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 *
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()
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();
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
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;
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 }
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()) {
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();
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 {
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()) {
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.
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
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
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
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
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
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 }
|
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/objectMonitor.inline.hpp"
59 #include "runtime/orderAccess.hpp"
60 #include "runtime/prefetch.inline.hpp"
61 #include "runtime/smallRegisterMap.inline.hpp"
62 #include "runtime/sharedRuntime.hpp"
63 #include "runtime/stackChunkFrameStream.inline.hpp"
64 #include "runtime/stackFrameStream.inline.hpp"
65 #include "runtime/stackOverflow.hpp"
66 #include "runtime/stackWatermarkSet.inline.hpp"
67 #include "utilities/debug.hpp"
68 #include "utilities/exceptions.hpp"
69 #include "utilities/macros.hpp"
70 #include "utilities/vmError.hpp"
71 #if INCLUDE_ZGC
72 #include "gc/z/zStackChunkGCData.inline.hpp"
73 #endif
74
75 #include <type_traits>
76
77 /*
78 * This file contains the implementation of continuation freezing (yield) and thawing (run).
79 *
80 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
81 * would likely call these operations many thousands of times per second second, on every core.
82 *
83 * Freeze might be called every time the application performs any I/O operation, every time it
84 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
85 * multiple times in each of those cases, as it is called by the return barrier, which may be
86 * invoked on method return.
87 *
88 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
89 * example, every effort is made to avoid Java-VM transitions as much as possible.
90 *
183 static void log_frames(JavaThread* thread);
184 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
185
186 #define assert_pfl(p, ...) \
187 do { \
188 if (!(p)) { \
189 JavaThread* t = JavaThread::active(); \
190 if (t->has_last_Java_frame()) { \
191 tty->print_cr("assert(" #p ") failed:"); \
192 t->print_frame_layout(); \
193 } \
194 } \
195 vmassert(p, __VA_ARGS__); \
196 } while(0)
197
198 #else
199 static void verify_continuation(oop continuation) { }
200 #define assert_pfl(p, ...)
201 #endif
202
203 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
204 template<typename ConfigT, bool preempt> static inline int freeze_internal(JavaThread* thread, intptr_t* const sp);
205
206 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
207 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
208
209
210 // Entry point to freeze. Transitions are handled manually
211 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
212 template<typename ConfigT>
213 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
214 assert(sp == current->frame_anchor()->last_Java_sp(), "");
215
216 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
217 current->set_cont_fastpath(nullptr);
218 }
219
220 return ConfigT::freeze(current, sp);
221 JRT_END
222
223 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
224 return prepare_thaw_internal(thread, return_barrier);
225 JRT_END
226
227 template<typename ConfigT>
228 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
229 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
230 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
231 // Also the preemption case with JVMTI events enabled might safepoint so
232 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
233 // JRT_ENTRY instead?
234 ResetNoHandleMark rnhm;
235 debug_only(PauseNoSafepointVerifier pnsv(&__nsv);)
236
237 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
238 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
239 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
240 JRT_END
241
242 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
243 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
244 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
245 }
246 JVM_END
247
248 ///////////
249
250 enum class oop_kind { NARROW, WIDE };
251 template <oop_kind oops, typename BarrierSetT>
252 class Config {
253 public:
254 typedef Config<oops, BarrierSetT> SelfT;
255 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
256
257 static int freeze(JavaThread* thread, intptr_t* const sp) {
258 return freeze_internal<SelfT, false>(thread, sp);
259 }
260
261 static int freeze_preempt(JavaThread* thread, intptr_t* const sp) {
262 return freeze_internal<SelfT, true>(thread, sp);
263 }
264
265 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
266 return thaw_internal<SelfT>(thread, kind);
267 }
268 };
269
270 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
271 const size_t page_size = os::vm_page_size();
272 if (size > page_size) {
273 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
274 return false;
275 }
276 }
277 return true;
278 }
279
280 #ifdef ASSERT
281 static oop get_continuation(JavaThread* thread) {
282 assert(thread != nullptr, "");
283 assert(thread->threadObj() != nullptr, "");
284 return java_lang_Thread::continuation(thread->threadObj());
285 }
286 #endif // ASSERT
287
288 inline void clear_anchor(JavaThread* thread) {
289 thread->frame_anchor()->clear();
290 }
291
292 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc = nullptr) {
293 if (pc == nullptr) {
294 pc = ContinuationHelper::return_address_at(
295 sp - frame::sender_sp_ret_address_offset());
296 }
297 assert(pc != nullptr, "");
298
299 JavaFrameAnchor* anchor = thread->frame_anchor();
300 anchor->set_last_Java_sp(sp);
301 anchor->set_last_Java_pc(pc);
302 ContinuationHelper::set_anchor_pd(anchor, sp);
303
304 assert(thread->has_last_Java_frame(), "");
305 assert(thread->last_frame().cb() != nullptr, "");
306 }
307
308 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
309 JavaFrameAnchor* anchor = thread->frame_anchor();
310 anchor->set_last_Java_sp(entry->entry_sp());
311 anchor->set_last_Java_pc(entry->entry_pc());
312 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
313
314 assert(thread->has_last_Java_frame(), "");
315 assert(thread->last_frame().cb() != nullptr, "");
316 }
317
318 #ifdef ASSERT
319 static int monitors_to_fix_on_stack(JavaThread* thread) {
320 ResourceMark rm(JavaThread::current());
321 ContinuationEntry* ce = thread->last_continuation();
322 RegisterMap map(thread,
323 RegisterMap::UpdateMap::include,
324 RegisterMap::ProcessFrames::include,
325 RegisterMap::WalkContinuation::include);
326 map.set_include_argument_oops(false);
327 ResourceHashtable<oopDesc*, bool> rhtable;
328 int monitor_count = 0;
329 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(thread, f); f = f.sender(&map)) {
330 if (f.is_interpreted_frame()) {
331 frame abs = !f.is_heap_frame() ? f : map.stack_chunk()->derelativize(f);
332 monitor_count += ContinuationHelper::InterpretedFrame::monitors_to_fix(thread, abs, rhtable, map.stack_chunk()());
333 } else if (f.is_compiled_frame()) {
334 monitor_count += ContinuationHelper::CompiledFrame::monitors_to_fix(thread, &map, f, rhtable);
335 } else if (f.is_native_frame()) {
336 monitor_count += ContinuationHelper::NativeFrame::monitors_to_fix(thread, f, rhtable);
337 }
338 }
339 return monitor_count;
340 }
341 #endif
342
343 #if CONT_JFR
344 class FreezeThawJfrInfo : public StackObj {
345 short _e_size;
346 short _e_num_interpreted_frames;
347 public:
348
349 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
350 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
351 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
352 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
353 };
354
355 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
356 if (e->should_commit()) {
357 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
358 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
359 e->set_continuationClass(continuation->klass());
360 e->set_interpretedFrames(_e_num_interpreted_frames);
361 e->set_size(_e_size);
362 e->commit();
363 }
364 }
365 #endif // CONT_JFR
366
367 /////////////// FREEZE ////
368
369 class FreezeBase : public StackObj {
370 protected:
371 JavaThread* const _thread;
372 ContinuationWrapper& _cont;
373 bool _barriers; // only set when we allocate a chunk
374
375 intptr_t* _bottom_address;
376
377 const bool _preempt;
378 // Used on preemption only
379 frame _last_frame;
380 oop _monitorenter_obj;
381
382 // Used to support freezing with held monitors
383 int _monitors_to_fix;
384 int _monitors_in_lockstack;
385
386 int _freeze_size; // total size of all frames plus metadata in words.
