1 /* 2 * Copyright (c) 2018, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/javaClasses.inline.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "code/codeCache.inline.hpp" 29 #include "code/nmethod.inline.hpp" 30 #include "code/vmreg.inline.hpp" 31 #include "compiler/oopMap.inline.hpp" 32 #include "gc/shared/continuationGCSupport.inline.hpp" 33 #include "gc/shared/gc_globals.hpp" 34 #include "gc/shared/barrierSet.hpp" 35 #include "gc/shared/memAllocator.hpp" 36 #include "gc/shared/threadLocalAllocBuffer.inline.hpp" 37 #include "interpreter/interpreter.hpp" 38 #include "jfr/jfrEvents.hpp" 39 #include "logging/log.hpp" 40 #include "logging/logStream.hpp" 41 #include "oops/access.inline.hpp" 42 #include "oops/method.inline.hpp" 43 #include "oops/oopsHierarchy.hpp" 44 #include "oops/objArrayOop.inline.hpp" 45 #include "oops/stackChunkOop.inline.hpp" 46 #include "prims/jvmtiThreadState.hpp" 47 #include "runtime/arguments.hpp" 48 #include "runtime/continuation.hpp" 49 #include "runtime/continuationEntry.inline.hpp" 50 #include "runtime/continuationHelper.inline.hpp" 51 #include "runtime/continuationJavaClasses.inline.hpp" 52 #include "runtime/continuationWrapper.inline.hpp" 53 #include "runtime/frame.inline.hpp" 54 #include "runtime/interfaceSupport.inline.hpp" 55 #include "runtime/javaThread.inline.hpp" 56 #include "runtime/jniHandles.inline.hpp" 57 #include "runtime/keepStackGCProcessed.hpp" 58 #include "runtime/orderAccess.hpp" 59 #include "runtime/prefetch.inline.hpp" 60 #include "runtime/smallRegisterMap.inline.hpp" 61 #include "runtime/sharedRuntime.hpp" 62 #include "runtime/stackChunkFrameStream.inline.hpp" 63 #include "runtime/stackFrameStream.inline.hpp" 64 #include "runtime/stackOverflow.hpp" 65 #include "runtime/stackWatermarkSet.inline.hpp" 66 #include "utilities/debug.hpp" 67 #include "utilities/exceptions.hpp" 68 #include "utilities/macros.hpp" 69 #if INCLUDE_ZGC 70 #include "gc/z/zStackChunkGCData.inline.hpp" 71 #endif 72 73 #include <type_traits> 74 75 /* 76 * This file contains the implementation of continuation freezing (yield) and thawing (run). 77 * 78 * This code is very latency-critical and very hot. An ordinary and well-behaved server application 79 * would likely call these operations many thousands of times per second second, on every core. 80 * 81 * Freeze might be called every time the application performs any I/O operation, every time it 82 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called 83 * multiple times in each of those cases, as it is called by the return barrier, which may be 84 * invoked on method return. 85 * 86 * The amortized budget for each of those two operations is ~100-150ns. That is why, for 87 * example, every effort is made to avoid Java-VM transitions as much as possible. 88 * 89 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers 90 * and so frames simply copied, and the bottom-most one is patched. 91 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets 92 * and absolute pointers, and barriers invoked. 93 */ 94 95 /************************************************ 96 97 Thread-stack layout on freeze/thaw. 98 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp 99 100 +----------------------------+ 101 | . | 102 | . | 103 | . | 104 | carrier frames | 105 | | 106 |----------------------------| 107 | | 108 | Continuation.run | 109 | | 110 |============================| 111 | enterSpecial frame | 112 | pc | 113 | rbp | 114 | ----- | 115 ^ | int argsize | = ContinuationEntry 116 | | oopDesc* cont | 117 | | oopDesc* chunk | 118 | | ContinuationEntry* parent | 119 | | ... | 120 | |============================| <------ JavaThread::_cont_entry = entry->sp() 121 | | ? alignment word ? | 122 | |----------------------------| <--\ 123 | | | | 124 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words 125 Address | | | | Caller is still in the chunk. 126 | |----------------------------| | 127 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame 128 | | rbp | | isn't the last one in the continuation. 129 | | | | 130 | | frame | | 131 | | | | 132 +----------------------------| \__ Continuation frames to be frozen/thawed 133 | | / 134 | frame | | 135 | | | 136 |----------------------------| | 137 | | | 138 | frame | | 139 | | | 140 |----------------------------| <--/ 141 | | 142 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub 143 | | instead of a doYield stub 144 |============================| <- the sp passed to freeze 145 | | 146 | Native freeze/thaw frames | 147 | . | 148 | . | 149 | . | 150 +----------------------------+ 151 152 ************************************************/ 153 154 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing 155 156 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only 157 #if CONT_JFR 158 #define CONT_JFR_ONLY(code) code 159 #else 160 #define CONT_JFR_ONLY(code) 161 #endif 162 163 // TODO: See AbstractAssembler::generate_stack_overflow_check, 164 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size() 165 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point. 166 167 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk 168 169 // Used to just annotatate cold/hot branches 170 #define LIKELY(condition) (condition) 171 #define UNLIKELY(condition) (condition) 172 173 // debugging functions 174 #ifdef ASSERT 175 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue 176 177 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); } 178 179 static void do_deopt_after_thaw(JavaThread* thread); 180 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st); 181 static void log_frames(JavaThread* thread); 182 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty); 183 184 #define assert_pfl(p, ...) \ 185 do { \ 186 if (!(p)) { \ 187 JavaThread* t = JavaThread::active(); \ 188 if (t->has_last_Java_frame()) { \ 189 tty->print_cr("assert(" #p ") failed:"); \ 190 t->print_frame_layout(); \ 191 } \ 192 } \ 193 vmassert(p, __VA_ARGS__); \ 194 } while(0) 195 196 #else 197 static void verify_continuation(oop continuation) { } 198 #define assert_pfl(p, ...) 199 #endif 200 201 // should match Continuation.preemptStatus() in Continuation.java 202 enum freeze_result { 203 freeze_ok = 0, 204 freeze_ok_bottom = 1, 205 freeze_pinned_cs = 2, 206 freeze_pinned_native = 3, 207 freeze_pinned_monitor = 4, 208 freeze_exception = 5 209 }; 210 211 const char* freeze_result_names[6] = { 212 "freeze_ok", 213 "freeze_ok_bottom", 214 "freeze_pinned_cs", 215 "freeze_pinned_native", 216 "freeze_pinned_monitor", 217 "freeze_exception" 218 }; 219 220 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint); 221 template<typename ConfigT> static inline int freeze_internal(JavaThread* current, intptr_t* const sp); 222 223 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier); 224 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind); 225 226 227 // Entry point to freeze. Transitions are handled manually 228 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry(); 229 template<typename ConfigT> 230 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp)) 231 assert(sp == current->frame_anchor()->last_Java_sp(), ""); 232 233 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) { 234 current->set_cont_fastpath(nullptr); 235 } 236 237 return ConfigT::freeze(current, sp); 238 JRT_END 239 240 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier)) 241 return prepare_thaw_internal(thread, return_barrier); 242 JRT_END 243 244 template<typename ConfigT> 245 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind)) 246 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events. 247 // vFrameStreamCommon allocates Handles in RegisterMap for continuations. 248 // JRT_ENTRY instead? 249 ResetNoHandleMark rnhm; 250 251 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier 252 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread)); 253 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind); 254 JRT_END 255 256 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) { 257 JavaThread* thread = JavaThread::thread_from_jni_environment(env); 258 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false); 259 } 260 JVM_END 261 262 /////////// 263 264 enum class oop_kind { NARROW, WIDE }; 265 template <oop_kind oops, typename BarrierSetT> 266 class Config { 267 public: 268 typedef Config<oops, BarrierSetT> SelfT; 269 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>; 270 271 static int freeze(JavaThread* thread, intptr_t* const sp) { 272 return freeze_internal<SelfT>(thread, sp); 273 } 274 275 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) { 276 return thaw_internal<SelfT>(thread, kind); 277 } 278 }; 279 280 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) { 281 const size_t page_size = os::vm_page_size(); 282 if (size > page_size) { 283 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) { 284 return false; 285 } 286 } 287 return true; 288 } 289 290 #ifdef ASSERT 291 static oop get_continuation(JavaThread* thread) { 292 assert(thread != nullptr, ""); 293 assert(thread->threadObj() != nullptr, ""); 294 return java_lang_Thread::continuation(thread->threadObj()); 295 } 296 297 inline void clear_anchor(JavaThread* thread) { 298 thread->frame_anchor()->clear(); 299 } 300 301 static void set_anchor(JavaThread* thread, intptr_t* sp) { 302 address pc = ContinuationHelper::return_address_at( 303 sp - frame::sender_sp_ret_address_offset()); 304 assert(pc != nullptr, ""); 305 306 JavaFrameAnchor* anchor = thread->frame_anchor(); 307 anchor->set_last_Java_sp(sp); 308 anchor->set_last_Java_pc(pc); 309 ContinuationHelper::set_anchor_pd(anchor, sp); 310 311 assert(thread->has_last_Java_frame(), ""); 312 assert(thread->last_frame().cb() != nullptr, ""); 313 } 314 #endif // ASSERT 315 316 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) { 317 JavaFrameAnchor* anchor = thread->frame_anchor(); 318 anchor->set_last_Java_sp(entry->entry_sp()); 319 anchor->set_last_Java_pc(entry->entry_pc()); 320 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry); 321 322 assert(thread->has_last_Java_frame(), ""); 323 assert(thread->last_frame().cb() != nullptr, ""); 324 } 325 326 #if CONT_JFR 327 class FreezeThawJfrInfo : public StackObj { 328 short _e_size; 329 short _e_num_interpreted_frames; 330 public: 331 332 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {} 333 inline void record_interpreted_frame() { _e_num_interpreted_frames++; } 334 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; } 335 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt); 336 }; 337 338 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) { 339 if (e->should_commit()) { 340 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size); 341 e->set_carrierThread(JFR_JVM_THREAD_ID(jt)); 342 e->set_continuationClass(continuation->klass()); 343 e->set_interpretedFrames(_e_num_interpreted_frames); 344 e->set_size(_e_size); 345 e->commit(); 346 } 347 } 348 #endif // CONT_JFR 349 350 /////////////// FREEZE //// 351 352 class FreezeBase : public StackObj { 353 protected: 354 JavaThread* const _thread; 355 ContinuationWrapper& _cont; 356 bool _barriers; // only set when we allocate a chunk 357 const bool _preempt; // used only on the slow path 358 const intptr_t * const _frame_sp; // Top frame sp for this freeze 359 360 intptr_t* _bottom_address; 361 362 int _freeze_size; // total size of all frames plus metadata in words. 