1 /*
2 * Copyright (c) 2018, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "classfile/javaClasses.inline.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "code/codeCache.inline.hpp"
28 #include "code/nmethod.inline.hpp"
29 #include "code/vmreg.inline.hpp"
30 #include "compiler/oopMap.inline.hpp"
31 #include "cppstdlib/type_traits.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/continuationGCSupport.inline.hpp"
34 #include "gc/shared/gc_globals.hpp"
35 #include "gc/shared/memAllocator.hpp"
36 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
37 #include "interpreter/bytecodeStream.hpp"
38 #include "interpreter/interpreter.hpp"
39 #include "interpreter/interpreterRuntime.hpp"
40 #include "jfr/jfrEvents.hpp"
41 #include "logging/log.hpp"
42 #include "logging/logStream.hpp"
43 #include "oops/access.inline.hpp"
44 #include "oops/constantPool.inline.hpp"
45 #include "oops/method.inline.hpp"
46 #include "oops/objArrayOop.inline.hpp"
47 #include "oops/oopsHierarchy.hpp"
48 #include "oops/stackChunkOop.inline.hpp"
49 #include "prims/jvmtiThreadState.hpp"
50 #include "runtime/arguments.hpp"
51 #include "runtime/continuation.hpp"
52 #include "runtime/continuationEntry.inline.hpp"
53 #include "runtime/continuationHelper.inline.hpp"
54 #include "runtime/continuationJavaClasses.inline.hpp"
55 #include "runtime/continuationWrapper.inline.hpp"
56 #include "runtime/frame.inline.hpp"
57 #include "runtime/interfaceSupport.inline.hpp"
58 #include "runtime/javaThread.inline.hpp"
59 #include "runtime/jniHandles.inline.hpp"
60 #include "runtime/keepStackGCProcessed.hpp"
61 #include "runtime/mountUnmountDisabler.hpp"
62 #include "runtime/objectMonitor.inline.hpp"
63 #include "runtime/orderAccess.hpp"
64 #include "runtime/prefetch.inline.hpp"
65 #include "runtime/sharedRuntime.hpp"
66 #include "runtime/smallRegisterMap.inline.hpp"
67 #include "runtime/stackChunkFrameStream.inline.hpp"
68 #include "runtime/stackFrameStream.inline.hpp"
69 #include "runtime/stackOverflow.hpp"
70 #include "runtime/stackWatermarkSet.inline.hpp"
71 #include "runtime/vframe.inline.hpp"
72 #include "runtime/vframe_hp.hpp"
73 #include "utilities/debug.hpp"
74 #include "utilities/exceptions.hpp"
75 #include "utilities/macros.hpp"
76 #include "utilities/vmError.hpp"
77 #if INCLUDE_SHENANDOAHGC
78 #include "gc/shenandoah/shenandoahStackChunkGCData.inline.hpp"
79 #endif
80 #if INCLUDE_ZGC
81 #include "gc/z/zStackChunkGCData.inline.hpp"
82 #endif
83 #if INCLUDE_JFR
84 #include "jfr/jfr.inline.hpp"
85 #endif
86 #ifdef COMPILER1
87 #include "c1/c1_Runtime1.hpp"
88 #endif
89 #ifdef COMPILER2
90 #include "opto/runtime.hpp"
91 #endif
92
93 /*
94 * This file contains the implementation of continuation freezing (yield) and thawing (run).
95 *
96 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
97 * would likely call these operations many thousands of times per second second, on every core.
98 *
99 * Freeze might be called every time the application performs any I/O operation, every time it
100 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
101 * multiple times in each of those cases, as it is called by the return barrier, which may be
102 * invoked on method return.
103 *
104 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
105 * example, every effort is made to avoid Java-VM transitions as much as possible.
106 *
107 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
108 * and so frames simply copied, and the bottom-most one is patched.
109 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
110 * and absolute pointers, and barriers invoked.
111 */
112
113 /************************************************
114
115 Thread-stack layout on freeze/thaw.
116 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
117
118 +----------------------------+
119 | . |
120 | . |
121 | . |
122 | carrier frames |
123 | |
124 |----------------------------|
125 | |
126 | Continuation.run |
127 | |
128 |============================|
129 | enterSpecial frame |
130 | pc |
131 | rbp |
132 | ----- |
133 ^ | int argsize | = ContinuationEntry
134 | | oopDesc* cont |
135 | | oopDesc* chunk |
136 | | ContinuationEntry* parent |
137 | | ... |
138 | |============================| <------ JavaThread::_cont_entry = entry->sp()
139 | | ? alignment word ? |
140 | |----------------------------| <--\
141 | | | |
142 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
143 Address | | | | Caller is still in the chunk.
144 | |----------------------------| |
145 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
146 | | rbp | | isn't the last one in the continuation.
147 | | | |
148 | | frame | |
149 | | | |
150 +----------------------------| \__ Continuation frames to be frozen/thawed
151 | | /
152 | frame | |
153 | | |
154 |----------------------------| |
155 | | |
156 | frame | |
157 | | |
158 |----------------------------| <--/
159 | |
160 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
161 | | instead of a doYield stub
162 |============================| <- the sp passed to freeze
163 | |
164 | Native freeze/thaw frames |
165 | . |
166 | . |
167 | . |
168 +----------------------------+
169
170 ************************************************/
171
172 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
173 #if CONT_JFR
174 #define CONT_JFR_ONLY(code) code
175 #else
176 #define CONT_JFR_ONLY(code)
177 #endif
178
179 // TODO: See AbstractAssembler::generate_stack_overflow_check,
180 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
181 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
182
183 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
184
185 // Used to just annotatate cold/hot branches
186 #define LIKELY(condition) (condition)
187 #define UNLIKELY(condition) (condition)
188
189 // debugging functions
190 #ifdef ASSERT
191 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
192
193 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
194
195 static void do_deopt_after_thaw(JavaThread* thread);
196 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
197 static bool verify_deopt_state(const frame& f);
198 static void log_frames(JavaThread* thread);
199 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp);
200 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
201 static void verify_frame_kind(frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr = nullptr, const char** code_name_ptr = nullptr, int* bci_ptr = nullptr, stackChunkOop chunk = nullptr);
202
203 #define assert_pfl(p, ...) \
204 do { \
205 if (!(p)) { \
206 JavaThread* t = JavaThread::active(); \
207 if (t->has_last_Java_frame()) { \
208 tty->print_cr("assert(" #p ") failed:"); \
209 t->print_frame_layout(); \
210 } \
211 } \
212 vmassert(p, __VA_ARGS__); \
213 } while(0)
214
215 #else
216 static void verify_continuation(oop continuation) { }
217 #define assert_pfl(p, ...)
218 #endif
219
220 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
221 template<typename ConfigT, bool preempt> static inline freeze_result 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 // Entry point to freeze. Transitions are handled manually
227 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
228 template<typename ConfigT>
229 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
230 assert(sp == current->frame_anchor()->last_Java_sp(), "");
231
232 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
233 current->set_cont_fastpath(nullptr);
234 }
235
236 return checked_cast<int>(ConfigT::freeze(current, sp));
237 JRT_END
238
239 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
240 return prepare_thaw_internal(thread, return_barrier);
241 JRT_END
242
243 template<typename ConfigT>
244 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
245 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
246 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
247 // Also the preemption case with JVMTI events enabled might safepoint so
248 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
249 // JRT_ENTRY instead?
250 ResetNoHandleMark rnhm;
251 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
252
253 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
254 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
255 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
256 JRT_END
257
258 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
259 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
260 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
261 }
262 JVM_END
263
264 ///////////
265
266 enum class oop_kind { NARROW, WIDE };
267 template <oop_kind oops, typename BarrierSetT>
268 class Config {
269 public:
270 typedef Config<oops, BarrierSetT> SelfT;
271 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
272
273 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
274 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
275 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
276 return res;
277 }
278
279 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
280 return freeze_internal<SelfT, true>(thread, sp);
281 }
282
283 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
284 return thaw_internal<SelfT>(thread, kind);
285 }
286 };
287
288 #ifdef _WINDOWS
289 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
290 address new_sp = sp - size;
291 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
292
293 if (new_sp < watermark) {
294 size_t page_size = os::vm_page_size();
295 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
296 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
297 while (pages_to_touch-- > 0) {
298 last_touched_page -= page_size;
299 *last_touched_page = 0;
300 }
301 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
302 }
303 }
304 #endif
305
306 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
307 const size_t page_size = os::vm_page_size();
308 if (size > page_size) {
309 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
310 return false;
311 }
312 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
313 }
314 return true;
315 }
316
317 #ifdef ASSERT
318 static oop get_continuation(JavaThread* thread) {
319 assert(thread != nullptr, "");
320 assert(thread->threadObj() != nullptr, "");
321 return java_lang_Thread::continuation(thread->threadObj());
322 }
323 #endif // ASSERT
324
325 inline void clear_anchor(JavaThread* thread) {
326 thread->frame_anchor()->clear();
327 }
328
329 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
330 assert(pc != nullptr, "");
331
332 JavaFrameAnchor* anchor = thread->frame_anchor();
333 anchor->set_last_Java_sp(sp);
334 anchor->set_last_Java_pc(pc);
335 ContinuationHelper::set_anchor_pd(anchor, sp);
336
337 assert(thread->has_last_Java_frame(), "");
338 assert(thread->last_frame().cb() != nullptr, "");
339 }
340
341 static void set_anchor(JavaThread* thread, intptr_t* sp) {
342 address pc = ContinuationHelper::return_address_at(
343 sp - frame::sender_sp_ret_address_offset());
344 set_anchor(thread, sp, pc);
345 }
346
347 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
348 JavaFrameAnchor* anchor = thread->frame_anchor();
349 anchor->set_last_Java_sp(entry->entry_sp());
350 anchor->set_last_Java_pc(entry->entry_pc());
351 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
352
353 assert(thread->has_last_Java_frame(), "");
354 assert(thread->last_frame().cb() != nullptr, "");
355 }
356
357 #if CONT_JFR
358 class FreezeThawJfrInfo : public StackObj {
359 short _e_size;
360 short _e_num_interpreted_frames;
361 public:
362
363 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
364 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
365 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
366 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
367 };
368
369 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
370 if (e->should_commit()) {
371 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
372 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
373 e->set_continuationClass(continuation->klass());
374 e->set_interpretedFrames(_e_num_interpreted_frames);
375 e->set_size(_e_size);
376 e->commit();
377 }
378 }
379 #endif // CONT_JFR
380
381 /////////////// FREEZE ////
382
383 class FreezeBase : public StackObj {
384 protected:
385 JavaThread* const _thread;
386 ContinuationWrapper& _cont;
387 bool _barriers; // only set when we allocate a chunk
388
389 intptr_t* _bottom_address;
390
391 // Used for preemption only
392 const bool _preempt;
393 frame _last_frame;
394
395 // Used to support freezing with held monitors
396 int _monitors_in_lockstack;
397
398 int _freeze_size; // total size of all frames plus metadata in words.