387 int _total_align_size;
388
389 intptr_t* _cont_stack_top;
390 intptr_t* _cont_stack_bottom;
391
392 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
393
394 #ifdef ASSERT
395 intptr_t* _orig_chunk_sp;
396 int _fast_freeze_size;
397 bool _empty;
398 #endif
399
400 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
401
402 NOT_PRODUCT(int _frames;)
403 DEBUG_ONLY(intptr_t* _last_write;)
404
405 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
406
407 public:
408 NOINLINE freeze_result freeze_slow();
409 void freeze_fast_existing_chunk();
410
411 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
412 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
413
414 inline int size_if_fast_freeze_available();
415
416 inline frame& last_frame() { return _last_frame; }
417 inline void set_last_frame() { _last_frame = _thread->last_frame(); }
418
419 #ifdef ASSERT
420 bool check_valid_fast_path();
421 #endif
422
423 protected:
424 inline void init_rest();
425 void throw_stack_overflow_on_humongous_chunk();
426
427 // fast path
428 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
429 inline void unwind_frames();
430 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
431
432 // slow path
433 virtual stackChunkOop allocate_chunk_slow(size_t stack_size) = 0;
434
435 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
436
437 private:
438 // slow path
439 frame freeze_start_frame();
440 frame freeze_start_frame_on_preempt();
441 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
442 inline frame freeze_start_frame_yield_stub();
443 template<typename FKind>
444 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
445 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
446 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
447 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
448 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
449 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
450 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
451 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
452 NOINLINE void finish_freeze(const frame& f, const frame& top);
453
454 void fix_monitors_in_interpreted_frame(frame& f);
455 template <typename RegisterMapT>
456 void fix_monitors_in_compiled_frame(frame& f, RegisterMapT* map);
457 void fix_monitors_in_fast_path();
458
459 inline bool stack_overflow();
460
461 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
462 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
463 template<typename FKind> static inline frame sender(const frame& f);
464 template<typename FKind> frame new_heap_frame(frame& f, frame& caller);
465 inline void set_top_frame_metadata_pd(const frame& hf);
466 inline void patch_pd(frame& callee, const frame& caller);
467 void adjust_interpreted_frame_unextended_sp(frame& f);
468 static inline void prepare_freeze_interpreted_top_frame(const frame& f);
469 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
470
471 protected:
472 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
473 bool freeze_fast_new_chunk(stackChunkOop chunk);
474
475 #ifdef ASSERT
476 bool is_empty(stackChunkOop chunk) {
477 // during freeze, the chunk is in an intermediate state (after setting the chunk's argsize but before setting its
478 // ultimate sp) so we use this instead of stackChunkOopDesc::is_empty
479 return chunk->sp() >= chunk->stack_size() - chunk->argsize() - frame::metadata_words_at_top;
480 }
481 #endif
482 };
483
484 template <typename ConfigT>
485 class Freeze : public FreezeBase {
486 private:
487 stackChunkOop allocate_chunk(size_t stack_size);
488
489 public:
490 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
491 : FreezeBase(thread, cont, frame_sp, preempt) {}
492
493 freeze_result try_freeze_fast();
494
495 protected:
496 virtual stackChunkOop allocate_chunk_slow(size_t stack_size) override { return allocate_chunk(stack_size); }
497 };
498
499 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
500 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
501 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
502
503 assert(_thread != nullptr, "");
504 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
505
506 DEBUG_ONLY(_cont.entry()->verify_cookie();)
507
508 assert(!Interpreter::contains(_cont.entryPC()), "");
509
510 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
511 #ifdef _LP64
512 if (((intptr_t)_bottom_address & 0xf) != 0) {
513 _bottom_address--;
514 }
515 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
516 #endif
517
518 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
519 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
520 assert(_bottom_address != nullptr, "");
521 assert(_bottom_address <= _cont.entrySP(), "");
522 DEBUG_ONLY(_last_write = nullptr;)
523
524 assert(_cont.chunk_invariant(), "");
525 assert(!Interpreter::contains(_cont.entryPC()), "");
526 #if !defined(PPC64) || defined(ZERO)
527 static const int doYield_stub_frame_size = frame::metadata_words;
528 #else
529 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
530 #endif
531 // With preemption doYield() might not have been resolved yet
532 assert(_preempt || SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size, "");
533
534 // properties of the continuation on the stack; all sizes are in words
535 _cont_stack_top = frame_sp + (!preempt ? doYield_stub_frame_size : 0); // we don't freeze the doYield stub frame
536 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
537 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
538
539 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
540 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
541 assert(cont_size() > 0, "");
542
543 if (LockingMode == LM_LEGACY) {
544 _monitors_to_fix = thread->held_monitor_count();
545 assert(_monitors_to_fix >= 0, "invariant");
546 _monitors_in_lockstack = 0;
547 assert(_thread->lock_stack().monitor_count() == 0, "should be set only for LM_LIGHTWEIGHT");
548 } else {
549 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
550 assert(_monitors_in_lockstack >= 0 && _monitors_in_lockstack <= 8, "_monitors_in_lockstack=%d", _monitors_in_lockstack);
551 _monitors_to_fix = _monitors_in_lockstack;
552 assert(thread->held_monitor_count() == 0, "should be set only for LM_LEGACY");
553 }
554 DEBUG_ONLY(int verified_cnt = monitors_to_fix_on_stack(_thread);)
555 assert(verified_cnt == _monitors_to_fix || thread->obj_locker_count() > 0 ||
556 (LockingMode == LM_LIGHTWEIGHT && verified_cnt == _thread->lock_stack().unique_count()),
557 "wrong monitor count. Found %d in the stack but counter is %d", verified_cnt, _monitors_to_fix);
558 }
559
560 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
561 _freeze_size = 0;
562 _total_align_size = 0;
563 NOT_PRODUCT(_frames = 0;)
564 }
565
566 void FreezeBase::fix_monitors_in_interpreted_frame(frame& f) {
567 assert(LockingMode == LM_LEGACY, "invariant");
568 BasicObjectLock* first_mon = f.interpreter_frame_monitor_begin();
569 BasicObjectLock* last_mon = f.interpreter_frame_monitor_end();
570 assert(last_mon <= first_mon, "must be");
571
572 if (first_mon == last_mon) {
573 // No monitors in this frame
574 return;
575 }
576
577 for (BasicObjectLock* current = f.previous_monitor_in_interpreter_frame(first_mon);
578 current >= last_mon; current = f.previous_monitor_in_interpreter_frame(current)) {
579 oop obj = current->obj();
580 if (obj == nullptr) {
581 continue;
582 }
583 if (obj == _monitorenter_obj) {
584 assert(_preempt, "should be preemption on monitorenter case");
585 assert(obj->mark().monitor() != nullptr, "failed to acquire the lock but its not inflated");
586 continue;
587 }
588 markWord mark = obj->mark();
589 if (mark.has_monitor() && !mark.monitor()->is_owner_anonymous()) {
590 // Good for freezing, nothing to do.
591 assert(mark.monitor()->is_owner(_thread), "invariant");
592 continue;
593 }
594 // Found locked monitor that needs fixing. Inflate will fix the owner for stack-locked case or inflated but anonymously owned.
595 ObjectMonitor* om = ObjectSynchronizer::inflate(_thread, obj, ObjectSynchronizer::InflateCause::inflate_cause_cont_freeze);
596 assert(om != nullptr, "invariant");
597 assert(om->is_owner(_thread), "invariant");
598 if (--_monitors_to_fix == 0) break;
599 }
600 assert(_monitors_to_fix >= 0, "invariant");
601 }
602
603 template <typename RegisterMapT>
604 void FreezeBase::fix_monitors_in_compiled_frame(frame& f, RegisterMapT* map) {
605 assert(LockingMode == LM_LEGACY, "invariant");
606 assert(!f.is_interpreted_frame(), "");
607 assert(ContinuationHelper::CompiledFrame::is_instance(f), "");
608
609 nmethod* nm = f.cb()->as_nmethod();
610
611 if (!nm->has_monitors()) {
612 // No monitors in this frame
613 return;
614 }
615
616 for (ScopeDesc* scope = nm->scope_desc_at(f.pc()); scope != nullptr; scope = scope->sender()) {
617 GrowableArray<MonitorValue*>* mons = scope->monitors();
618 if (mons == nullptr || mons->is_empty()) {
619 continue;
620 }
621
622 for (int index = (mons->length()-1); index >= 0; index--) { // see compiledVFrame::monitors()
623 MonitorValue* mon = mons->at(index);
624 if (mon->eliminated()) {
625 continue; // we ignore eliminated monitors
626 }
627 ScopeValue* ov = mon->owner();
628 StackValue* owner_sv = StackValue::create_stack_value(&f, map, ov); // it is an oop
629 oop obj = owner_sv->get_obj()();
630 if (obj == nullptr) {
631 continue;
632 }
633 if (obj == _monitorenter_obj) {
634 assert(_preempt, "should be preemption on monitorenter case");
635 assert(obj->mark().monitor() != nullptr, "failed to acquire the lock but its not inflated");
636 continue;
637 }
638 markWord mark = obj->mark();
639 if (mark.has_monitor() && !mark.monitor()->is_owner_anonymous()) {
640 // Good for freezing, nothing to do.
641 assert(mark.monitor()->is_owner(_thread), "invariant");
642 continue;
643 }
644 // Found locked monitor that needs fixing. Inflate will fix the owner for stack-locked case or inflated but anonymously owned.
645 ObjectMonitor* om = ObjectSynchronizer::inflate(_thread, obj, ObjectSynchronizer::InflateCause::inflate_cause_cont_freeze);
646 assert(om != nullptr, "invariant");
647 assert(om->is_owner(_thread), "invariant");
648 if (--_monitors_to_fix == 0) break;
649 }
650 }
651 assert(_monitors_to_fix >= 0, "invariant");
652 }
653
654 void FreezeBase::fix_monitors_in_fast_path() {
655 if (LockingMode == LM_LIGHTWEIGHT) {
656 stackChunkOop chunk = _cont.tail();
657 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
658 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
659
660 chunk->set_lockStackSize((uint8_t)_monitors_in_lockstack);
661 chunk->set_has_lockStack(true);
662
663 DEBUG_ONLY(_monitors_to_fix -= _monitors_in_lockstack;)
664 } else {
665 assert(LockingMode == LM_LEGACY, "invariant");
666 _monitorenter_obj = _thread->is_on_monitorenter() ? _thread->current_pending_monitor()->object() : nullptr;
667
668 ResourceMark rm(_thread);
669 RegisterMap map(JavaThread::current(),
670 RegisterMap::UpdateMap::include,
671 RegisterMap::ProcessFrames::skip, // already processed in unwind_frames()
672 RegisterMap::WalkContinuation::skip);
673 map.set_include_argument_oops(false);
674 frame freezed_top(_cont_stack_top);
675 frame first = !_preempt ? freezed_top : freezed_top.sender(&map);
676 for (frame f = first; _monitors_to_fix > 0 && Continuation::is_frame_in_continuation(_cont.entry(), f); f = f.sender(&map)) {
677 fix_monitors_in_compiled_frame(f, &map);
678 }
679 }
680 assert(_monitors_to_fix == 0, "missing monitors in stack");
681 assert(_thread->held_monitor_count() == 0, "should be 0 now");
682 }
683
684 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
685 stackChunkOop chunk = _cont.tail();
686 chunk->copy_from_stack_to_chunk(from, to, size);
687 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
688
689 #ifdef ASSERT
690 if (_last_write != nullptr) {
691 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
692 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
693 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
694 _last_write = to;
695 }
696 #endif
697 }
698
699 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
700 void FreezeBase::unwind_frames() {
701 ContinuationEntry* entry = _cont.entry();
702 entry->flush_stack_processing(_thread);
703 set_anchor_to_entry(_thread, entry);
704 }
705
706 template <typename ConfigT>
707 freeze_result Freeze<ConfigT>::try_freeze_fast() {
708 assert(_thread->thread_state() == _thread_in_vm, "");
709 assert(_thread->cont_fastpath(), "");
710
711 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
712 assert(_fast_freeze_size == 0, "");
713
714 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack);
715 if (freeze_fast_new_chunk(chunk)) {
716 return freeze_ok;
717 }
718 if (_thread->has_pending_exception()) {
719 return freeze_exception;
720 }
721
722 // TODO R REMOVE when deopt change is fixed
723 assert(!_thread->cont_fastpath() || _barriers, "");
724 log_develop_trace(continuations)("-- RETRYING SLOW --");
725 return freeze_slow();
726 }
727
728 // Returns size needed if the continuation fits, otherwise 0.