363 int _total_align_size; 364 365 intptr_t* _cont_stack_top; 366 intptr_t* _cont_stack_bottom; 367 368 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;) 369 370 #ifdef ASSERT 371 intptr_t* _orig_chunk_sp; 372 int _fast_freeze_size; 373 bool _empty; 374 #endif 375 376 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector; 377 378 NOT_PRODUCT(int _frames;) 379 DEBUG_ONLY(intptr_t* _last_write;) 380 381 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp); 382 383 public: 384 NOINLINE freeze_result freeze_slow(); 385 void freeze_fast_existing_chunk(); 386 387 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; }) 388 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; } 389 390 inline int size_if_fast_freeze_available(); 391 392 #ifdef ASSERT 393 bool check_valid_fast_path(); 394 #endif 395 396 protected: 397 inline void init_rest(); 398 void throw_stack_overflow_on_humongous_chunk(); 399 400 // fast path 401 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size); 402 inline void unwind_frames(); 403 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp); 404 405 // slow path 406 virtual stackChunkOop allocate_chunk_slow(size_t stack_size) = 0; 407 408 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); } 409 410 private: 411 // slow path 412 frame freeze_start_frame(); 413 frame freeze_start_frame_safepoint_stub(frame f); 414 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top); 415 inline frame freeze_start_frame_yield_stub(frame f); 416 template<typename FKind> 417 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize); 418 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame); 419 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame); 420 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize); 421 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame); 422 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted); 423 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted); 424 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller); 425 NOINLINE void finish_freeze(const frame& f, const frame& top); 426 427 inline bool stack_overflow(); 428 429 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f) 430 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); } 431 template<typename FKind> static inline frame sender(const frame& f); 432 template<typename FKind> frame new_heap_frame(frame& f, frame& caller); 433 inline void set_top_frame_metadata_pd(const frame& hf); 434 inline void patch_pd(frame& callee, const frame& caller); 435 void adjust_interpreted_frame_unextended_sp(frame& f); 436 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf); 437 438 protected: 439 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)); 440 bool freeze_fast_new_chunk(stackChunkOop chunk); 441 442 #ifdef ASSERT 443 bool is_empty(stackChunkOop chunk) { 444 // during freeze, the chunk is in an intermediate state (after setting the chunk's argsize but before setting its 445 // ultimate sp) so we use this instead of stackChunkOopDesc::is_empty 446 return chunk->sp() >= chunk->stack_size() - chunk->argsize() - frame::metadata_words_at_top; 447 } 448 #endif 449 }; 450 451 template <typename ConfigT> 452 class Freeze : public FreezeBase { 453 private: 454 stackChunkOop allocate_chunk(size_t stack_size); 455 456 public: 457 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp) 458 : FreezeBase(thread, cont, frame_sp) {} 459 460 freeze_result try_freeze_fast(); 461 462 protected: 463 virtual stackChunkOop allocate_chunk_slow(size_t stack_size) override { return allocate_chunk(stack_size); } 464 }; 465 466 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp) : 467 _thread(thread), _cont(cont), _barriers(false), _preempt(false), _frame_sp(frame_sp) { 468 DEBUG_ONLY(_jvmti_event_collector = nullptr;) 469 470 assert(_thread != nullptr, ""); 471 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), ""); 472 473 DEBUG_ONLY(_cont.entry()->verify_cookie();) 474 475 assert(!Interpreter::contains(_cont.entryPC()), ""); 476 477 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension(); 478 #ifdef _LP64 479 if (((intptr_t)_bottom_address & 0xf) != 0) { 480 _bottom_address--; 481 } 482 assert(is_aligned(_bottom_address, frame::frame_alignment), ""); 483 #endif 484 485 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT, 486 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord); 487 assert(_bottom_address != nullptr, ""); 488 assert(_bottom_address <= _cont.entrySP(), ""); 489 DEBUG_ONLY(_last_write = nullptr;) 490 491 assert(_cont.chunk_invariant(), ""); 492 assert(!Interpreter::contains(_cont.entryPC()), ""); 493 #if !defined(PPC64) || defined(ZERO) 494 static const int doYield_stub_frame_size = frame::metadata_words; 495 #else 496 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord; 497 #endif 498 assert(SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size, ""); 499 500 // properties of the continuation on the stack; all sizes are in words 501 _cont_stack_top = frame_sp + doYield_stub_frame_size; // we don't freeze the doYield stub frame 502 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0) 503 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw 504 505 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT, 506 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom)); 507 assert(cont_size() > 0, ""); 508 } 509 510 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling 511 _freeze_size = 0; 512 _total_align_size = 0; 513 NOT_PRODUCT(_frames = 0;) 514 } 515 516 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) { 517 stackChunkOop chunk = _cont.tail(); 518 chunk->copy_from_stack_to_chunk(from, to, size); 519 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);) 520 521 #ifdef ASSERT 522 if (_last_write != nullptr) { 523 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT 524 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size), 525 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address()); 526 _last_write = to; 527 } 528 #endif 529 } 530 531 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation) 532 void FreezeBase::unwind_frames() { 533 ContinuationEntry* entry = _cont.entry(); 534 entry->flush_stack_processing(_thread); 535 set_anchor_to_entry(_thread, entry); 536 } 537 538 template <typename ConfigT> 539 freeze_result Freeze<ConfigT>::try_freeze_fast() { 540 assert(_thread->thread_state() == _thread_in_vm, ""); 541 assert(_thread->cont_fastpath(), ""); 542 543 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();) 544 assert(_fast_freeze_size == 0, ""); 545 546 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words); 547 if (freeze_fast_new_chunk(chunk)) { 548 return freeze_ok; 549 } 550 if (_thread->has_pending_exception()) { 551 return freeze_exception; 552 } 553 554 // TODO R REMOVE when deopt change is fixed 555 assert(!_thread->cont_fastpath() || _barriers, ""); 556 log_develop_trace(continuations)("-- RETRYING SLOW --"); 557 return freeze_slow(); 558 } 559 560 // Returns size needed if the continuation fits, otherwise 0. 561 int FreezeBase::size_if_fast_freeze_available() { 562 stackChunkOop chunk = _cont.tail(); 563 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) { 564 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers"); 565 return 0; 566 } 567 568 int total_size_needed = cont_size(); 569 const int chunk_sp = chunk->sp(); 570 571 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk, 572 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty. 573 // Consider leaving the chunk's argsize set when emptying it and removing the following branch, 574 // although that would require changing stackChunkOopDesc::is_empty 575 if (chunk_sp < chunk->stack_size()) { 576 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top; 577 } 578 579 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom; 580 bool available = chunk_free_room >= total_size_needed; 581 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT, 582 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom)); 583 return available ? total_size_needed : 0; 584 } 585 586 void FreezeBase::freeze_fast_existing_chunk() { 587 stackChunkOop chunk = _cont.tail(); 588 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();) 589 590 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();) 591 assert(_fast_freeze_size > 0, ""); 592 593 if (chunk->sp() < chunk->stack_size()) { // we are copying into a non-empty chunk 594 DEBUG_ONLY(_empty = false;) 595 assert(chunk->sp() < (chunk->stack_size() - chunk->argsize()), ""); 596 #ifdef ASSERT 597 { 598 intptr_t* retaddr_slot = (chunk->sp_address() 599 - frame::sender_sp_ret_address_offset()); 600 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(), 601 "unexpected saved return address"); 602 } 603 #endif 604 605 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame 606 // we overlap; we'll overwrite the chunk's top frame's callee arguments 607 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top; 608 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack"); 609 610 // increase max_size by what we're freezing minus the overlap 611 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top); 612 613 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top; 614 assert(bottom_sp == _bottom_address, ""); 615 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame 616 // should have a return barrier (installed back when we thawed it). 617 #ifdef ASSERT 618 { 619 intptr_t* retaddr_slot = (bottom_sp 620 - frame::sender_sp_ret_address_offset()); 621 assert(ContinuationHelper::return_address_at(retaddr_slot) 622 == StubRoutines::cont_returnBarrier(), 623 "should be the continuation return barrier"); 624 } 625 #endif 626 // We copy the fp from the chunk back to the stack because it contains some caller data, 627 // including, possibly, an oop that might have gone stale since we thawed. 628 patch_stack_pd(bottom_sp, chunk->sp_address()); 629 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks 630 631 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false)); 632 } else { // the chunk is empty 633 DEBUG_ONLY(_empty = true;) 634 const int chunk_start_sp = chunk->sp(); 635 636 assert(chunk_start_sp == chunk->stack_size(), ""); 637 638 chunk->set_max_thawing_size(cont_size()); 639 chunk->set_argsize(_cont.argsize()); 640 641 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false)); 642 } 643 } 644 645 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) { 646 DEBUG_ONLY(_empty = true;) 647 648 // Install new chunk 649 _cont.set_tail(chunk); 650 651 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous 652 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers); 653 return false; 654 } 655 656 chunk->set_max_thawing_size(cont_size()); 657 chunk->set_argsize(_cont.argsize()); 658 659 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments. 660 // They'll then be stored twice: in the chunk and in the parent chunk's top frame 661 const int chunk_start_sp = cont_size() + frame::metadata_words; 662 assert(chunk_start_sp == chunk->stack_size(), ""); 663 664 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;) 665 666 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true)); 667 668 return true; 669 } 670 671 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) { 672 assert(chunk != nullptr, ""); 673 assert(!chunk->has_mixed_frames(), ""); 674 assert(!chunk->is_gc_mode(), ""); 675 assert(!chunk->has_bitmap(), ""); 676 assert(!chunk->requires_barriers(), ""); 677 assert(chunk == _cont.tail(), ""); 678 679 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before 680 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here 681 // will either see no continuation on the stack, or a consistent chunk. 682 unwind_frames(); 683 684 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d", 685 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize()); 686 assert(chunk_start_sp <= chunk->stack_size(), ""); 687 assert(chunk_start_sp >= cont_size(), "no room in the chunk"); 688 689 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze 690 assert(!