Merge pull request #14501 from JuliaLang/yyc/stack_rng

Don't use pthread_*_np functions to get stack master thread

src/init.c

@@ -88,46 +88,10 @@ jl_options_t jl_options = { 0,    // quiet
 int jl_boot_file_loaded = 0;
 size_t jl_page_size;
 
-void jl_init_stack_limits(void)
+void jl_init_stack_limits(int ismaster)
 {
-#if defined(_OS_LINUX_)
-    pthread_attr_t attr;
-    if (pthread_getattr_np(pthread_self(), &attr)) {
-        // On linux `pthread_getattr_np` can fail on master thread if
-        // `/proc/self/maps` is not readable, in which case we fall back to
-        // `rlimit`
-        struct rlimit rl;
-        getrlimit(RLIMIT_STACK, &rl);
-        jl_stack_hi = (char*)&attr;
-        jl_stack_lo = jl_stack_hi - rl.rlim_cur;
-    }
-    else {
-        void *stackaddr;
-        size_t stacksize;
-        pthread_attr_getstack(&attr, &stackaddr, &stacksize);
-        pthread_attr_destroy(&attr);
-        jl_stack_lo = (char*)stackaddr;
-        jl_stack_hi = (char*)stackaddr + stacksize;
-    }
-#elif defined(_OS_DARWIN_)
-    extern void *pthread_get_stackaddr_np(pthread_t thread);
-    extern size_t pthread_get_stacksize_np(pthread_t thread);
-    pthread_t thread = pthread_self();
-    void *stackaddr = pthread_get_stackaddr_np(thread);
-    size_t stacksize = pthread_get_stacksize_np(thread);
-    jl_stack_lo = (char*)stackaddr;
-    jl_stack_hi = (char*)stackaddr + stacksize;
-#elif defined(_OS_FREEBSD_)
-    pthread_attr_t attr;
-    pthread_attr_init(&attr);
-    pthread_attr_get_np(pthread_self(), &attr);
-    void *stackaddr;
-    size_t stacksize;
-    pthread_attr_getstack(&attr, &stackaddr, &stacksize);
-    pthread_attr_destroy(&attr);
-    jl_stack_lo = (char*)stackaddr;
-    jl_stack_hi = (char*)stackaddr + stacksize;
-#elif defined(_OS_WINDOWS_)
+#ifdef _OS_WINDOWS_
+    (void)ismaster;
 #  ifdef _COMPILER_MICROSOFT_
 #    ifdef _P64
     void **tib = (void**)__readgsqword(0x30);
@@ -147,7 +111,46 @@ void jl_init_stack_limits(void)
     jl_stack_lo = (char*)tib[2]; // Stack Limit / Ceiling of stack (low address)
 #else
 #  ifdef JULIA_ENABLE_THREADING
-#    warning "Getting stack size for thread is not supported."
+    // Only use pthread_*_np functions to get stack address for non-master
+    // threads since it seems to return bogus values for master thread on Linux
+    // and possibly OSX.
+    if (!ismaster) {
+#    if defined(_OS_LINUX_)
+        pthread_attr_t attr;
+        pthread_getattr_np(pthread_self(), &attr);
+        void *stackaddr;
+        size_t stacksize;
+        pthread_attr_getstack(&attr, &stackaddr, &stacksize);
+        pthread_attr_destroy(&attr);
+        jl_stack_lo = (char*)stackaddr;
+        jl_stack_hi = (char*)stackaddr + stacksize;
+        return;
+#    elif defined(_OS_DARWIN_)
+        extern void *pthread_get_stackaddr_np(pthread_t thread);
+        extern size_t pthread_get_stacksize_np(pthread_t thread);
+        pthread_t thread = pthread_self();
+        void *stackaddr = pthread_get_stackaddr_np(thread);
+        size_t stacksize = pthread_get_stacksize_np(thread);
+        jl_stack_lo = (char*)stackaddr;
+        jl_stack_hi = (char*)stackaddr + stacksize;
+        return;
+#    elif defined(_OS_FREEBSD_)
+        pthread_attr_t attr;
+        pthread_attr_init(&attr);
+        pthread_attr_get_np(pthread_self(), &attr);
+        void *stackaddr;
+        size_t stacksize;
+        pthread_attr_getstack(&attr, &stackaddr, &stacksize);
+        pthread_attr_destroy(&attr);
+        jl_stack_lo = (char*)stackaddr;
+        jl_stack_hi = (char*)stackaddr + stacksize;
+        return;
+#    else
+#      warning "Getting stack size for thread is not supported."
+#    endif
+    }
+#  else
+    (void)ismaster;
 #  endif
     struct rlimit rl;
     getrlimit(RLIMIT_STACK, &rl);
@@ -180,7 +183,7 @@ static void jl_find_stack_bottom(void)
     }
 #endif
 #endif
-    jl_init_stack_limits();
+    jl_init_stack_limits(1);
 }
 