729 int FreezeBase::size_if_fast_freeze_available() {
730 stackChunkOop chunk = _cont.tail();
731 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
732 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
733 return 0;
734 }
735
736 int total_size_needed = cont_size();
737 const int chunk_sp = chunk->sp();
738
739 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
740 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
741 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
742 // although that would require changing stackChunkOopDesc::is_empty
743 if (chunk_sp < chunk->stack_size()) {
744 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
745 }
746
747 total_size_needed += _monitors_in_lockstack;
748
749 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
750 bool available = chunk_free_room >= total_size_needed;
751 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
752 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
753 return available ? total_size_needed : 0;
754 }
755
756 void FreezeBase::freeze_fast_existing_chunk() {
757 stackChunkOop chunk = _cont.tail();
758 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
759
760 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
761 assert(_fast_freeze_size > 0, "");
762
763 if (chunk->sp() < chunk->stack_size()) { // we are copying into a non-empty chunk
764 DEBUG_ONLY(_empty = false;)
765 assert(chunk->sp() < (chunk->stack_size() - chunk->argsize()), "");
766 #ifdef ASSERT
767 {
768 intptr_t* retaddr_slot = (chunk->sp_address()
811 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
812 }
813 }
814
815 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
816 DEBUG_ONLY(_empty = true;)
817
818 // Install new chunk
819 _cont.set_tail(chunk);
820
821 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
822 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
823 return false;
824 }
825
826 chunk->set_max_thawing_size(cont_size());
827 chunk->set_argsize(_cont.argsize());
828
829 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
830 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
831 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
832 assert(chunk_start_sp == chunk->stack_size(), "");
833
834 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
835
836 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
837
838 return true;
839 }
840
841 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
842 assert(chunk != nullptr, "");
843 assert(!chunk->has_mixed_frames(), "");
844 assert(!chunk->is_gc_mode(), "");
845 assert(!chunk->has_bitmap(), "");
846 assert(!chunk->requires_barriers(), "");
847 assert(chunk == _cont.tail(), "");
848
849 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
850 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
851 // will either see no continuation on the stack, or a consistent chunk.
852 unwind_frames();
853
854 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
855 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
856 assert(chunk_start_sp <= chunk->stack_size(), "");
857 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
858
859 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
860 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
861
862 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
863 #ifdef ASSERT
864 if (!_empty) {
865 intptr_t* retaddr_slot = (_orig_chunk_sp
866 - frame::sender_sp_ret_address_offset());
867 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
868 "unexpected saved return address");
869 }
870 #endif
871
872 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
873 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
874 intptr_t* from = _cont_stack_top - frame::metadata_words_at_bottom;
875 intptr_t* to = chunk_top - frame::metadata_words_at_bottom;
876 copy_to_chunk(from, to, cont_size() + frame::metadata_words_at_bottom);
877 // Because we're not patched yet, the chunk is now in a bad state
878
879 // patch return pc of the bottom-most frozen frame (now in the chunk)
880 // with the actual caller's return address
881 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
882 - _cont.argsize()
883 - frame::metadata_words_at_top
884 - frame::sender_sp_ret_address_offset());
885 #ifdef ASSERT
886 if (!_empty) {
887 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
888 == StubRoutines::cont_returnBarrier(),
889 "should be the continuation return barrier");
890 }
891 #endif
892 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
893 chunk->pc());
894
895 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
896 chunk->set_sp(chunk_new_sp);
897 // set chunk->pc to the return address of the topmost frame in the chunk
898 chunk->set_pc(ContinuationHelper::return_address_at(
899 _cont_stack_top - frame::sender_sp_ret_address_offset()));
900
901 // Fix monitors after unwinding frames to make sure oops are valid.
902 if (_monitors_to_fix > 0) {
903 fix_monitors_in_fast_path();
904 }
905
906 _cont.write();
907
908 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
909 LogTarget(Trace, continuations) lt;
910 if (lt.develop_is_enabled()) {
911 LogStream ls(lt);
912 chunk->print_on(true, &ls);
913 }
914
915 // Verification
916 assert(_cont.chunk_invariant(), "");
917 chunk->verify();
918
919 #if CONT_JFR
920 EventContinuationFreezeFast e;
921 if (e.should_commit()) {
922 e.set_id(cast_from_oop<u8>(chunk));
923 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
924 e.set_size(cont_size() << LogBytesPerWord);
925 e.commit();
944 #endif
945
946 init_rest();
947
948 HandleMark hm(Thread::current());
949
950 frame f = freeze_start_frame();
951
952 LogTarget(Debug, continuations) lt;
953 if (lt.develop_is_enabled()) {
954 LogStream ls(lt);
955 f.print_on(&ls);
956 }
957
958 frame caller; // the frozen caller in the chunk
959 freeze_result res = recurse_freeze(f, caller, 0, false, true);
960
961 if (res == freeze_ok) {
962 finish_freeze(f, caller);
963 _cont.write();
964 assert(_monitors_to_fix == 0, "missing monitors in stack");
965 assert(_thread->held_monitor_count() == 0, "should be 0 now");
966 }
967
968 return res;
969 }
970
971 frame FreezeBase::freeze_start_frame() {
972 if (LIKELY(!_preempt)) {
973 return freeze_start_frame_yield_stub();
974 } else {
975 return freeze_start_frame_on_preempt();
976 }
977 }
978
979 frame FreezeBase::freeze_start_frame_yield_stub() {
980 frame f = _thread->last_frame();
981 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be");
982 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
983 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
984 return f;
985 }
986
987 frame FreezeBase::freeze_start_frame_on_preempt() {
988 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
989 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
990 return _last_frame;
991 }
992
993 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
994 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
995 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
996 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
997 || ((top && _preempt) == f.is_native_frame()), "");
998
999 if (stack_overflow()) {
1000 return freeze_exception;
1001 }
1002
1003 if (f.is_compiled_frame()) {
1004 if (UNLIKELY(f.oop_map() == nullptr)) {
1005 // special native frame
1006 return freeze_pinned_native;
1007 }
1008 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
1009 } else if (f.is_interpreted_frame()) {
1010 assert(!f.interpreter_frame_method()->is_native() || (_preempt && top && _thread->is_on_monitorenter()), "");
1011 if (_preempt && top && f.interpreter_frame_method()->is_native()) {
1012 // TODO: Allow preemption for this case too
1013 return freeze_pinned_native;
1014 }
1015 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
1016 } else if (_preempt && top) {
1017 assert(ContinuationHelper::Frame::is_stub(f.cb()) || (f.is_native_frame() && _thread->is_on_monitorenter()), "invariant");
1018 if (f.is_native_frame()) {
1019 // TODO: Allow preemption for this case too
1020 return freeze_pinned_native;
1021 } else {
1022 return recurse_freeze_stub_frame(f, caller);
1023 }
1024 } else {
1025 return freeze_pinned_native;
1026 }
1027 }
1028
1029 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1030 // See also StackChunkFrameStream<frame_kind>::frame_size()
1031 template<typename FKind>
1032 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
1033 assert(FKind::is_instance(f), "");
1034
1035 assert(fsize > 0, "");
1036 assert(argsize >= 0, "");
1037 _freeze_size += fsize;
1038 NOT_PRODUCT(_frames++;)
1039
1040 assert(FKind::frame_bottom(f) <= _bottom_address, "");
1041
1042 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
1043 // enterSpecial and an interpreted frame
1064
1065 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
1066 LogTarget(Trace, continuations) lt;
1067 if (lt.develop_is_enabled()) {
1068 LogStream ls(lt);
1069 DEBUG_ONLY(hf.print_value_on(&ls, nullptr);)
1070 assert(hf.is_heap_frame(), "should be");
1071 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
1072 if (is_bottom_frame) {
1073 ls.print_cr("bottom h-frame:");
1074 hf.print_on(&ls);
1075 }
1076 }
1077 }
1078
1079 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
1080 // See also StackChunkFrameStream<frame_kind>::frame_size()
1081 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
1082 int argsize = argsize_md - frame::metadata_words_at_top;
1083 assert(callee.is_interpreted_frame()
1084 || ContinuationHelper::Frame::is_stub(callee.cb())
1085 || callee.cb()->as_nmethod()->is_osr_method()
1086 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
1087 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
1088 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
1089
1090 LogTarget(Trace, continuations) lt;
1091
1092 #ifdef ASSERT
1093 bool empty = _cont.is_empty();
1094 log_develop_trace(continuations)("empty: %d", empty);
1095 #endif
1096
1097 stackChunkOop chunk = _cont.tail();
1098
1099 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1100
1101 _freeze_size += frame::metadata_words; // for top frame's metadata
1102
1103 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1104 int unextended_sp = -1;
1106 unextended_sp = chunk->sp();
1107 if (!chunk->is_empty()) {
1108 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1109 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1110 bool top_interpreted = Interpreter::contains(chunk->pc());
1111 if (callee.is_interpreted_frame() == top_interpreted) {
1112 overlap = argsize_md;
1113 }
1114 }
1115 }
1116
1117 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1118
1119 _freeze_size -= overlap;
1120 assert(_freeze_size >= 0, "");
1121
1122 assert(chunk == nullptr || chunk->is_empty()
1123 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1124 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1125
1126 _freeze_size += _monitors_in_lockstack;
1127
1128 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1129 bool allocated_old_in_freeze_fast = _barriers;
1130 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1131 "Chunk allocated in freeze_fast is of insufficient size "
1132 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1133 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1134
1135 DEBUG_ONLY(bool empty_chunk = true);
1136 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1137 // ALLOCATE NEW CHUNK
1138
1139 if (lt.develop_is_enabled()) {
1140 LogStream ls(lt);
1141 if (chunk == nullptr) {
1142 ls.print_cr("no chunk");
1143 } else {
1144 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1145 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1146 chunk->print_on(&ls);
1147 }
1166 // REUSE EXISTING CHUNK
1167 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1168 if (chunk->is_empty()) {
1169 int sp = chunk->stack_size() - argsize_md;
1170 chunk->set_sp(sp);
1171 chunk->set_argsize(argsize);
1172 _freeze_size += overlap;
1173 assert(chunk->max_thawing_size() == 0, "");
1174 } DEBUG_ONLY(else empty_chunk = false;)
1175 }
1176 assert(!chunk->is_gc_mode(), "");
1177 assert(!chunk->has_bitmap(), "");
1178 chunk->set_has_mixed_frames(true);
1179
1180 assert(chunk->requires_barriers() == _barriers, "");
1181 assert(!_barriers || is_empty(chunk), "");
1182
1183 assert(!is_empty(chunk) || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1184 assert(!is_empty(chunk) || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1185
1186 if (_preempt) {
1187 frame f = _thread->last_frame();
1188 if (f.is_interpreted_frame()) {
1189 // Do it now that we know freezing will be successful.
