From 550f865685c3517102f30dac4c3c1666005a9904 Mon Sep 17 00:00:00 2001 From: Jameson Nash Date: Mon, 25 Sep 2023 13:20:19 +0200 Subject: [PATCH] add support for async backtraces of Tasks on any thread (#51430) --- src/Makefile | 2 +- src/gc-stacks.c | 80 +++++--- src/gc.c | 8 +- src/interpreter.c | 3 +- src/julia.h | 5 + src/julia_internal.h | 13 +- src/julia_threads.h | 8 +- src/mtarraylist.c | 81 ++++++++ src/signals-mach.c | 29 ++- src/signals-unix.c | 54 +++-- src/signals-win.c | 109 ++++++---- src/stackwalk.c | 470 +++++++++++++++++++++++-------------------- src/threading.c | 6 +- 13 files changed, 536 insertions(+), 332 deletions(-) create mode 100644 src/mtarraylist.c diff --git a/src/Makefile b/src/Makefile index 51d8bf1f2de65a..42ff9aa3a9b361 100644 --- a/src/Makefile +++ b/src/Makefile @@ -43,7 +43,7 @@ endif SRCS := \ jltypes gf typemap smallintset ast builtins module interpreter symbol \ dlload sys init task array staticdata toplevel jl_uv datatype \ - simplevector runtime_intrinsics precompile jloptions \ + simplevector runtime_intrinsics precompile jloptions mtarraylist \ threading partr stackwalk gc gc-debug gc-pages gc-stacks gc-alloc-profiler method \ jlapi signal-handling safepoint timing subtype rtutils gc-heap-snapshot \ crc32c APInt-C processor ircode opaque_closure codegen-stubs coverage runtime_ccall diff --git a/src/gc-stacks.c b/src/gc-stacks.c index 20e283afb2b862..693cb8d0eadf01 100644 --- a/src/gc-stacks.c +++ b/src/gc-stacks.c @@ -119,7 +119,7 @@ static void _jl_free_stack(jl_ptls_t ptls, void *stkbuf, size_t bufsz) if (bufsz <= pool_sizes[JL_N_STACK_POOLS - 1]) { unsigned pool_id = select_pool(bufsz); if (pool_sizes[pool_id] == bufsz) { - arraylist_push(&ptls->heap.free_stacks[pool_id], stkbuf); + small_arraylist_push(&ptls->heap.free_stacks[pool_id], stkbuf); return; } } @@ -148,7 +148,7 @@ void jl_release_task_stack(jl_ptls_t ptls, jl_task_t *task) #ifdef _COMPILER_ASAN_ENABLED_ __asan_unpoison_stack_memory((uintptr_t)stkbuf, bufsz); #endif - arraylist_push(&ptls->heap.free_stacks[pool_id], stkbuf); + small_arraylist_push(&ptls->heap.free_stacks[pool_id], stkbuf); } } } @@ -163,9 +163,9 @@ JL_DLLEXPORT void *jl_malloc_stack(size_t *bufsz, jl_task_t *owner) JL_NOTSAFEPO if (ssize <= pool_sizes[JL_N_STACK_POOLS - 1]) { unsigned pool_id = select_pool(ssize); ssize = pool_sizes[pool_id]; - arraylist_t *pool = &ptls->heap.free_stacks[pool_id]; + small_arraylist_t *pool = &ptls->heap.free_stacks[pool_id]; if (pool->len > 0) { - stk = arraylist_pop(pool); + stk = small_arraylist_pop(pool); } } else { @@ -184,8 +184,8 @@ JL_DLLEXPORT void *jl_malloc_stack(size_t *bufsz, jl_task_t *owner) JL_NOTSAFEPO } *bufsz = ssize; if (owner) { - arraylist_t *live_tasks = &ptls->heap.live_tasks; - arraylist_push(live_tasks, owner); + small_arraylist_t *live_tasks = &ptls->heap.live_tasks; + mtarraylist_push(live_tasks, owner); } return stk; } @@ -209,7 +209,7 @@ void sweep_stack_pools(void) // free half of stacks that remain unused since last sweep for (int p = 0; p < JL_N_STACK_POOLS; p++) { - arraylist_t *al = &ptls2->heap.free_stacks[p]; + small_arraylist_t *al = &ptls2->heap.free_stacks[p]; size_t n_to_free; if (al->len > MIN_STACK_MAPPINGS_PER_POOL) { n_to_free = al->len / 2; @@ -220,12 +220,12 @@ void sweep_stack_pools(void) n_to_free = 0; } for (int n = 0; n < n_to_free; n++) { - void *stk = arraylist_pop(al); + void *stk = small_arraylist_pop(al); free_stack(stk, pool_sizes[p]); } } - arraylist_t *live_tasks = &ptls2->heap.live_tasks; + small_arraylist_t *live_tasks = &ptls2->heap.live_tasks; size_t n = 0; size_t ndel = 0; size_t l = live_tasks->len; @@ -268,24 +268,52 @@ void sweep_stack_pools(void) JL_DLLEXPORT jl_array_t *jl_live_tasks(void) { - jl_task_t *ct = jl_current_task; - jl_ptls_t ptls = ct->ptls; - 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, changing the number of tasks - } 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]; + size_t nthreads = jl_atomic_load_acquire(&jl_n_threads); + jl_ptls_t *allstates = jl_atomic_load_relaxed(&jl_all_tls_states); + size_t l = 0; // l is not reset on restart, so we keep getting more aggressive at making a big enough list everything it fails +restart: + for (size_t i = 0; i < nthreads; i++) { + // skip GC threads since they don't have tasks + if (gc_first_tid <= i && i < gc_first_tid + jl_n_gcthreads) { + continue; + } + jl_ptls_t ptls2 = allstates[i]; + if (ptls2 == NULL) + continue; + small_arraylist_t *live_tasks = &ptls2->heap.live_tasks; + size_t n = mtarraylist_length(live_tasks); + l += n + (ptls2->root_task->stkbuf != NULL); + } + l += l / 20; // add 5% for margin of estimation error + jl_array_t *a = jl_alloc_vec_any(l); // may gc, changing the number of tasks and forcing us to reload everything + nthreads = jl_atomic_load_acquire(&jl_n_threads); + allstates = jl_atomic_load_relaxed(&jl_all_tls_states); + size_t j = 0; + for (size_t i = 0; i < nthreads; i++) { + // skip GC threads since they don't have tasks + if (gc_first_tid <= i && i < gc_first_tid + jl_n_gcthreads) { + continue; + } + jl_ptls_t ptls2 = allstates[i]; + if (ptls2 == NULL) + continue; + jl_task_t *t = ptls2->root_task; + if (t->stkbuf != NULL) { + if (j == l) + goto restart; + ((void**)jl_array_data(a))[j++] = t; + } + small_arraylist_t *live_tasks = &ptls2->heap.live_tasks; + size_t n = mtarraylist_length(live_tasks); + for (size_t i = 0; i < n; i++) { + jl_task_t *t = (jl_task_t*)mtarraylist_get(live_tasks, i); + if (t->stkbuf != NULL) { + if (j == l) + goto restart; + ((void**)jl_array_data(a))[j++] = t; + } + } } - l = jl_array_len(a); if (j < l) { JL_GC_PUSH1(&a); jl_array_del_end(a, l - j); diff --git a/src/gc.c b/src/gc.c index d606f922906323..d9e5c7906cce32 100644 --- a/src/gc.c +++ b/src/gc.c @@ -1084,7 +1084,7 @@ JL_DLLEXPORT jl_weakref_t *jl_gc_new_weakref_th(jl_ptls_t ptls, jl_weakref_t *wr = (jl_weakref_t*)jl_gc_alloc(ptls, sizeof(void*), jl_weakref_type); wr->value = value; // NOTE: wb not needed here - arraylist_push(&ptls->heap.weak_refs, wr); + small_arraylist_push(&ptls->heap.weak_refs, wr); return wr; } @@ -3679,8 +3679,10 @@ void jl_init_thread_heap(jl_ptls_t ptls) p[i].freelist = NULL; p[i].newpages = NULL; } - arraylist_new(&heap->weak_refs, 0); - arraylist_new(&heap->live_tasks, 0); + small_arraylist_new(&heap->weak_refs, 0); + small_arraylist_new(&heap->live_tasks, 0); + for (int i = 0; i < JL_N_STACK_POOLS; i++) + small_arraylist_new(&heap->free_stacks[i], 0); heap->mallocarrays = NULL; heap->mafreelist = NULL; heap->big_objects = NULL; diff --git a/src/interpreter.c b/src/interpreter.c index 1f9c416d99b1b1..6f546db9bbbb10 100644 --- a/src/interpreter.c +++ b/src/interpreter.c @@ -65,7 +65,8 @@ extern void JL_GC_ENABLEFRAME(interpreter_state*) JL_NOTSAFEPOINT; // we define this separately so that we can populate the frame before we add it to the backtrace // it's recommended to mark the containing function with NOINLINE, though not essential #define JL_GC_ENABLEFRAME(frame) \ - ((void**)&frame[1])[0] = __builtin_frame_address(0); + jl_signal_fence(); \ + ((void**)&frame[1])[0] = __builtin_frame_address(0); #endif diff --git a/src/julia.h b/src/julia.h index caebdf450ed758..bb747a77a518d7 100644 --- a/src/julia.h +++ b/src/julia.h @@ -968,6 +968,11 @@ JL_DLLEXPORT void *jl_gc_managed_realloc(void *d, size_t sz, size_t oldsz, int isaligned, jl_value_t *owner); JL_DLLEXPORT void jl_gc_safepoint(void); +void *mtarraylist_get(small_arraylist_t *_a, size_t idx) JL_NOTSAFEPOINT; +size_t mtarraylist_length(small_arraylist_t *_a) JL_NOTSAFEPOINT; +void mtarraylist_add(small_arraylist_t *_a, void *elt, size_t idx) JL_NOTSAFEPOINT; +void mtarraylist_push(small_arraylist_t *_a, void *elt) JL_NOTSAFEPOINT; + // object accessors ----------------------------------------------------------- #define jl_svec_len(t) (((jl_svec_t*)(t))->length) diff --git a/src/julia_internal.h b/src/julia_internal.h index 9ecc97cc646833..149eebf21ada0f 100644 --- a/src/julia_internal.h +++ b/src/julia_internal.h @@ -195,10 +195,12 @@ JL_DLLEXPORT void jl_set_profile_peek_duration(double); JL_DLLEXPORT void jl_init_profile_lock(void); JL_DLLEXPORT uintptr_t jl_lock_profile_rd_held(void) JL_NOTSAFEPOINT; -JL_DLLEXPORT void jl_lock_profile(void) JL_NOTSAFEPOINT; -JL_DLLEXPORT void jl_unlock_profile(void) JL_NOTSAFEPOINT; -JL_DLLEXPORT void jl_lock_profile_wr(void) JL_NOTSAFEPOINT; -JL_DLLEXPORT void jl_unlock_profile_wr(void) JL_NOTSAFEPOINT; +JL_DLLEXPORT void jl_lock_profile(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER; +JL_DLLEXPORT void jl_unlock_profile(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_LEAVE; +JL_DLLEXPORT void jl_lock_profile_wr(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER; +JL_DLLEXPORT void jl_unlock_profile_wr(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_LEAVE; +int jl_lock_stackwalk(void) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_ENTER; +void jl_unlock_stackwalk(int lockret) JL_NOTSAFEPOINT JL_NOTSAFEPOINT_LEAVE; // number of cycles since power-on static inline uint64_t cycleclock(void) JL_NOTSAFEPOINT @@ -1181,6 +1183,9 @@ void jl_print_bt_entry_codeloc(int sig, jl_bt_element_t *bt_data) JL_NOTSAFEPOIN #ifdef _OS_WINDOWS_ JL_DLLEXPORT void jl_refresh_dbg_module_list(void); #endif +int jl_thread_suspend_and_get_state(int tid, int timeout, bt_context_t *ctx) JL_NOTSAFEPOINT; +void jl_thread_resume(int tid) JL_NOTSAFEPOINT; + // *to is NULL or malloc'd pointer, from is allowed to be NULL STATIC_INLINE char *jl_copy_str(char **to, const char *from) JL_NOTSAFEPOINT { diff --git a/src/julia_threads.h b/src/julia_threads.h index 97a45c356656bf..8570e069383d5e 100644 --- a/src/julia_threads.h +++ b/src/julia_threads.h @@ -107,7 +107,7 @@ typedef struct { // handle to reference an OS thread #ifdef _OS_WINDOWS_ -typedef DWORD jl_thread_t; +typedef HANDLE jl_thread_t; #else typedef pthread_t jl_thread_t; #endif @@ -138,10 +138,10 @@ typedef struct { typedef struct { // variable for tracking weak references - arraylist_t weak_refs; + small_arraylist_t weak_refs; // live tasks started on this thread // that are holding onto a stack from the pool - arraylist_t live_tasks; + small_arraylist_t live_tasks; // variables for tracking malloc'd arrays struct _mallocarray_t *mallocarrays; @@ -169,7 +169,7 @@ typedef struct { jl_gc_pool_t norm_pools[JL_GC_N_POOLS]; #define JL_N_STACK_POOLS 16 - arraylist_t free_stacks[JL_N_STACK_POOLS]; + small_arraylist_t free_stacks[JL_N_STACK_POOLS]; } jl_thread_heap_t; typedef struct { diff --git a/src/mtarraylist.c b/src/mtarraylist.c new file mode 100644 index 00000000000000..8bad44797dab43 --- /dev/null +++ b/src/mtarraylist.c @@ -0,0 +1,81 @@ +// This file is a part of Julia. License is MIT: https://julialang.org/license + +#include "julia.h" +#include "julia_internal.h" +#include "julia_assert.h" + +#ifdef __cplusplus +extern "C" { +#endif + +// this file provides some alternate API functions for small_arraylist (push and add) +// which can be safely observed from other threads concurrently +// there is only permitted to be a single writer thread (or a mutex) +// but there can be any number of observers + +typedef struct { + _Atomic(uint32_t) len; + uint32_t max; + _Atomic(_Atomic(void*)*) items; + _Atomic(void*) _space[SMALL_AL_N_INLINE]; +} small_mtarraylist_t; + +// change capacity to at least newlen +static void mtarraylist_resizeto(small_mtarraylist_t *a, size_t len, size_t newlen) JL_NOTSAFEPOINT +{ + size_t max = a->max; + if (newlen > max) { + size_t nm = max * 2; + if (nm == 0) + nm = 1; + while (newlen > nm) + nm *= 2; + void *olditems = (void*)jl_atomic_load_relaxed(&a->items); + void *p = calloc_s(nm * sizeof(void*)); + memcpy(p, olditems, len * sizeof(void*)); + jl_atomic_store_release(&a->items, (_Atomic(void*)*)p); + a->max = nm; + if (olditems != (void*)&a->_space[0]) { + jl_task_t *ct = jl_current_task; + jl_gc_add_quiescent(ct->ptls, (void**)olditems, free); + } + } +} + +// single-threaded +void mtarraylist_push(small_arraylist_t *_a, void *elt) +{ + small_mtarraylist_t *a = (small_mtarraylist_t*)_a; + size_t len = jl_atomic_load_relaxed(&a->len); + mtarraylist_resizeto(a, len, len + 1); + jl_atomic_store_release(&jl_atomic_load_relaxed(&a->items)[len], elt); + jl_atomic_store_release(&a->len, len + 1); +} + +// single-threaded +void mtarraylist_add(small_arraylist_t *_a, void *elt, size_t idx) +{ + small_mtarraylist_t *a = (small_mtarraylist_t*)_a; + size_t len = jl_atomic_load_relaxed(&a->len); + mtarraylist_resizeto(a, len, idx + 1); + jl_atomic_store_release(&jl_atomic_load_relaxed(&a->items)[idx], elt); + if (jl_atomic_load_relaxed(&a->len) < idx + 1) + jl_atomic_store_release(&a->len, idx + 1); +} + +// concurrent-safe +size_t mtarraylist_length(small_arraylist_t *_a) +{ + small_mtarraylist_t *a = (small_mtarraylist_t*)_a; + return jl_atomic_load_relaxed(&a->len); +} + +// concurrent-safe +void *mtarraylist_get(small_arraylist_t *_a, size_t idx) +{ + small_mtarraylist_t *a = (small_mtarraylist_t*)_a; + size_t len = jl_atomic_load_acquire(&a->len); + if (idx >= len) + return NULL; + return jl_atomic_load_relaxed(&jl_atomic_load_relaxed(&a->items)[idx]); +} diff --git a/src/signals-mach.c b/src/signals-mach.c index 8ac7e5301d7ad5..32f0354cf17151 100644 --- a/src/signals-mach.c +++ b/src/signals-mach.c @@ -381,12 +381,12 @@ static void attach_exception_port(thread_port_t thread, int segv_only) HANDLE_MACH_ERROR("thread_set_exception_ports", ret); } -static int jl_thread_suspend_and_get_state2(int tid, host_thread_state_t *ctx) +static int jl_thread_suspend_and_get_state2(int tid, host_thread_state_t *ctx) JL_NOTSAFEPOINT { jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[tid]; if (ptls2 == NULL) // this thread is not alive return 0; - jl_task_t *ct2 = ptls2 ? jl_atomic_load_relaxed(&ptls2->current_task) : NULL; + jl_task_t *ct2 = jl_atomic_load_relaxed(&ptls2->current_task); if (ct2 == NULL) // this thread is already dead return 0; @@ -404,18 +404,18 @@ static int jl_thread_suspend_and_get_state2(int tid, host_thread_state_t *ctx) return 1; } -static void jl_thread_suspend_and_get_state(int tid, int timeout, unw_context_t **ctx) +int jl_thread_suspend_and_get_state(int tid, int timeout, bt_context_t *ctx) { (void)timeout; - static host_thread_state_t state; + host_thread_state_t state; if (!jl_thread_suspend_and_get_state2(tid, &state)) { - *ctx = NULL; - return; + return 0; } - *ctx = (unw_context_t*)&state; + *ctx = *(unw_context_t*)&state; + return 1; } -static void jl_thread_resume(int tid, int sig) +void jl_thread_resume(int tid) { jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[tid]; mach_port_t thread = pthread_mach_thread_np(ptls2->system_id); @@ -583,8 +583,15 @@ static void jl_unlock_profile_mach(int dlsymlock, int keymgr_locked) jl_unlock_profile(); } -#define jl_lock_profile() int keymgr_locked = jl_lock_profile_mach(1) -#define jl_unlock_profile() jl_unlock_profile_mach(1, keymgr_locked) +int jl_lock_stackwalk(void) +{ + return jl_lock_profile_mach(1); +} + +void jl_unlock_stackwalk(int lockret) +{ + jl_unlock_profile_mach(1, lockret); +} void *mach_profile_listener(void *arg) { @@ -677,7 +684,7 @@ void *mach_profile_listener(void *arg) bt_data_prof[bt_size_cur++].uintptr = 0; } // We're done! Resume the thread. - jl_thread_resume(i, 0); + jl_thread_resume(i); } jl_unlock_profile_mach(0, keymgr_locked); if (running) { diff --git a/src/signals-unix.c b/src/signals-unix.c index 165ef07abb3fe2..f38389913aa596 100644 --- a/src/signals-unix.c +++ b/src/signals-unix.c @@ -293,6 +293,18 @@ int exc_reg_is_write_fault(uintptr_t esr) { #include "signals-mach.c" #else +int jl_lock_stackwalk(void) +{ + jl_lock_profile(); + return 0; +} + +void jl_unlock_stackwalk(int lockret) +{ + (void)lockret; + jl_unlock_profile(); +} + #if defined(_OS_LINUX_) && (defined(_CPU_X86_64_) || defined(_CPU_X86_)) int is_write_fault(void *context) { @@ -386,12 +398,12 @@ JL_NO_ASAN static void segv_handler(int sig, siginfo_t *info, void *context) } #if !defined(JL_DISABLE_LIBUNWIND) -static unw_context_t *signal_context; +static bt_context_t *signal_context; pthread_mutex_t in_signal_lock; static pthread_cond_t exit_signal_cond; static pthread_cond_t signal_caught_cond; -static void jl_thread_suspend_and_get_state(int tid, int timeout, unw_context_t **ctx) +int jl_thread_suspend_and_get_state(int tid, int timeout, bt_context_t *ctx) { struct timespec ts; clock_gettime(CLOCK_REALTIME, &ts); @@ -401,9 +413,8 @@ static void jl_thread_suspend_and_get_state(int tid, int timeout, unw_context_t jl_task_t *ct2 = ptls2 ? jl_atomic_load_relaxed(&ptls2->current_task) : NULL; if (ct2 == NULL) { // this thread is not alive or already dead - *ctx = NULL; pthread_mutex_unlock(&in_signal_lock); - return; + return 0; } jl_atomic_store_release(&ptls2->signal_request, 1); pthread_kill(ptls2->system_id, SIGUSR2); @@ -412,9 +423,8 @@ static void jl_thread_suspend_and_get_state(int tid, int timeout, unw_context_t if (err == ETIMEDOUT) { sig_atomic_t request = 1; if (jl_atomic_cmpswap(&ptls2->signal_request, &request, 0)) { - *ctx = NULL; pthread_mutex_unlock(&in_signal_lock); - return; + return 0; } // Request is either now 0 (meaning the other thread is waiting for // exit_signal_cond already), @@ -431,15 +441,16 @@ static void jl_thread_suspend_and_get_state(int tid, int timeout, unw_context_t // checking it is 0, and add an acquire barrier for good measure) int request = jl_atomic_load_acquire(&ptls2->signal_request); assert(request == 0); (void) request; - *ctx = signal_context; + jl_atomic_store_release(&ptls2->signal_request, 1); // prepare to resume normally + *ctx = *signal_context; + return 1; } -static void jl_thread_resume(int tid, int sig) +void jl_thread_resume(int tid) { jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[tid]; - jl_atomic_store_release(&ptls2->signal_request, sig == -1 ? 