399 int _total_align_size;
400
401 intptr_t* _cont_stack_top;
402 intptr_t* _cont_stack_bottom;
403
404 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
405
406 #ifdef ASSERT
407 intptr_t* _orig_chunk_sp;
408 int _fast_freeze_size;
409 bool _empty;
410 #endif
411
412 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
413
414 NOT_PRODUCT(int _frames;)
415 DEBUG_ONLY(intptr_t* _last_write;)
416
417 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
418
419 public:
420 NOINLINE freeze_result freeze_slow();
421 void freeze_fast_existing_chunk();
422
423 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
424 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
425
426 inline int size_if_fast_freeze_available();
427
428 inline frame& last_frame() { return _last_frame; }
429
430 #ifdef ASSERT
431 bool check_valid_fast_path();
432 #endif
433
434 protected:
435 inline void init_rest();
436 void throw_stack_overflow_on_humongous_chunk();
437
438 // fast path
439 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
440 inline void unwind_frames();
441 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
442
443 // slow path
444 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
445
446 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
447
448 private:
449 // slow path
450 frame freeze_start_frame();
451 frame freeze_start_frame_on_preempt();
452 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
453 inline frame freeze_start_frame_yield_stub();
454 template<typename FKind>
455 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
456 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
457 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
458 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
459 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
460 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
461 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
462 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
463 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
464 NOINLINE void finish_freeze(const frame& f, const frame& top);
465
466 void freeze_lockstack(stackChunkOop chunk);
467
468 inline bool stack_overflow();
469
470 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
471 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
472 template<typename FKind> static inline frame sender(const frame& f);
473 template<typename FKind> frame new_heap_frame(frame& f, frame& caller, int size_adjust = 0);
474 inline void set_top_frame_metadata_pd(const frame& hf);
475 inline void patch_pd(frame& callee, const frame& caller, bool is_bottom_frame);
476 inline void patch_pd_unused(intptr_t* sp);
477 void adjust_interpreted_frame_unextended_sp(frame& f);
478 inline void prepare_freeze_interpreted_top_frame(frame& f);
479 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
480
481 protected:
482 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
483 bool freeze_fast_new_chunk(stackChunkOop chunk);
484 };
485
486 template <typename ConfigT>
487 class Freeze : public FreezeBase {
488 private:
489 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
490
491 public:
492 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
493 : FreezeBase(thread, cont, frame_sp, preempt) {}
494
495 freeze_result try_freeze_fast();
496
497 protected:
498 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
499 };
500
501 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
502 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
503 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
504
505 assert(_thread != nullptr, "");
506 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
507
508 DEBUG_ONLY(_cont.entry()->verify_cookie();)
509
510 assert(!Interpreter::contains(_cont.entryPC()), "");
511
512 _bottom_address = align_down(_cont.entrySP() - _cont.entry_frame_extension(), frame::frame_alignment);
513
514 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
515 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
516 assert(_bottom_address != nullptr, "");
517 assert(_bottom_address <= _cont.entrySP(), "");
518 DEBUG_ONLY(_last_write = nullptr;)
519
520 assert(_cont.chunk_invariant(), "");
521 assert(!Interpreter::contains(_cont.entryPC()), "");
522 #if defined(PPC64) && !defined(ZERO)
523 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
524 #elif defined(S390) && !defined(ZERO)
525 static const int doYield_stub_frame_size = frame::z_abi_160_base_size >> LogBytesPerWord;
526 #else
527 static const int doYield_stub_frame_size = frame::metadata_words;
528 #endif
529 // With preemption doYield() might not have been resolved yet
530 assert(_preempt || SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size,
531 "_preempt = %d, cont_doYield_stub()->frame_size() = %d, doYield_stub_frame_size = %d",
532 (_preempt ? 1 : 0), SharedRuntime::cont_doYield_stub()->frame_size(), doYield_stub_frame_size);
533
534 if (preempt) {
535 _last_frame = _thread->last_frame();
536 }
537
538 // properties of the continuation on the stack; all sizes are in words
539 _cont_stack_top = frame_sp + (!preempt ? doYield_stub_frame_size : 0); // we don't freeze the doYield stub frame
540 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
541 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
542
543 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
544 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
545 assert(cont_size() > 0, "");
546
547 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
548 }
549
550 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
551 _freeze_size = 0;
552 _total_align_size = 0;
553 NOT_PRODUCT(_frames = 0;)
554 }
555
556 void FreezeBase::freeze_lockstack(stackChunkOop chunk) {
557 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
558
559 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
560 chunk->set_lockstack_size(checked_cast<uint8_t>(_monitors_in_lockstack));
561 chunk->set_has_lockstack(true);
562 }
563
564 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
565 stackChunkOop chunk = _cont.tail();
566 chunk->copy_from_stack_to_chunk(from, to, size);
567 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
568
569 #ifdef ASSERT
570 if (_last_write != nullptr) {
571 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
572 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
573 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
574 _last_write = to;
575 }
576 #endif
577 }
578
579 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
580 #ifdef ASSERT
581 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
582 if (watermark == nullptr) {
583 return;
584 }
585 ContinuationEntry* ce = thread->last_continuation();
586 RegisterMap map(thread,
587 RegisterMap::UpdateMap::include,
588 RegisterMap::ProcessFrames::skip,
589 RegisterMap::WalkContinuation::skip);
590 map.set_include_argument_oops(false);
591 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
592 watermark->assert_is_frame_safe(f);
593 }
594 #endif // ASSERT
595 }
596
597 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
598 void FreezeBase::unwind_frames() {
599 ContinuationEntry* entry = _cont.entry();
600 entry->flush_stack_processing(_thread);
601 assert_frames_in_continuation_are_safe(_thread);
602 JFR_ONLY(Jfr::check_and_process_sample_request(_thread);)
603 set_anchor_to_entry(_thread, entry);
604 }
605
606 template <typename ConfigT>
607 freeze_result Freeze<ConfigT>::try_freeze_fast() {
608 assert(_thread->thread_state() == _thread_in_vm, "");
609 assert(_thread->cont_fastpath(), "");
610
611 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
612 assert(_fast_freeze_size == 0, "");
613
614 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
615 if (freeze_fast_new_chunk(chunk)) {
616 return freeze_ok;
617 }
618 if (_thread->has_pending_exception()) {
619 return freeze_exception;
620 }
621
622 // TODO R REMOVE when deopt change is fixed
623 assert(!_thread->cont_fastpath() || _barriers, "");
624 log_develop_trace(continuations)("-- RETRYING SLOW --");
625 return freeze_slow();
626 }
627
628 // Returns size needed if the continuation fits, otherwise 0.
629 int FreezeBase::size_if_fast_freeze_available() {
630 stackChunkOop chunk = _cont.tail();
631 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
632 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
633 return 0;
634 }
635
636 int total_size_needed = cont_size();
637 const int chunk_sp = chunk->sp();
638
639 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
640 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
641 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
642 // although that would require changing stackChunkOopDesc::is_empty
643 if (!chunk->is_empty()) {
644 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
645 }
646
647 total_size_needed += _monitors_in_lockstack;
648
649 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
650 bool available = chunk_free_room >= total_size_needed;
651 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
652 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
653 return available ? total_size_needed : 0;
654 }
655
656 void FreezeBase::freeze_fast_existing_chunk() {
657 stackChunkOop chunk = _cont.tail();
658
659 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
660 assert(_fast_freeze_size > 0, "");
661
662 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
663 DEBUG_ONLY(_empty = false;)
664 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
665 #ifdef ASSERT
666 {
667 intptr_t* retaddr_slot = (chunk->sp_address()
668 - frame::sender_sp_ret_address_offset());
669 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
670 "unexpected saved return address");
671 }
672 #endif
673
674 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
675 // we overlap; we'll overwrite the chunk's top frame's callee arguments
676 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
677 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
678
679 // increase max_size by what we're freezing minus the overlap
680 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
681
682 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
683 assert(bottom_sp == _bottom_address, "");
684 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
685 // should have a return barrier (installed back when we thawed it).
686 #ifdef ASSERT
687 {
688 intptr_t* retaddr_slot = (bottom_sp
689 - frame::sender_sp_ret_address_offset());
690 assert(ContinuationHelper::return_address_at(retaddr_slot)
691 == StubRoutines::cont_returnBarrier(),
692 "should be the continuation return barrier");
693 }
694 #endif
695 // We copy the fp from the chunk back to the stack because it contains some caller data,
696 // including, possibly, an oop that might have gone stale since we thawed.
697 patch_stack_pd(bottom_sp, chunk->sp_address());
698 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
699
700 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
701 } else { // the chunk is empty
702 const int chunk_start_sp = chunk->stack_size();
703
704 DEBUG_ONLY(_empty = true;)
705 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
706
707 chunk->set_max_thawing_size(cont_size());
708 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
709 chunk->set_sp(chunk->bottom());
710
711 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
712 }
713 }
714
715 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
716 DEBUG_ONLY(_empty = true;)
717
718 // Install new chunk
719 _cont.set_tail(chunk);
720
721 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
722 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
723 return false;
724 }
725
726 chunk->set_max_thawing_size(cont_size());
727
728 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
729 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
730 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
731 assert(chunk_start_sp == chunk->stack_size(), "");
732
733 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
734
735 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
736
737 return true;
738 }
739
740 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
741 assert(chunk != nullptr, "");
742 assert(!chunk->has_mixed_frames(), "");
743 assert(!chunk->is_gc_mode(), "");
744 assert(!chunk->has_bitmap(), "");
745 assert(!chunk->requires_barriers(), "");
746 assert(chunk == _cont.tail(), "");
747
748 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
749 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
750 // will either see no continuation on the stack, or a consistent chunk.