(_fast_freeze_size > 0) || _orig_chunk_sp - (chunk->start_address() + chunk_new_sp) == _fast_freeze_size, ""); 691 692 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp; 693 #ifdef ASSERT 694 if (!_empty) { 695 intptr_t* retaddr_slot = (_orig_chunk_sp 696 - frame::sender_sp_ret_address_offset()); 697 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(), 698 "unexpected saved return address"); 699 } 700 #endif 701 702 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT, 703 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top)); 704 intptr_t* from = _cont_stack_top - frame::metadata_words_at_bottom; 705 intptr_t* to = chunk_top - frame::metadata_words_at_bottom; 706 copy_to_chunk(from, to, cont_size() + frame::metadata_words_at_bottom); 707 // Because we're not patched yet, the chunk is now in a bad state 708 709 // patch return pc of the bottom-most frozen frame (now in the chunk) 710 // with the actual caller's return address 711 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size() 712 - _cont.argsize() 713 - frame::metadata_words_at_top 714 - frame::sender_sp_ret_address_offset()); 715 #ifdef ASSERT 716 if (!_empty) { 717 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot) 718 == StubRoutines::cont_returnBarrier(), 719 "should be the continuation return barrier"); 720 } 721 #endif 722 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot, 723 chunk->pc()); 724 725 // We're always writing to a young chunk, so the GC can't see it until the next safepoint. 726 chunk->set_sp(chunk_new_sp); 727 // set chunk->pc to the return address of the topmost frame in the chunk 728 chunk->set_pc(ContinuationHelper::return_address_at( 729 _cont_stack_top - frame::sender_sp_ret_address_offset())); 730 731 _cont.write(); 732 733 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash()); 734 LogTarget(Trace, continuations) lt; 735 if (lt.develop_is_enabled()) { 736 LogStream ls(lt); 737 chunk->print_on(true, &ls); 738 } 739 740 // Verification 741 assert(_cont.chunk_invariant(), ""); 742 chunk->verify(); 743 744 #if CONT_JFR 745 EventContinuationFreezeFast e; 746 if (e.should_commit()) { 747 e.set_id(cast_from_oop<u8>(chunk)); 748 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);) 749 e.set_size(cont_size() << LogBytesPerWord); 750 e.commit(); 751 } 752 #endif 753 } 754 755 NOINLINE freeze_result FreezeBase::freeze_slow() { 756 #ifdef ASSERT 757 ResourceMark rm; 758 #endif 759 760 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash()); 761 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, ""); 762 763 #if CONT_JFR 764 EventContinuationFreezeSlow e; 765 if (e.should_commit()) { 766 e.set_id(cast_from_oop<u8>(_cont.continuation())); 767 e.commit(); 768 } 769 #endif 770 771 init_rest(); 772 773 HandleMark hm(Thread::current()); 774 775 frame f = freeze_start_frame(); 776 777 LogTarget(Debug, continuations) lt; 778 if (lt.develop_is_enabled()) { 779 LogStream ls(lt); 780 f.print_on(&ls); 781 } 782 783 frame caller; // the frozen caller in the chunk 784 freeze_result res = recurse_freeze(f, caller, 0, false, true); 785 786 if (res == freeze_ok) { 787 finish_freeze(f, caller); 788 _cont.write(); 789 } 790 791 return res; 792 } 793 794 frame FreezeBase::freeze_start_frame() { 795 frame f = _thread->last_frame(); 796 if (LIKELY(!_preempt)) { 797 return freeze_start_frame_yield_stub(f); 798 } else { 799 return freeze_start_frame_safepoint_stub(f); 800 } 801 } 802 803 frame FreezeBase::freeze_start_frame_yield_stub(frame f) { 804 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be"); 805 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); 806 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), ""); 807 return f; 808 } 809 810 frame FreezeBase::freeze_start_frame_safepoint_stub(frame f) { 811 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO) 812 f.set_fp(f.real_fp()); // f.set_fp(*Frame::callee_link_address(f)); // ???? 813 #else 814 Unimplemented(); 815 #endif 816 if (!Interpreter::contains(f.pc())) { 817 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be"); 818 assert(f.oop_map() != nullptr, "must be"); 819 820 if (Interpreter::contains(ContinuationHelper::StubFrame::return_pc(f))) { 821 f = sender<ContinuationHelper::StubFrame>(f); // Safepoint stub in interpreter 822 } 823 } 824 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), ""); 825 return f; 826 } 827 828 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap. 829 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) { 830 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame 831 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb())), ""); 832 833 if (stack_overflow()) { 834 return freeze_exception; 835 } 836 837 if (f.is_compiled_frame()) { 838 if (UNLIKELY(f.oop_map() == nullptr)) { 839 // special native frame 840 return freeze_pinned_native; 841 } 842 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted); 843 } else if (f.is_interpreted_frame()) { 844 assert((_preempt && top) || !f.interpreter_frame_method()->is_native(), ""); 845 if (_preempt && top && f.interpreter_frame_method()->is_native()) { 846 // int native entry 847 return freeze_pinned_native; 848 } 849 850 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted); 851 } else if (_preempt && top && ContinuationHelper::Frame::is_stub(f.cb())) { 852 return recurse_freeze_stub_frame(f, caller); 853 } else { 854 return freeze_pinned_native; 855 } 856 } 857 858 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap. 859 // See also StackChunkFrameStream<frame_kind>::frame_size() 860 template<typename FKind> 861 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) { 862 assert(FKind::is_instance(f), ""); 863 864 assert(fsize > 0, ""); 865 assert(argsize >= 0, ""); 866 _freeze_size += fsize; 867 NOT_PRODUCT(_frames++;) 868 869 assert(FKind::frame_bottom(f) <= _bottom_address, ""); 870 871 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between 872 // enterSpecial and an interpreted frame 873 if (FKind::frame_bottom(f) >= _bottom_address - 1) { 874 return finalize_freeze(f, caller, argsize); // recursion end 875 } else { 876 frame senderf = sender<FKind>(f); 877 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), ""); 878 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call 879 return result; 880 } 881 } 882 883 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) { 884 LogTarget(Trace, continuations) lt; 885 if (lt.develop_is_enabled()) { 886 LogStream ls(lt); 887 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame); 888 ls.print_cr("fsize: %d argsize: %d", fsize, argsize); 889 f.print_value_on(&ls, nullptr); 890 } 891 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), ""); 892 } 893 894 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) { 895 LogTarget(Trace, continuations) lt; 896 if (lt.develop_is_enabled()) { 897 LogStream ls(lt); 898 DEBUG_ONLY(hf.print_value_on(&ls, nullptr);) 899 assert(hf.is_heap_frame(), "should be"); 900 DEBUG_ONLY(print_frame_layout(hf, false, &ls);) 901 if (is_bottom_frame) { 902 ls.print_cr("bottom h-frame:"); 903 hf.print_on(&ls); 904 } 905 } 906 } 907 908 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap. 909 // See also StackChunkFrameStream<frame_kind>::frame_size() 910 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) { 911 int argsize = argsize_md - frame::metadata_words_at_top; 912 assert(callee.is_interpreted_frame() 913 || callee.cb()->as_nmethod()->is_osr_method() 914 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize()); 915 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d", 916 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize); 917 918 LogTarget(Trace, continuations) lt; 919 920 #ifdef ASSERT 921 bool empty = _cont.is_empty(); 922 log_develop_trace(continuations)("empty: %d", empty); 923 #endif 924 925 stackChunkOop chunk = _cont.tail(); 926 927 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), ""); 928 929 _freeze_size += frame::metadata_words; // for top frame's metadata 930 931 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind 932 int unextended_sp = -1; 933 if (chunk != nullptr) { 934 unextended_sp = chunk->sp(); 935 if (!chunk->is_empty()) { 936 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk); 937 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()); 938 bool top_interpreted = Interpreter::contains(chunk->pc()); 939 if (callee.is_interpreted_frame() == top_interpreted) { 940 overlap = argsize_md; 941 } 942 } 943 } 944 945 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp); 946 947 _freeze_size -= overlap; 948 assert(_freeze_size >= 0, ""); 949 950 assert(chunk == nullptr || chunk->is_empty() 951 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), ""); 952 assert(chunk != nullptr || unextended_sp < _freeze_size, ""); 953 954 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available 955 bool allocated_old_in_freeze_fast = _barriers; 956 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()), 957 "Chunk allocated in freeze_fast is of insufficient size " 958 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty()); 959 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation"); 960 961 DEBUG_ONLY(bool empty_chunk = true); 962 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) { 963 // ALLOCATE NEW CHUNK 964 965 if (lt.develop_is_enabled()) { 966 LogStream ls(lt); 967 if (chunk == nullptr) { 968 ls.print_cr("no chunk"); 969 } else { 970 ls.print_cr("chunk barriers: %d _size: %d free size: %d", 971 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words); 972 chunk->print_on(&ls); 973 } 974 } 975 976 _freeze_size += overlap; // we're allocating a new chunk, so no overlap 977 // overlap = 0; 978 979 chunk = allocate_chunk_slow(_freeze_size); 980 if (chunk == nullptr) { 981 return freeze_exception; 982 } 983 984 // Install new chunk 985 _cont.set_tail(chunk); 986 987 int sp = chunk->stack_size() - argsize_md; 988 chunk->set_sp(sp); 989 chunk->set_argsize(argsize); 990 assert(is_empty(chunk), ""); 991 } else { 992 // REUSE EXISTING CHUNK 993 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty()); 994 if (chunk->is_empty()) { 995 int sp = chunk->stack_size() - argsize_md; 996 chunk->set_sp(sp); 997 chunk->set_argsize(argsize); 998 _freeze_size += overlap; 999 assert(chunk->max_thawing_size() == 0, ""); 1000 } DEBUG_ONLY(else empty_chunk = false;) 1001 } 1002 assert(!chunk->is_gc_mode(), ""); 1003 assert(!chunk->has_bitmap(), ""); 1004 chunk->set_has_mixed_frames(true); 1005 1006 assert(chunk->requires_barriers() == _barriers, ""); 1007 assert(!_barriers || is_empty(chunk), ""); 1008 1009 assert(!is_empty(chunk) || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), ""); 1010 assert(!is_empty(chunk) || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), ""); 1011 1012 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before 1013 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here 1014 // will either see no continuation or a consistent chunk. 