.h>
 #include <llvm/Transforms/Vectorize.h>
-#if defined(JL_ASAN_ENABLED)
 #include <llvm/Transforms/Instrumentation/AddressSanitizer.h>
-#endif
-#if defined(JL_TSAN_ENABLED)
 #include <llvm/Transforms/Instrumentation/ThreadSanitizer.h>
-#endif
 #include <llvm/Transforms/Scalar/GVN.h>
 #include <llvm/Transforms/IPO/AlwaysInliner.h>
 #include <llvm/Transforms/InstCombine/InstCombine.h>

src/gc-stacks.c

@@ -226,6 +226,12 @@ void sweep_stack_pools(void)
  t->stkbuf = NULL;
  _jl_free_stack(ptls2, stkbuf, bufsz);
  }
+#ifdef JL_TSAN_ENABLED
+ if (t->ctx.tsan_state) {
+ __tsan_destroy_fiber(t->ctx.tsan_state);
+ t->ctx.tsan_state = NULL;
+ }
+#endif
  }
  if (n >= l - ndel)
  break;

src/julia.h

@@ -65,7 +65,6 @@
 # define JL_THREAD_LOCAL
 #endif
 
-// Duplicated from options.h
 #if defined(__has_feature) // Clang flavor
 #if __has_feature(address_sanitizer)
 #define JL_ASAN_ENABLED
@@ -74,6 +73,9 @@
 #define JL_MSAN_ENABLED
 #endif
 #if __has_feature(thread_sanitizer)
+#if __clang_major__ < 11
+#error Thread sanitizer runtime libraries in clang < 11 leak memory and cannot be used
+#endif
 #define JL_TSAN_ENABLED
 #endif
 #else // GCC flavor
@@ -1810,10 +1812,6 @@ typedef struct _jl_task_t {
  unsigned int copy_stack:31; // sizeof stack for copybuf
  unsigned int started:1;
 
-#if defined(JL_TSAN_ENABLED)
- void *tsan_state;
-#endif
-
  // current exception handler
  jl_handler_t *eh;
  // saved gc stack top for context switches

src/julia_internal.h

@@ -21,7 +21,8 @@ void __sanitizer_finish_switch_fiber(void*, const void**, size_t*);
 #endif
 #ifdef JL_TSAN_ENABLED
 void *__tsan_create_fiber(unsigned flags);
-// void __tsan_destroy_fiber(void *fiber);
+void *__tsan_get_current_fiber(void);
+void __tsan_destroy_fiber(void *fiber);
 void __tsan_switch_to_fiber(void *fiber, unsigned flags);
 #endif
 #ifdef __cplusplus

src/julia_threads.h

@@ -47,9 +47,15 @@ typedef win32_ucontext_t jl_ucontext_t;
 #if defined(JL_HAVE_ASM) || defined(JL_HAVE_SIGALTSTACK)
 typedef struct {
  jl_jmp_buf uc_mcontext;
+#if defined(JL_TSAN_ENABLED)
+ void *tsan_state;
+#endif
 } jl_ucontext_t;
 #endif
 #if defined(JL_HAVE_ASYNCIFY)
+#if defined(JL_TSAN_ENABLED)
+#error TSAN not currently supported with asyncify
+#endif
 typedef struct {
  // This is the extent of the asyncify stack, but because the top of the
  // asyncify stack (stacktop) is also the bottom of the C stack, we can
@@ -62,7 +68,12 @@ typedef struct {
 #if defined(JL_HAVE_UCONTEXT) || defined(JL_HAVE_UNW_CONTEXT)
 #define UNW_LOCAL_ONLY
 #include <libunwind.h>
-typedef ucontext_t jl_ucontext_t;
+typedef struct {
+ ucontext_t ctx;
+#if defined(JL_TSAN_ENABLED)
+ void *tsan_state;
+#endif
+} jl_ucontext_t;
 #endif
 #endif
 

src/support/win32_ucontext.h

@@ -16,6 +16,9 @@ typedef struct {
  size_t ss_size;
  } uc_stack;
  jmp_buf uc_mcontext;
+#ifdef JL_TSAN_ENABLED
+ void *tsan_state;
+#endif
 } win32_ucontext_t;
 void jl_makecontext(win32_ucontext_t *ucp, void (*func)(void));
 void jl_swapcontext(win32_ucontext_t *oucp, const win32_ucontext_t *ucp);