751 unwind_frames();
752
753 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
754 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
755 assert(chunk_start_sp <= chunk->stack_size(), "");
756 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
757
758 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
759 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
760
761 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
762 #ifdef ASSERT
763 if (!_empty) {
764 intptr_t* retaddr_slot = (_orig_chunk_sp
765 - frame::sender_sp_ret_address_offset());
766 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
767 "unexpected saved return address");
768 }
769 #endif
770
771 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
772 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
773
774 int adjust = frame::metadata_words_at_bottom;
775 #if INCLUDE_ASAN && defined(AARCH64)
776 // Reading at offset frame::metadata_words_at_bottom from _cont_stack_top
777 // will accesss memory at the callee frame, which on preemption cases will
778 // be the VM native method being called. The Arm 64-bit ABI doesn't specify
779 // a location where the frame record (returnpc+fp) has to be stored within
780 // a stack frame, and GCC currently chooses to save it at the top of the
781 // frame (lowest address). ASan treats this memory access in the callee as
782 // an overflow access to one of the locals stored in that frame. For these
783 // preemption cases we don't need to read these words anyways so we avoid it.
784 if (_preempt) {
785 adjust = 0;
786 }
787 #endif
788 intptr_t* from = _cont_stack_top - adjust;
789 intptr_t* to = chunk_top - adjust;
790 copy_to_chunk(from, to, cont_size() + adjust);
791 // Because we're not patched yet, the chunk is now in a bad state
792
793 // patch return pc of the bottom-most frozen frame (now in the chunk)
794 // with the actual caller's return address
795 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
796 - _cont.argsize()
797 - frame::metadata_words_at_top
798 - frame::sender_sp_ret_address_offset());
799 #ifdef ASSERT
800 if (!_empty) {
801 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
802 == StubRoutines::cont_returnBarrier(),
803 "should be the continuation return barrier");
804 }
805 #endif
806 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
807 chunk->pc());
808
809 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
810 chunk->set_sp(chunk_new_sp);
811
812 // set chunk->pc to the return address of the topmost frame in the chunk
813 if (_preempt) {
814 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
815 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
816 // be the pc we used when creating the oopmap. Get the top's frame last pc from
817 // the anchor instead.
818 address last_pc = _last_frame.pc();
819 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
820 chunk->set_pc(last_pc);
821 // For stub/native frames the fp is not used while frozen, and will be constructed
822 // again when thawing the frame (see ThawBase::handle_preempted_continuation). We
823 // patch it with a special bad address to help with debugging, particularly when
824 // inspecting frames and identifying invalid accesses.
825 patch_pd_unused(chunk_top);
826 } else {
827 chunk->set_pc(ContinuationHelper::return_address_at(
828 _cont_stack_top - frame::sender_sp_ret_address_offset()));
829 }
830
831 if (_monitors_in_lockstack > 0) {
832 freeze_lockstack(chunk);
833 }
834
835 _cont.write();
836
837 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
838 LogTarget(Trace, continuations) lt;
839 if (lt.develop_is_enabled()) {
840 LogStream ls(lt);
841 chunk->print_on(true, &ls);
842 }
843
844 // Verification
845 assert(_cont.chunk_invariant(), "");
846 chunk->verify();
847
848 #if CONT_JFR
849 EventContinuationFreezeFast e;
850 if (e.should_commit()) {
851 e.set_id(cast_from_oop<u8>(chunk));
852 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
853 e.set_size(cont_size() << LogBytesPerWord);
854 e.commit();
855 }
856 #endif
857 }
858
859 NOINLINE freeze_result FreezeBase::freeze_slow() {
860 #ifdef ASSERT
861 ResourceMark rm;
862 #endif
863
864 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
865 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
866
867 #if CONT_JFR
868 EventContinuationFreezeSlow e;
869 if (e.should_commit()) {
870 e.set_id(cast_from_oop<u8>(_cont.continuation()));
871 e.commit();
872 }
873 #endif
874
875 init_rest();
876
877 HandleMark hm(Thread::current());
878
879 frame f = freeze_start_frame();
880
881 LogTarget(Debug, continuations) lt;
882 if (lt.develop_is_enabled()) {
883 LogStream ls(lt);
884 f.print_on(&ls);
885 }
886
887 frame caller; // the frozen caller in the chunk
888 freeze_result res = recurse_freeze(f, caller, 0, false, true);
889
890 if (res == freeze_ok) {
891 finish_freeze(f, caller);
892 _cont.write();
893 }
894
895 return res;
896 }
897
898 frame FreezeBase::freeze_start_frame() {
899 if (LIKELY(!_preempt)) {
900 return freeze_start_frame_yield_stub();
901 } else {
902 return freeze_start_frame_on_preempt();
903 }
904 }
905
906 frame FreezeBase::freeze_start_frame_yield_stub() {
907 frame f = _thread->last_frame();
908 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be");
909 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
910 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
911 return f;
912 }
913
914 frame FreezeBase::freeze_start_frame_on_preempt() {
915 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
916 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
917 return _last_frame;
918 }
919
920 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
921 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
922 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
923 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
924 || ((top && _preempt) == f.is_native_frame()), "");
925
926 if (stack_overflow()) {
927 return freeze_exception;
928 }
929
930 if (f.is_compiled_frame()) {
931 if (UNLIKELY(f.oop_map() == nullptr)) {
932 // special native frame
933 return freeze_pinned_native;
934 }
935 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
936 } else if (f.is_interpreted_frame()) {
937 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
938 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
939 } else if (top && _preempt) {
940 assert(f.is_native_frame() || f.is_runtime_frame(), "");
941 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
942 } else {
943 // Frame can't be frozen. Most likely the call_stub or upcall_stub
944 // which indicates there are further natives frames up the stack.
945 return freeze_pinned_native;
946 }
947 }
948
949 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
950 // See also StackChunkFrameStream<frame_kind>::frame_size()
951 template<typename FKind>
952 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
953 assert(FKind::is_instance(f), "");
954
955 assert(fsize > 0, "");
956 assert(argsize >= 0, "");
957 _freeze_size += fsize;
958 NOT_PRODUCT(_frames++;)
959
960 assert(FKind::frame_bottom(f) <= _bottom_address, "");
961
962 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
963 // enterSpecial and an interpreted frame
964 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
965 return finalize_freeze(f, caller, argsize); // recursion end
966 } else {
967 frame senderf = sender<FKind>(f);
968 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
969 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
970 return result;
971 }
972 }
973
974 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
975 LogTarget(Trace, continuations) lt;
976 if (lt.develop_is_enabled()) {
977 LogStream ls(lt);
978 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
979 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
980 f.print_value_on(&ls);
981 }
982 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
983 }
984
985 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
986 LogTarget(Trace, continuations) lt;
987 if (lt.develop_is_enabled()) {
988 LogStream ls(lt);
989 DEBUG_ONLY(hf.print_value_on(&ls);)
990 assert(hf.is_heap_frame(), "should be");
991 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
992 if (is_bottom_frame) {
993 ls.print_cr("bottom h-frame:");
994 hf.print_on(&ls);
995 }
996 }
997 }
998
999 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
1000 // See also StackChunkFrameStream<frame_kind>::frame_size()
1001 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
1002 int argsize = argsize_md - frame::metadata_words_at_top;
1003 assert(callee.is_interpreted_frame()
1004 || ContinuationHelper::Frame::is_stub(callee.cb())
1005 || callee.cb()->as_nmethod()->is_osr_method()
1006 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
1007 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
1008 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
1009
1010 LogTarget(Trace, continuations) lt;
1011
1012 #ifdef ASSERT
1013 bool empty = _cont.is_empty();
1014 log_develop_trace(continuations)("empty: %d", empty);
1015 #endif
1016
1017 stackChunkOop chunk = _cont.tail();
1018
1019 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1020
1021 _freeze_size += frame::metadata_words; // for top frame's metadata
1022
1023 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1024 int unextended_sp = -1;
1025 if (chunk != nullptr) {
1026 if (!chunk->is_empty()) {
1027 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1028 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1029 bool top_interpreted = Interpreter::contains(chunk->pc());
1030 if (callee.is_interpreted_frame() == top_interpreted) {
1031 overlap = argsize_md;
1032 }
1033 } else {
1034 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1035 }
1036 }
1037
1038 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1039
1040 _freeze_size -= overlap;
1041 assert(_freeze_size >= 0, "");
1042
1043 assert(chunk == nullptr || chunk->is_empty()
1044 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1045 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1046
1047 _freeze_size += _monitors_in_lockstack;
1048
1049 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1050 bool allocated_old_in_freeze_fast = _barriers;
1051 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1052 "Chunk allocated in freeze_fast is of insufficient size "
1053 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1054 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1055
1056 DEBUG_ONLY(bool empty_chunk = true);
1057 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1058 // ALLOCATE NEW CHUNK
1059
1060 if (lt.develop_is_enabled()) {
1061 LogStream ls(lt);
1062 if (chunk == nullptr) {
1063 ls.print_cr("no chunk");
1064 } else {
1065 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1066 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1067 chunk->print_on(&ls);
1068 }
1069 }
1070
1071 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1072 // overlap = 0;
1073
1074 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1075 if (chunk == nullptr) {
1076 return freeze_exception;
1077 }
1078
1079 // Install new chunk
1080 _cont.set_tail(chunk);
1081 assert(chunk->is_empty(), "");
1082 } else {
1083 // REUSE EXISTING CHUNK
1084 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1085 if (chunk->is_empty()) {
1086 int sp = chunk->stack_size() - argsize_md;
1087 chunk->set_sp(sp);
1088 chunk->set_bottom(sp);
1089 _freeze_size += overlap;
1090 assert(chunk->max_thawing_size() == 0, "");
1091 } DEBUG_ONLY(else empty_chunk = false;)
1092 }
1093 assert(!chunk->is_gc_mode(), "");
1094 assert(!chunk->has_bitmap(), "");
1095 chunk->set_has_mixed_frames(true);
1096
1097 assert(chunk->requires_barriers() == _barriers, "");
1098 assert(!_barriers || chunk->is_empty(), "");
1099
1100 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1101 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1102
1103 if (_preempt) {
1104 frame top_frame = _thread->last_frame();
1105 if (top_frame.is_interpreted_frame()) {
1106 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1107 // We need it so that on resume we can restore the sp to the right place, since
1108 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1109 // We do it now that we know freezing will be successful.