1190 prepare_freeze_interpreted_top_frame(f);
1191 }
1192 }
1193
1194 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1195 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1196 // will either see no continuation or a consistent chunk.
1197 unwind_frames();
1198
1199 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1200
1201 if (lt.develop_is_enabled()) {
1202 LogStream ls(lt);
1203 ls.print_cr("top chunk:");
1204 chunk->print_on(&ls);
1205 }
1206
1207 // Fix monitors after unwinding frames to make sure oops are valid. Also do it now to avoid unnecessary
1208 // calls to fix_monitors_in_interpreted_frame/fix_monitors_in_compiled_frame() later.
1209 if (_monitors_to_fix > 0) {
1210 if (_monitors_in_lockstack > 0) {
1211 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
1212 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
1213
1214 chunk->set_lockStackSize((uint8_t)_monitors_in_lockstack);
1215 chunk->set_has_lockStack(true);
1216
1217 _monitors_to_fix -= _monitors_in_lockstack;
1218 assert(_monitors_to_fix == 0, "invariant");
1219 }
1220 _monitorenter_obj = _thread->is_on_monitorenter() ? _thread->current_pending_monitor()->object() : nullptr;
1221 }
1222
1223 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1224 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1225
1226 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1227
1228 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1229 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1230 #ifdef ASSERT
1231 if (lt.develop_is_enabled()) {
1232 LogStream ls(lt);
1233 ls.print_cr("top hframe before (freeze):");
1234 assert(caller.is_heap_frame(), "should be");
1235 caller.print_on(&ls);
1236 }
1237
1238 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1239
1240 frame entry = sender(callee);
1241
1242 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1243 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1244 #endif
1245
1246 return freeze_ok_bottom;
1247 }
1248
1249 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1250 if (is_bottom_frame) {
1251 // If we're the bottom frame, we need to replace the return barrier with the real
1252 // caller's pc.
1253 address last_pc = caller.pc();
1254 assert((last_pc == nullptr) == is_empty(_cont.tail()), "");
1255 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1256 } else {
1257 assert(!caller.is_empty(), "");
1258 }
1259
1260 patch_pd(hf, caller);
1261
1262 if (f.is_interpreted_frame()) {
1289 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1290 // See also StackChunkFrameStream<frame_kind>::frame_size()
1291 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1292 int callee_argsize /* incl. metadata */,
1293 bool callee_interpreted) {
1294 adjust_interpreted_frame_unextended_sp(f);
1295
1296 // The frame's top never includes the stack arguments to the callee
1297 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1298 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1299 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1300
1301 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1302
1303 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1304 // including metadata between f and its args
1305 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1306
1307 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d",
1308 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize);
1309 // we'd rather not yield inside methods annotated with @JvmtiMountTransition. In the preempt case
1310 // we already checked it is safe to do so in Continuation::is_safe_vthread_to_preempt().
1311 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition() || _preempt, "");
1312
1313 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1314 if (UNLIKELY(result > freeze_ok_bottom)) {
1315 return result;
1316 }
1317
1318 bool is_bottom_frame = result == freeze_ok_bottom;
1319 assert(!caller.is_empty() || is_bottom_frame, "");
1320
1321 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1322
1323 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1324 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1325
1326 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1327 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1328 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1329
1330 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1331 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1332 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1333 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1334
1335 relativize_interpreted_frame_metadata(f, hf);
1336
1337 patch(f, hf, caller, is_bottom_frame);
1338
1339 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1340 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1341 caller = hf;
1342
1343 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1344 frame_method->record_gc_epoch();
1345
1346 if (_monitors_to_fix > 0) {
1347 // Check if we have monitors in this frame
1348 fix_monitors_in_interpreted_frame(f);
1349 }
1350
1351 return freeze_ok;
1352 }
1353
1354 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1355 // See also StackChunkFrameStream<frame_kind>::frame_size()
1356 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1357 int callee_argsize /* incl. metadata */,
1358 bool callee_interpreted) {
1359 // The frame's top never includes the stack arguments to the callee
1360 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1361 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1362 // including metadata between f and its stackargs
1363 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1364 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1365
1366 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d",
1367 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1368 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1369 _freeze_size, fsize, argsize);
1370 // we'd rather not yield inside methods annotated with @JvmtiMountTransition. In the preempt case
1371 // we already checked it is safe to do so in Continuation::is_safe_vthread_to_preempt().
1372 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition() || _preempt, "");
1373
1374 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1375 if (UNLIKELY(result > freeze_ok_bottom)) {
1376 return result;
1377 }
1378
1379 bool is_bottom_frame = result == freeze_ok_bottom;
1380 assert(!caller.is_empty() || is_bottom_frame, "");
1381
1382 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1383
1384 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller);
1385
1386 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1387
1388 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1389 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1390
1391 if (caller.is_interpreted_frame()) {
1392 _total_align_size += frame::align_wiggle; // See Thaw::align
1393 }
1394
1395 patch(f, hf, caller, is_bottom_frame);
1396
1397 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1398
1399 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1400 caller = hf;
1401
1402 if (_monitors_to_fix > 0) {
1403 // Check if we have monitors in this frame
1404 fix_monitors_in_compiled_frame(f, SmallRegisterMap::instance);
1405 }
1406
1407 return freeze_ok;
1408 }
1409
1410 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1411 DEBUG_ONLY(frame fsender = sender(f);)
1412 assert(!fsender.is_native_frame() || (Continuation::is_continuation_enterSpecial(fsender) && !_cont.is_empty()), "sender should't be native except for enterSpecial case");
1413
1414 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f, 0, 0);
1415 const int fsize = f.cb()->frame_size();
1416
1417 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1418 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1419
1420 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1421 if (UNLIKELY(result > freeze_ok_bottom)) {
1422 return result;
1423 }
1424
1425 bool is_bottom_frame = result == freeze_ok_bottom;
1426 assert(!caller.is_empty() || (is_bottom_frame && !_cont.is_empty()), "");
1427
1428 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1429
1430 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1431 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf, 0, 0);
1432
1433 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1434
1435 if (caller.is_interpreted_frame()) {
1436 _total_align_size += frame::align_wiggle;
1437 }
1438
1439 patch(f, hf, caller, is_bottom_frame);
1440
1441 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1442
1443 caller = hf;
1444 return freeze_ok;
1445 }
1446
1447 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1448 stackChunkOop chunk = _cont.tail();
1449
1450 LogTarget(Trace, continuations) lt;
1451 if (lt.develop_is_enabled()) {
1452 LogStream ls(lt);
1453 assert(top.is_heap_frame(), "should be");
1454 top.print_on(&ls);
1455 }
1456
1457 set_top_frame_metadata_pd(top);
1458
1459 chunk->set_sp(chunk->to_offset(top.sp()));
1460 chunk->set_pc(top.pc());
1461
1462 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1463
1464 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1465
1466 // At this point the chunk is consistent
1467
1468 if (UNLIKELY(_barriers)) {
1469 log_develop_trace(continuations)("do barriers on old chunk");
1470 // Serial and Parallel GC can allocate objects directly into the old generation.