1015 unwind_frames(); 1016 1017 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - frame::metadata_words); 1018 1019 if (lt.develop_is_enabled()) { 1020 LogStream ls(lt); 1021 ls.print_cr("top chunk:"); 1022 chunk->print_on(&ls); 1023 } 1024 1025 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty 1026 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame(); 1027 1028 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);) 1029 1030 assert(chunk->is_in_chunk(_last_write - _freeze_size), 1031 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address())); 1032 #ifdef ASSERT 1033 if (lt.develop_is_enabled()) { 1034 LogStream ls(lt); 1035 ls.print_cr("top hframe before (freeze):"); 1036 assert(caller.is_heap_frame(), "should be"); 1037 caller.print_on(&ls); 1038 } 1039 1040 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), ""); 1041 1042 frame entry = sender(callee); 1043 1044 assert(Continuation::is_return_barrier_entry(entry.pc()) || Continuation::is_continuation_enterSpecial(entry), ""); 1045 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), ""); 1046 #endif 1047 1048 return freeze_ok_bottom; 1049 } 1050 1051 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) { 1052 if (is_bottom_frame) { 1053 // If we're the bottom frame, we need to replace the return barrier with the real 1054 // caller's pc. 1055 address last_pc = caller.pc(); 1056 assert((last_pc == nullptr) == is_empty(_cont.tail()), ""); 1057 ContinuationHelper::Frame::patch_pc(caller, last_pc); 1058 } else { 1059 assert(!caller.is_empty(), ""); 1060 } 1061 1062 patch_pd(hf, caller); 1063 1064 if (f.is_interpreted_frame()) { 1065 assert(hf.is_heap_frame(), "should be"); 1066 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller); 1067 } 1068 1069 #ifdef ASSERT 1070 if (hf.is_compiled_frame()) { 1071 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc 1072 log_develop_trace(continuations)("Freezing deoptimized frame"); 1073 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), ""); 1074 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), ""); 1075 } 1076 } 1077 #endif 1078 } 1079 1080 #ifdef ASSERT 1081 static void verify_frame_top(const frame& f, intptr_t* top) { 1082 ResourceMark rm; 1083 InterpreterOopMap mask; 1084 f.interpreted_frame_oop_map(&mask); 1085 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask), 1086 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT, 1087 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask))); 1088 } 1089 #endif // ASSERT 1090 1091 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap. 1092 // See also StackChunkFrameStream<frame_kind>::frame_size() 1093 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller, 1094 int callee_argsize /* incl. metadata */, 1095 bool callee_interpreted) { 1096 adjust_interpreted_frame_unextended_sp(f); 1097 1098 // The frame's top never includes the stack arguments to the callee 1099 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted); 1100 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f); 1101 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top); 1102 1103 DEBUG_ONLY(verify_frame_top(f, stack_frame_top)); 1104 1105 Method* frame_method = ContinuationHelper::Frame::frame_method(f); 1106 // including metadata between f and its args 1107 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top; 1108 1109 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d", 1110 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize); 1111 // we'd rather not yield inside methods annotated with @JvmtiMountTransition 1112 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), ""); 1113 1114 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize); 1115 if (UNLIKELY(result > freeze_ok_bottom)) { 1116 return result; 1117 } 1118 1119 bool is_bottom_frame = result == freeze_ok_bottom; 1120 assert(!caller.is_empty() || is_bottom_frame, ""); 1121 1122 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);) 1123 1124 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller); 1125 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64 1126 1127 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted); 1128 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf); 1129 assert(heap_frame_bottom == heap_frame_top + fsize, ""); 1130 1131 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned. 1132 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame. 1133 copy_to_chunk(stack_frame_top, heap_frame_top, fsize); 1134 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), ""); 1135 1136 relativize_interpreted_frame_metadata(f, hf); 1137 1138 patch(f, hf, caller, is_bottom_frame); 1139 1140 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();) 1141 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);) 1142 caller = hf; 1143 1144 // Mark frame_method's GC epoch for class redefinition on_stack calculation. 1145 frame_method->record_gc_epoch(); 1146 1147 return freeze_ok; 1148 } 1149 1150 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap. 1151 // See also StackChunkFrameStream<frame_kind>::frame_size() 1152 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller, 1153 int callee_argsize /* incl. metadata */, 1154 bool callee_interpreted) { 1155 // The frame's top never includes the stack arguments to the callee 1156 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted); 1157 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f); 1158 // including metadata between f and its stackargs 1159 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top; 1160 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top); 1161 1162 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d", 1163 ContinuationHelper::Frame::frame_method(f) != nullptr ? 1164 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "", 1165 _freeze_size, fsize, argsize); 1166 // we'd rather not yield inside methods annotated with @JvmtiMountTransition 1167 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), ""); 1168 1169 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize); 1170 if (UNLIKELY(result > freeze_ok_bottom)) { 1171 return result; 1172 } 1173 1174 bool is_bottom_frame = result == freeze_ok_bottom; 1175 assert(!caller.is_empty() || is_bottom_frame, ""); 1176 1177 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);) 1178 1179 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller); 1180 1181 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted); 1182 1183 copy_to_chunk(stack_frame_top, heap_frame_top, fsize); 1184 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), ""); 1185 1186 if (caller.is_interpreted_frame()) { 1187 _total_align_size += frame::align_wiggle; // See Thaw::align 1188 } 1189 1190 patch(f, hf, caller, is_bottom_frame); 1191 1192 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), ""); 1193 1194 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);) 1195 caller = hf; 1196 return freeze_ok; 1197 } 1198 1199 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) { 1200 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f, 0, 0); 1201 const int fsize = f.cb()->frame_size(); 1202 1203 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT, 1204 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize)); 1205 1206 // recurse_freeze_java_frame and freeze inlined here because we need to use a full RegisterMap for lock ownership 1207 NOT_PRODUCT(_frames++;) 1208 _freeze_size += fsize; 1209 1210 RegisterMap map(_cont.thread(), 1211 RegisterMap::UpdateMap::include, 1212 RegisterMap::ProcessFrames::skip, 1213 RegisterMap::WalkContinuation::skip); 1214 map.set_include_argument_oops(false); 1215 ContinuationHelper::update_register_map<ContinuationHelper::StubFrame>(f, &map); 1216 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case 1217 frame senderf = sender<ContinuationHelper::StubFrame>(f); 1218 assert(senderf.unextended_sp() < _bottom_address - 1, ""); 1219 assert(senderf.is_compiled_frame(), ""); 1220 1221 if (UNLIKELY(senderf.oop_map() == nullptr)) { 1222 // native frame 1223 return freeze_pinned_native; 1224 } 1225 1226 freeze_result result = recurse_freeze_compiled_frame(senderf, caller, 0, 0); // This might be deoptimized 1227 if (UNLIKELY(result > freeze_ok_bottom)) { 1228 return result; 1229 } 1230 assert(result != freeze_ok_bottom, ""); 1231 assert(!caller.is_interpreted_frame(), ""); 1232 1233 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false);) 1234 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller); 1235 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf, 0, 0); 1236 copy_to_chunk(stack_frame_top, heap_frame_top, fsize); 1237 DEBUG_ONLY(after_freeze_java_frame(hf, false);) 1238 1239 caller = hf; 1240 return freeze_ok; 1241 } 1242 1243 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) { 1244 stackChunkOop chunk = _cont.tail(); 1245 1246 LogTarget(Trace, continuations) lt; 1247 if (lt.develop_is_enabled()) { 1248 LogStream ls(lt); 1249 assert(top.is_heap_frame(), "should be"); 1250 top.print_on(&ls); 1251 } 1252 1253 set_top_frame_metadata_pd(top); 1254 1255 chunk->set_sp(chunk->to_offset(top.sp())); 1256 chunk->set_pc(top.pc()); 1257 1258 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size); 1259 1260 // At this point the chunk is consistent 1261 1262 if (UNLIKELY(_barriers)) { 1263 log_develop_trace(continuations)("do barriers on old chunk"); 1264 // Serial and Parallel GC can allocate objects directly into the old generation. 1265 // Then we want to relativize the derived pointers eagerly so that 1266 // old chunks are all in GC mode. 1267 assert(!UseG1GC, "G1 can not deal with allocating outside of eden"); 1268 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking"); 1269 if (UseShenandoahGC) { 1270 _cont.tail()->relativize_derived_pointers_concurrently(); 1271 } else { 1272 ContinuationGCSupport::transform_stack_chunk(_cont.tail()); 1273 } 1274 // For objects in the old generation we must maintain the remembered set 1275 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(); 1276 } 1277 1278 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames()); 1279 if (lt.develop_is_enabled()) { 1280 LogStream ls(lt); 1281 chunk->print_on(true, &ls); 1282 } 1283 1284 if (lt.develop_is_enabled()) { 1285 LogStream ls(lt); 1286 ls.print_cr("top hframe after (freeze):"); 1287 assert(_cont.last_frame().is_heap_frame(), "should be"); 1288 _cont.last_frame().print_on(&ls); 1289 } 1290 1291 assert(_cont.chunk_invariant(), ""); 1292 } 1293 1294 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code 1295 JavaThread* t = !_preempt ? _thread : JavaThread::current(); 1296 assert(t == JavaThread::current(), ""); 1297 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) { 1298 if (!_preempt) { 1299 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead 1300 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing"); 1301 } 1302 return true; 1303 } 1304 return false; 1305 } 1306 1307 class StackChunkAllocator : public MemAllocator { 1308 const size_t _stack_size; 1309 ContinuationWrapper& _continuation_wrapper; 1310 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector; 1311 mutable bool _took_slow_path; 1312 1313 // Does the minimal amount of initialization needed for a TLAB allocation. 1314 // We don't need to do a full initialization, as such an allocation need not be immediately walkable. 1315 virtual oop initialize(HeapWord* mem) const override { 1316 assert(_stack_size > 0, ""); 1317 assert(_stack_size <= max_jint, ""); 1318 assert(_word_size > _stack_size, ""); 1319 1320 // zero out fields (but not the stack) 1321 const size_t hs = oopDesc::header_size(); 1322 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs); 1323 1324 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size); 1325 jdk_internal_vm_StackChunk::set_sp(mem, (int)_stack_size); 1326 1327 return finish(mem); 1328 } 1329 1330 stackChunkOop allocate_fast() const { 1331 if (!UseTLAB) { 1332 return nullptr; 1333 } 1334 1335 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast(); 1336 if (mem == nullptr) { 1337 return nullptr; 1338 } 1339 1340 oop obj = initialize(mem); 1341 return stackChunkOopDesc::cast(obj); 1342 } 1343 1344 public: 1345 StackChunkAllocator(Klass* klass, 1346 size_t word_size, 1347 Thread* thread, 1348 size_t stack_size, 1349 ContinuationWrapper& continuation_wrapper, 1350 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector) 1351 : MemAllocator(klass, word_size, thread), 1352 _stack_size(stack_size), 1353 _continuation_wrapper(continuation_wrapper), 1354 _jvmti_event_collector(jvmti_event_collector), 1355 _took_slow_path(false) {} 1356 1357 // Provides it's own, specialized allocation which skips instrumentation 1358 // if the memory can be allocated without going to a slow-path. 1359 stackChunkOop allocate() const { 1360 // First try to allocate without any slow-paths or instrumentation. 