1110 prepare_freeze_interpreted_top_frame(top_frame);
1111 }
1112
1113 // Do this now so should_process_args_at_top() is set before calling finish_freeze
1114 // in case we might need to apply GC barriers to frames in this stackChunk.
1115 if (_thread->at_preemptable_init()) {
1116 assert(top_frame.is_interpreted_frame(), "only InterpreterRuntime::_new/resolve_from_cache allowed");
1117 chunk->set_at_klass_init(true);
1118 methodHandle m(_thread, top_frame.interpreter_frame_method());
1119 Bytecode_invoke call = Bytecode_invoke_check(m, top_frame.interpreter_frame_bci());
1120 assert(!call.is_valid() || call.is_invokestatic(), "only invokestatic allowed");
1121 if (call.is_invokestatic() && call.size_of_parameters() > 0) {
1122 assert(top_frame.interpreter_frame_expression_stack_size() > 0, "should have parameters in exp stack");
1123 chunk->set_has_args_at_top(true);
1124 }
1125 }
1126 }
1127
1128 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1129 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1130 // will either see no continuation or a consistent chunk.
1131 unwind_frames();
1132
1133 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1134
1135 if (lt.develop_is_enabled()) {
1136 LogStream ls(lt);
1137 ls.print_cr("top chunk:");
1138 chunk->print_on(&ls);
1139 }
1140
1141 if (_monitors_in_lockstack > 0) {
1142 freeze_lockstack(chunk);
1143 }
1144
1145 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1146 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1147
1148 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1149
1150 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1151 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1152 #ifdef ASSERT
1153 if (lt.develop_is_enabled()) {
1154 LogStream ls(lt);
1155 ls.print_cr("top hframe before (freeze):");
1156 assert(caller.is_heap_frame(), "should be");
1157 caller.print_on(&ls);
1158 }
1159
1160 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1161
1162 frame entry = sender(callee);
1163
1164 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1165 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1166 #endif
1167
1168 return freeze_ok_bottom;
1169 }
1170
1171 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1172 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1173 if (is_bottom_frame) {
1174 // If we're the bottom frame, we need to replace the return barrier with the real
1175 // caller's pc.
1176 address last_pc = caller.pc();
1177 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1178 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1179 } else {
1180 assert(!caller.is_empty(), "");
1181 }
1182
1183 patch_pd(hf, caller, is_bottom_frame);
1184
1185 if (f.is_interpreted_frame()) {
1186 assert(hf.is_heap_frame(), "should be");
1187 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1188 }
1189
1190 #ifdef ASSERT
1191 if (hf.is_compiled_frame()) {
1192 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1193 log_develop_trace(continuations)("Freezing deoptimized frame");
1194 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1195 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1196 }
1197 }
1198 #endif
1199 }
1200
1201 #ifdef ASSERT
1202 static void verify_frame_top(const frame& f, intptr_t* top) {
1203 ResourceMark rm;
1204 InterpreterOopMap mask;
1205 f.interpreted_frame_oop_map(&mask);
1206 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1207 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1208 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1209 }
1210 #endif // ASSERT
1211
1212 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1213 // See also StackChunkFrameStream<frame_kind>::frame_size()
1214 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1215 int callee_argsize /* incl. metadata */,
1216 bool callee_interpreted) {
1217 adjust_interpreted_frame_unextended_sp(f);
1218
1219 // The frame's top never includes the stack arguments to the callee
1220 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1221 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1222 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1223
1224 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1225
1226 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1227 // including metadata between f and its args
1228 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1229
1230 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1231 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1232 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1233 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1234
1235 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1236 if (UNLIKELY(result > freeze_ok_bottom)) {
1237 return result;
1238 }
1239
1240 bool is_bottom_frame = result == freeze_ok_bottom;
1241 assert(!caller.is_empty() || is_bottom_frame, "");
1242
1243 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1244
1245 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1246 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1247
1248 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1249 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1250 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1251
1252 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1253 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1254 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1255 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1256
1257 relativize_interpreted_frame_metadata(f, hf);
1258
1259 patch(f, hf, caller, is_bottom_frame);
1260
1261 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1262 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1263 caller = hf;
1264
1265 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1266 frame_method->record_gc_epoch();
1267
1268 return freeze_ok;
1269 }
1270
1271 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1272 // See also StackChunkFrameStream<frame_kind>::frame_size()
1273 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1274 int callee_argsize /* incl. metadata */,
1275 bool callee_interpreted) {
1276 // The frame's top never includes the stack arguments to the callee
1277 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1278 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1279 // including metadata between f and its stackargs
1280 int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1281 int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1282
1283 int real_frame_size = 0;
1284 bool augmented = f.was_augmented_on_entry(real_frame_size);
1285 if (augmented) {
1286 assert(f.cb()->as_nmethod()->is_compiled_by_c2(), "should be c2 compiled");
1287 // The args reside inside the frame so clear argsize. If the caller is compiled,
1288 // this will cause the stack arguments passed by the caller to be freezed when
1289 // freezing the caller frame itself. If the caller is interpreted this will have
1290 // the effect of discarding the arg area created in the i2c stub.
1291 argsize = 0;
1292 fsize = real_frame_size - (callee_interpreted ? 0 : callee_argsize);
1293 }
1294
1295 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d augmented: %d",
1296 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1297 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1298 _freeze_size, fsize, argsize, augmented);
1299 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1300 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1301
1302 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1303 if (UNLIKELY(result > freeze_ok_bottom)) {
1304 return result;
1305 }
1306
1307 bool is_bottom_frame = result == freeze_ok_bottom;
1308 assert(!caller.is_empty() || is_bottom_frame, "");
1309 assert(!is_bottom_frame || !augmented, "thaw extended frame without caller?");
1310
1311 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1312
1313 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller, augmented ? real_frame_size - f.cb()->as_nmethod()->frame_size() : 0);
1314
1315 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1316
1317 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1318 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1319
1320 if (caller.is_interpreted_frame()) {
1321 // When thawing the frame we might need to add alignment (see Thaw::align)
1322 _total_align_size += frame::align_wiggle;
1323 }
1324
1325 patch(f, hf, caller, is_bottom_frame);
1326
1327 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1328
1329 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1330 caller = hf;
1331 return freeze_ok;
1332 }
1333
1334 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1335 DEBUG_ONLY(frame fsender = sender(f);)
1336 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1337
1338 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1339 const int fsize = f.cb()->frame_size();
1340
1341 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1342 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1343
1344 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1345 if (UNLIKELY(result > freeze_ok_bottom)) {
1346 return result;
1347 }
1348
1349 assert(result == freeze_ok, "should have caller");
1350 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1351
1352 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1353 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1354
1355 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1356
1357 patch(f, hf, caller, false /*is_bottom_frame*/);
1358
1359 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1360
1361 caller = hf;
1362 return freeze_ok;
1363 }
1364
1365 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1366 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1367 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1368 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1369 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1370 // caller into the native convention. This is so that we have a valid oopMap in case of
1371 // having to block in the slow path. But that would require freezing those registers too
1372 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1373 // given that this would be a rare case anyways just pin the vthread to the carrier.
1374 return freeze_pinned_native;
1375 }
1376
1377 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1378 // There are no stackargs but argsize must include the metadata
1379 const int argsize = frame::metadata_words_at_top;
1380 const int fsize = f.cb()->frame_size() + argsize;
1381
1382 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1383 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1384
1385 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1386 if (UNLIKELY(result > freeze_ok_bottom)) {
1387 return result;
1388 }
1389
1390 assert(result == freeze_ok, "should have caller frame");
1391 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1392
1393 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1394 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1395
1396 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1397
1398 if (caller.is_interpreted_frame()) {
1399 // When thawing the frame we might need to add alignment (see Thaw::align)
1400 _total_align_size += frame::align_wiggle;
1401 }
1402
1403 patch(f, hf, caller, false /* is_bottom_frame */);
1404
1405 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1406
1407 caller = hf;
1408 return freeze_ok;
1409 }
1410
1411 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1412 stackChunkOop chunk = _cont.tail();
1413
1414 LogTarget(Trace, continuations) lt;
1415 if (lt.develop_is_enabled()) {
1416 LogStream ls(lt);
1417 assert(top.is_heap_frame(), "should be");
1418 top.print_on(&ls);
1419 }
1420
1421 set_top_frame_metadata_pd(top);
1422
1423 chunk->set_sp(chunk->to_offset(top.sp()));
1424 chunk->set_pc(top.pc());
1425
1426 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1427
1428 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1429
1430 // At this point the chunk is consistent
1431
1432 if (UNLIKELY(_barriers)) {
1433 log_develop_trace(continuations)("do barriers on old chunk");
1434 // Serial and Parallel GC can allocate objects directly into the old generation.
1435 // Then we want to relativize the derived pointers eagerly so that
1436 // old chunks are all in GC mode.