1471 // Then we want to relativize the derived pointers eagerly so that
1472 // old chunks are all in GC mode.
1473 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1474 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1475 if (UseShenandoahGC) {
1476 _cont.tail()->relativize_derived_pointers_concurrently();
1477 } else {
1478 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1479 }
1480 // For objects in the old generation we must maintain the remembered set
1481 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1482 }
1483
1484 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1485 if (lt.develop_is_enabled()) {
1486 LogStream ls(lt);
1487 chunk->print_on(true, &ls);
1488 }
1489
1490 if (lt.develop_is_enabled()) {
1491 LogStream ls(lt);
1492 ls.print_cr("top hframe after (freeze):");
1493 assert(_cont.last_frame().is_heap_frame(), "should be");
1494 _cont.last_frame().print_on(&ls);
1495 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1496 }
1497
1498 assert(_cont.chunk_invariant(), "");
1499 }
1500
1501 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1502 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1503 assert(t == JavaThread::current(), "");
1504 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1505 if (!_preempt) {
1506 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1507 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1508 }
1509 return true;
1510 }
1511 return false;
1512 }
1513
1514 class StackChunkAllocator : public MemAllocator {
1515 const size_t _stack_size;
1516 ContinuationWrapper& _continuation_wrapper;
1517 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1518 JavaThread* const _target;
1519 mutable bool _took_slow_path;
1520
1521 // Does the minimal amount of initialization needed for a TLAB allocation.
1522 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1523 virtual oop initialize(HeapWord* mem) const override {
1524 assert(_stack_size > 0, "");
1525 assert(_stack_size <= max_jint, "");
1526 assert(_word_size > _stack_size, "");
1527
1528 // zero out fields (but not the stack)
1529 const size_t hs = oopDesc::header_size();
1530 oopDesc::set_klass_gap(mem, 0);
1531 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1532
1533 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1534 jdk_internal_vm_StackChunk::set_sp(mem, (int)_stack_size);
1535
1536 return finish(mem);
1537 }
1538
1539 stackChunkOop allocate_fast() const {
1540 if (!UseTLAB) {
1541 return nullptr;
1542 }
1543
1544 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1545 if (mem == nullptr) {
1546 return nullptr;
1547 }
1548
1549 oop obj = initialize(mem);
1550 return stackChunkOopDesc::cast(obj);
1551 }
1552
1553 bool is_preempt() const { return _thread != _target; }
1554
1555 public:
1556 StackChunkAllocator(Klass* klass,
1557 size_t word_size,
1558 Thread* thread,
1559 size_t stack_size,
1560 ContinuationWrapper& continuation_wrapper,
1561 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector,
1562 JavaThread* target)
1563 : MemAllocator(klass, word_size, thread),
1564 _stack_size(stack_size),
1565 _continuation_wrapper(continuation_wrapper),
1566 _jvmti_event_collector(jvmti_event_collector),
1567 _target(target),
1568 _took_slow_path(false) {}
1569
1570 // Provides it's own, specialized allocation which skips instrumentation
1571 // if the memory can be allocated without going to a slow-path.
1572 stackChunkOop allocate() const {
1573 // First try to allocate without any slow-paths or instrumentation.
1574 stackChunkOop obj = allocate_fast();
1575 if (obj != nullptr) {
1576 return obj;
1577 }
1578
1579 // Now try full-blown allocation with all expensive operations,
1580 // including potentially safepoint operations.
1581 _took_slow_path = true;
1582
1583 // Protect unhandled Loom oops
1584 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1585
1586 // Can safepoint
1587 _jvmti_event_collector->start();
1601
1602 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1603 size_t size_in_words = klass->instance_size(stack_size);
1604
1605 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1606 if (!_preempt) {
1607 throw_stack_overflow_on_humongous_chunk();
1608 }
1609 return nullptr;
1610 }
1611
1612 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1613 assert(current == JavaThread::current(), "should be current");
1614
1615 // Allocate the chunk.
1616 //
1617 // This might safepoint while allocating, but all safepointing due to
1618 // instrumentation have been deferred. This property is important for
1619 // some GCs, as this ensures that the allocated object is in the young
1620 // generation / newly allocated memory.
1621 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, _cont, _jvmti_event_collector, _thread);
1622 stackChunkOop chunk = allocator.allocate();
1623
1624 if (chunk == nullptr) {
1625 return nullptr; // OOME
1626 }
1627
1628 // assert that chunk is properly initialized
1629 assert(chunk->stack_size() == (int)stack_size, "");
1630 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1631 assert(chunk->sp() == chunk->stack_size(), "");
1632 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1633 assert(chunk->max_thawing_size() == 0, "");
1634 assert(chunk->pc() == nullptr, "");
1635 assert(chunk->argsize() == 0, "");
1636 assert(chunk->flags() == 0, "");
1637 assert(chunk->is_gc_mode() == false, "");
1638 assert(chunk->lockStackSize() == 0, "");
1639 assert(chunk->objectMonitor() == nullptr, "");
1640
1641 // fields are uninitialized
1642 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1643 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1644
1645 #if INCLUDE_ZGC
1646 if (UseZGC) {
1647 if (ZGenerational) {
1648 ZStackChunkGCData::initialize(chunk);
1649 }
1650 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1651 _barriers = false;
1652 } else
1653 #endif
1654 #if INCLUDE_SHENANDOAHGC
1655 if (UseShenandoahGC) {
1656 _barriers = chunk->requires_barriers();
1657 } else
1658 #endif
1659 {
1692 return count;
1693 }
1694
1695 static void invalidate_jvmti_stack(JavaThread* thread) {
1696 if (thread->is_interp_only_mode()) {
1697 JvmtiThreadState *state = thread->jvmti_thread_state();
1698 if (state != nullptr)
1699 state->invalidate_cur_stack_depth();
1700 }
1701 }
1702
1703 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1704 if (JvmtiExport::can_post_frame_pop()) {
1705 int num_frames = num_java_frames(cont);
1706
1707 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1708 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames);
1709 }
1710 invalidate_jvmti_stack(thread);
1711 }
1712
1713 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, ObjectMonitor* mon, bool post_mount_event) {
1714 assert(current->vthread() != nullptr, "must be");
1715
1716 HandleMarkCleaner hm(current);
1717 Handle vth(current, current->vthread());
1718
1719 ContinuationWrapper::SafepointOp so(current, cont);
1720
1721 // Since we might safepoint set the anchor so that the stack can we walked.
1722 set_anchor(current, top.sp());
1723
1724 JRT_BLOCK
1725 current->rebind_to_jvmti_thread_state_of(vth());
1726 {
1727 MutexLocker mu(JvmtiThreadState_lock);
1728 JvmtiThreadState* state = current->jvmti_thread_state();
1729 if (state != NULL && state->is_pending_interp_only_mode()) {
1730 JvmtiEventController::enter_interp_only_mode();
1731 }
1732 }
1733 assert(current->is_in_VTMS_transition(), "sanity check");
1734 assert(!current->is_in_tmp_VTMS_transition(), "sanity check");
1735 JvmtiVTMSTransitionDisabler::finish_VTMS_transition((jthread)vth.raw_value(), /* is_mount */ true);
1736
1737 if (post_mount_event && JvmtiExport::should_post_vthread_mount()) {
1738 JvmtiExport::post_vthread_mount((jthread)vth.raw_value());
1739 } else if (!post_mount_event) {
1740 // Preemption cancelled case. Clear the unmount event pending flag. The
1741 // flag might actually not be set if _VTMS_notify_jvmti_events was enabled
1742 // after preemption happened (late binding agents). But since this would
1743 // be a rare case we just do it unconditionally.
1744 current->set_jvmti_unmount_event_pending(false);
1745 }
1746
1747 if (mon != nullptr) {
1748 JvmtiExport::post_monitor_contended_entered(current, mon);
1749 }
1750 JRT_BLOCK_END
1751
1752 clear_anchor(current);
1753 }
1754 #endif // INCLUDE_JVMTI
1755
1756 #ifdef ASSERT
1757
1758 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1759 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1760 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1761 bool FreezeBase::check_valid_fast_path() {
1762 ContinuationEntry* ce = _thread->last_continuation();
1763 RegisterMap map(_thread,
1764 RegisterMap::UpdateMap::skip,
1765 RegisterMap::ProcessFrames::skip,
1766 RegisterMap::WalkContinuation::skip);
1767 map.set_include_argument_oops(false);
1768 int i = 0;
1769 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), i++) {
1770 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (i == 0 && f.is_runtime_frame()))) {
1771 return false;
1772 }
1773 }
1774 return true;
1775 }
1776 #endif // ASSERT
1777
1778 static inline int freeze_epilog(ContinuationWrapper& cont) {
1779 verify_continuation(cont.continuation());
1780 assert(!cont.is_empty(), "");
1781
1782 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1783 return 0;
1784 }
1785
1786 static int freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1787 if (UNLIKELY(res != freeze_ok)) {
1788 verify_continuation(cont.continuation());
1789 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1790 return res;
1791 }
1792
1793 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1794 return freeze_epilog(cont);
1795 }
1796
1797 static int preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1798 if (UNLIKELY(res != freeze_ok)) {
1799 verify_continuation(cont.continuation());
1800 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1801 return res;
1802 }
1803
1804 patch_return_pc_with_preempt_stub(old_last_frame);
1805 cont.set_preempted(true);
1806 cont.tail()->set_is_preempted(true);
1807
1808 if (cont.thread()->is_on_monitorenter()) {
1809 cont.tail()->set_objectMonitor(cont.thread()->current_pending_monitor());
1810 }
1811
1812 return freeze_epilog(cont);
1813 }
1814
1815 template<typename ConfigT, bool preempt>
1816 static inline int freeze_internal(JavaThread* current, intptr_t* const sp) {
1817 assert(!current->has_pending_exception(), "");
1818
1819 #ifdef ASSERT
1820 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1821 log_frames(current);
1822 #endif
1823
1824 CONT_JFR_ONLY(EventContinuationFreeze event;)
1825
1826 ContinuationEntry* entry = current->last_continuation();
1827
1828 oop oopCont = entry->cont_oop(current);
1829 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1830 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1831
1832 verify_continuation(oopCont);
1833 ContinuationWrapper cont(current, oopCont);
1834 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1835
1836 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1837
1838 const bool pinned_monitor = NOT_LOOM_MONITOR_SUPPORT(current->held_monitor_count() > 0) LOOM_MONITOR_SUPPORT_ONLY(current->jni_monitor_count() > 0);
1839 if (entry->is_pinned() || pinned_monitor) {
1840 log_develop_debug(continuations)("PINNED due to critical section/hold monitor");
1841 verify_continuation(cont.continuation());
1842 freeze_result res = entry->is_pinned() ? freeze_pinned_cs : freeze_pinned_monitor;
1843 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1844 return res;
1845 }
1846
1847 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1848
1849 if (preempt) {
1850 freeze.set_last_frame(); // remember last frame
1851 #ifdef AARCH64
1852 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1853 freeze.set_jvmti_event_collector(&jsoaec);
1854
1855 // Force aarch64 to slow path always for now. It needs extra instructions to correctly set
1856 // the last pc we copy into the stackChunk, since it will not necessarily be at sp[-1]).
