Merge pull request #14190 from JuliaLang/yyc/gc/safepoint

GC safepoint and transition support

Author: JeffBezanson

base/locks.jl

@@ -23,6 +23,11 @@ function lock!(l::TatasLock)
             end
         end
         ccall(:jl_cpu_pause, Void, ())
+        # Temporary solution before we have gc transition support in codegen.
+        # This could mess up gc state when we add codegen support.
+        # Use these as a safe point
+        gc_state = ccall(:jl_gc_safe_enter, Int8, ())
+        ccall(:jl_gc_safe_leave, Void, (Int8,), gc_state)
     end
 end
 
@@ -61,6 +66,11 @@ function lock!(l::RecursiveTatasLock)
             end
         end
         ccall(:jl_cpu_pause, Void, ())
+        # Temporary solution before we have gc transition support in codegen.
+        # This could mess up gc state when we add codegen support.
+        # Use these as a safe point
+        gc_state = ccall(:jl_gc_safe_enter, Int8, ())
+        ccall(:jl_gc_safe_leave, Void, (Int8,), gc_state)
     end
 end
 
@@ -116,7 +126,11 @@ function lock!(m::Mutex)
     if m.ownertid == threadid()
         return 0
     end
+    # Temporary solution before we have gc transition support in codegen.
+    # This could mess up gc state when we add codegen support.
+    gc_state = ccall(:jl_gc_safe_enter, Int8, ())
     ccall(:uv_mutex_lock, Void, (Ptr{Void},), m.handle)
+    ccall(:jl_gc_safe_leave, Void, (Int8,), gc_state)
     m.ownertid = threadid()
     return 0
 end

doc/devdocs/debuggingtips.rst

@@ -24,7 +24,7 @@ Similarly, if you're debugging some of julia's internals (e.g.,
 
 This is a good way to circumvent problems that arise from the order in which julia's output streams are initialized.
 
-Julia's flisp interpreter uses ``value_t*`` objects; these can be displayed
+Julia's flisp interpreter uses ``value_t`` objects; these can be displayed
 with ``call fl_print(ios_stdout, obj)``.
 
 Useful Julia variables for Inspecting
@@ -74,7 +74,7 @@ Another useful frame is ``to_function(jl_lambda_info_t *li, bool cstyle)``. The
 
    #2  0x00007ffff7928bf7 in to_function (li=0x2812060, cstyle=false) at codegen.cpp:584
    584	        abort();
-   (gdb) p jl_(jl_uncompress_ast(li, li.ast))
+   (gdb) p jl_(jl_uncompress_ast(li, li->ast))
 
 Inserting breakpoints upon certain conditions
 ---------------------------------------------
@@ -91,10 +91,31 @@ Calling a particular method
 
 ::
 
-   (gdb) break jl_apply_generic if strcmp(F->name->name, "method_to_break")==0
+   (gdb) break jl_apply_generic if strcmp((char*)(jl_symbol_name)(jl_gf_mtable(F)->name), "method_to_break")==0
 
 Since this function is used for every call, you will make everything 1000x slower if you do this.
 
+Dealing with signals
+--------------------
+
+Julia requires a few signal to function property. The profiler uses ``SIGUSR2``
+for sampling and the garbage collector uses ``SIGSEGV`` for threads
+synchronization. If you are debugging some code that uses the profiler or
+multiple julia threads, you may want to let the debugger ignore these signals
+since they can be triggered very often during normal operations. The command to
+do this in GDB is (replace ``SIGSEGV`` with ``SIGUSRS`` or other signals you
+want to ignore)::
+
+   (gdb) handle SIGSEGV noprint nostop pass
+
+The corresponding LLDB command is (after the process is started)::
+
+   (lldb) pro hand -p true -s false -n false SIGSEGV
+
+If you are debugging a segfault with threaded code, you can set a breakpoint on
+``jl_critical_error`` (``sigdie_handler`` should also work on Linux and BSD) in
+order to only catch the actual segfault rather than the GC synchronization points.
+
 Debugging during julia's build process (bootstrap)
 --------------------------------------------------
 

src/ast.c

@@ -202,7 +202,7 @@ static jl_value_t *resolve_globals(jl_value_t *expr, jl_lambda_info_t *lam);
 
 static jl_value_t *scm_to_julia(value_t e, int expronly)
 {
-    int en = jl_gc_enable(0);
+    int en = jl_gc_enable(0); // Might GC
     jl_value_t *v;
     JL_TRY {
         v = scm_to_julia_(e, expronly);

src/builtins.c

@@ -194,6 +194,9 @@ JL_DLLEXPORT void jl_enter_handler(jl_handler_t *eh)
     JL_SIGATOMIC_BEGIN();
     eh->prev = jl_current_task->eh;
     eh->gcstack = jl_pgcstack;
+#ifdef JULIA_ENABLE_THREADING
+    eh->gc_state = jl_get_ptls_states()->gc_state;
+#endif
     jl_current_task->eh = eh;
     // TODO: this should really go after setjmp(). see comment in
     // ctx_switch in task.c.