1437 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1438 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1439 if (UseShenandoahGC) {
1440 _cont.tail()->relativize_derived_pointers_concurrently();
1441 } else {
1442 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1443 }
1444 // For objects in the old generation we must maintain the remembered set
1445 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1446 }
1447
1448 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1449 if (lt.develop_is_enabled()) {
1450 LogStream ls(lt);
1451 chunk->print_on(true, &ls);
1452 }
1453
1454 if (lt.develop_is_enabled()) {
1455 LogStream ls(lt);
1456 ls.print_cr("top hframe after (freeze):");
1457 assert(_cont.last_frame().is_heap_frame(), "should be");
1458 _cont.last_frame().print_on(&ls);
1459 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1460 }
1461
1462 assert(_cont.chunk_invariant(), "");
1463 }
1464
1465 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1466 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1467 assert(t == JavaThread::current(), "");
1468 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1469 if (!_preempt) {
1470 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1471 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1472 }
1473 return true;
1474 }
1475 return false;
1476 }
1477
1478 class StackChunkAllocator : public MemAllocator {
1479 const size_t _stack_size;
1480 int _argsize_md;
1481 ContinuationWrapper& _continuation_wrapper;
1482 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1483 mutable bool _took_slow_path;
1484
1485 // Does the minimal amount of initialization needed for a TLAB allocation.
1486 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1487 virtual oop initialize(HeapWord* mem) const override {
1488 assert(_stack_size > 0, "");
1489 assert(_stack_size <= max_jint, "");
1490 assert(_word_size > _stack_size, "");
1491
1492 // zero out fields (but not the stack)
1493 const size_t hs = oopDesc::header_size();
1494 if (oopDesc::has_klass_gap()) {
1495 oopDesc::set_klass_gap(mem, 0);
1496 }
1497 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1498
1499 int bottom = (int)_stack_size - _argsize_md;
1500
1501 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1502 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1503 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1504
1505 return finish(mem);
1506 }
1507
1508 stackChunkOop allocate_fast() const {
1509 if (!UseTLAB) {
1510 return nullptr;
1511 }
1512
1513 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1514 if (mem == nullptr) {
1515 return nullptr;
1516 }
1517
1518 oop obj = initialize(mem);
1519 return stackChunkOopDesc::cast(obj);
1520 }
1521
1522 public:
1523 StackChunkAllocator(Klass* klass,
1524 size_t word_size,
1525 Thread* thread,
1526 size_t stack_size,
1527 int argsize_md,
1528 ContinuationWrapper& continuation_wrapper,
1529 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1530 : MemAllocator(klass, word_size, thread),
1531 _stack_size(stack_size),
1532 _argsize_md(argsize_md),
1533 _continuation_wrapper(continuation_wrapper),
1534 _jvmti_event_collector(jvmti_event_collector),
1535 _took_slow_path(false) {}
1536
1537 // Provides it's own, specialized allocation which skips instrumentation
1538 // if the memory can be allocated without going to a slow-path.
1539 stackChunkOop allocate() const {
1540 // First try to allocate without any slow-paths or instrumentation.
1541 stackChunkOop obj = allocate_fast();
1542 if (obj != nullptr) {
1543 return obj;
1544 }
1545
1546 // Now try full-blown allocation with all expensive operations,
1547 // including potentially safepoint operations.
1548 _took_slow_path = true;
1549
1550 // Protect unhandled Loom oops
1551 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1552
1553 // Can safepoint
1554 _jvmti_event_collector->start();
1555
1556 // Can safepoint
1557 return stackChunkOopDesc::cast(MemAllocator::allocate());
1558 }
1559
1560 bool took_slow_path() const {
1561 return _took_slow_path;
1562 }
1563 };
1564
1565 template <typename ConfigT>
1566 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1567 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1568
1569 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1570 size_t size_in_words = klass->instance_size(stack_size);
1571
1572 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1573 if (!_preempt) {
1574 throw_stack_overflow_on_humongous_chunk();
1575 }
1576 return nullptr;
1577 }
1578
1579 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1580 assert(current == JavaThread::current(), "should be current");
1581
1582 // Allocate the chunk.
1583 //
1584 // This might safepoint while allocating, but all safepointing due to
1585 // instrumentation have been deferred. This property is important for
1586 // some GCs, as this ensures that the allocated object is in the young
1587 // generation / newly allocated memory.
1588 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1589 stackChunkOop chunk = allocator.allocate();
1590
1591 if (chunk == nullptr) {
1592 return nullptr; // OOME
1593 }
1594
1595 // assert that chunk is properly initialized
1596 assert(chunk->stack_size() == (int)stack_size, "");
1597 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1598 assert(chunk->sp() == chunk->bottom(), "");
1599 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1600 assert(chunk->max_thawing_size() == 0, "");
1601 assert(chunk->pc() == nullptr, "");
1602 assert(chunk->is_empty(), "");
1603 assert(chunk->flags() == 0, "");
1604 assert(chunk->is_gc_mode() == false, "");
1605 assert(chunk->lockstack_size() == 0, "");
1606
1607 // fields are uninitialized
1608 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1609 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1610
1611 #if INCLUDE_ZGC
1612 if (UseZGC) {
1613 ZStackChunkGCData::initialize(chunk);
1614 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1615 _barriers = false;
1616 } else
1617 #endif
1618 #if INCLUDE_SHENANDOAHGC
1619 if (UseShenandoahGC) {
1620 ShenandoahStackChunkGCData::initialize(chunk);
1621 _barriers = chunk->requires_barriers();
1622 } else
1623 #endif
1624 {
1625 if (!allocator.took_slow_path()) {
1626 // Guaranteed to be in young gen / newly allocated memory
1627 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1628 _barriers = false;
1629 } else {
1630 // Some GCs could put direct allocations in old gen for slow-path
1631 // allocations; need to explicitly check if that was the case.
1632 _barriers = chunk->requires_barriers();
1633 }
1634 }
1635
1636 if (_barriers) {
1637 log_develop_trace(continuations)("allocation requires barriers");
1638 }
1639
1640 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1641
1642 return chunk;
1643 }
1644
1645 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1646 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1647 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1648 }
1649
1650 class AnchorMark : public StackObj {
1651 JavaThread* _current;
1652 frame& _top_frame;
1653 intptr_t* _last_sp_from_frame;
1654 bool _is_interpreted;
1655
1656 public:
1657 AnchorMark(JavaThread* current, frame& f) : _current(current), _top_frame(f), _is_interpreted(false) {
1658 intptr_t* sp = anchor_mark_set_pd();
1659 set_anchor(_current, sp);
1660 }
1661 ~AnchorMark() {
1662 clear_anchor(_current);
1663 anchor_mark_clear_pd();
1664 }
1665 inline intptr_t* anchor_mark_set_pd();
1666 inline void anchor_mark_clear_pd();
1667 };
1668
1669 #if INCLUDE_JVMTI
1670 static int num_java_frames(ContinuationWrapper& cont) {
1671 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1672 int count = 0;
1673 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1674 count += chunk->num_java_frames();
1675 }
1676 return count;
1677 }
1678
1679 static void invalidate_jvmti_stack(JavaThread* thread) {
1680 JvmtiThreadState *state = thread->jvmti_thread_state();
1681 if (state != nullptr) {
1682 state->invalidate_cur_stack_depth();
1683 }
1684 }
1685
1686 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1687 if (!cont.entry()->is_virtual_thread()) {
1688 if (JvmtiExport::has_frame_pops(thread)) {
1689 int num_frames = num_java_frames(cont);
1690
1691 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1692 JvmtiExport::continuation_yield_cleanup(thread, num_frames);
1693 }
1694 invalidate_jvmti_stack(thread);
1695 }
1696 }
1697
1698 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, Continuation::preempt_kind pk) {
1699 assert(current->vthread() != nullptr, "must be");
1700
1701 HandleMarkCleaner hm(current); // Cleanup all handles (including so._conth) before returning to Java.
1702 Handle vth(current, current->vthread());
1703 ContinuationWrapper::SafepointOp so(current, cont);
1704 AnchorMark am(current, top); // Set anchor so that the stack is walkable.
1705
1706 JRT_BLOCK
1707 MountUnmountDisabler::end_transition(current, vth(), true /*is_mount*/, false /*is_thread_start*/);
1708
1709 if (current->pending_contended_entered_event()) {
1710 // No monitor JVMTI events for ObjectLocker case.
1711 if (pk != Continuation::object_locker) {
1712 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1713 }
1714 current->set_contended_entered_monitor(nullptr);
1715 }
1716 JRT_BLOCK_END
1717 }
1718 #endif // INCLUDE_JVMTI
1719
1720 #ifdef ASSERT
1721 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1722 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1723 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1724 bool FreezeBase::check_valid_fast_path() {
1725 ContinuationEntry* ce = _thread->last_continuation();
1726 RegisterMap map(_thread,
1727 RegisterMap::UpdateMap::skip,
1728 RegisterMap::ProcessFrames::skip,
1729 RegisterMap::WalkContinuation::skip);
1730 map.set_include_argument_oops(false);
1731 bool is_top_frame = true;
1732 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1733 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1734 return false;
1735 }
1736 }
1737 return true;
1738 }
1739
1740 static void verify_frame_kind(frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr, const char** code_name_ptr, int* bci_ptr, stackChunkOop chunk) {
1741 Method* m;
1742 const char* code_name;
1743 int bci;
1744 ResourceMark rm;
1745
1746 if (preempt_kind == Continuation::monitorenter) {
1747 assert(top.is_interpreted_frame() || top.is_runtime_frame(), "unexpected %sframe",
1748 top.is_compiled_frame() ? "compiled " : top.is_native_frame() ? "native " : "");
1749 bool at_sync_method;
1750 if (top.is_interpreted_frame()) {
1751 m = top.interpreter_frame_method();
1752 assert(!m->is_native() || m->is_synchronized(), "invalid method %s", m->external_name());
1753 address bcp = top.interpreter_frame_bcp();
1754 assert(bcp != nullptr || m->is_native(), "");
1755 at_sync_method = m->is_synchronized() && (bcp == nullptr || bcp == m->code_base());
1756 // bcp is advanced on monitorenter before making the VM call, adjust for that.
