tasks-debugging: make it possible to get the backtrace of a task

src/gc-stacks.c

@@ -216,8 +216,12 @@ void sweep_stack_pools(void)
             continue;
         while (1) {
             jl_task_t *t = (jl_task_t*)lst[n];
+            assert(jl_is_task(t));
             if (gc_marked(jl_astaggedvalue(t)->bits.gc)) {
-                n++;
+                if (t->stkbuf == NULL)
+                    ndel++; // jl_release_task_stack called
+                else
+                    n++;
             }
             else {
                 ndel++;
@@ -243,3 +247,30 @@ void sweep_stack_pools(void)
         live_tasks->len -= ndel;
     }
 }
+
+JL_DLLEXPORT jl_array_t *jl_live_tasks(void)
+{
+    jl_ptls_t ptls = jl_get_ptls_states();
+    arraylist_t *live_tasks = &ptls->heap.live_tasks;
+    size_t i, j, l;
+    jl_array_t *a;
+    do {
+        l = live_tasks->len;
+        a = jl_alloc_vec_any(l + 1); // may gc
+    } while (l + 1 < live_tasks->len);
+    l = live_tasks->len;
+    void **lst = live_tasks->items;
+    j = 0;
+    ((void**)jl_array_data(a))[j++] = ptls->root_task;
+    for (i = 0; i < l; i++) {
+        if (((jl_task_t*)lst[i])->stkbuf != NULL)
+            ((void**)jl_array_data(a))[j++] = lst[i];
+    }
+    l = jl_array_len(a);
+    if (j < l) {
+        JL_GC_PUSH1(&a);
+        jl_array_del_end(a, l - j);
+        JL_GC_POP();
+    }
+    return a;
+}

src/julia.expmap

@@ -32,6 +32,7 @@
     add_library_mapping;
     utf8proc_*;
     jlbacktrace;
+    jlbacktracet;
     julia_type_to_llvm;
     _IO_stdin_used;
     __ZN4llvm23createLowerSimdLoopPassEv;

src/julia.h

@@ -1762,7 +1762,17 @@ typedef struct _jl_task_t {
     // current exception handler
     jl_handler_t *eh;
 
-    jl_ucontext_t ctx; // saved thread state
+    union {
+        jl_ucontext_t ctx; // saved thread state
+#ifdef _OS_WINDOWS_
+        jl_ucontext_t copy_stack_ctx;
+#else
+        struct jl_stack_context_t copy_stack_ctx;
+#endif
+    };
+#if defined(JL_TSAN_ENABLED)
+    void *tsan_state;
+#endif
     void *stkbuf; // malloc'd memory (either copybuf or stack)
     size_t bufsz; // actual sizeof stkbuf
     unsigned int copy_stack:31; // sizeof stack for copybuf

src/julia_threads.h

@@ -16,8 +16,9 @@
 
 //  Options for task switching algorithm (in order of preference):
 // JL_HAVE_ASM -- mostly setjmp
-// JL_HAVE_ASYNCIFY -- task switching based on the binaryen asyncify transform
-// JL_HAVE_UNW_CONTEXT -- hybrid of libunwind for start, setjmp for resume
+// JL_HAVE_ASM && JL_HAVE_UNW_CONTEXT -- libunwind-based
+// JL_HAVE_UNW_CONTEXT -- libunwind-based
+// JL_HAVE_ASYNCIFY -- task switching based on the binary asyncify transform
 // JL_HAVE_UCONTEXT -- posix standard API, requires syscall for resume
 // JL_HAVE_SIGALTSTACK -- requires several syscall for start, setjmp for resume
 
@@ -33,24 +34,25 @@ typedef win32_ucontext_t jl_ucontext_t;
 #if (defined(_CPU_X86_64_) || defined(_CPU_X86_) || defined(_CPU_AARCH64_) ||  \
      defined(_CPU_ARM_) || defined(_CPU_PPC64_))
 #define JL_HAVE_ASM
-#elif defined(_OS_DARWIN_)
+#endif
+#if defined(_OS_DARWIN_)
 #define JL_HAVE_UNW_CONTEXT
-#elif defined(_OS_LINUX_)
-#define JL_HAVE_UCONTEXT
+//#elif defined(_OS_LINUX_)
+//#define JL_HAVE_UNW_CONTEXT // very slow, but more debugging
 #elif defined(_OS_EMSCRIPTEN_)
 #define JL_HAVE_ASYNCIFY
-#else
-#define JL_HAVE_UNW_CONTEXT
+#elif !defined(JL_HAVE_ASM)
+#define JL_HAVE_UNW_CONTEXT // optimistically?
 #endif
 #endif
 