src/debuginfo.cpp

@@ -194,7 +194,9 @@ class JuliaJITEventListener: public JITEventListener
     virtual void NotifyFunctionEmitted(const Function &F, void *Code,
                                        size_t Size, const EmittedFunctionDetails &Details)
     {
+        int8_t gc_state = jl_gc_safe_enter();
         uv_rwlock_wrlock(&threadsafe);
+        jl_gc_safe_leave(gc_state);
 #if defined(_OS_WINDOWS_)
         create_PRUNTIME_FUNCTION((uint8_t*)Code, Size, F.getName(), (uint8_t*)Code, Size, NULL);
 #endif
@@ -205,7 +207,9 @@ class JuliaJITEventListener: public JITEventListener
 
     std::map<size_t, FuncInfo, revcomp>& getMap()
     {
+        int8_t gc_state = jl_gc_safe_enter();
         uv_rwlock_rdlock(&threadsafe);
+        jl_gc_safe_leave(gc_state);
         return info;
     }
 #endif // ifndef USE_MCJIT
@@ -225,7 +229,9 @@ class JuliaJITEventListener: public JITEventListener
     virtual void NotifyObjectEmitted(const ObjectImage &obj)
 #endif
     {
+        int8_t gc_state = jl_gc_safe_enter();
         uv_rwlock_wrlock(&threadsafe);
+        jl_gc_safe_leave(gc_state);
 #ifdef LLVM36
         object::section_iterator Section = obj.section_begin();
         object::section_iterator EndSection = obj.section_end();
@@ -458,7 +464,9 @@ class JuliaJITEventListener: public JITEventListener
 
     std::map<size_t, ObjectInfo, revcomp>& getObjectMap()
     {
+        int8_t gc_state = jl_gc_safe_enter();
         uv_rwlock_rdlock(&threadsafe);
+        jl_gc_safe_leave(gc_state);
         return objectmap;
     }
 #endif // USE_MCJIT
@@ -477,6 +485,8 @@ JL_DLLEXPORT void ORCNotifyObjectEmitted(JITEventListener *Listener,
 extern "C"
 char *jl_demangle(const char *name)
 {
+    // This function is not allowed to reference any TLS variables since
+    // it can be called from an unmanaged thread on OSX.
     const char *start = name + 6;
     const char *end = name + strlen(name);
     char *ret;
@@ -508,6 +518,8 @@ void lookup_pointer(DIContext *context, char **name, size_t *line,
                     char **inlinedat_file, size_t pointer,
                     int demangle, int *fromC)
 {
+    // This function is not allowed to reference any TLS variables since
+    // it can be called from an unmanaged thread on OSX.
     DILineInfo info, topinfo;
     DIInliningInfo inlineinfo;
     if (demangle && *name != NULL) {
@@ -629,6 +641,8 @@ void jl_getDylibFunctionInfo(char **name, char **filename, size_t *line,
                              char** inlinedat_file, size_t *inlinedat_line,
                              size_t pointer, int *fromC, int skipC, int skipInline)
 {
+    // This function is not allowed to reference any TLS variables since
+    // it can be called from an unmanaged thread on OSX.
 #ifdef _OS_WINDOWS_
     IMAGEHLP_MODULE64 ModuleInfo;
     BOOL isvalid;
@@ -838,6 +852,8 @@ void jl_getFunctionInfo(char **name, char **filename, size_t *line,
                         char **inlinedat_file, size_t *inlinedat_line,
                         size_t pointer, int *fromC, int skipC, int skipInline)
 {
+    // This function is not allowed to reference any TLS variables since
+    // it can be called from an unmanaged thread on OSX.
     *name = NULL;
     *line = -1;
     *filename = NULL;

src/dump.c

@@ -1974,7 +1974,7 @@ JL_DLLEXPORT jl_value_t *jl_ast_rettype(jl_lambda_info_t *li, jl_value_t *ast)
     ios_mem(&src, 0);
     ios_setbuf(&src, (char*)bytes->data, jl_array_len(bytes), 0);
     src.size = jl_array_len(bytes);
-    int en = jl_gc_enable(0);
+    int en = jl_gc_enable(0); // Might GC
     jl_value_t *rt = jl_deserialize_value(&src, NULL);
     jl_gc_enable(en);
     tree_literal_values = NULL;
@@ -1994,7 +1994,7 @@ JL_DLLEXPORT jl_value_t *jl_compress_ast(jl_lambda_info_t *li, jl_value_t *ast)
     ios_mem(&dest, 0);
     jl_array_t *last_tlv = tree_literal_values;
     jl_module_t *last_tem = tree_enclosing_module;
-    int en = jl_gc_enable(0);
+    int en = jl_gc_enable(0); // Might GC
 
     if (li->module->constant_table == NULL) {
         li->module->constant_table = jl_alloc_cell_1d(0);
@@ -2038,7 +2038,7 @@ JL_DLLEXPORT jl_value_t *jl_uncompress_ast(jl_lambda_info_t *li, jl_value_t *dat
     ios_mem(&src, 0);
     ios_setbuf(&src, (char*)bytes->data, jl_array_len(bytes), 0);
     src.size = jl_array_len(bytes);
-    int en = jl_gc_enable(0);
+    int en = jl_gc_enable(0); // Might GC
     (void)jl_deserialize_value(&src, NULL); // skip ret type
     jl_value_t *v = jl_deserialize_value(&src, NULL);
     jl_gc_enable(en);