1757 bool at_sync_bytecode = bcp > m->code_base() && Bytecode(m, bcp - 1).code() == Bytecodes::Code::_monitorenter;
1758 assert(at_sync_method || at_sync_bytecode, "");
1759 bci = at_sync_method ? -1 : top.interpreter_frame_bci();
1760 } else {
1761 JavaThread* current = JavaThread::current();
1762 CodeBlob* cb = top.cb();
1763 RegisterMap reg_map(current,
1764 RegisterMap::UpdateMap::skip,
1765 RegisterMap::ProcessFrames::skip,
1766 RegisterMap::WalkContinuation::include);
1767 if (top.is_heap_frame()) {
1768 assert(chunk != nullptr, "");
1769 reg_map.set_stack_chunk(chunk);
1770 top = chunk->relativize(top);
1771 top.set_frame_index(0);
1772 }
1773 frame fr = top.sender(®_map);
1774 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1775 compiledVFrame* cvf = compiledVFrame::cast(vf);
1776 m = cvf->method();
1777 bci = cvf->scope()->bci();
1778 at_sync_method = bci == SynchronizationEntryBCI;
1779 assert(!at_sync_method || m->is_synchronized(), "bci is %d but method %s is not synchronized", bci, m->external_name());
1780 bool is_c1_monitorenter = false, is_c2_monitorenter = false;
1781 COMPILER1_PRESENT(is_c1_monitorenter = cb == Runtime1::blob_for(StubId::c1_monitorenter_id) ||
1782 cb == Runtime1::blob_for(StubId::c1_monitorenter_nofpu_id);)
1783 COMPILER2_PRESENT(is_c2_monitorenter = cb == CodeCache::find_blob(OptoRuntime::complete_monitor_locking_Java());)
1784 assert(is_c1_monitorenter || is_c2_monitorenter, "wrong runtime stub frame");
1785 }
1786 code_name = at_sync_method ? "synchronized method" : "monitorenter";
1787 } else if (preempt_kind == Continuation::object_wait) {
1788 assert(top.is_interpreted_frame() || top.is_native_frame(), "");
1789 m = top.is_interpreted_frame() ? top.interpreter_frame_method() : top.cb()->as_nmethod()->method();
1790 assert(m->is_object_wait0(), "");
1791 bci = 0;
1792 code_name = "";
1793 } else {
1794 assert(preempt_kind == Continuation::object_locker, "invalid preempt kind");
1795 assert(top.is_interpreted_frame(), "");
1796 m = top.interpreter_frame_method();
1797 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
1798 Bytecodes::Code code = current_bytecode.code();
1799 assert(code == Bytecodes::Code::_new || code == Bytecodes::Code::_invokestatic ||
1800 (code == Bytecodes::Code::_getstatic || code == Bytecodes::Code::_putstatic), "invalid bytecode");
1801 bci = top.interpreter_frame_bci();
1802 code_name = Bytecodes::name(current_bytecode.code());
1803 }
1804 assert(bci >= 0 || m->is_synchronized(), "invalid bci:%d at method %s", bci, m->external_name());
1805
1806 if (m_ptr != nullptr) {
1807 *m_ptr = m;
1808 *code_name_ptr = code_name;
1809 *bci_ptr = bci;
1810 }
1811 }
1812
1813 static void log_preempt_after_freeze(const ContinuationWrapper& cont) {
1814 JavaThread* current = cont.thread();
1815 int64_t tid = current->monitor_owner_id();
1816
1817 StackChunkFrameStream<ChunkFrames::Mixed> sfs(cont.tail());
1818 frame top_frame = sfs.to_frame();
1819 bool at_init = current->at_preemptable_init();
1820 bool at_enter = current->current_pending_monitor() != nullptr;
1821 bool at_wait = current->current_waiting_monitor() != nullptr;
1822 assert((at_enter && !at_wait) || (!at_enter && at_wait), "");
1823 Continuation::preempt_kind pk = at_init ? Continuation::object_locker : at_enter ? Continuation::monitorenter : Continuation::object_wait;
1824
1825 Method* m = nullptr;
1826 const char* code_name = nullptr;
1827 int bci = InvalidFrameStateBci;
1828 verify_frame_kind(top_frame, pk, &m, &code_name, &bci, cont.tail());
1829 assert(m != nullptr && code_name != nullptr && bci != InvalidFrameStateBci, "should be set");
1830
1831 ResourceMark rm(current);
1832 if (bci < 0) {
1833 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " while synchronizing on %smethod %s", tid, m->is_native() ? "native " : "", m->external_name());
1834 } else if (m->is_object_wait0()) {
1835 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at native method %s", tid, m->external_name());
1836 } else {
1837 Klass* k = current->preempt_init_klass();
1838 assert(k != nullptr || !at_init, "");
1839 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at %s(bci:%d) in method %s %s%s", tid, code_name, bci,
1840 m->external_name(), at_init ? "trying to initialize klass " : "", at_init ? k->external_name() : "");
1841 }
1842 }
1843 #endif // ASSERT
1844
1845 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1846 verify_continuation(cont.continuation());
1847 assert(!cont.is_empty(), "");
1848
1849 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1850 return freeze_ok;
1851 }
1852
1853 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1854 if (UNLIKELY(res != freeze_ok)) {
1855 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1856 verify_continuation(cont.continuation());
1857 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1858 return res;
1859 }
1860
1861 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1862 return freeze_epilog(cont);
1863 }
1864
1865 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1866 if (UNLIKELY(res != freeze_ok)) {
1867 verify_continuation(cont.continuation());
1868 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1869 return res;
1870 }
1871
1872 // Set up things so that on return to Java we jump to preempt stub.
1873 patch_return_pc_with_preempt_stub(old_last_frame);
1874 cont.tail()->set_preempted(true);
1875 DEBUG_ONLY(log_preempt_after_freeze(cont);)
1876 return freeze_epilog(cont);
1877 }
1878
1879 template<typename ConfigT, bool preempt>
1880 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1881 assert(!current->has_pending_exception(), "");
1882
1883 #ifdef ASSERT
1884 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1885 log_frames(current);
1886 #endif
1887
1888 CONT_JFR_ONLY(EventContinuationFreeze event;)
1889
1890 ContinuationEntry* entry = current->last_continuation();
1891
1892 oop oopCont = entry->cont_oop(current);
1893 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1894 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1895
1896 verify_continuation(oopCont);
1897 ContinuationWrapper cont(current, oopCont);
1898 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1899
1900 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1901
1902 if (entry->is_pinned()) {
1903 log_develop_debug(continuations)("PINNED due to critical section");
1904 verify_continuation(cont.continuation());
1905 const freeze_result res = freeze_pinned_cs;
1906 if (!preempt) {
1907 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1908 }
1909 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1910 // Avoid Thread.yield() loops without safepoint polls.
1911 if (SafepointMechanism::should_process(current) && !preempt) {
1912 cont.done(); // allow safepoint
1913 ThreadInVMfromJava tivmfj(current);
1914 }
1915 return res;
1916 }
1917
1918 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1919
1920 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1921 bool fast = UseContinuationFastPath && current->cont_fastpath();
1922 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1923 freeze.freeze_fast_existing_chunk();
1924 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1925 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1926 }
1927
1928 if (preempt) {
1929 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1930 freeze.set_jvmti_event_collector(&jsoaec);
1931
1932 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1933
1934 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1935 preempt_epilog(cont, res, freeze.last_frame());
1936 return res;
1937 }
1938
1939 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1940 assert(current == JavaThread::current(), "must be current thread");
1941 JRT_BLOCK
1942 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1943 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1944 freeze.set_jvmti_event_collector(&jsoaec);
1945
1946 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1947
1948 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1949 freeze_epilog(current, cont, res);
1950 cont.done(); // allow safepoint in the transition back to Java
1951 return res;
1952 JRT_BLOCK_END
1953 }
1954
1955 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1956 ContinuationEntry* entry = thread->last_continuation();
1957 if (entry == nullptr) {
1958 return freeze_ok;
1959 }
1960 if (entry->is_pinned()) {
1961 return freeze_pinned_cs;
1962 }
1963
1964 RegisterMap map(thread,
1965 RegisterMap::UpdateMap::include,
1966 RegisterMap::ProcessFrames::skip,
1967 RegisterMap::WalkContinuation::skip);
1968 map.set_include_argument_oops(false);
1969 frame f = thread->last_frame();
1970
1971 if (!safepoint) {
1972 f = f.sender(&map); // this is the yield frame
1973 } else { // safepoint yield
1974 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1975 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1976 #else
1977 Unimplemented();
1978 #endif
1979 if (!Interpreter::contains(f.pc())) {
1980 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1981 assert(f.oop_map() != nullptr, "must be");
1982 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1983 }
1984 }
1985
1986 while (true) {
1987 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1988 return freeze_pinned_native;
1989 }
1990
1991 f = f.sender(&map);
1992 if (!Continuation::is_frame_in_continuation(entry, f)) {
1993 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
1994 if (scope == cont_scope) {
1995 break;
1996 }
1997 entry = entry->parent();
1998 if (entry == nullptr) {
1999 break;
2000 }
2001 if (entry->is_pinned()) {
2002 return freeze_pinned_cs;
2003 }
2004 }
2005 }
2006 return freeze_ok;
2007 }
2008
2009 /////////////// THAW ////
2010
2011 static int thaw_size(stackChunkOop chunk) {
2012 int size = chunk->max_thawing_size();
2013 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
2014 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
2015 return size;
2016 }
2017
2018 // make room on the stack for thaw
2019 // returns the size in bytes, or 0 on failure
2020 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
2021 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
2022
2023 assert(thread == JavaThread::current(), "");
2024
2025 ContinuationEntry* ce = thread->last_continuation();
2026 assert(ce != nullptr, "");
2027 oop continuation = ce->cont_oop(thread);
2028 assert(continuation == get_continuation(thread), "");
2029 verify_continuation(continuation);
2030
2031 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
2032 assert(chunk != nullptr, "");
2033
2034 // The tail can be empty because it might still be available for another freeze.
2035 // However, here we want to thaw, so we get rid of it (it will be GCed).