3 : 1); pthread_cond_broadcast(&exit_signal_cond); - pthread_cond_wait(&signal_caught_cond, &in_signal_lock); // wait for thread to acknowledge + pthread_cond_wait(&signal_caught_cond, &in_signal_lock); // wait for thread to acknowledge (so that signal_request doesn't get mixed up) // The other thread is waiting to leave exit_signal_cond (verify that here by // checking it is 0, and add an acquire barrier for good measure) int request = jl_atomic_load_acquire(&ptls2->signal_request); @@ -474,14 +485,14 @@ CFI_NORETURN static void jl_exit_thread0(int signo, jl_bt_element_t *bt_data, size_t bt_size) { jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[0]; - unw_context_t *signal_context; + bt_context_t signal_context; // This also makes sure `sleep` is aborted. - jl_thread_suspend_and_get_state(0, 30, &signal_context); - if (signal_context != NULL) { + if (jl_thread_suspend_and_get_state(0, 30, &signal_context)) { thread0_exit_signo = signo; ptls2->bt_size = bt_size; // <= JL_MAX_BT_SIZE memcpy(ptls2->bt_data, bt_data, ptls2->bt_size * sizeof(bt_data[0])); - jl_thread_resume(0, -1); // resume with message 3 (call jl_exit_thread0_cb) + jl_atomic_store_release(&ptls2->signal_request, 3); + jl_thread_resume(0); // resume with message 3 (call jl_exit_thread0_cb) } else { // thread 0 is gone? just do the exit ourself @@ -877,11 +888,11 @@ static void *signal_listener(void *arg) int nthreads = jl_atomic_load_acquire(&jl_n_threads); bt_size = 0; #if !defined(JL_DISABLE_LIBUNWIND) - unw_context_t *signal_context; + bt_context_t signal_context; // sample each thread, round-robin style in reverse order // (so that thread zero gets notified last) if (critical || profile) { - jl_lock_profile(); + int lockret = jl_lock_stackwalk(); int *randperm; if (profile) randperm = profile_get_randperm(nthreads); @@ -889,8 +900,7 @@ static void *signal_listener(void *arg) // Stop the threads in the random or reverse round-robin order. int i = profile ? randperm[idx] : idx; // notify thread to stop - jl_thread_suspend_and_get_state(i, 1, &signal_context); - if (signal_context == NULL) + if (!jl_thread_suspend_and_get_state(i, 1, &signal_context)) continue; // do backtrace on thread contexts for critical signals @@ -898,7 +908,7 @@ static void *signal_listener(void *arg) if (critical) { bt_size += rec_backtrace_ctx(bt_data + bt_size, JL_MAX_BT_SIZE / nthreads - 1, - signal_context, NULL); + &signal_context, NULL); bt_data[bt_size++].uintptr = 0; } @@ -920,7 +930,7 @@ static void *signal_listener(void *arg) } else { // Get backtrace data bt_size_cur += rec_backtrace_ctx((jl_bt_element_t*)bt_data_prof + bt_size_cur, - bt_size_max - bt_size_cur - 1, signal_context, NULL); + bt_size_max - bt_size_cur - 1, &signal_context, NULL); } jl_set_safe_restore(old_buf); @@ -945,9 +955,9 @@ static void *signal_listener(void *arg) } // notify thread to resume - jl_thread_resume(i, sig); + jl_thread_resume(i); } - jl_unlock_profile(); + jl_unlock_stackwalk(lockret); } #ifndef HAVE_MACH if (profile && running) { diff --git a/src/signals-win.c b/src/signals-win.c index cca0af52ace53c..d1f83d6bfdcc46 100644 --- a/src/signals-win.c +++ b/src/signals-win.c @@ -344,6 +344,54 @@ JL_DLLEXPORT void jl_install_sigint_handler(void) static volatile HANDLE hBtThread = 0; +int jl_thread_suspend_and_get_state(int tid, int timeout, bt_context_t *ctx) +{ + (void)timeout; + jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[tid]; + if (ptls2 == NULL) // this thread is not alive + return 0; + jl_task_t *ct2 = jl_atomic_load_relaxed(&ptls2->current_task); + if (ct2 == NULL) // this thread is already dead + return 0; + HANDLE hThread = ptls2->system_id; + if ((DWORD)-1 == SuspendThread(hThread)) + return 0; + assert(sizeof(*ctx) == sizeof(CONTEXT)); + memset(ctx, 0, sizeof(CONTEXT)); + ctx->ContextFlags = CONTEXT_CONTROL | CONTEXT_INTEGER; + if (!GetThreadContext(hThread, ctx)) { + if ((DWORD)-1 == ResumeThread(hThread)) + abort(); + return 0; + } + return 1; +} + +void jl_thread_resume(int tid) +{ + jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[tid]; + HANDLE hThread = ptls2->system_id; + if ((DWORD)-1 == ResumeThread(hThread)) { + fputs("failed to resume main thread! aborting.", stderr); + abort(); + } +} + +int jl_lock_stackwalk(void) +{ + uv_mutex_lock(&jl_in_stackwalk); + jl_lock_profile(); + return 0; +} + +void jl_unlock_stackwalk(int lockret) +{ + (void)lockret; + jl_unlock_profile(); + uv_mutex_unlock(&jl_in_stackwalk); +} + + static DWORD WINAPI profile_bt( LPVOID lparam ) { // Note: illegal to use jl_* functions from this thread except for profiling-specific functions @@ -357,58 +405,45 @@ static DWORD WINAPI profile_bt( LPVOID lparam ) continue; } else { - uv_mutex_lock(&jl_in_stackwalk); - jl_lock_profile(); - if ((DWORD)-1 == SuspendThread(hMainThread)) { - fputs("failed to suspend main thread. aborting profiling.", stderr); - break; - } + // TODO: bring this up to parity with other OS by adding loop over tid here + int lockret = jl_lock_stackwalk(); CONTEXT ctxThread; - memset(&ctxThread, 0, sizeof(CONTEXT)); - ctxThread.ContextFlags = CONTEXT_CONTROL | CONTEXT_INTEGER; - if (!GetThreadContext(hMainThread, &ctxThread)) { - fputs("failed to get context from main thread. aborting profiling.", stderr); + if (!jl_thread_suspend_and_get_state(0, 0, &ctxThread)) { + jl_unlock_stackwalk(lockret); + fputs("failed to suspend main thread. aborting profiling.", stderr); jl_profile_stop_timer(); + break; } - else { - // Get backtrace data - bt_size_cur += rec_backtrace_ctx((jl_bt_element_t*)bt_data_prof + bt_size_cur, - bt_size_max - bt_size_cur - 1, &ctxThread, NULL); + // Get backtrace data + bt_size_cur += rec_backtrace_ctx((jl_bt_element_t*)bt_data_prof + bt_size_cur, + bt_size_max - bt_size_cur - 1, &ctxThread, NULL); - jl_ptls_t ptls = jl_atomic_load_relaxed(&jl_all_tls_states)[0]; // given only