1361 stackChunkOop obj = allocate_fast(); 1362 if (obj != nullptr) { 1363 return obj; 1364 } 1365 1366 // Now try full-blown allocation with all expensive operations, 1367 // including potentially safepoint operations. 1368 _took_slow_path = true; 1369 1370 // Protect unhandled Loom oops 1371 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper); 1372 1373 // Can safepoint 1374 _jvmti_event_collector->start(); 1375 1376 // Can safepoint 1377 return stackChunkOopDesc::cast(MemAllocator::allocate()); 1378 } 1379 1380 bool took_slow_path() const { 1381 return _took_slow_path; 1382 } 1383 }; 1384 1385 template <typename ConfigT> 1386 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size) { 1387 log_develop_trace(continuations)("allocate_chunk allocating new chunk"); 1388 1389 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass()); 1390 size_t size_in_words = klass->instance_size(stack_size); 1391 1392 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) { 1393 if (!_preempt) { 1394 throw_stack_overflow_on_humongous_chunk(); 1395 } 1396 return nullptr; 1397 } 1398 1399 JavaThread* current = _preempt ? JavaThread::current() : _thread; 1400 assert(current == JavaThread::current(), "should be current"); 1401 1402 // Allocate the chunk. 1403 // 1404 // This might safepoint while allocating, but all safepointing due to 1405 // instrumentation have been deferred. This property is important for 1406 // some GCs, as this ensures that the allocated object is in the young 1407 // generation / newly allocated memory. 1408 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, _cont, _jvmti_event_collector); 1409 stackChunkOop chunk = allocator.allocate(); 1410 1411 if (chunk == nullptr) { 1412 return nullptr; // OOME 1413 } 1414 1415 // assert that chunk is properly initialized 1416 assert(chunk->stack_size() == (int)stack_size, ""); 1417 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size); 1418 assert(chunk->sp() == chunk->stack_size(), ""); 1419 assert((intptr_t)chunk->start_address() % 8 == 0, ""); 1420 assert(chunk->max_thawing_size() == 0, ""); 1421 assert(chunk->pc() == nullptr, ""); 1422 assert(chunk->argsize() == 0, ""); 1423 assert(chunk->flags() == 0, ""); 1424 assert(chunk->is_gc_mode() == false, ""); 1425 1426 // fields are uninitialized 1427 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk()); 1428 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation()); 1429 1430 #if INCLUDE_ZGC 1431 if (UseZGC) { 1432 if (ZGenerational) { 1433 ZStackChunkGCData::initialize(chunk); 1434 } 1435 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation"); 1436 _barriers = false; 1437 } else 1438 #endif 1439 #if INCLUDE_SHENANDOAHGC 1440 if (UseShenandoahGC) { 1441 _barriers = chunk->requires_barriers(); 1442 } else 1443 #endif 1444 { 1445 if (!allocator.took_slow_path()) { 1446 // Guaranteed to be in young gen / newly allocated memory 1447 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation"); 1448 _barriers = false; 1449 } else { 1450 // Some GCs could put direct allocations in old gen for slow-path 1451 // allocations; need to explicitly check if that was the case. 1452 _barriers = chunk->requires_barriers(); 1453 } 1454 } 1455 1456 if (_barriers) { 1457 log_develop_trace(continuations)("allocation requires barriers"); 1458 } 1459 1460 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), ""); 1461 1462 return chunk; 1463 } 1464 1465 void FreezeBase::throw_stack_overflow_on_humongous_chunk() { 1466 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead 1467 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk"); 1468 } 1469 1470 #if INCLUDE_JVMTI 1471 static int num_java_frames(ContinuationWrapper& cont) { 1472 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address) 1473 int count = 0; 1474 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) { 1475 count += chunk->num_java_frames(); 1476 } 1477 return count; 1478 } 1479 1480 static void invalidate_jvmti_stack(JavaThread* thread) { 1481 if (thread->is_interp_only_mode()) { 1482 JvmtiThreadState *state = thread->jvmti_thread_state(); 1483 if (state != nullptr) 1484 state->invalidate_cur_stack_depth(); 1485 } 1486 } 1487 1488 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) { 1489 if (JvmtiExport::can_post_frame_pop()) { 1490 int num_frames = num_java_frames(cont); 1491 1492 ContinuationWrapper::SafepointOp so(Thread::current(), cont); 1493 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames); 1494 } 1495 invalidate_jvmti_stack(thread); 1496 } 1497 #endif // INCLUDE_JVMTI 1498 1499 #ifdef ASSERT 1500 static bool monitors_on_stack(JavaThread* thread) { 1501 ContinuationEntry* ce = thread->last_continuation(); 1502 RegisterMap map(thread, 1503 RegisterMap::UpdateMap::include, 1504 RegisterMap::ProcessFrames::include, 1505 RegisterMap::WalkContinuation::skip); 1506 map.set_include_argument_oops(false); 1507 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) { 1508 if ((f.is_interpreted_frame() && ContinuationHelper::InterpretedFrame::is_owning_locks(f)) || 1509 (f.is_compiled_frame() && ContinuationHelper::CompiledFrame::is_owning_locks(map.thread(), &map, f))) { 1510 return true; 1511 } 1512 } 1513 return false; 1514 } 1515 1516 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c 1517 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the 1518 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath. 1519 bool FreezeBase::check_valid_fast_path() { 1520 ContinuationEntry* ce = _thread->last_continuation(); 1521 RegisterMap map(_thread, 1522 RegisterMap::UpdateMap::skip, 1523 RegisterMap::ProcessFrames::skip, 1524 RegisterMap::WalkContinuation::skip); 1525 map.set_include_argument_oops(false); 1526 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) { 1527 if (!f.is_compiled_frame() || f.is_deoptimized_frame()) { 1528 return false; 1529 } 1530 } 1531 return true; 1532 } 1533 #endif // ASSERT 1534 1535 static inline int freeze_epilog(JavaThread* thread, ContinuationWrapper& cont) { 1536 verify_continuation(cont.continuation()); 1537 assert(!cont.is_empty(), ""); 1538 // This is done for the sake of the enterSpecial frame 1539 StackWatermarkSet::after_unwind(thread); 1540 1541 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash()); 1542 1543 return 0; 1544 } 1545 1546 static int freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) { 1547 if (UNLIKELY(res != freeze_ok)) { 1548 verify_continuation(cont.continuation()); 1549 log_develop_trace(continuations)("=== end of freeze (fail %d)", res); 1550 return res; 1551 } 1552 1553 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint 1554 return freeze_epilog(thread, cont); 1555 } 1556 1557 template<typename ConfigT> 1558 static inline int freeze_internal(JavaThread* current, intptr_t* const sp) { 1559 assert(!current->has_pending_exception(), ""); 1560 1561 #ifdef ASSERT 1562 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp())); 1563 log_frames(current); 1564 #endif 1565 1566 CONT_JFR_ONLY(EventContinuationFreeze event;) 1567 1568 ContinuationEntry* entry = current->last_continuation(); 1569 1570 oop oopCont = entry->cont_oop(current); 1571 assert(oopCont == current->last_continuation()->cont_oop(current), ""); 1572 assert(ContinuationEntry::assert_entry_frame_laid_out(current), ""); 1573 1574 verify_continuation(oopCont); 1575 ContinuationWrapper cont(current, oopCont); 1576 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont)); 1577 1578 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), ""); 1579 1580 assert(monitors_on_stack(current) == ((current->held_monitor_count() - current->jni_monitor_count()) > 0), 1581 "Held monitor count and locks on stack invariant: " INT64_FORMAT " JNI: " INT64_FORMAT, (int64_t)current->held_monitor_count(), (int64_t)current->jni_monitor_count()); 1582 1583 if (entry->is_pinned() || current->held_monitor_count() > 0) { 1584 log_develop_debug(continuations)("PINNED due to critical section/hold monitor"); 1585 verify_continuation(cont.continuation()); 1586 freeze_result res = entry->is_pinned() ? freeze_pinned_cs : freeze_pinned_monitor; 1587 log_develop_trace(continuations)("=== end of freeze (fail %d)", res); 1588 return res; 1589 } 1590 1591 Freeze<ConfigT> freeze(current, cont, sp); 1592 1593 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), ""); 1594 bool fast = UseContinuationFastPath && current->cont_fastpath(); 1595 if (fast && freeze.size_if_fast_freeze_available() > 0) { 1596 freeze.freeze_fast_existing_chunk(); 1597 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);) 1598 freeze_epilog(current, cont); 1599 return 0; 1600 } 1601 1602 log_develop_trace(continuations)("chunk unavailable; transitioning to VM"); 1603 assert(current == JavaThread::current(), "must be current thread except for preempt"); 1604 JRT_BLOCK 1605 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk 1606 JvmtiSampledObjectAllocEventCollector jsoaec(false); 1607 freeze.set_jvmti_event_collector(&jsoaec); 1608 1609 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow(); 1610 1611 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);) 1612 freeze_epilog(current, cont, res); 1613 cont.done(); // allow safepoint in the transition back to Java 1614 return res; 1615 JRT_BLOCK_END 1616 } 1617 1618 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) { 1619 ContinuationEntry* entry = thread->last_continuation(); 1620 if (entry == nullptr) { 1621 return freeze_ok; 1622 } 1623 if (entry->is_pinned()) { 1624 return freeze_pinned_cs; 1625 } else if (thread->held_monitor_count() > 0) { 1626 return freeze_pinned_monitor; 1627 } 1628 1629 RegisterMap map(thread, 1630 RegisterMap::UpdateMap::include, 1631 RegisterMap::ProcessFrames::skip, 1632 RegisterMap::WalkContinuation::skip); 1633 map.set_include_argument_oops(false); 1634 frame f = thread->last_frame(); 1635 1636 if (!safepoint) { 1637 f = f.sender(&map); // this is the yield frame 1638 } else { // safepoint yield 1639 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO) 1640 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp? 1641 #else 1642 Unimplemented(); 1643 #endif 1644 if (!Interpreter::contains(f.pc())) { 1645 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be"); 1646 assert(f.oop_map() != nullptr, "must be"); 1647 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case 1648 } 1649 } 1650 1651 while (true) { 1652 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) { 1653 return freeze_pinned_native; 1654 } 1655 1656 f = f.sender(&map); 1657 if (!Continuation::is_frame_in_continuation(entry, f)) { 1658 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread)); 1659 if (scope == cont_scope) { 1660 break; 1661 } 1662 intx monitor_count = entry->parent_held_monitor_count(); 1663 entry = entry->parent(); 1664 if (entry == nullptr) { 1665 break; 1666 } 1667 if (entry->is_pinned()) { 1668 return freeze_pinned_cs; 1669 } else if (monitor_count > 0) { 1670 return freeze_pinned_monitor; 1671 } 1672 } 1673 } 1674 return freeze_ok; 1675 } 1676 1677 /////////////// THAW //// 1678 1679 static int thaw_size(stackChunkOop chunk) { 1680 int size = chunk->max_thawing_size(); 1681 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast 1682 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom 1683 return size; 1684 } 1685 1686 // make room on the stack for thaw 1687 // returns the size in bytes, or 0 on failure 1688 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) { 1689 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier); 1690 1691 assert(thread == JavaThread::current(), ""); 1692 1693 ContinuationEntry* ce = thread->last_continuation(); 1694 assert(ce != nullptr, ""); 1695 oop continuation = ce->cont_oop(thread); 1696 assert(continuation == get_continuation(thread), ""); 1697 verify_continuation(continuation); 1698 1699 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation); 1700 assert(chunk != nullptr, ""); 1701 1702 // The tail can be empty because it might still be available for another freeze. 