1857 freeze_result res = freeze.freeze_slow();
1858 CONT_JFR_ONLY(cont.post_jfr_event(&event, oopCont, current);)
1859 return preempt_epilog(cont, res, freeze.last_frame());
1860 #endif
1861 }
1862
1863 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1864 bool fast = UseContinuationFastPath && current->cont_fastpath();
1865 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1866 freeze.freeze_fast_existing_chunk();
1867 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1868 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1869 }
1870
1871 if (preempt) {
1872 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1873 freeze.set_jvmti_event_collector(&jsoaec);
1874
1875 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1876
1877 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1878 preempt_epilog(cont, res, freeze.last_frame());
1879 return res;
1880 }
1881
1882 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1883 assert(current == JavaThread::current(), "must be current thread");
1884 JRT_BLOCK
1885 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1886 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1887 freeze.set_jvmti_event_collector(&jsoaec);
1888
1889 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1890
1891 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1892 freeze_epilog(current, cont, res);
1893 cont.done(); // allow safepoint in the transition back to Java
1894 return res;
1895 JRT_BLOCK_END
1896 }
1897
1898 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1899 ContinuationEntry* entry = thread->last_continuation();
1900 if (entry == nullptr) {
1901 return freeze_ok;
1902 }
1903 if (entry->is_pinned()) {
1943 entry = entry->parent();
1944 if (entry == nullptr) {
1945 break;
1946 }
1947 if (entry->is_pinned()) {
1948 return freeze_pinned_cs;
1949 } else if (monitor_count > 0) {
1950 return freeze_pinned_monitor;
1951 }
1952 }
1953 }
1954 return freeze_ok;
1955 }
1956
1957 /////////////// THAW ////
1958
1959 static int thaw_size(stackChunkOop chunk) {
1960 int size = chunk->max_thawing_size();
1961 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
1962 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
1963 size += frame::metadata_words; // for preemption case (see push_preempt_rerun_adapter)
1964 return size;
1965 }
1966
1967 // make room on the stack for thaw
1968 // returns the size in bytes, or 0 on failure
1969 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
1970 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
1971
1972 assert(thread == JavaThread::current(), "");
1973
1974 ContinuationEntry* ce = thread->last_continuation();
1975 assert(ce != nullptr, "");
1976 oop continuation = ce->cont_oop(thread);
1977 assert(continuation == get_continuation(thread), "");
1978 verify_continuation(continuation);
1979
1980 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
1981 assert(chunk != nullptr, "");
1982
1983 // The tail can be empty because it might still be available for another freeze.
2024
2025 NOT_PRODUCT(int _frames;)
2026
2027 protected:
2028 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2029 _thread(thread), _cont(cont),
2030 _fastpath(nullptr) {
2031 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2032 assert (cont.tail() != nullptr, "no last chunk");
2033 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2034 }
2035
2036 void clear_chunk(stackChunkOop chunk);
2037 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2038 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2039
2040 // fast path
2041 inline void prefetch_chunk_pd(void* start, int size_words);
2042 void patch_return(intptr_t* sp, bool is_last);
2043
2044 intptr_t* handle_preempted_continuation(stackChunkOop original_chunk, intptr_t* sp, bool fast_case);
2045 inline intptr_t* push_preempt_rerun_adapter(frame top, bool is_interpreted_frame);
2046 inline intptr_t* push_preempt_monitorenter_redo(stackChunkOop chunk);
2047
2048 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top);
2049 void finish_thaw(frame& f);
2050
2051 private:
2052 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2053 void finalize_thaw(frame& entry, int argsize);
2054
2055 inline bool seen_by_gc();
2056
2057 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2058 inline void after_thaw_java_frame(const frame& f, bool bottom);
2059 inline void patch(frame& f, const frame& caller, bool bottom);
2060 void clear_bitmap_bits(address start, address end);
2061
2062 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames);
2063 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2064 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2065
2066 void push_return_frame(frame& f);
2067 inline frame new_entry_frame();
2068 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom);
2069 inline void patch_pd(frame& f, const frame& sender);
2070 inline void patch_pd(frame& f, intptr_t* caller_sp);
2071 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2072
2073 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2074
2075 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2076
2077 public:
2078 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2079 };
2080
2081 template <typename ConfigT>
2082 class Thaw : public ThawBase {
2083 public:
2084 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2085
2086 inline bool can_thaw_fast(stackChunkOop chunk) {
2087 return !_barriers
2088 && _thread->cont_fastpath_thread_state()
2089 && !chunk->has_thaw_slowpath_condition()
2090 && !PreserveFramePointer;
2091 }
2092
2093 inline intptr_t* thaw(Continuation::thaw_kind kind);
2094 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2095 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2096 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2097 };
2098
2099 template <typename ConfigT>
2100 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2101 verify_continuation(_cont.continuation());
2102 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2103 assert(!_cont.is_empty(), "");
2104
2105 stackChunkOop chunk = _cont.tail();
2106 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2107 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2108
2109 _barriers = chunk->requires_barriers();
2110 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2111 : thaw_slow(chunk, kind);
2112 }
2113
2114 class ReconstructedStack : public StackObj {
2115 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2116 int _thaw_size;
2117 int _argsize;
2118 public:
2119 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2120 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2121 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2122 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2123 // We're at most one alignment word away from entrySP
2124 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2125 }
2126
2127 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2128
2129 // top and bottom stack pointers
2130 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2131 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2132
2133 // several operations operate on the totality of the stack being reconstructed,
2134 // including the metadata words
2135 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2136 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2137 };
2138
2139 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2140 chunk->set_sp(chunk->stack_size());
2141 chunk->set_argsize(0);
2142 chunk->set_max_thawing_size(0);
2143 }
2144
2145 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2146 bool empty = false;
2147 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2148 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2149 assert(chunk_sp == f.sp(), "");
2150 assert(chunk_sp == f.unextended_sp(), "");
2151
2152 int frame_size = f.cb()->frame_size();
2153 argsize = f.stack_argsize();
2154 bool is_stub = f.is_stub();
2155
2156 f.next(SmallRegisterMap::instance, true /* stop */);
2157 empty = f.is_done();
2158 assert(!empty || argsize == chunk->argsize(), "");
2159
2160 assert(!is_stub || !empty, "runtime stub should have caller frame");
2161 if (is_stub) {
2162 // If we don't thaw the top compiled frame too, after restoring the saved
2163 // registers back in Java, we would hit the return barrier to thaw one more
2164 // frame effectively overwritting the restored registers during that call.
2165 f.get_cb();
2166 frame_size += f.cb()->frame_size();
2167 argsize = f.stack_argsize();
2168 f.next(SmallRegisterMap::instance, true /* stop */);
2169 empty = f.is_done();
2170 assert(!empty || argsize == chunk->argsize(), "");
2171 }
2172
2173 if (empty) {
2174 clear_chunk(chunk);
2175 } else {
2176 chunk->set_sp(chunk->sp() + frame_size);
2177 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2178 // We set chunk->pc to the return pc into the next frame
2179 chunk->set_pc(f.pc());
2180 #ifdef ASSERT
2181 {
2182 intptr_t* retaddr_slot = (chunk_sp
2183 + frame_size
2184 - frame::sender_sp_ret_address_offset());
2185 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2186 "unexpected pc");
2187 }
2188 #endif
2189 }
2190 assert(empty == chunk->is_empty(), "");
2191 // returns the size required to store the frame on stack, and because it is a
2192 // compiled frame, it must include a copy of the arguments passed by the caller
2283 e.commit();
2284 }
2285 #endif
2286
2287 #ifdef ASSERT
2288 set_anchor(_thread, rs.sp());
2289 log_frames(_thread);
2290 if (LoomDeoptAfterThaw) {
2291 do_deopt_after_thaw(_thread);
2292 }
2293 clear_anchor(_thread);
2294 #endif
2295
2296 return rs.sp();
2297 }
2298
2299 inline bool ThawBase::seen_by_gc() {
2300 return _barriers || _cont.tail()->is_gc_mode();
2301 }
2302
2303 template <typename ConfigT>
2304 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2305 bool retry_fast_path = false;
2306
2307 bool preempted_case = chunk->is_preempted();
2308 if (preempted_case) {
2309 assert(_cont.is_preempted(), "must be");
2310 assert(chunk->objectMonitor() != nullptr, "must be");
2311
2312 ObjectMonitor* mon = chunk->objectMonitor();
2313 if (!mon->is_owner(_thread)) {
2314 return push_preempt_monitorenter_redo(chunk);
2315 }
2316 chunk->set_is_preempted(false);
2317 retry_fast_path = true;
2318 }
2319
2320 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2321 if (UseZGC || UseShenandoahGC) {
2322 _cont.tail()->relativize_derived_pointers_concurrently();
2323 }
2324 #endif
2325
2326 // First thaw after freeze. If there were oops in the stacklock
2327 // during freeze, restore them now.