profiling hMainThread + jl_ptls_t ptls = jl_atomic_load_relaxed(&jl_all_tls_states)[0]; // given only profiling hMainThread - // store threadid but add 1 as 0 is preserved to indicate end of block - bt_data_prof[bt_size_cur++].uintptr = ptls->tid + 1; + // store threadid but add 1 as 0 is preserved to indicate end of block + bt_data_prof[bt_size_cur++].uintptr = ptls->tid + 1; - // store task id (never null) - bt_data_prof[bt_size_cur++].jlvalue = (jl_value_t*)jl_atomic_load_relaxed(&ptls->current_task); + // store task id (never null) + bt_data_prof[bt_size_cur++].jlvalue = (jl_value_t*)jl_atomic_load_relaxed(&ptls->current_task); - // store cpu cycle clock - bt_data_prof[bt_size_cur++].uintptr = cycleclock(); + // store cpu cycle clock + bt_data_prof[bt_size_cur++].uintptr = cycleclock(); - // store whether thread is sleeping but add 1 as 0 is preserved to indicate end of block - bt_data_prof[bt_size_cur++].uintptr = jl_atomic_load_relaxed(&ptls->sleep_check_state) + 1; + // store whether thread is sleeping but add 1 as 0 is preserved to indicate end of block + bt_data_prof[bt_size_cur++].uintptr = jl_atomic_load_relaxed(&ptls->sleep_check_state) + 1; - // Mark the end of this block with two 0's - bt_data_prof[bt_size_cur++].uintptr = 0; - bt_data_prof[bt_size_cur++].uintptr = 0; - } - jl_unlock_profile(); - uv_mutex_unlock(&jl_in_stackwalk); - if ((DWORD)-1 == ResumeThread(hMainThread)) { - jl_profile_stop_timer(); - fputs("failed to resume main thread! aborting.", stderr); - jl_gc_debug_critical_error(); - abort(); - } + // Mark the end of this block with two 0's + bt_data_prof[bt_size_cur++].uintptr = 0; + bt_data_prof[bt_size_cur++].uintptr = 0; + jl_unlock_stackwalk(lockret); + jl_thread_resume(0); jl_check_profile_autostop(); } } } - jl_unlock_profile(); uv_mutex_unlock(&jl_in_stackwalk); jl_profile_stop_timer(); - hBtThread = 0; + hBtThread = NULL; return 0; } diff --git a/src/stackwalk.c b/src/stackwalk.c index bd48b27b0cacd0..21a858e55b6944 100644 --- a/src/stackwalk.c +++ b/src/stackwalk.c @@ -865,7 +865,7 @@ _os_ptr_munge(uintptr_t ptr) extern bt_context_t *jl_to_bt_context(void *sigctx); -void jl_rec_backtrace(jl_task_t *t) JL_NOTSAFEPOINT +static void jl_rec_backtrace(jl_task_t *t) JL_NOTSAFEPOINT { jl_task_t *ct = jl_current_task; jl_ptls_t ptls = ct->ptls; @@ -874,222 +874,242 @@ void jl_rec_backtrace(jl_task_t *t) JL_NOTSAFEPOINT ptls->bt_size = rec_backtrace(ptls->bt_data, JL_MAX_BT_SIZE, 0); return; } - if (t->copy_stack || !t->started || t->stkbuf == NULL) - return; - int16_t old = -1; - if (!jl_atomic_cmpswap(&t->tid, &old, ptls->tid) && 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.ctx.uc_mcontext; + int16_t old = -1; + while (!jl_atomic_cmpswap(&t->tid, &old, ptls->tid) && old != ptls->tid) { + int lockret = jl_lock_stackwalk(); + // if this task is already running somewhere, we need to stop the thread it is running on and query its state + if (!jl_thread_suspend_and_get_state(old, 0, &c)) { + jl_unlock_stackwalk(lockret); + return; + } + jl_unlock_stackwalk(lockret); + if (jl_atomic_load_relaxed(&t->tid) == old) { + jl_ptls_t ptls2 = jl_atomic_load_relaxed(&jl_all_tls_states)[old]; + if (ptls2->previous_task == t || // we might print the wrong stack here, since we can't know whether we executed the swapcontext yet or not, but it at least avoids trying to access the state inside uc_mcontext which might not be set yet + (ptls2->previous_task == NULL && jl_atomic_load_relaxed(&ptls2->current_task) == t)) { // this case should be always accurate + // use the thread context for the unwind state + context = &c; + } + break; + } + // got the wrong thread stopped, try again + jl_thread_resume(old); + } + if (context == NULL && (!t->copy_stack && t->started && t->stkbuf != NULL)) { + // need to read the context from the task stored state +#if defined(_OS_WINDOWS_) + memset(&c, 0, sizeof(c)); + _JUMP_BUFFER *mctx = (_JUMP_BUFFER*)&t->ctx.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 + 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; + c.Eip = mctx->Eip; + c.Esp = mctx->Esp; + c.Ebp = mctx->Ebp; #endif - context = &c; + context = &c; #elif defined(JL_HAVE_UNW_CONTEXT) - context = &t->ctx.ctx; + context = &t->ctx.ctx; #elif defined(JL_HAVE_UCONTEXT) - context = jl_to_bt_context(&t->ctx.ctx); + context = jl_to_bt_context(&t->ctx.ctx); #elif defined(JL_HAVE_ASM) - bt_context_t c; - memset(&c, 0, sizeof(c)); - #if defined(_OS_LINUX_) && defined(__GLIBC__) - __jmp_buf *mctx = &t->ctx.ctx.uc_mcontext->__jmpbuf; - mcontext_t *mc = &c.uc_mcontext; - #if defined(_CPU_X86_) - // https://github.com/bminor/glibc/blame/master/sysdeps/i386/__longjmp.S - // https://github.com/bminor/glibc/blame/master/sysdeps/i386/jmpbuf-offsets.h - // https://github.com/bminor/musl/blame/master/src/setjmp/i386/longjmp.s - mc->gregs[REG_EBX] = (*mctx)[0]; - mc->gregs[REG_ESI] = (*mctx)[1]; - mc->gregs[REG_EDI] = (*mctx)[2]; - mc->gregs[REG_EBP] = (*mctx)[3]; - mc->gregs[REG_ESP] = (*mctx)[4]; - mc->gregs[REG_EIP] = (*mctx)[5]; - // ifdef PTR_DEMANGLE ? - mc->gregs[REG_ESP] = ptr_demangle(mc->gregs[REG_ESP]); - mc->gregs[REG_EIP] = ptr_demangle(mc->gregs[REG_EIP]); - context = &c; - #elif defined(_CPU_X86_64_) - // https://github.com/bminor/glibc/blame/master/sysdeps/x86_64/__longjmp.S - // https://github.com/bminor/glibc/blame/master/sysdeps/x86_64/jmpbuf-offsets.h - // https://github.com/bminor/musl/blame/master/src/setjmp/x86_64/setjmp.s - mc->gregs[REG_RBX] = (*mctx)[0]; - mc->gregs[REG_RBP] = (*mctx)[1]; - mc->gregs[REG_R12] = (*mctx)[2]; - mc->gregs[REG_R13] = (*mctx)[3]; - mc->gregs[REG_R14] = (*mctx)[4]; - mc->gregs[REG_R15] = (*mctx)[5]; - mc->gregs[REG_RSP] = (*mctx)[6]; - mc->gregs[REG_RIP] = (*mctx)[7]; - // ifdef PTR_DEMANGLE ? - mc->gregs[REG_RBP] = ptr_demangle(mc->gregs[REG_RBP]); - mc->gregs[REG_RSP] = ptr_demangle(mc->gregs[REG_RSP]); - mc->gregs[REG_RIP] = ptr_demangle(mc->gregs[REG_RIP]); - context = &c; - #elif