1703 // However, here we want to thaw, so we get rid of it (it will be GCed). 1704 if (UNLIKELY(chunk->is_empty())) { 1705 chunk = chunk->parent(); 1706 assert(chunk != nullptr, ""); 1707 assert(!chunk->is_empty(), ""); 1708 jdk_internal_vm_Continuation::set_tail(continuation, chunk); 1709 } 1710 1711 // Verification 1712 chunk->verify(); 1713 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty"); 1714 1715 // Only make space for the last chunk because we only thaw from the last chunk 1716 int size = thaw_size(chunk) << LogBytesPerWord; 1717 1718 const address bottom = (address)thread->last_continuation()->entry_sp(); 1719 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages 1720 // for the Java frames in the check below. 1721 if (!stack_overflow_check(thread, size + 300, bottom)) { 1722 return 0; 1723 } 1724 1725 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d", 1726 p2i(bottom), p2i(bottom - size), size); 1727 return size; 1728 } 1729 1730 class ThawBase : public StackObj { 1731 protected: 1732 JavaThread* _thread; 1733 ContinuationWrapper& _cont; 1734 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;) 1735 1736 intptr_t* _fastpath; 1737 bool _barriers; 1738 intptr_t* _top_unextended_sp_before_thaw; 1739 int _align_size; 1740 DEBUG_ONLY(intptr_t* _top_stack_address); 1741 1742 StackChunkFrameStream<ChunkFrames::Mixed> _stream; 1743 1744 NOT_PRODUCT(int _frames;) 1745 1746 protected: 1747 ThawBase(JavaThread* thread, ContinuationWrapper& cont) : 1748 _thread(thread), _cont(cont), 1749 _fastpath(nullptr) { 1750 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;) 1751 assert (cont.tail() != nullptr, "no last chunk"); 1752 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());) 1753 } 1754 1755 void clear_chunk(stackChunkOop chunk); 1756 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize); 1757 void copy_from_chunk(intptr_t* from, intptr_t* to, int size); 1758 1759 // fast path 1760 inline void prefetch_chunk_pd(void* start, int size_words); 1761 void patch_return(intptr_t* sp, bool is_last); 1762 1763 // slow path 1764 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, bool return_barrier); 1765 1766 private: 1767 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top); 1768 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames); 1769 void finalize_thaw(frame& entry, int argsize); 1770 1771 inline bool seen_by_gc(); 1772 1773 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame); 1774 inline void after_thaw_java_frame(const frame& f, bool bottom); 1775 inline void patch(frame& f, const frame& caller, bool bottom); 1776 void clear_bitmap_bits(address start, address end); 1777 1778 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames); 1779 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller); 1780 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames); 1781 void finish_thaw(frame& f); 1782 1783 void push_return_frame(frame& f); 1784 inline frame new_entry_frame(); 1785 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom); 1786 inline void patch_pd(frame& f, const frame& sender); 1787 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom); 1788 1789 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; } 1790 1791 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f); 1792 1793 public: 1794 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; }) 1795 }; 1796 1797 template <typename ConfigT> 1798 class Thaw : public ThawBase { 1799 public: 1800 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {} 1801 1802 inline bool can_thaw_fast(stackChunkOop chunk) { 1803 return !_barriers 1804 && _thread->cont_fastpath_thread_state() 1805 && !chunk->has_thaw_slowpath_condition() 1806 && !PreserveFramePointer; 1807 } 1808 1809 inline intptr_t* thaw(Continuation::thaw_kind kind); 1810 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk); 1811 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom); 1812 }; 1813 1814 template <typename ConfigT> 1815 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) { 1816 verify_continuation(_cont.continuation()); 1817 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), ""); 1818 assert(!_cont.is_empty(), ""); 1819 1820 stackChunkOop chunk = _cont.tail(); 1821 assert(chunk != nullptr, "guaranteed by prepare_thaw"); 1822 assert(!chunk->is_empty(), "guaranteed by prepare_thaw"); 1823 1824 _barriers = chunk->requires_barriers(); 1825 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk) 1826 : thaw_slow(chunk, kind != Continuation::thaw_top); 1827 } 1828 1829 class ReconstructedStack : public StackObj { 1830 intptr_t* _base; // _cont.entrySP(); // top of the entry frame 1831 int _thaw_size; 1832 int _argsize; 1833 public: 1834 ReconstructedStack(intptr_t* base, int thaw_size, int argsize) 1835 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) { 1836 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned. 1837 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), ""); 1838 // We're at most one alignment word away from entrySP 1839 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame"); 1840 } 1841 1842 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); } 1843 1844 // top and bottom stack pointers 1845 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); } 1846 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); } 1847 1848 // several operations operate on the totality of the stack being reconstructed, 1849 // including the metadata words 1850 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; } 1851 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; } 1852 }; 1853 1854 inline void ThawBase::clear_chunk(stackChunkOop chunk) { 1855 chunk->set_sp(chunk->stack_size()); 1856 chunk->set_argsize(0); 1857 chunk->set_max_thawing_size(0); 1858 } 1859 1860 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) { 1861 bool empty = false; 1862 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk); 1863 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();) 1864 assert(chunk_sp == f.sp(), ""); 1865 assert(chunk_sp == f.unextended_sp(), ""); 1866 1867 const int frame_size = f.cb()->frame_size(); 1868 argsize = f.stack_argsize(); 1869 1870 f.next(SmallRegisterMap::instance, true /* stop */); 1871 empty = f.is_done(); 1872 assert(!empty || argsize == chunk->argsize(), ""); 1873 1874 if (empty) { 1875 clear_chunk(chunk); 1876 } else { 1877 chunk->set_sp(chunk->sp() + frame_size); 1878 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size); 1879 // We set chunk->pc to the return pc into the next frame 1880 chunk->set_pc(f.pc()); 1881 #ifdef ASSERT 1882 { 1883 intptr_t* retaddr_slot = (chunk_sp 1884 + frame_size 1885 - frame::sender_sp_ret_address_offset()); 1886 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot), 1887 "unexpected pc"); 1888 } 1889 #endif 1890 } 1891 assert(empty == chunk->is_empty(), ""); 1892 // returns the size required to store the frame on stack, and because it is a 1893 // compiled frame, it must include a copy of the arguments passed by the caller 1894 return frame_size + argsize + frame::metadata_words_at_top; 1895 } 1896 1897 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) { 1898 assert(to >= _top_stack_address, "overwrote past thawing space" 1899 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address)); 1900 assert(to + size <= _cont.entrySP(), "overwrote past thawing space"); 1901 _cont.tail()->copy_from_chunk_to_stack(from, to, size); 1902 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);) 1903 } 1904 1905 void ThawBase::patch_return(intptr_t* sp, bool is_last) { 1906 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp)); 1907 1908 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC(); 1909 ContinuationHelper::patch_return_address_at( 1910 sp - frame::sender_sp_ret_address_offset(), 1911 pc); 1912 } 1913 1914 template <typename ConfigT> 1915 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) { 1916 assert(chunk == _cont.tail(), ""); 1917 assert(!chunk->has_mixed_frames(), ""); 1918 assert(!chunk->requires_barriers(), ""); 1919 assert(!chunk->has_bitmap(), ""); 1920 assert(!_thread->is_interp_only_mode(), ""); 1921 1922 LogTarget(Trace, continuations) lt; 1923 if (lt.develop_is_enabled()) { 1924 LogStream ls(lt); 1925 ls.print_cr("thaw_fast"); 1926 chunk->print_on(true, &ls); 1927 } 1928 1929 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame. 1930 static const int threshold = 500; // words 1931 1932 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw 1933 int argsize, thaw_size; 1934 1935 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp(); 1936 1937 bool partial, empty; 1938 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) { 1939 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address 1940 1941 partial = false; 1942 argsize = chunk->argsize(); // must be called *before* clearing the chunk 1943 clear_chunk(chunk); 1944 thaw_size = full_chunk_size; 1945 empty = true; 1946 } else { // thaw a single frame 1947 partial = true; 1948 thaw_size = remove_top_compiled_frame_from_chunk(chunk, argsize); 1949 empty = chunk->is_empty(); 1950 } 1951 1952 // Are we thawing the last frame(s) in the continuation 1953 const bool is_last = empty && chunk->parent() == nullptr; 1954 assert(!is_last || argsize == 0, ""); 1955 1956 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT, 1957 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP())); 1958 1959 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize); 1960 1961 // also copy metadata words at frame bottom 1962 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size()); 1963 1964 // update the ContinuationEntry 1965 _cont.set_argsize(argsize); 1966 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize()); 1967 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), ""); 1968 1969 // install the return barrier if not last frame, or the entry's pc if last 1970 patch_return(rs.bottom_sp(), is_last); 1971 1972 // insert the back links from callee to caller frames 1973 patch_caller_links(rs.top(), rs.top() + rs.total_size()); 1974 1975 assert(is_last == _cont.is_empty(), ""); 1976 assert(_cont.chunk_invariant(), ""); 1977 1978 #if CONT_JFR 1979 EventContinuationThawFast e; 1980 if (e.should_commit()) { 1981 e.set_id(cast_from_oop<u8>(chunk)); 1982 e.set_size(thaw_size << LogBytesPerWord); 1983 e.set_full(!partial); 1984 e.commit(); 1985 } 1986 #endif 1987 1988 #ifdef ASSERT 1989 set_anchor(_thread, rs.sp()); 1990 log_frames(_thread); 1991 if (LoomDeoptAfterThaw) { 1992 do_deopt_after_thaw(_thread); 1993 } 1994 clear_anchor(_thread); 1995 #endif 1996 1997 return rs.sp(); 1998 } 1999 2000 inline bool ThawBase::seen_by_gc() { 2001 return _barriers || _cont.tail()->is_gc_mode(); 2002 } 2003 2004 NOINLINE intptr_t* ThawBase::thaw_slow(stackChunkOop chunk, bool return_barrier) { 2005 LogTarget(Trace, continuations) lt; 2006 if (lt.develop_is_enabled()) { 2007 LogStream ls(lt); 2008 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, return_barrier, p2i(chunk)); 2009 chunk->print_on(true, &ls); 2010 } 2011 2012 #if CONT_JFR 2013 EventContinuationThawSlow e; 2014 if (e.should_commit()) { 2015 e.set_id(cast_from_oop<u8>(_cont.continuation())); 2016 e.commit(); 2017 } 2018 #endif 2019 2020 DEBUG_ONLY(_frames = 0;) 2021 _align_size = 0; 2022 int num_frames = (return_barrier ? 