2328 if (chunk->lockStackSize() > 0) {
2329 int lockStackSize = chunk->lockStackSize();
2330 assert(lockStackSize > 0, "should be");
2331
2332 oop tmp_lockstack[8];
2333 chunk->copy_lockstack(tmp_lockstack);
2334 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2335
2336 chunk->set_lockStackSize(0);
2337 chunk->set_has_lockStack(false);
2338 retry_fast_path = true;
2339 }
2340
2341 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2342 // and FLAG_PREEMPTED flags from the stackChunk.
2343 if (retry_fast_path && can_thaw_fast(chunk)) {
2344 intptr_t* sp = thaw_fast(chunk);
2345 if (preempted_case) {
2346 assert(_cont.is_preempted(), "must be");
2347 return handle_preempted_continuation(chunk, sp, true /* fast_case */);
2348 }
2349 return sp;
2350 }
2351
2352 LogTarget(Trace, continuations) lt;
2353 if (lt.develop_is_enabled()) {
2354 LogStream ls(lt);
2355 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2356 chunk->print_on(true, &ls);
2357 }
2358
2359 #if CONT_JFR
2360 EventContinuationThawSlow e;
2361 if (e.should_commit()) {
2362 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2363 e.commit();
2364 }
2365 #endif
2366
2367 DEBUG_ONLY(_frames = 0;)
2368 _align_size = 0;
2369 bool is_return_barrier = kind != Continuation::thaw_top;
2370 int num_frames = (is_return_barrier ? 1 : 2);
2371
2372 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2373 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2374
2375 stackChunkOop original_chunk = chunk;
2376
2377 frame heap_frame = _stream.to_frame();
2378 if (lt.develop_is_enabled()) {
2379 LogStream ls(lt);
2380 ls.print_cr("top hframe before (thaw):");
2381 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2382 heap_frame.print_value_on(&ls, nullptr);
2383 }
2384
2385 frame caller; // the thawed caller on the stack
2386 recurse_thaw(heap_frame, caller, num_frames, true);
2387 finish_thaw(caller); // caller is now the topmost thawed frame
2388 _cont.write();
2389
2390 assert(_cont.chunk_invariant(), "");
2391
2392 JVMTI_ONLY(if (!is_return_barrier) invalidate_jvmti_stack(_thread));
2393
2394 _thread->set_cont_fastpath(_fastpath);
2395
2396 intptr_t* sp = caller.sp();
2397
2398 if (preempted_case) {
2399 assert(_cont.is_preempted(), "must be");
2400 return handle_preempted_continuation(original_chunk, sp, false /* fast_case */);
2401 }
2402 return sp;
2403 }
2404
2405 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top) {
2406 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2407 assert(!_cont.is_empty(), "no more frames");
2408 assert(num_frames > 0, "");
2409 assert(!heap_frame.is_empty(), "");
2410
2411 if (top && ContinuationHelper::Frame::is_stub(heap_frame.cb())) {
2412 recurse_thaw_stub_frame(heap_frame, caller, num_frames);
2413 } else if (!heap_frame.is_interpreted_frame()) {
2414 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2415 } else {
2416 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames);
2417 }
2418 }
2419
2420 template<typename FKind>
2421 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2422 assert(num_frames > 0, "");
2423
2424 DEBUG_ONLY(_frames++;)
2425
2426 int argsize = _stream.stack_argsize();
2427
2428 _stream.next(SmallRegisterMap::instance);
2429 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2430
2431 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk
2513 }
2514
2515 void ThawBase::clear_bitmap_bits(address start, address end) {
2516 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2517 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2518
2519 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2520 // or they will keep objects that are otherwise unreachable alive.
2521
2522 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2523 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2524 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2525 // anyways and we assert that before returning.
2526 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2527 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2528 stackChunkOop chunk = _cont.tail();
2529 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2530 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2531 }
2532
2533 intptr_t* ThawBase::handle_preempted_continuation(stackChunkOop original_chunk, intptr_t* sp, bool fast_case) {
2534 frame top(sp);
2535 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2536 bool top_is_interpreted = Interpreter::contains(top.pc());
2537
2538 if (fast_case) {
2539 assert(ContinuationHelper::Frame::is_stub(top.cb()), "invariant");
2540 int fsize = ContinuationHelper::StubFrame::size(top);
2541 patch_pd(top, sp + fsize);
2542 }
2543
2544 _cont.set_preempted(false);
2545 bool same_chunk = original_chunk == _cont.tail();
2546
2547 ObjectMonitor* mon = original_chunk->objectMonitor();
2548 if (same_chunk && mon != nullptr) {
2549 original_chunk->set_objectMonitor(nullptr);
2550 }
2551
2552 bool post_mount_event = true;
2553 if (_thread->preemption_cancelled()) {
2554 // Since we never actually unmounted don't post the mount event.
2555 post_mount_event = false;
2556 _thread->set_preemption_cancelled(false);
2557 }
2558
2559 #if INCLUDE_JVMTI
2560 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2561 if (is_vthread) {
2562 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2563 jvmti_mount_end(_thread, _cont, top, mon, post_mount_event);
2564 } else {
2565 _thread->set_is_in_VTMS_transition(false);
2566 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2567 }
2568 }
2569 #endif
2570
2571 if (!top_is_interpreted) {
2572 assert(ContinuationHelper::Frame::is_stub(top.cb()), "invariant");
2573 // The continuation might now run on a different platform thread than the previous time so
2574 // we need to adjust the current thread saved in the stub frame before restoring registers.
2575 JavaThread** thread_addr = frame::saved_thread_address(top);
2576 if (thread_addr != nullptr) *thread_addr = _thread;
2577 }
2578 sp = push_preempt_rerun_adapter(top, top_is_interpreted /* is_interpreted_frame */);
2579 return sp;
2580 }
2581
2582 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames) {
2583 assert(hf.is_interpreted_frame(), "");
2584
2585 if (UNLIKELY(seen_by_gc())) {
2586 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance);
2587 }
2588
2589 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2590
2591 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2592
2593 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2594
2595 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2596
2597 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2598 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2599 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2600 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2601
2622
2623 maybe_set_fastpath(f.sp());
2624
2625 const int locals = hf.interpreter_frame_method()->max_locals();
2626
2627 if (!is_bottom_frame) {
2628 // can only fix caller once this frame is thawed (due to callee saved regs)
2629 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance);
2630 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2631 assert(hf.is_heap_frame(), "should be");
2632 address start = (address)(heap_frame_bottom - locals);
2633 address end = (address)heap_frame_bottom;
2634 clear_bitmap_bits(start, end);
2635 }
2636
2637 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2638 caller = f;
2639 }
2640
2641 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2642 assert(hf.is_compiled_frame(), "");
2643 assert(_cont.is_preempted() || !stub_caller, "stub caller not at preemption");
2644
2645 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2646 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance);
2647 }
2648
2649 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2650
2651 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2652
2653 assert(caller.sp() == caller.unextended_sp(), "");
2654
2655 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2656 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2657 }
2658
2659 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2660 // yet laid out in the stack, and so the original_pc is not stored in it.
2661 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2662 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame);
2663 intptr_t* const stack_frame_top = f.sp();
2664 intptr_t* const heap_frame_top = hf.unextended_sp();
2665
2666 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2667 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize;
2668 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2669
2670 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2671 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2672 // copy metadata, except the metadata at the top of the (unextended) entry frame
2673 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2674
2675 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2676 // (we might have one padding word for alignment)
2677 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2678 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2679
2680 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2681
2682 patch(f, caller, is_bottom_frame);
2683
2684 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2685 assert(!f.is_deoptimized_frame(), "");
2686 if (hf.is_deoptimized_frame()) {
2687 maybe_set_fastpath(f.sp());
2688 } else if (_thread->is_interp_only_mode()
2689 || (_cont.is_preempted() && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2690 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2691 // cannot rely on nmethod patching for deopt.
2692
2693 log_develop_trace(continuations)("Deoptimizing thawed frame");
2694 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2695
2696 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2697 assert(f.is_deoptimized_frame(), "");
2698 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2699 maybe_set_fastpath(f.sp());
2700 }
2701
2702 if (!is_bottom_frame) {
2703 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2704 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance);
2705 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2706 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2707 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2708 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2709 clear_bitmap_bits(start, start + argsize_in_bytes);
2710 }
2711
2712 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2713 caller = f;
2714 }
2715
2716 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2717 DEBUG_ONLY(_frames++;)
2718 bool is_bottom_frame = false;
2719
2720 if (UNLIKELY(seen_by_gc())) {
2721 // Process the stub's caller here since we might need the full map (if the stub was
2722 // generated on a poll on return we shouldn't need a full map).