defined(_CPU_ARM_) - // https://github.com/bminor/glibc/blame/master/sysdeps/arm/__longjmp.S - // https://github.com/bminor/glibc/blame/master/sysdeps/arm/include/bits/setjmp.h - // https://github.com/bminor/musl/blame/master/src/setjmp/arm/longjmp.S - mc->arm_sp = (*mctx)[0]; - mc->arm_lr = (*mctx)[1]; - mc->arm_r4 = (*mctx)[2]; // aka v1 - mc->arm_r5 = (*mctx)[3]; // aka v2 - mc->arm_r6 = (*mctx)[4]; // aka v3 - mc->arm_r7 = (*mctx)[5]; // aka v4 - mc->arm_r8 = (*mctx)[6]; // aka v5 - mc->arm_r9 = (*mctx)[7]; // aka v6 aka sb - mc->arm_r10 = (*mctx)[8]; // aka v7 aka sl - mc->arm_fp = (*mctx)[10]; // aka v8 aka r11 - // ifdef PTR_DEMANGLE ? - mc->arm_sp = ptr_demangle(mc->arm_sp); - mc->arm_lr = ptr_demangle(mc->arm_lr); - mc->arm_pc = mc->arm_lr; - context = &c; - #elif defined(_CPU_AARCH64_) - // https://github.com/bminor/glibc/blame/master/sysdeps/aarch64/__longjmp.S - // https://github.com/bminor/glibc/blame/master/sysdeps/aarch64/jmpbuf-offsets.h - // https://github.com/bminor/musl/blame/master/src/setjmp/aarch64/longjmp.s - // https://github.com/libunwind/libunwind/blob/ec171c9ba7ea3abb2a1383cee2988a7abd483a1f/src/aarch64/unwind_i.h#L62 - unw_fpsimd_context_t *mcfp = (unw_fpsimd_context_t*)&mc->__reserved; - mc->regs[19] = (*mctx)[0]; - mc->regs[20] = (*mctx)[1]; - mc->regs[21] = (*mctx)[2]; - mc->regs[22] = (*mctx)[3]; - mc->regs[23] = (*mctx)[4]; - mc->regs[24] = (*mctx)[5]; - mc->regs[25] = (*mctx)[6]; - mc->regs[26] = (*mctx)[7]; - mc->regs[27] = (*mctx)[8]; - mc->regs[28] = (*mctx)[9]; - mc->regs[29] = (*mctx)[10]; // aka fp - mc->regs[30] = (*mctx)[11]; // aka lr - // Yes, they did skip 12 why writing the code originally; and, no, I do not know why. - mc->sp = (*mctx)[13]; - mcfp->vregs[7] = (*mctx)[14]; // aka d8 - mcfp->vregs[8] = (*mctx)[15]; // aka d9 - mcfp->vregs[9] = (*mctx)[16]; // aka d10 - mcfp->vregs[10] = (*mctx)[17]; // aka d11 - mcfp->vregs[11] = (*mctx)[18]; // aka d12 - mcfp->vregs[12] = (*mctx)[19]; // aka d13 - mcfp->vregs[13] = (*mctx)[20]; // aka d14 - mcfp->vregs[14] = (*mctx)[21]; // aka d15 - // ifdef PTR_DEMANGLE ? - mc->sp = ptr_demangle(mc->sp); - mc->regs[30] = ptr_demangle(mc->regs[30]); - mc->pc = mc->regs[30]; - context = &c; - #else - #pragma message("jl_rec_backtrace not defined for ASM/SETJMP on unknown linux") - (void)mc; - (void)c; - (void)mctx; - #endif - #elif defined(_OS_DARWIN_) - sigjmp_buf *mctx = &t->ctx.ctx.uc_mcontext; - #if defined(_CPU_X86_64_) - // from https://github.com/apple/darwin-libplatform/blob/main/src/setjmp/x86_64/_setjmp.s - x86_thread_state64_t *mc = (x86_thread_state64_t*)&c; - mc->__rbx = ((uint64_t*)mctx)[0]; - mc->__rbp = ((uint64_t*)mctx)[1]; - mc->__rsp = ((uint64_t*)mctx)[2]; - mc->__r12 = ((uint64_t*)mctx)[3]; - mc->__r13 = ((uint64_t*)mctx)[4]; - mc->__r14 = ((uint64_t*)mctx)[5]; - mc->__r15 = ((uint64_t*)mctx)[6]; - mc->__rip = ((uint64_t*)mctx)[7]; - // added in libsystem_plaform 177.200.16 (macOS Mojave 10.14.3) - // prior to that _os_ptr_munge_token was (hopefully) typically 0, - // so x ^ 0 == x and this is a no-op - mc->__rbp = _OS_PTR_UNMUNGE(mc->__rbp); - mc->__rsp = _OS_PTR_UNMUNGE(mc->__rsp); - mc->__rip = _OS_PTR_UNMUNGE(mc->__rip); - context = &c; - #elif defined(_CPU_AARCH64_) - // from https://github.com/apple/darwin-libplatform/blob/main/src/setjmp/arm64/setjmp.s - // https://github.com/apple/darwin-xnu/blob/main/osfmk/mach/arm/_structs.h - // https://github.com/llvm/llvm-project/blob/7714e0317520207572168388f22012dd9e152e9e/libunwind/src/Registers.hpp -> Registers_arm64 - arm_thread_state64_t *mc = (arm_thread_state64_t*)&c; - mc->__x[19] = ((uint64_t*)mctx)[0]; - mc->__x[20] = ((uint64_t*)mctx)[1]; - mc->__x[21] = ((uint64_t*)mctx)[2]; - mc->__x[22] = ((uint64_t*)mctx)[3]; - mc->__x[23] = ((uint64_t*)mctx)[4]; - mc->__x[24] = ((uint64_t*)mctx)[5]; - mc->__x[25] = ((uint64_t*)mctx)[6]; - mc->__x[26] = ((uint64_t*)mctx)[7]; - mc->__x[27] = ((uint64_t*)mctx)[8]; - mc->__x[28] = ((uint64_t*)mctx)[9]; - mc->__x[10] = ((uint64_t*)mctx)[10]; - mc->__x[11] = ((uint64_t*)mctx)[11]; - mc->__x[12] = ((uint64_t*)mctx)[12]; - // 13 is reserved/unused - double *mcfp = (double*)&mc[1]; - mcfp[7] = ((uint64_t*)mctx)[14]; // aka d8 - mcfp[8] = ((uint64_t*)mctx)[15]; // aka d9 - mcfp[9] = ((uint64_t*)mctx)[16]; // aka d10 - mcfp[10] = ((uint64_t*)mctx)[17]; // aka d11 - mcfp[11] = ((uint64_t*)mctx)[18]; // aka d12 - mcfp[12] = ((uint64_t*)mctx)[19]; // aka d13 - mcfp[13] = ((uint64_t*)mctx)[20]; // aka d14 - mcfp[14] = ((uint64_t*)mctx)[21]; // aka d15 - mc->__fp = _OS_PTR_UNMUNGE(mc->__x[10]); - mc->__lr = _OS_PTR_UNMUNGE(mc->__x[11]); - mc->__x[12] = _OS_PTR_UNMUNGE(mc->__x[12]); - mc->__sp = mc->__x[12]; - // libunwind is broken for signed-pointers, but perhaps best not to leave the signed pointer lying around either - mc->__pc = ptrauth_strip(mc->__lr, 0); - mc->__pad = 0; // aka __ra_sign_state = not signed - context = &c; - #else - #pragma message("jl_rec_backtrace not defined for ASM/SETJMP on unknown darwin") - (void)mctx; - (void)c; - #endif - #elif defined(_OS_FREEBSD_) && defined(_CPU_X86_64_) - sigjmp_buf *mctx = &t->ctx.ctx.uc_mcontext; - mcontext_t *mc = &c.uc_mcontext; - // https://github.com/freebsd/freebsd-src/blob/releng/13.1/lib/libc/amd64/gen/_setjmp.S - mc->mc_rip = ((long*)mctx)[0]; - mc->mc_rbx = ((long*)mctx)[1]; - mc->mc_rsp = ((long*)mctx)[2]; - mc->mc_rbp = ((long*)mctx)[3]; - mc->mc_r12 = ((long*)mctx)[4]; - mc->mc_r13 = ((long*)mctx)[5]; - mc->mc_r14 = ((long*)mctx)[6]; - mc->mc_r15 = ((long*)mctx)[7]; - context = &c; - #else - #pragma message("jl_rec_backtrace not defined for ASM/SETJMP on unknown system") - (void)c; - #endif + memset(&c, 0, sizeof(c)); + #if defined(_OS_LINUX_) && defined(__GLIBC__) + __jmp_buf *mctx = &t->ctx.ctx.uc_mcontext->__jmpbuf; + mcontext_t *mc = &c.uc_mcontext; + #if defined(_CPU_X86_) + // https://github.com/bminor/glibc/blame/master/sysdeps/i386/__longjmp.S + // https://github.com/bminor/glibc/blame/master/sysdeps/i386/jmpbuf-offsets.h + // https://github.com/bminor/musl/blame/master/src/setjmp/i386/longjmp.s + mc->gregs[REG_EBX] = (*mctx)[0]; + mc->gregs[REG_ESI] = (*mctx)[1]; + mc->gregs[REG_EDI] = (*mctx)[2]; + mc->gregs[REG_EBP] = (*mctx)[3]; + mc->gregs[REG_ESP] = (*mctx)[4]; + mc->gregs[REG_EIP] = (*mctx)[5]; + // ifdef PTR_DEMANGLE ? + mc->gregs[REG_ESP] = ptr_demangle(mc->gregs[REG_ESP]); + mc->gregs[REG_EIP] = ptr_demangle(mc->gregs[REG_EIP]); + context = &c; + #elif defined(_CPU_X86_64_) + // https://github.com/bminor/glibc/blame/master/sysdeps/x86_64/__longjmp.S + // https://github.com/bminor/glibc/blame/master/sysdeps/x86_64/jmpbuf-offsets.h + // https://github.com/bminor/musl/blame/master/src/setjmp/x86_64/setjmp.s + mc->gregs[REG_RBX] = (*mctx)[0]; + mc->gregs[REG_RBP] = (*mctx)[1]; + mc->gregs[REG_R12] = (*mctx)[2]; + mc->gregs[REG_R13] = (*mctx)[3]; + mc->gregs[REG_R14] = (*mctx)[4]; + mc->gregs[REG_R15] = (*mctx)[5]; + mc->gregs[REG_RSP] = (*mctx)[6]; + mc->gregs[REG_RIP] = (*mctx)[7]; + // ifdef PTR_DEMANGLE ? + mc->gregs[REG_RBP] = ptr_demangle(mc->gregs[REG_RBP]); + mc->gregs[REG_RSP] = ptr_demangle(mc->gregs[REG_RSP]); + mc->gregs[REG_RIP] = ptr_demangle(mc->gregs[REG_RIP]); + context = &c; + #elif defined(_CPU_ARM_) + // https://github.com/bminor/glibc/blame/master/sysdeps/arm/__longjmp.S + // https://github.com/bminor/glibc/blame/master/sysdeps/arm/include/bits/setjmp.h + // https://github.com/bminor/musl/blame/master/src/setjmp/arm/longjmp.S + mc->arm_sp = (*mctx)[0]; + mc->arm_lr = (*mctx)[1]; + mc->arm_r4 = (*mctx)[2]; // aka v1 + mc->arm_r5 = (*mctx)[3]; // aka v2 + mc->arm_r6 = (*mctx)[4]; // aka v3 + mc->arm_r7 = (*mctx)[5]; // aka v4 + mc->arm_r8 = (*mctx)[6]; // aka v5 + mc->arm_r9 = (*mctx)[7]; // aka v6 aka sb + mc->arm_r10 = (*mctx)[8]; // aka v7 aka sl + mc->arm_fp = (*mctx)[10]; // aka v8 aka r11 + // ifdef PTR_DEMANGLE ? + mc->arm_sp = ptr_demangle(mc->arm_sp); + mc->arm_lr = ptr_demangle(mc->arm_lr); + mc->arm_pc = mc->arm_lr; + context = &c; + #elif defined(_CPU_AARCH64_) + // https://github.com/bminor/glibc/blame/master/sysdeps/aarch64/__longjmp.S + // https://github.com/bminor/glibc/blame/master/sysdeps/aarch64/jmpbuf-offsets.h + // https://github.com/bminor/musl/blame/master/src/setjmp/aarch64/longjmp.s + // https://github.com/libunwind/libunwind/blob/ec171c9ba7ea3abb2a1383cee2988a7abd483a1f/src/aarch64/unwind_i.h#L62 + unw_fpsimd_context_t *mcfp = (unw_fpsimd_context_t*)&mc->__reserved; + mc->regs[19] = (*mctx)[0]; + mc->regs[20] = (*mctx)[1]; + mc->regs[21] = (*mctx)[2]; + mc->regs[22] = (*mctx)[3]; + mc->regs[23] = (*mctx)[4]; + mc->regs[24] = (*mctx)[5]; + mc->regs[25] = (*mctx)[6]; + mc->regs[26] = (*mctx)[7]; + mc->regs[27] = (*mctx)[8]; + mc->regs[28] = (*mctx)[9]; + mc->regs[29] = (*mctx)[10]; // aka fp + mc->regs[30] = (*mctx)[11]; // aka lr + // Yes, they did skip 12 why writing the code originally; and, no, I do not know why. + mc->sp = (*mctx)[13]; + mcfp->vregs[7] = (*mctx)[14]; // aka d8 + mcfp->vregs[8] = (*mctx)[15]; // aka d9 + mcfp->vregs[9] = (*mctx)[16]; // aka d10 + mcfp->vregs[10] = (*mctx)[17]; // aka d11 + mcfp->vregs[11] = (*mctx)[18]; // aka d12 + mcfp->vregs[12] = (*mctx)[19]; // aka d13 + mcfp->vregs[13] = (*mctx)[20]; // aka d14 + mcfp->vregs[14] = (*mctx)[21]; // aka d15 + // ifdef PTR_DEMANGLE ? + mc->sp = ptr_demangle(mc->sp); + mc->regs[30] = ptr_demangle(mc->regs[30]); + mc->pc = mc->regs[30]; + context = &c; + #else + #pragma message("jl_rec_backtrace not defined for ASM/SETJMP on unknown linux") + (void)mc; + (void)c; + (void)mctx; + #endif + #elif defined(_OS_DARWIN_) + sigjmp_buf *mctx = &t->ctx.ctx.uc_mcontext; + #if defined(_CPU_X86_64_) + // from https://github.com/apple/darwin-libplatform/blob/main/src/setjmp/x86_64/_setjmp.s + x86_thread_state64_t *mc = (x86_thread_state64_t*)&c; + mc->__rbx = ((uint64_t*)mctx)[0]; + mc->__rbp = ((uint64_t*)mctx)[1]; + mc->__rsp = ((uint64_t*)mctx)[2]; + mc->__r12 = ((uint64_t*)mctx)[3]; + mc->__r13 = ((uint64_t*)mctx)[4]; + mc->__r14 = ((uint64_t*)mctx)[5]; + mc->__r15 = ((uint64_t*)mctx)[6]; + mc->__rip = ((uint64_t*)mctx)[7]; + // added in libsystem_plaform 177.200.16 (macOS Mojave 10.14.3) + // prior to that _os_ptr_munge_token was (hopefully) typically 0, + // so x ^ 0 == x and this is a no-op + mc->__rbp = _OS_PTR_UNMUNGE(mc->__rbp); + mc->__rsp = _OS_PTR_UNMUNGE(mc->__rsp); + mc->__rip = _OS_PTR_UNMUNGE(mc->__rip); + context = &c; + #elif defined(_CPU_AARCH64_) + // from https://github.com/apple/darwin-libplatform/blob/main/src/setjmp/arm64/setjmp.s + // https://github.com/apple/darwin-xnu/blob/main/osfmk/mach/arm/_structs.h + // https://github.com/llvm/llvm-project/blob/7714e0317520207572168388f22012dd9e152e9e/libunwind/src/Registers.hpp -> Registers_arm64 + arm_thread_state64_t *mc = (arm_thread_state64_t*)&c; + mc->__x[19] = ((uint64_t*)mctx)[0]; + mc->__x[20] = ((uint64_t*)mctx)[1]; + mc->__x[21] = ((uint64_t*)mctx)[2]; + mc->__x[22] = ((uint64_t*)mctx)[3]; + mc->__x[23] = ((uint64_t*)mctx)[4]; + mc->__x[24] = ((uint64_t*)mctx)[5]; + mc->__x[25] = ((uint64_t*)mctx)[6]; + mc->__x[26] = ((uint64_t*)mctx)[7]; + mc->__x[27] = ((uint64_t*)mctx)[8]; + mc->__x[28] = ((uint64_t*)mctx)[9]; + mc->__x[10] = ((uint64_t*)mctx)[10]; + mc->__x[11] = ((uint64_t*)mctx)[11]; + mc->__x[12] = ((uint64_t*)mctx)[12]; + // 13 is reserved/unused + double *mcfp = (double*)&mc[1]; + mcfp[7] = ((uint64_t*)mctx)[14]; // aka d8 + mcfp[8] = ((uint64_t*)mctx)[15]; // aka d9 + mcfp[9] = ((uint64_t*)mctx)[16]; // aka d10 + mcfp[10] = ((uint64_t*)mctx)[17]; // aka d11 + mcfp[11] = ((uint64_t*)mctx)[18]; // aka d12 + mcfp[12] = ((uint64_t*)mctx)[19]; // aka d13 + mcfp[13] = ((uint64_t*)mctx)[20]; // aka d14 + mcfp[14] = ((uint64_t*)mctx)[21]; // aka d15 + mc->__fp = _OS_PTR_UNMUNGE(mc->__x[10]); + mc->__lr = _OS_PTR_UNMUNGE(mc->__x[11]); + mc->__x[12] = _OS_PTR_UNMUNGE(mc->__x[12]); + mc->__sp = mc->__x[12]; + // libunwind