-#if defined(JL_HAVE_ASM) || defined(JL_HAVE_SIGALTSTACK)
-typedef struct {
+
+struct jl_stack_context_t {
     jl_jmp_buf uc_mcontext;
-#if defined(JL_TSAN_ENABLED)
-    void *tsan_state;
-#endif
-} jl_ucontext_t;
+};
+
+#if (!defined(JL_HAVE_UNW_CONTEXT) && defined(JL_HAVE_ASM)) || defined(JL_HAVE_SIGALTSTACK)
+typedef struct jl_stack_context_t jl_ucontext_t;
 #endif
 #if defined(JL_HAVE_ASYNCIFY)
 #if defined(JL_TSAN_ENABLED)
@@ -65,15 +67,14 @@ typedef struct {
     void *stacktop;
 } jl_ucontext_t;
 #endif
-#if defined(JL_HAVE_UCONTEXT) || defined(JL_HAVE_UNW_CONTEXT)
+#if defined(JL_HAVE_UNW_CONTEXT)
 #define UNW_LOCAL_ONLY
 #include <libunwind.h>
-typedef struct {
-    ucontext_t ctx;
-#if defined(JL_TSAN_ENABLED)
-    void *tsan_state;
+typedef unw_context_t jl_ucontext_t;
 #endif
-} jl_ucontext_t;
+#if defined(JL_HAVE_UCONTEXT)
+#include <ucontext.h>
+typedef ucontext_t jl_ucontext_t;
 #endif
 #endif
 
@@ -210,7 +211,15 @@ struct _jl_tls_states_t {
     struct _jl_timing_block_t *timing_stack;
     void *stackbase;
     size_t stacksize;
-    jl_ucontext_t base_ctx; // base context of stack
+    union {
+        jl_ucontext_t base_ctx; // base context of stack
+        // This hack is needed to support always_copy_stacks:
+#ifdef _OS_WINDOWS_
+        jl_ucontext_t copy_stack_ctx;
+#else
+        struct jl_stack_context_t copy_stack_ctx;
+#endif
+    };
     jl_jmp_buf *safe_restore;
     // Temp storage for exception thrown in signal handler. Not rooted.
     struct _jl_value_t *sig_exception;

src/partr.c

@@ -40,7 +40,7 @@ uint64_t io_wakeup_leave;
 JL_DLLEXPORT int jl_set_task_tid(jl_task_t *task, int tid) JL_NOTSAFEPOINT
 {
     // Try to acquire the lock on this task.
-    int16_t was = task->tid;
+    int16_t was = jl_atomic_load_relaxed(&task->tid);
     if (was == tid)
         return 1;
     if (was == -1)

src/signals-unix.c

@@ -43,6 +43,7 @@
 
 #include "julia_assert.h"
 
+// helper function for returning the unw_context_t inside a ucontext_t
 static bt_context_t *jl_to_bt_context(void *sigctx)
 {
 #ifdef __APPLE__

src/stackwalk.c

@@ -14,11 +14,11 @@
 // returning from the callee function will invalidate the context
 #ifdef _OS_WINDOWS_
 jl_mutex_t jl_in_stackwalk;
-#define jl_unw_get(context) RtlCaptureContext(context)
+#define jl_unw_get(context) (RtlCaptureContext(context), 0)
 #elif !defined(JL_DISABLE_LIBUNWIND)
 #define jl_unw_get(context) unw_getcontext(context)
 #else
-void jl_unw_get(void *context) {};
+int jl_unw_get(void *context) { return -1; }
 #endif
 