2723 RegisterMap map(nullptr,
2724 RegisterMap::UpdateMap::include,
2725 RegisterMap::ProcessFrames::skip,
2726 RegisterMap::WalkContinuation::skip);
2727 map.set_include_argument_oops(false);
2728 _stream.next(&map);
2729 if (!_stream.is_done()) {
2730 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2731 }
2732 } else {
2733 _stream.next(SmallRegisterMap::instance);
2734 }
2735
2736 if (_stream.is_done()) {
2737 finalize_thaw(caller, 0);
2738 is_bottom_frame = true;
2739 } else {
2740 frame f = _stream.to_frame();
2741 if (f.is_interpreted_frame()) {
2742 recurse_thaw_interpreted_frame(f, caller, num_frames);
2743 } else {
2744 recurse_thaw_compiled_frame(f, caller, num_frames, true);
2745 }
2746 }
2747
2748 assert(!is_bottom_frame || !_cont.is_empty(), "");
2749 assert(caller.sp() == caller.unextended_sp(), "");
2750 assert(!caller.is_native_frame() || (is_bottom_frame && Continuation::is_continuation_enterSpecial(caller)), "caller should't be native except for enterSpecial case");
2751
2752 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2753
2754 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2755 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_stub_frame
2756 }
2757
2758 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2759 intptr_t* stack_frame_top = f.sp();
2760 intptr_t* heap_frame_top = hf.sp();
2761 int fsize = ContinuationHelper::StubFrame::size(hf);
2762
2763 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2764 fsize + frame::metadata_words);
2765
2766 patch(f, caller, is_bottom_frame);
2767
2768 if (!is_bottom_frame) {
2769 // can only fix caller once this frame is thawed (due to callee saved regs)
2770 RegisterMap map(nullptr,
2771 RegisterMap::UpdateMap::include,
2772 RegisterMap::ProcessFrames::skip,
2773 RegisterMap::WalkContinuation::skip);
2774 map.set_include_argument_oops(false);
2775 f.oop_map()->update_register_map(&f, &map);
2776 ContinuationHelper::update_register_map_with_callee(caller, &map);
2777 _cont.tail()->fix_thawed_frame(caller, &map);
2778 }
2779
2780 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2781 caller = f;
2782 }
2783
2784 void ThawBase::finish_thaw(frame& f) {
2785 stackChunkOop chunk = _cont.tail();
2786
2787 if (chunk->is_empty()) {
2788 // Only remove chunk from list if it can't be reused for another freeze
2789 if (seen_by_gc()) {
2790 _cont.set_tail(chunk->parent());
2791 } else {
2792 chunk->set_has_mixed_frames(false);
2793 }
2794 chunk->set_max_thawing_size(0);
2795 assert(chunk->argsize() == 0, "");
2796 } else {
2797 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
2798 }
2799 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
2800
2852
2853 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
2854 assert(oopCont == get_continuation(thread), "");
2855 verify_continuation(oopCont);
2856
2857 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
2858
2859 ContinuationWrapper cont(thread, oopCont);
2860 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
2861
2862 #ifdef ASSERT
2863 set_anchor_to_entry(thread, cont.entry());
2864 log_frames(thread);
2865 clear_anchor(thread);
2866 #endif
2867
2868 Thaw<ConfigT> thw(thread, cont);
2869 intptr_t* const sp = thw.thaw(kind);
2870 assert(is_aligned(sp, frame::frame_alignment), "");
2871
2872 #ifdef ASSERT
2873 intptr_t* sp0 = sp;
2874 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
2875
2876 bool sent_entry_pc = false;
2877 if (pc0 == Interpreter::cont_preempt_rerun_interpreter_adapter() ||
2878 pc0 == StubRoutines::cont_preempt_rerun_compiler_adapter()) {
2879 sp0 += frame::metadata_words; // see push_preempt_rerun_adapter
2880 #ifdef AARCH64
2881 if (pc0 == StubRoutines::cont_preempt_rerun_compiler_adapter()) {
2882 address pc1 = *(address*)(sp0 - frame::sender_sp_ret_address_offset());
2883 CodeBlob* cb = CodeCache::find_blob(pc1);
2884 assert(cb != nullptr, "should be either c1 or c2 runtime stub");
2885 if (cb->frame_size() == 2) {
2886 sp0 += frame::metadata_words;
2887 }
2888 }
2889 #endif
2890 } else if (pc0 == StubRoutines::cont_preempt_monitorenter_redo()) {
2891 sp0 += 2 * frame::metadata_words; // see push_preempt_monitorenter_redo
2892 sent_entry_pc = true;
2893 }
2894 set_anchor(thread, sp0, sent_entry_pc ? cont.entryPC() : nullptr);
2895 log_frames(thread);
2896 if (LoomVerifyAfterThaw) {
2897 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
2898 }
2899 assert(ContinuationEntry::assert_entry_frame_laid_out(thread), "");
2900 clear_anchor(thread);
2901
2902 LogTarget(Trace, continuations) lt;
2903 if (lt.develop_is_enabled()) {
2904 LogStream ls(lt);
2905 ls.print_cr("Jumping to frame (thaw):");
2906 frame(sp).print_value_on(&ls, nullptr);
2907 }
2908 #endif
2909
2910 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
2911
2912 verify_continuation(cont.continuation());
2913 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
2914
2986 st->print_cr("size: %d argsize: %d",
2987 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
2988 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
2989 }
2990 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
2991 if (reg != nullptr) {
2992 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
2993 }
2994 cl.reset();
2995 DEBUG_ONLY(thread->print_frame_layout();)
2996 if (chunk != nullptr) {
2997 chunk->print_on(true, st);
2998 }
2999 return false;
3000 }
3001 }
3002 return true;
3003 }
3004
3005 static void log_frames(JavaThread* thread) {
3006 const static int show_entry_callers = 100;
3007 LogTarget(Trace, continuations) lt;
3008 if (!lt.develop_is_enabled()) {
3009 return;
3010 }
3011 LogStream ls(lt);
3012
3013 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3014 if (!thread->has_last_Java_frame()) {
3015 ls.print_cr("NO ANCHOR!");
3016 }
3017
3018 RegisterMap map(thread,
3019 RegisterMap::UpdateMap::include,
3020 RegisterMap::ProcessFrames::include,
3021 RegisterMap::WalkContinuation::skip);
3022 map.set_include_argument_oops(false);
3023
3024 if (false) {
3025 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3026 f.print_on(&ls);
3027 }
3028 } else {
3029 map.set_skip_missing(true);
3030 ResetNoHandleMark rnhm;
3031 ResourceMark rm;
3032 HandleMark hm(Thread::current());
3033 FrameValues values;
3034
3035 int i = 0;
3036 int post_entry = -1;
3037 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3038 f.describe(values, i, &map, i == 0);
3039 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3040 post_entry++;
3041 if (post_entry >= show_entry_callers)
3042 break;
3043 }
3044 values.print_on(thread, &ls);
3045 }
3046
3047 ls.print_cr("======= end frames =========");
3048 }
3049 #endif // ASSERT
3050
3051 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3052
3053 #ifdef ASSERT
3054 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3055 ResourceMark rm;
3056 FrameValues values;
3057 assert(f.get_cb() != nullptr, "");
3058 RegisterMap map(f.is_heap_frame() ?
3059 nullptr :
3060 JavaThread::current(),
3061 RegisterMap::UpdateMap::include,
3062 RegisterMap::ProcessFrames::skip,
3063 RegisterMap::WalkContinuation::skip);
3064 map.set_include_argument_oops(false);
3065 map.set_skip_missing(true);
3066 if (callee_complete) {
3067 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3068 }
3069 const_cast<frame&>(f).describe(values, 0, &map, true);
3070 values.print_on(static_cast<JavaThread*>(nullptr), st);
3071 }
3072 #endif
3073
3074 static address thaw_entry = nullptr;
3075 static address freeze_entry = nullptr;
3076 static address freeze_preempt_entry = nullptr;
3077
3078 address Continuation::thaw_entry() {
3079 return ::thaw_entry;
3080 }
3081
3082 address Continuation::freeze_entry() {
3083 return ::freeze_entry;
3084 }
3085
3086 address Continuation::freeze_preempt_entry() {
3087 return ::freeze_preempt_entry;
3088 }
3089
3090 class ConfigResolve {
3091 public:
3092 static void resolve() { resolve_compressed(); }
3093
3094 static void resolve_compressed() {
3095 UseCompressedOops ? resolve_gc<true>()
3096 : resolve_gc<false>();
3097 }
3098
3099 private:
3100 template <bool use_compressed>
3101 static void resolve_gc() {
3102 BarrierSet* bs = BarrierSet::barrier_set();
3103 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3104 switch (bs->kind()) {
3105 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3106 case BarrierSet::bs_name: { \
3107 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3108 } \
3109 break;
3110 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3111 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3112
3113 default:
3114 fatal("BarrierSet resolving not implemented");
3115 };
3116 }
3117
3118 template <bool use_compressed, typename BarrierSetT>
3119 static void resolve() {
3120 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3121
3122 freeze_entry = (address)freeze<SelectedConfigT>;
3123 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3124
3125 // If we wanted, we could templatize by kind and have three different thaw entries
3126 thaw_entry = (address)thaw<SelectedConfigT>;
3127 }
3128 };
3129
3130 void Continuation::init() {
3131 ConfigResolve::resolve();
3132 }
|