is broken for signed-pointers, but perhaps best not to leave the signed pointer lying around either + mc->__pc = ptrauth_strip(mc->__lr, 0); + mc->__pad = 0; // aka __ra_sign_state = not signed + context = &c; + #else + #pragma message("jl_rec_backtrace not defined for ASM/SETJMP on unknown darwin") + (void)mctx; + (void)c; + #endif + #elif defined(_OS_FREEBSD_) && defined(_CPU_X86_64_) + sigjmp_buf *mctx = &t->ctx.ctx.uc_mcontext; + mcontext_t *mc = &c.uc_mcontext; + // https://github.com/freebsd/freebsd-src/blob/releng/13.1/lib/libc/amd64/gen/_setjmp.S + mc->mc_rip = ((long*)mctx)[0]; + mc->mc_rbx = ((long*)mctx)[1]; + mc->mc_rsp = ((long*)mctx)[2]; + mc->mc_rbp = ((long*)mctx)[3]; + mc->mc_r12 = ((long*)mctx)[4]; + mc->mc_r13 = ((long*)mctx)[5]; + mc->mc_r14 = ((long*)mctx)[6]; + mc->mc_r15 = ((long*)mctx)[7]; + context = &c; + #else + #pragma message("jl_rec_backtrace not defined for ASM/SETJMP on unknown system") + (void)c; + #endif #elif defined(JL_HAVE_ASYNCIFY) - #pragma message("jl_rec_backtrace not defined for ASYNCIFY") + #pragma message("jl_rec_backtrace not defined for ASYNCIFY") #elif defined(JL_HAVE_SIGALTSTACK) - #pragma message("jl_rec_backtrace not defined for SIGALTSTACK") + #pragma message("jl_rec_backtrace not defined for SIGALTSTACK") #else - #pragma message("jl_rec_backtrace not defined for unknown task system") + #pragma message("jl_rec_backtrace not defined for unknown task system") #endif + } if (context) - ptls->bt_size = rec_backtrace_ctx(ptls->bt_data, JL_MAX_BT_SIZE, context, t->gcstack); + 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); + else if (old != ptls->tid) + jl_thread_resume(old); } //-------------------------------------------------- @@ -1118,7 +1138,7 @@ JL_DLLEXPORT void jlbacktrace(void) JL_NOTSAFEPOINT } } -// Print backtrace for specified task +// Print backtrace for specified task to jl_safe_printf stderr JL_DLLEXPORT void jlbacktracet(jl_task_t *t) JL_NOTSAFEPOINT { jl_task_t *ct = jl_current_task; @@ -1138,9 +1158,7 @@ JL_DLLEXPORT void jl_print_backtrace(void) JL_NOTSAFEPOINT extern int gc_first_tid; -// Print backtraces for all live tasks, for all threads. -// WARNING: this is dangerous and can crash if used outside of gdb, if -// all of Julia's threads are not stopped! +// Print backtraces for all live tasks, for all threads, to jl_safe_printf stderr JL_DLLEXPORT void jl_print_task_backtraces(int show_done) JL_NOTSAFEPOINT { size_t nthreads = jl_atomic_load_acquire(&jl_n_threads); @@ -1151,25 +1169,35 @@ JL_DLLEXPORT void jl_print_task_backtraces(int show_done) JL_NOTSAFEPOINT continue; } jl_ptls_t ptls2 = allstates[i]; - arraylist_t *live_tasks = &ptls2->heap.live_tasks; - size_t n = live_tasks->len; + if (ptls2 == NULL) + continue; + small_arraylist_t *live_tasks = &ptls2->heap.live_tasks; + size_t n = mtarraylist_length(live_tasks); + jl_task_t *t = ptls2->root_task; + int t_state = jl_atomic_load_relaxed(&t->_state); jl_safe_printf("==== Thread %d created %zu live tasks\n", - ptls2->tid + 1, n + 1); - jl_safe_printf(" ---- Root task (%p)\n", ptls2->root_task); - jl_safe_printf(" (sticky: %d, started: %d, state: %d, tid: %d)\n", - ptls2->root_task->sticky, ptls2->root_task->started, - jl_atomic_load_relaxed(&ptls2->root_task->_state), - jl_atomic_load_relaxed(&ptls2->root_task->tid) + 1); - jlbacktracet(ptls2->root_task); - - void **lst = live_tasks->items; - for (size_t j = 0; j < live_tasks->len; j++) { - jl_task_t *t = (jl_task_t *)lst[j]; + ptls2->tid + 1, n + (t_state != JL_TASK_STATE_DONE)); + if (show_done || t_state != JL_TASK_STATE_DONE) { + jl_safe_printf(" ---- Root task (%p)\n", ptls2->root_task); + jl_safe_printf(" (sticky: %d, started: %d, state: %d, tid: %d)\n", + t->sticky, t->started, t_state, + jl_atomic_load_relaxed(&t->tid) + 1); + if (t->stkbuf != NULL) + jlbacktracet(t); + else + jl_safe_printf(" no stack\n"); + jl_safe_printf(" ---- End root task\n"); + } + + for (size_t j = 0; j < n; j++) { + jl_task_t *t = (jl_task_t*)mtarraylist_get(live_tasks, j); + if (t == NULL) + continue; int t_state = jl_atomic_load_relaxed(&t->_state); - if (!show_done && t_state == JL_TASK_STATE_DONE) { + if (!show_done && t_state == JL_TASK_STATE_DONE) continue; - } jl_safe_printf(" ---- Task %zu (%p)\n", j + 1, t); + // n.b. this information might not be consistent with the stack printing after it, since it could start running or change tid, etc. jl_safe_printf(" (sticky: %d, started: %d, state: %d, tid: %d)\n", t->sticky, t->started, t_state, jl_atomic_load_relaxed(&t->tid) + 1); diff --git a/src/threading.c b/src/threading.c index 7ed8e3b6e7dc99..6b7e14d3b8db42 100644 --- a/src/threading.c +++ b/src/threading.c @@ -305,6 +305,8 @@ static uv_mutex_t tls_lock; // controls write-access to these variables: _Atomic(jl_ptls_t*) jl_all_tls_states JL_GLOBALLY_ROOTED; int jl_all_tls_states_size; static uv_cond_t cond; +// concurrent reads are permitted, using the same pattern as mtsmall_arraylist +// it is implemented separately because the API of direct jl_all_tls_states use is already widely prevalent // return calling thread's ID JL_DLLEXPORT int16_t jl_threadid(void) @@ -338,7 +340,7 @@ jl_ptls_t jl_init_threadtls(int16_t tid) #ifndef _OS_WINDOWS_ pthread_setspecific(jl_task_exit_key, (void*)ptls); #endif - ptls->system_id = (jl_thread_t)(uintptr_t)uv_thread_self(); + ptls->system_id = uv_thread_self(); ptls->rngseed = jl_rand(); if (tid == 0) ptls->disable_gc = 1; @@ -373,10 +375,10 @@ jl_ptls_t jl_init_threadtls(int16_t tid) uv_cond_init(&ptls->wake_signal); uv_mutex_lock(&tls_lock); - jl_ptls_t *allstates = jl_atomic_load_relaxed(&jl_all_tls_states); if (tid == -1) tid = jl_atomic_load_relaxed(&jl_n_threads); ptls->tid = tid; + jl_ptls_t *allstates = jl_atomic_load_relaxed(&jl_all_tls_states); if (jl_all_tls_states_size <= tid) { int i, newsize = jl_all_tls_states_size + tid + 2; jl_ptls_t *newpptls = (jl_ptls_t*)calloc(newsize, sizeof(jl_ptls_t));