 #ifdef __cplusplus
@@ -204,7 +204,9 @@ NOINLINE size_t rec_backtrace(jl_bt_element_t *bt_data, size_t maxsize, int skip
 {
     bt_context_t context;
     memset(&context, 0, sizeof(context));
-    jl_unw_get(&context);
+    int r = jl_unw_get(&context);
+    if (r < 0)
+        return 0;
     jl_gcframe_t *pgcstack = jl_pgcstack;
     bt_cursor_t cursor;
     if (!jl_unw_init(&cursor, &context))
@@ -239,9 +241,9 @@ JL_DLLEXPORT jl_value_t *jl_backtrace_from_here(int returnsp, int skip)
     bt_context_t context;
     bt_cursor_t cursor;
     memset(&context, 0, sizeof(context));
-    jl_unw_get(&context);
+    int r = jl_unw_get(&context);
     jl_gcframe_t *pgcstack = jl_pgcstack;
-    if (jl_unw_init(&cursor, &context)) {
+    if (r == 0 && jl_unw_init(&cursor, &context)) {
         // Skip frame for jl_backtrace_from_here itself
         skip += 1;
         size_t offset = 0;
@@ -688,8 +690,59 @@ void jl_print_bt_entry_codeloc(jl_bt_element_t *bt_entry) JL_NOTSAFEPOINT
     }
 }
 
+extern bt_context_t *jl_to_bt_context(void *sigctx);
+
+void jl_rec_backtrace(jl_task_t *t)
+{
+    jl_ptls_t ptls = jl_get_ptls_states();
+    ptls->bt_size = 0;
+    if (t == ptls->current_task) {
+        ptls->bt_size = rec_backtrace(ptls->bt_data, JL_MAX_BT_SIZE, 0);
+        return;
+    }
+    if (t->copy_stack || !t->started || t->stkbuf == NULL)
+        return;
+    int old = jl_atomic_compare_exchange(&t->tid, -1, ptls->tid);
+    if (old != -1 && old != ptls->tid)
+        return;
+    bt_context_t *context = NULL;
+#if defined(_OS_WINDOWS_)
+    bt_context_t c;
+    memset(&c, 0, sizeof(c));
+    _JUMP_BUFFER *mctx = (_JUMP_BUFFER*)&t->ctx.uc_mcontext;
+#if defined(_CPU_X86_64_)
+    c.Rbx = mctx->Rbx;
+    c.Rsp = mctx->Rsp;
+    c.Rbp = mctx->Rbp;
+    c.Rsi = mctx->Rsi;
+    c.Rdi = mctx->Rdi;
+    c.R12 = mctx->R12;
+    c.R13 = mctx->R13;
+    c.R14 = mctx->R14;
+    c.R15 = mctx->R15;
+    c.Rip = mctx->Rip;
+    memcpy(&c.Xmm6, &mctx->Xmm6, 10 * sizeof(mctx->Xmm6)); // Xmm6-Xmm15
+#else
+    c.Eip = mctx->Eip;
+    c.Esp = mctx->Esp;
+    c.Ebp = mctx->Ebp;
+#endif
+    context = &c;
+#elif defined(JL_HAVE_UNW_CONTEXT)
+    context = &t->ctx;
+#elif defined(JL_HAVE_UCONTEXT)
+    context = jl_to_bt_context(&t->ctx);
+#else
+#endif
+    if (context)
+        ptls->bt_size = rec_backtrace_ctx(ptls->bt_data, JL_MAX_BT_SIZE, context, t->gcstack);
+    if (old == -1)
+        jl_atomic_store_relaxed(&t->tid, old);
+}
+
 //--------------------------------------------------
 // Tools for interactive debugging in gdb
+
 JL_DLLEXPORT void jl_gdblookup(void* ip)
 {
     jl_print_native_codeloc((uintptr_t)ip);
@@ -701,9 +754,19 @@ JL_DLLEXPORT void jlbacktrace(void) JL_NOTSAFEPOINT
     jl_excstack_t *s = jl_get_ptls_states()->current_task->excstack;
     if (!s)
         return;
-    size_t bt_size = jl_excstack_bt_size(s, s->top);
+    size_t i, bt_size = jl_excstack_bt_size(s, s->top);
     jl_bt_element_t *bt_data = jl_excstack_bt_data(s, s->top);
-    for (size_t i = 0; i < bt_size; i += jl_bt_entry_size(bt_data + i)) {
+    for (i = 0; i < bt_size; i += jl_bt_entry_size(bt_data + i)) {
+        jl_print_bt_entry_codeloc(bt_data + i);
+    }
+}
+JL_DLLEXPORT void jlbacktracet(jl_task_t *t)
+{
+    jl_ptls_t ptls = jl_get_ptls_states();
+    jl_rec_backtrace(t);
+    size_t i, bt_size = ptls->bt_size;
+    jl_bt_element_t *bt_data = ptls->bt_data;
+    for (i = 0; i < bt_size; i += jl_bt_entry_size(bt_data + i)) {
         jl_print_bt_entry_codeloc(bt_data + i);
     }
 }