2036 if (UNLIKELY(chunk->is_empty())) {
2037 chunk = chunk->parent();
2038 assert(chunk != nullptr, "");
2039 assert(!chunk->is_empty(), "");
2040 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
2041 }
2042
2043 // Verification
2044 chunk->verify();
2045 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
2046
2047 // Only make space for the last chunk because we only thaw from the last chunk
2048 int size = thaw_size(chunk) << LogBytesPerWord;
2049
2050 const address bottom = (address)thread->last_continuation()->entry_sp();
2051 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
2052 // for the Java frames in the check below.
2053 if (!stack_overflow_check(thread, size + 300, bottom)) {
2054 return 0;
2055 }
2056
2057 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
2058 p2i(bottom), p2i(bottom - size), size);
2059 return size;
2060 }
2061
2062 class ThawBase : public StackObj {
2063 protected:
2064 JavaThread* _thread;
2065 ContinuationWrapper& _cont;
2066 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
2067
2068 intptr_t* _fastpath;
2069 bool _barriers;
2070 bool _preempted_case;
2071 bool _should_patch_caller_pc;
2072 bool _process_args_at_top;
2073 intptr_t* _top_unextended_sp_before_thaw;
2074 int _align_size;
2075 DEBUG_ONLY(intptr_t* _top_stack_address;)
2076 DEBUG_ONLY(address _caller_raw_pc;)
2077
2078 // Only used for preemption on ObjectLocker
2079 ObjectMonitor* _init_lock;
2080
2081 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
2082
2083 NOT_PRODUCT(int _frames;)
2084
2085 protected:
2086 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2087 _thread(thread), _cont(cont),
2088 _fastpath(nullptr) {
2089 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2090 assert (cont.tail() != nullptr, "no last chunk");
2091 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2092 }
2093
2094 void clear_chunk(stackChunkOop chunk);
2095 template<bool check_stub>
2096 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2097 int remove_scalarized_frames(StackChunkFrameStream<ChunkFrames::CompiledOnly>& scfs, int &argsize);
2098 void check_top_for_deoptimization(stackChunkOop chunk);
2099 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2100
2101 void thaw_lockstack(stackChunkOop chunk);
2102
2103 // fast path
2104 inline void prefetch_chunk_pd(void* start, int size_words);
2105 void patch_return(intptr_t* sp, bool is_last);
2106
2107 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
2108 inline intptr_t* push_cleanup_continuation();
2109 inline intptr_t* push_preempt_adapter();
2110 intptr_t* redo_vmcall(JavaThread* current, frame& top);
2111 void throw_interrupted_exception(JavaThread* current, frame& top);
2112
2113 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
2114 void finish_thaw(frame& f);
2115
2116 private:
2117 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2118 void finalize_thaw(frame& entry, int argsize);
2119
2120 inline bool seen_by_gc();
2121
2122 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2123 inline void after_thaw_java_frame(const frame& f, bool bottom);
2124 inline void patch(frame& f, const frame& caller, bool bottom);
2125 void clear_bitmap_bits(address start, address end);
2126
2127 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top);
2128 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2129 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2130 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
2131
2132 void push_return_frame(const frame& f);
2133 inline frame new_entry_frame();
2134 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom, int size_adjust = 0);
2135 inline void patch_pd(frame& f, const frame& sender);
2136 inline void patch_pd(frame& f, intptr_t* caller_sp);
2137 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2138
2139 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2140
2141 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2142
2143 public:
2144 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2145 };
2146
2147 template <typename ConfigT>
2148 class Thaw : public ThawBase {
2149 public:
2150 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2151
2152 inline bool can_thaw_fast(stackChunkOop chunk) {
2153 return !_barriers
2154 && _thread->cont_fastpath_thread_state()
2155 && !chunk->has_thaw_slowpath_condition()
2156 && !PreserveFramePointer;
2157 }
2158
2159 inline intptr_t* thaw(Continuation::thaw_kind kind);
2160 template<bool check_stub = false>
2161 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2162 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2163 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2164 };
2165
2166 template <typename ConfigT>
2167 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2168 verify_continuation(_cont.continuation());
2169 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2170 assert(!_cont.is_empty(), "");
2171
2172 stackChunkOop chunk = _cont.tail();
2173 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2174 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2175
2176 _barriers = chunk->requires_barriers();
2177 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2178 : thaw_slow(chunk, kind);
2179 }
2180
2181 class ReconstructedStack : public StackObj {
2182 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2183 int _thaw_size;
2184 int _argsize;
2185 public:
2186 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2187 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2188 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2189 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2190 // We're at most one alignment word away from entrySP
2191 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2192 }
2193
2194 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2195
2196 // top and bottom stack pointers
2197 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2198 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2199
2200 // several operations operate on the totality of the stack being reconstructed,
2201 // including the metadata words
2202 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2203 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2204 };
2205
2206 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2207 chunk->set_sp(chunk->bottom());
2208 chunk->set_max_thawing_size(0);
2209 }
2210
2211 void ThawBase::check_top_for_deoptimization(stackChunkOop chunk) {
2212 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2213 if (f.is_stub()) {
2214 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2215 assert(!f.is_done(), "");
2216
2217 f.get_cb();
2218 assert(f.is_compiled(), "");
2219 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2220 // The caller of the runtime stub when the continuation is preempted is not at a
2221 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2222 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2223 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2224 }
2225 }
2226 }
2227
2228 int ThawBase::remove_scalarized_frames(StackChunkFrameStream<ChunkFrames::CompiledOnly>& f, int &argsize) {
2229 intptr_t* top = f.sp();
2230
2231 while (f.cb()->as_nmethod()->needs_stack_repair()) {
2232 f.next(SmallRegisterMap::instance_no_args(), false /* stop */);
2233 }
2234 assert(!f.is_done(), "");
2235 assert(f.is_compiled(), "");
2236
2237 intptr_t* bottom = f.sp() + f.cb()->frame_size();
2238 argsize = f.stack_argsize();
2239 return bottom - top;
2240 }
2241
2242 template<bool check_stub>
2243 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2244 bool empty = false;
2245 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2246 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2247 assert(chunk_sp == f.sp(), "");
2248 assert(chunk_sp == f.unextended_sp(), "");
2249
2250 int frame_size = f.cb()->frame_size();
2251 argsize = f.stack_argsize();
2252
2253 assert(!f.is_stub() || check_stub, "");
2254 if (check_stub && f.is_stub()) {
2255 // If we don't thaw the top compiled frame too, after restoring the saved
2256 // registers back in Java, we would hit the return barrier to thaw one more
2257 // frame effectively overwriting the restored registers during that call.
2258 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2259 assert(!f.is_done(), "");
2260
2261 f.get_cb();
2262 assert(f.is_compiled(), "");
2263
2264 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2265 // The caller of the runtime stub when the continuation is preempted is not at a
2266 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2267 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2268 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2269 }
2270
2271 if (f.cb()->as_nmethod()->needs_stack_repair()) {
2272 frame_size += remove_scalarized_frames(f, argsize);
2273 } else {
2274 frame_size += f.cb()->frame_size();
2275 argsize = f.stack_argsize();
2276 }
2277 } else if (f.cb()->as_nmethod()->needs_stack_repair()) {
2278 frame_size = remove_scalarized_frames(f, argsize);
2279 }
2280
2281 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2282 empty = f.is_done();
2283 assert(!empty || argsize == chunk->argsize(), "");
2284
2285 if (empty) {
2286 clear_chunk(chunk);
2287 } else {
2288 chunk->set_sp(chunk->sp() + frame_size);
2289 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2290 // We set chunk->pc to the return pc into the next frame
2291 chunk->set_pc(f.pc());
2292 #ifdef ASSERT
2293 {
2294 intptr_t* retaddr_slot = (chunk_sp
2295 + frame_size
2296 - frame::sender_sp_ret_address_offset());
2297 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2298 "unexpected pc");
2299 }
2300 #endif
2301 }
2302 assert(empty == chunk->is_empty(), "");
2303 // returns the size required to store the frame on stack, and because it is a
2304 // compiled frame, it must include a copy of the arguments passed by the caller
2305 return frame_size + argsize + frame::metadata_words_at_top;
2306 }
2307
2308 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2309 int lockStackSize = chunk->lockstack_size();
2310 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2311
2312 oop tmp_lockstack[LockStack::CAPACITY];
2313 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2314 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2315
2316 chunk->set_lockstack_size(0);
2317 chunk->set_has_lockstack(false);
2318 }
2319
2320 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2321 assert(to >= _top_stack_address, "overwrote past thawing space"
2322 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2323 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2324 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2325 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2326 }
2327
2328 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2329 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2330
2331 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2332 ContinuationHelper::patch_return_address_at(
2333 sp - frame::sender_sp_ret_address_offset(),
2334 pc);
2335 }
2336
2337 template <typename ConfigT>
2338 template<bool check_stub>
2339 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2340 assert(chunk == _cont.tail(), "");
2341 assert(!chunk->has_mixed_frames(), "");
2342 assert(!chunk->requires_barriers(), "");
2343 assert(!chunk->has_bitmap(), "");
2344 assert(!_thread->is_interp_only_mode(), "");
2345
2346 LogTarget(Trace, continuations) lt;
2347 if (lt.develop_is_enabled()) {
2348 LogStream ls(lt);
2349 ls.print_cr("thaw_fast");
2350 chunk->print_on(true, &ls);
2351 }
2352
2353 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2354 static const int threshold = 500; // words
2355
2356 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2357 int argsize, thaw_size;
2358
2359 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2360
2361 bool partial, empty;
2362 if (LIKELY(!ForceSingleFrameThaw && (full_chunk_size < threshold))) {
2363 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2364
2365 if (check_stub) {
2366 check_top_for_deoptimization(chunk);
2367 }
2368 partial = false;
2369 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2370 clear_chunk(chunk);
2371 thaw_size = full_chunk_size;
2372 empty = true;
2373 } else { // thaw a single frame
2374 partial = true;
2375 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2376 empty = chunk->is_empty();
2377 }
2378
2379 // Are we thawing the last frame(s) in the continuation
2380 const bool is_last = empty && chunk->parent() == nullptr;
2381 assert(!is_last || argsize == 0, "");
2382
2383 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2384 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2385
2386 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2387
2388 // also copy metadata words at frame bottom
2389 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2390
2391 // update the ContinuationEntry
2392 _cont.set_argsize(argsize);
2393 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2394 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2395
2396 // install the return barrier if not last frame, or the entry's pc if last
2397 patch_return(rs.bottom_sp(), is_last);
2398
2399 // insert the back links from callee to caller frames
2400 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2401
2402 assert(is_last == _cont.is_empty(), "");
2403 assert(_cont.chunk_invariant(), "");
2404
2405 #if CONT_JFR
2406 EventContinuationThawFast e;
2407 if (e.should_commit()) {
2408 e.set_id(cast_from_oop<u8>(chunk));
2409 e.set_size(thaw_size << LogBytesPerWord);
2410 e.set_full(!partial);
2411 e.commit();
2412 }
2413 #endif
2414
2415 #ifdef ASSERT
2416 if (LoomDeoptAfterThaw) {
2417 frame top(rs.sp());
2418 AnchorMark am(_thread, top);
2419 log_frames(_thread);
2420 do_deopt_after_thaw(_thread);
2421 }
2422 #endif
2423
2424 return rs.sp();
2425 }
2426
2427 inline bool ThawBase::seen_by_gc() {
2428 return _barriers || _cont.tail()->is_gc_mode();
2429 }
2430
2431 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2432 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2433 if (UseZGC || UseShenandoahGC) {
2434 chunk->relativize_derived_pointers_concurrently();
2435 }
2436 #endif
2437 }
2438
2439 template <typename ConfigT>
2440 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2441 Continuation::preempt_kind preempt_kind;
2442 bool retry_fast_path = false;
2443
2444 _process_args_at_top = false;
2445 _preempted_case = chunk->preempted();
2446 if (_preempted_case) {
2447 ObjectMonitor* mon = nullptr;
2448 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2449 if (waiter != nullptr) {
2450 // Mounted again after preemption. Resume the pending monitor operation,
2451 // which will be either a monitorenter or Object.wait() call.