1 : 2); 2023 2024 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk); 2025 _top_unextended_sp_before_thaw = _stream.unextended_sp(); 2026 2027 frame heap_frame = _stream.to_frame(); 2028 if (lt.develop_is_enabled()) { 2029 LogStream ls(lt); 2030 ls.print_cr("top hframe before (thaw):"); 2031 assert(heap_frame.is_heap_frame(), "should have created a relative frame"); 2032 heap_frame.print_value_on(&ls, nullptr); 2033 } 2034 2035 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC 2036 if (UseZGC || UseShenandoahGC) { 2037 _cont.tail()->relativize_derived_pointers_concurrently(); 2038 } 2039 #endif 2040 2041 frame caller; // the thawed caller on the stack 2042 recurse_thaw(heap_frame, caller, num_frames, true); 2043 finish_thaw(caller); // caller is now the topmost thawed frame 2044 _cont.write(); 2045 2046 assert(_cont.chunk_invariant(), ""); 2047 2048 JVMTI_ONLY(if (!return_barrier) invalidate_jvmti_stack(_thread)); 2049 2050 _thread->set_cont_fastpath(_fastpath); 2051 2052 intptr_t* sp = caller.sp(); 2053 return sp; 2054 } 2055 2056 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top) { 2057 log_develop_debug(continuations)("thaw num_frames: %d", num_frames); 2058 assert(!_cont.is_empty(), "no more frames"); 2059 assert(num_frames > 0, ""); 2060 assert(!heap_frame.is_empty(), ""); 2061 2062 if (top && heap_frame.is_safepoint_blob_frame()) { 2063 assert(ContinuationHelper::Frame::is_stub(heap_frame.cb()), "cb: %s", heap_frame.cb()->name()); 2064 recurse_thaw_stub_frame(heap_frame, caller, num_frames); 2065 } else if (!heap_frame.is_interpreted_frame()) { 2066 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false); 2067 } else { 2068 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames); 2069 } 2070 } 2071 2072 template<typename FKind> 2073 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) { 2074 assert(num_frames > 0, ""); 2075 2076 DEBUG_ONLY(_frames++;) 2077 2078 int argsize = _stream.stack_argsize(); 2079 2080 _stream.next(SmallRegisterMap::instance); 2081 assert(_stream.to_frame().is_empty() == _stream.is_done(), ""); 2082 2083 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk 2084 // as it makes detecting that situation and adjusting unextended_sp tricky 2085 if (num_frames == 1 && !_stream.is_done() && FKind::interpreted && _stream.is_compiled()) { 2086 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top"); 2087 num_frames++; 2088 } 2089 2090 if (num_frames == 1 || _stream.is_done()) { // end recursion 2091 finalize_thaw(caller, FKind::interpreted ? 0 : argsize); 2092 return true; // bottom 2093 } else { // recurse 2094 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false); 2095 return false; 2096 } 2097 } 2098 2099 void ThawBase::finalize_thaw(frame& entry, int argsize) { 2100 stackChunkOop chunk = _cont.tail(); 2101 2102 if (!_stream.is_done()) { 2103 assert(_stream.sp() >= chunk->sp_address(), ""); 2104 chunk->set_sp(chunk->to_offset(_stream.sp())); 2105 chunk->set_pc(_stream.pc()); 2106 } else { 2107 chunk->set_argsize(0); 2108 chunk->set_sp(chunk->stack_size()); 2109 chunk->set_pc(nullptr); 2110 } 2111 assert(_stream.is_done() == chunk->is_empty(), ""); 2112 2113 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw); 2114 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed); 2115 2116 _cont.set_argsize(argsize); 2117 entry = new_entry_frame(); 2118 2119 assert(entry.sp() == _cont.entrySP(), ""); 2120 assert(Continuation::is_continuation_enterSpecial(entry), ""); 2121 assert(_cont.is_entry_frame(entry), ""); 2122 } 2123 2124 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) { 2125 LogTarget(Trace, continuations) lt; 2126 if (lt.develop_is_enabled()) { 2127 LogStream ls(lt); 2128 ls.print_cr("======== THAWING FRAME: %d", num_frame); 2129 assert(hf.is_heap_frame(), "should be"); 2130 hf.print_value_on(&ls, nullptr); 2131 } 2132 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf)); 2133 } 2134 2135 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) { 2136 #ifdef ASSERT 2137 LogTarget(Trace, continuations) lt; 2138 if (lt.develop_is_enabled()) { 2139 LogStream ls(lt); 2140 ls.print_cr("thawed frame:"); 2141 print_frame_layout(f, false, &ls); // f.print_on(&ls); 2142 } 2143 #endif 2144 } 2145 2146 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) { 2147 assert(!bottom || caller.fp() == _cont.entryFP(), ""); 2148 if (bottom) { 2149 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc() 2150 : StubRoutines::cont_returnBarrier()); 2151 } else { 2152 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap. 2153 // If the caller is not deoptimized, pc is unchanged. 2154 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc()); 2155 } 2156 2157 patch_pd(f, caller); 2158 2159 if (f.is_interpreted_frame()) { 2160 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller); 2161 } 2162 2163 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), ""); 2164 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), ""); 2165 } 2166 2167 void ThawBase::clear_bitmap_bits(address start, address end) { 2168 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start)); 2169 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end)); 2170 2171 // we need to clear the bits that correspond to arguments as they reside in the caller frame 2172 // or they will keep objects that are otherwise unreachable alive. 2173 2174 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since 2175 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned. 2176 // If that's the case the bit range corresponding to the last stack slot should not have bits set 2177 // anyways and we assert that before returning. 2178 address effective_end = UseCompressedOops ? end : align_down(end, wordSize); 2179 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end)); 2180 stackChunkOop chunk = _cont.tail(); 2181 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end)); 2182 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set"); 2183 } 2184 2185 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames) { 2186 assert(hf.is_interpreted_frame(), ""); 2187 2188 if (UNLIKELY(seen_by_gc())) { 2189 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance); 2190 } 2191 2192 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames); 2193 2194 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);) 2195 2196 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64 2197 2198 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame); 2199 2200 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top; 2201 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f); 2202 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top; 2203 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf); 2204 2205 assert(hf.is_heap_frame(), "should be"); 2206 assert(!f.is_heap_frame(), "should not be"); 2207 2208 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top); 2209 assert((stack_frame_bottom == stack_frame_top + fsize), ""); 2210 2211 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned. 2212 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame. 2213 copy_from_chunk(heap_frame_top, stack_frame_top, fsize); 2214 2215 // Make sure the relativized locals is already set. 2216 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom"); 2217 2218 derelativize_interpreted_frame_metadata(hf, f); 2219 patch(f, caller, is_bottom_frame); 2220 2221 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame"); 2222 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), ""); 2223 2224 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();) 2225 2226 maybe_set_fastpath(f.sp()); 2227 2228 const int locals = hf.interpreter_frame_method()->max_locals(); 2229 2230 if (!is_bottom_frame) { 2231 // can only fix caller once this frame is thawed (due to callee saved regs) 2232 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance); 2233 } else if (_cont.tail()->has_bitmap() && locals > 0) { 2234 assert(hf.is_heap_frame(), "should be"); 2235 address start = (address)(heap_frame_bottom - locals); 2236 address end = (address)heap_frame_bottom; 2237 clear_bitmap_bits(start, end); 2238 } 2239 2240 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);) 2241 caller = f; 2242 } 2243 2244 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) { 2245 assert(!hf.is_interpreted_frame(), ""); 2246 assert(_cont.is_preempted() || !stub_caller, "stub caller not at preemption"); 2247 2248 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap 2249 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance); 2250 } 2251 2252 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames); 2253 2254 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);) 2255 2256 assert(caller.sp() == caller.unextended_sp(), ""); 2257 2258 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) { 2259 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in freeze_interpreted_frame 2260 } 2261 2262 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not 2263 // yet laid out in the stack, and so the original_pc is not stored in it. 2264 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized. 2265 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame); 2266 intptr_t* const stack_frame_top = f.sp(); 2267 intptr_t* const heap_frame_top = hf.unextended_sp(); 2268 2269 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0; 2270 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize; 2271 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), ""); 2272 2273 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom; 2274 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom; 2275 // copy metadata, except the metadata at the top of the (unextended) entry frame 2276 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top); 2277 2278 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP 2279 // (we might have one padding word for alignment) 2280 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), ""); 2281 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), ""); 2282 2283 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above 2284 2285 patch(f, caller, is_bottom_frame); 2286 2287 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above) 2288 assert(!f.is_deoptimized_frame(), ""); 2289 if (hf.is_deoptimized_frame()) { 2290 maybe_set_fastpath(f.sp()); 2291 } else if (_thread->is_interp_only_mode() 2292 || (_cont.is_preempted() && f.cb()->as_nmethod()->is_marked_for_deoptimization())) { 2293 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so 2294 // cannot rely on nmethod patching for deopt. 2295 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller"); 2296 2297 log_develop_trace(continuations)("Deoptimizing thawed frame"); 2298 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr)); 2299 2300 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for 2301 assert(f.is_deoptimized_frame(), ""); 2302 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), ""); 2303 maybe_set_fastpath(f.sp()); 2304 } 2305 2306 if (!is_bottom_frame) { 2307 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack 2308 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance); 2309 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) { 2310 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top); 2311 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */); 2312 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size; 2313 clear_bitmap_bits(start, start + argsize_in_bytes); 2314 } 2315 2316 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);) 2317 caller = f; 2318 } 2319 2320 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) { 2321 DEBUG_ONLY(_frames++;) 2322 2323 { 2324 RegisterMap map(nullptr, 2325 RegisterMap::UpdateMap::include, 2326 RegisterMap::ProcessFrames::skip, 2327 RegisterMap::WalkContinuation::skip); 2328 map.set_include_argument_oops(false); 2329 _stream.next(&map); 2330 assert(!