 struct uv_shutdown_queue_item { uv_handle_t *h; struct uv_shutdown_queue_item *next; };

src/julia_internal.h

@@ -228,7 +228,7 @@ void jl_init_primitives(void);
 void jl_init_codegen(void);
 void jl_init_intrinsic_functions(void);
 void jl_init_tasks(void);
-void jl_init_stack_limits(void);
+void jl_init_stack_limits(int ismaster);
 void jl_init_root_task(void *stack, size_t ssize);
 void jl_init_serializer(void);
 void jl_gc_init(void);

src/signals-mach.c

@@ -116,6 +116,17 @@ kern_return_t catch_exception_raise(mach_port_t            exception_port,
     if (thread == mach_profiler_thread) {
         return profiler_segv_handler(exception_port, thread, task, exception, code, code_count);
     }
+#endif
+#ifdef JULIA_ENABLE_THREADING
+    jl_tls_states_t *ptls = NULL;
+    for (int16_t tid = 0;tid < jl_n_threads;tid++) {
+        if (pthread_mach_thread_np(jl_all_task_states[tid].system_id) == thread) {
+            ptls = jl_all_task_states[tid].ptls;
+            break;
+        }
+    }
+#else
+    jl_tls_states_t *ptls = &jl_tls_states;
 #endif
     kern_return_t ret = thread_get_state(thread, x86_EXCEPTION_STATE64, (thread_state_t)&exc_state, &exc_count);
     HANDLE_MACH_ERROR("thread_get_state", ret);
@@ -126,7 +137,7 @@ kern_return_t catch_exception_raise(mach_port_t            exception_port,
     if (msync((void*)(fault_addr & ~(jl_page_size - 1)), 1, MS_ASYNC) == 0) { // check if this was a valid address
 #endif
         jl_value_t *excpt;
-        if (is_addr_on_stack((void*)fault_addr)) {
+        if (is_addr_on_stack(ptls, (void*)fault_addr)) {
             excpt = jl_stackovf_exception;
         }
 #ifdef SEGV_EXCEPTION

src/signals-unix.c

@@ -35,14 +35,14 @@ unsigned sig_stack_size = SIGSTKSZ;
 static pthread_t signals_thread;
 static volatile int remote_sig;
 
-static int is_addr_on_stack(void *addr)
+static int is_addr_on_stack(jl_tls_states_t *ptls, void *addr)
 {
 #ifdef COPY_STACKS
-    return ((char*)addr > (char*)jl_stack_lo-3000000 &&
-            (char*)addr < (char*)jl_stack_hi);
+    return ((char*)addr > (char*)ptls->stack_lo-3000000 &&
+            (char*)addr < (char*)ptls->stack_hi);
 #else
-    return ((char*)addr > (char*)jl_current_task->stkbuf &&
-            (char*)addr < (char*)jl_current_task->stkbuf + jl_current_task->ssize);
+    return ((char*)addr > (char*)ptls->current_task->stkbuf &&
+            (char*)addr < (char*)ptls->current_task->stkbuf + ptls->current_task->ssize);
 #endif
 }
 
@@ -75,7 +75,7 @@ static void segv_handler(int sig, siginfo_t *info, void *context)
     sigset_t sset;
     assert(sig == SIGSEGV);
 
-    if (jl_in_jl_ || is_addr_on_stack(info->si_addr)) { // stack overflow, or restarting jl_
+    if (jl_in_jl_ || is_addr_on_stack(jl_get_ptls_states(), info->si_addr)) { // stack overflow, or restarting jl_
         sigemptyset(&sset);
         sigaddset(&sset, SIGSEGV);
         sigprocmask(SIG_UNBLOCK, &sset, NULL);

src/threading.c

@@ -216,7 +216,7 @@ void ti_threadfun(void *arg)
 
     // initialize this thread (set tid, create heap, etc.)
     ti_initthread(ta->tid);
-    jl_init_stack_limits();
+    jl_init_stack_limits(0);
 
     // set up tasking
     jl_init_root_task(jl_stack_lo, jl_stack_hi - jl_stack_lo);