_stream.is_done(), ""); 2331 if (UNLIKELY(seen_by_gc())) { // we're now doing this on the stub's caller 2332 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map); 2333 } 2334 assert(!_stream.is_done(), ""); 2335 } 2336 2337 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true); // this could be deoptimized 2338 2339 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false, num_frames);) 2340 2341 assert(ContinuationHelper::Frame::is_stub(hf.cb()), ""); 2342 assert(caller.sp() == caller.unextended_sp(), ""); 2343 assert(!caller.is_interpreted_frame(), ""); 2344 2345 int fsize = ContinuationHelper::StubFrame::size(hf); 2346 2347 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false); 2348 intptr_t* stack_frame_top = f.sp(); 2349 intptr_t* heap_frame_top = hf.sp(); 2350 2351 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words, 2352 fsize + frame::metadata_words); 2353 2354 { // can only fix caller once this frame is thawed (due to callee saved regs) 2355 RegisterMap map(nullptr, 2356 RegisterMap::UpdateMap::include, 2357 RegisterMap::ProcessFrames::skip, 2358 RegisterMap::WalkContinuation::skip); // map.clear(); 2359 map.set_include_argument_oops(false); 2360 f.oop_map()->update_register_map(&f, &map); 2361 ContinuationHelper::update_register_map_with_callee(caller, &map); 2362 _cont.tail()->fix_thawed_frame(caller, &map); 2363 } 2364 2365 DEBUG_ONLY(after_thaw_java_frame(f, false);) 2366 caller = f; 2367 } 2368 2369 void ThawBase::finish_thaw(frame& f) { 2370 stackChunkOop chunk = _cont.tail(); 2371 2372 if (chunk->is_empty()) { 2373 // Only remove chunk from list if it can't be reused for another freeze 2374 if (seen_by_gc()) { 2375 _cont.set_tail(chunk->parent()); 2376 } else { 2377 chunk->set_has_mixed_frames(false); 2378 } 2379 chunk->set_max_thawing_size(0); 2380 assert(chunk->argsize() == 0, ""); 2381 } else { 2382 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size); 2383 } 2384 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), ""); 2385 2386 if (!is_aligned(f.sp(), frame::frame_alignment)) { 2387 assert(f.is_interpreted_frame(), ""); 2388 f.set_sp(align_down(f.sp(), frame::frame_alignment)); 2389 } 2390 push_return_frame(f); 2391 chunk->fix_thawed_frame(f, SmallRegisterMap::instance); // can only fix caller after push_return_frame (due to callee saved regs) 2392 2393 assert(_cont.is_empty() == _cont.last_frame().is_empty(), ""); 2394 2395 log_develop_trace(continuations)("thawed %d frames", _frames); 2396 2397 LogTarget(Trace, continuations) lt; 2398 if (lt.develop_is_enabled()) { 2399 LogStream ls(lt); 2400 ls.print_cr("top hframe after (thaw):"); 2401 _cont.last_frame().print_value_on(&ls, nullptr); 2402 } 2403 } 2404 2405 void ThawBase::push_return_frame(frame& f) { // see generate_cont_thaw 2406 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), ""); 2407 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), ""); 2408 2409 LogTarget(Trace, continuations) lt; 2410 if (lt.develop_is_enabled()) { 2411 LogStream ls(lt); 2412 ls.print_cr("push_return_frame"); 2413 f.print_value_on(&ls, nullptr); 2414 } 2415 2416 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space" 2417 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address)); 2418 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked. 2419 ContinuationHelper::push_pd(f); 2420 2421 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), ""); 2422 } 2423 2424 // returns new top sp 2425 // called after preparations (stack overflow check and making room) 2426 template<typename ConfigT> 2427 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) { 2428 assert(thread == JavaThread::current(), "Must be current thread"); 2429 2430 CONT_JFR_ONLY(EventContinuationThaw event;) 2431 2432 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp())); 2433 2434 ContinuationEntry* entry = thread->last_continuation(); 2435 assert(entry != nullptr, ""); 2436 oop oopCont = entry->cont_oop(thread); 2437 2438 assert(!jdk_internal_vm_Continuation::done(oopCont), ""); 2439 assert(oopCont == get_continuation(thread), ""); 2440 verify_continuation(oopCont); 2441 2442 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), ""); 2443 2444 ContinuationWrapper cont(thread, oopCont); 2445 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont)); 2446 2447 #ifdef ASSERT 2448 set_anchor_to_entry(thread, cont.entry()); 2449 log_frames(thread); 2450 clear_anchor(thread); 2451 #endif 2452 2453 Thaw<ConfigT> thw(thread, cont); 2454 intptr_t* const sp = thw.thaw(kind); 2455 assert(is_aligned(sp, frame::frame_alignment), ""); 2456 2457 // All the frames have been thawed so we know they don't hold any monitors 2458 assert(thread->held_monitor_count() == 0, "Must be"); 2459 2460 #ifdef ASSERT 2461 intptr_t* sp0 = sp; 2462 set_anchor(thread, sp0); 2463 log_frames(thread); 2464 if (LoomVerifyAfterThaw) { 2465 assert(do_verify_after_thaw(thread, cont.tail(), tty), ""); 2466 } 2467 assert(ContinuationEntry::assert_entry_frame_laid_out(thread), ""); 2468 clear_anchor(thread); 2469 2470 LogTarget(Trace, continuations) lt; 2471 if (lt.develop_is_enabled()) { 2472 LogStream ls(lt); 2473 ls.print_cr("Jumping to frame (thaw):"); 2474 frame(sp).print_value_on(&ls, nullptr); 2475 } 2476 #endif 2477 2478 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);) 2479 2480 verify_continuation(cont.continuation()); 2481 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash()); 2482 2483 return sp; 2484 } 2485 2486 #ifdef ASSERT 2487 static void do_deopt_after_thaw(JavaThread* thread) { 2488 int i = 0; 2489 StackFrameStream fst(thread, true, false); 2490 fst.register_map()->set_include_argument_oops(false); 2491 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map()); 2492 for (; !fst.is_done(); fst.next()) { 2493 if (fst.current()->cb()->is_nmethod()) { 2494 nmethod* nm = fst.current()->cb()->as_nmethod(); 2495 if (!nm->method()->is_continuation_native_intrinsic()) { 2496 nm->make_deoptimized(); 2497 } 2498 } 2499 } 2500 } 2501 2502 class ThawVerifyOopsClosure: public OopClosure { 2503 intptr_t* _p; 2504 outputStream* _st; 2505 bool is_good_oop(oop o) { 2506 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass(); 2507 } 2508 public: 2509 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {} 2510 intptr_t* p() { return _p; } 2511 void reset() { _p = nullptr; } 2512 2513 virtual void do_oop(oop* p) { 2514 oop o = *p; 2515 if (o == nullptr || is_good_oop(o)) { 2516 return; 2517 } 2518 _p = (intptr_t*)p; 2519 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p)); 2520 } 2521 virtual void do_oop(narrowOop* p) { 2522 oop o = RawAccess<>::oop_load(p); 2523 if (o == nullptr || is_good_oop(o)) { 2524 return; 2525 } 2526 _p = (intptr_t*)p; 2527 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p)); 2528 } 2529 }; 2530 2531 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) { 2532 assert(thread->has_last_Java_frame(), ""); 2533 2534 ResourceMark rm; 2535 ThawVerifyOopsClosure cl(st); 2536 NMethodToOopClosure cf(&cl, false); 2537 2538 StackFrameStream fst(thread, true, false); 2539 fst.register_map()->set_include_argument_oops(false); 2540 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map()); 2541 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) { 2542 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) { 2543 st->print_cr(">>> do_verify_after_thaw deopt"); 2544 fst.current()->deoptimize(nullptr); 2545 fst.current()->print_on(st); 2546 } 2547 2548 fst.current()->oops_do(&cl, &cf, fst.register_map()); 2549 if (cl.p() != nullptr) { 2550 frame fr = *fst.current(); 2551 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers()); 2552 fr.print_on(st); 2553 if (!fr.is_interpreted_frame()) { 2554 st->print_cr("size: %d argsize: %d", 2555 ContinuationHelper::NonInterpretedUnknownFrame::size(fr), 2556 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr)); 2557 } 2558 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp()); 2559 if (reg != nullptr) { 2560 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99); 2561 } 2562 cl.reset(); 2563 DEBUG_ONLY(thread->print_frame_layout();) 2564 if (chunk != nullptr) { 2565 chunk->print_on(true, st); 2566 } 2567 return false; 2568 } 2569 } 2570 return true; 2571 } 2572 2573 static void log_frames(JavaThread* thread) { 2574 const static int show_entry_callers = 3; 2575 LogTarget(Trace, continuations) lt; 2576 if (!lt.develop_is_enabled()) { 2577 return; 2578 } 2579 LogStream ls(lt); 2580 2581 ls.print_cr("------- frames ---------"); 2582 if (!thread->has_last_Java_frame()) { 2583 ls.print_cr("NO ANCHOR!"); 2584 } 2585 2586 RegisterMap map(thread, 2587 RegisterMap::UpdateMap::include, 2588 RegisterMap::ProcessFrames::include, 2589 RegisterMap::WalkContinuation::skip); 2590 map.set_include_argument_oops(false); 2591 2592 if (false) { 2593 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) { 2594 f.print_on(&ls); 2595 } 2596 } else { 2597 map.set_skip_missing(true); 2598 ResetNoHandleMark rnhm; 2599 ResourceMark rm; 2600 HandleMark hm(Thread::current()); 2601 FrameValues values; 2602 2603 int i = 0; 2604 int post_entry = -1; 2605 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) { 2606 f.describe(values, i++, &map); 2607 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f)) 2608 post_entry++; 2609 if (post_entry >= show_entry_callers) 2610 break; 2611 } 2612 values.print_on(thread, &ls); 2613 } 2614 2615 ls.print_cr("======= end frames ========="); 2616 } 2617 #endif // ASSERT 2618 2619 #include CPU_HEADER_INLINE(continuationFreezeThaw) 2620 2621 #ifdef ASSERT 2622 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) { 2623 ResourceMark rm; 2624 FrameValues values; 2625 assert(f.get_cb() != nullptr, ""); 2626 RegisterMap map(f.is_heap_frame() ? 2627 nullptr : 2628 JavaThread::current(), 2629 RegisterMap::UpdateMap::include, 2630 RegisterMap::ProcessFrames::skip, 2631 RegisterMap::WalkContinuation::skip); 2632 map.set_include_argument_oops(false); 2633 map.set_skip_missing(true); 2634 if (callee_complete) { 2635 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f)); 2636 } 2637 const_cast<frame&>(f).describe(values, 0, &map); 2638 values.print_on(static_cast<JavaThread*>(nullptr), st); 2639 } 2640 #endif 2641 2642 static address thaw_entry = nullptr; 2643 static address freeze_entry = nullptr; 2644 2645 address Continuation::thaw_entry() { 2646 return ::thaw_entry; 2647 } 2648 2649 address Continuation::freeze_entry() { 2650 return ::freeze_entry; 2651 } 2652 2653 class ConfigResolve { 2654 public: 2655 static void resolve() { resolve_compressed(); } 2656 2657 static void resolve_compressed() { 2658 UseCompressedOops ? resolve_gc<true>() 2659 : resolve_gc<false>(); 2660 } 2661 2662 private: 2663 template <bool use_compressed> 2664 static void resolve_gc() { 2665 BarrierSet* bs = BarrierSet::barrier_set(); 2666 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set"); 2667 switch (bs->kind()) { 2668 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \ 2669 case BarrierSet::bs_name: { \ 2670 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \ 2671 } \ 2672 break; 2673 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE) 2674 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE 2675 2676 default: 2677 fatal("BarrierSet resolving not implemented"); 2678 }; 2679 } 2680 2681 template <bool use_compressed, typename BarrierSetT> 2682 static void resolve() { 2683 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT; 2684 2685 freeze_entry = (address)freeze<SelectedConfigT>; 2686 2687 // If we wanted, we could templatize by kind and have three different thaw entries 2688 thaw_entry = (address)thaw<SelectedConfigT>; 2689 } 2690 }; 2691 2692 void Continuation::init() { 2693 ConfigResolve::resolve(); 2694 }