2452 mon = waiter->monitor();
2453 preempt_kind = waiter->is_wait() ? Continuation::object_wait : Continuation::monitorenter;
2454
2455 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2456 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2457 if (!mon_acquired) {
2458 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2459 log_develop_trace(continuations, preempt)("Failed to acquire monitor, unmounting again");
2460 return push_cleanup_continuation();
2461 }
2462 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2463 JVMTI_ONLY(assert(_thread->contended_entered_monitor() == nullptr || _thread->contended_entered_monitor() == mon, ""));
2464 } else {
2465 // Preemption cancelled on moniterenter or ObjectLocker case. We
2466 // actually acquired the monitor after freezing all frames so no
2467 // need to call resume_operation. If this is the ObjectLocker case
2468 // we released the monitor already at ~ObjectLocker, so _init_lock
2469 // will be set to nullptr below since there is no monitor to release.
2470 preempt_kind = Continuation::monitorenter;
2471 }
2472
2473 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2474 relativize_chunk_concurrently(chunk);
2475
2476 if (chunk->at_klass_init()) {
2477 preempt_kind = Continuation::object_locker;
2478 chunk->set_at_klass_init(false);
2479 _process_args_at_top = chunk->has_args_at_top();
2480 if (_process_args_at_top) {
2481 // Only needed for the top frame which will be thawed.
2482 chunk->set_has_args_at_top(false);
2483 }
2484 assert(waiter == nullptr || mon != nullptr, "should have a monitor");
2485 _init_lock = mon; // remember monitor since we will need it on handle_preempted_continuation()
2486 }
2487 chunk->set_preempted(false);
2488 retry_fast_path = true;
2489 } else {
2490 relativize_chunk_concurrently(chunk);
2491 }
2492
2493 // On first thaw after freeze restore oops to the lockstack if any.
2494 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2495 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2496 thaw_lockstack(chunk);
2497 retry_fast_path = true;
2498 }
2499
2500 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2501 // and FLAG_PREEMPTED flags from the stackChunk.
2502 if (retry_fast_path && can_thaw_fast(chunk)) {
2503 intptr_t* sp = thaw_fast<true>(chunk);
2504 if (_preempted_case) {
2505 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2506 }
2507 return sp;
2508 }
2509
2510 LogTarget(Trace, continuations) lt;
2511 if (lt.develop_is_enabled()) {
2512 LogStream ls(lt);
2513 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2514 chunk->print_on(true, &ls);
2515 }
2516
2517 #if CONT_JFR
2518 EventContinuationThawSlow e;
2519 if (e.should_commit()) {
2520 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2521 e.commit();
2522 }
2523 #endif
2524
2525 DEBUG_ONLY(_frames = 0;)
2526 _align_size = 0;
2527 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2528
2529 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2530 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2531
2532 frame heap_frame = _stream.to_frame();
2533 if (lt.develop_is_enabled()) {
2534 LogStream ls(lt);
2535 ls.print_cr("top hframe before (thaw):");
2536 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2537 heap_frame.print_value_on(&ls);
2538 }
2539
2540 frame caller; // the thawed caller on the stack
2541 _should_patch_caller_pc = false;
2542 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2543 finish_thaw(caller); // caller is now the topmost thawed frame
2544 _cont.write();
2545
2546 assert(_cont.chunk_invariant(), "");
2547
2548 JVMTI_ONLY(if (!_cont.entry()->is_virtual_thread()) invalidate_jvmti_stack(_thread));
2549
2550 _thread->set_cont_fastpath(_fastpath);
2551
2552 intptr_t* sp = caller.sp();
2553
2554 if (_preempted_case) {
2555 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2556 }
2557 return sp;
2558 }
2559
2560 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2561 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2562 assert(!_cont.is_empty(), "no more frames");
2563 assert(num_frames > 0, "");
2564 assert(!heap_frame.is_empty(), "");
2565
2566 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2567 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2568 } else if (!heap_frame.is_interpreted_frame()) {
2569 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2570 } else {
2571 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames, top_on_preempt_case);
2572 }
2573 }
2574
2575 template<typename FKind>
2576 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2577 assert(num_frames > 0, "");
2578
2579 DEBUG_ONLY(_frames++;)
2580
2581 int argsize = _stream.stack_argsize();
2582 CodeBlob* cb = _stream.cb();
2583
2584 _stream.next(SmallRegisterMap::instance_no_args());
2585 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2586
2587 // We never leave a compiled caller of an interpreted frame as the top frame in the chunk
2588 // as it makes detecting that situation and adjusting unextended_sp tricky. We also always
2589 // thaw the caller of a frame that needs_stack_repair, as it would otherwise complicate things:
2590 // - Regardless of whether the frame was extended or not, we would need to copy the right arg
2591 // size if its greater than the one given by the normal method signature (non-scalarized).
2592 // - If the frame was indeed extended, leaving its caller as the top frame would complicate walking
2593 // the chunk (we need unextended_sp, but we only have sp).
2594 if (num_frames == 1 && !_stream.is_done() && ((FKind::interpreted && _stream.is_compiled()) || (FKind::compiled && cb->as_nmethod_or_null()->needs_stack_repair()))) {
2595 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2596 num_frames++;
2597 }
2598
2599 if (num_frames == 1 || _stream.is_done()) { // end recursion
2600 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2601 return true; // bottom
2602 } else { // recurse
2603 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2604 return false;
2605 }
2606 }
2607
2608 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2609 stackChunkOop chunk = _cont.tail();
2610
2611 if (!_stream.is_done()) {
2612 assert(_stream.sp() >= chunk->sp_address(), "");
2613 chunk->set_sp(chunk->to_offset(_stream.sp()));
2614 chunk->set_pc(_stream.pc());
2615 } else {
2616 chunk->set_sp(chunk->bottom());
2617 chunk->set_pc(nullptr);
2618 }
2619 assert(_stream.is_done() == chunk->is_empty(), "");
2620
2621 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2622 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2623
2624 _cont.set_argsize(argsize);
2625 entry = new_entry_frame();
2626
2627 assert(entry.sp() == _cont.entrySP(), "");
2628 assert(Continuation::is_continuation_enterSpecial(entry), "");
2629 assert(_cont.is_entry_frame(entry), "");
2630 assert(entry.pc() == entry.raw_pc(), "");
2631 DEBUG_ONLY(_caller_raw_pc = entry.pc();)
2632 }
2633
2634 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2635 LogTarget(Trace, continuations) lt;
2636 if (lt.develop_is_enabled()) {
2637 LogStream ls(lt);
2638 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2639 assert(hf.is_heap_frame(), "should be");
2640 hf.print_value_on(&ls);
2641 }
2642 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2643 }
2644
2645 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2646 #ifdef ASSERT
2647 LogTarget(Trace, continuations) lt;
2648 if (lt.develop_is_enabled()) {
2649 LogStream ls(lt);
2650 ls.print_cr("thawed frame:");
2651 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2652 }
2653 #endif
2654 }
2655
2656 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) {
2657 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2658 if (bottom) {
2659 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2660 : StubRoutines::cont_returnBarrier());
2661 } else if (_should_patch_caller_pc) {
2662 // Caller was deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2663 // Also, on some platforms, if the caller is interpreted but the callee not we also need to patch.
2664
2665 #if defined(PPC64) || defined(S390)
2666 assert(caller.is_deoptimized_frame() || caller.is_interpreted_frame(), "");
2667 #else
2668 assert(caller.is_deoptimized_frame(), "");
2669 #endif
2670
2671 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc());
2672 _should_patch_caller_pc = false;
2673 }
2674
2675 patch_pd(f, caller);
2676
2677 if (f.is_interpreted_frame()) {
2678 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2679 }
2680
2681 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2682 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2683 assert(!caller.is_compiled_frame() || verify_deopt_state(caller), "");
2684 }
2685
2686 void ThawBase::clear_bitmap_bits(address start, address end) {
2687 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2688 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2689
2690 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2691 // or they will keep objects that are otherwise unreachable alive.
2692
2693 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2694 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2695 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2696 // anyways and we assert that before returning.
2697 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2698 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2699 stackChunkOop chunk = _cont.tail();
2700 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2701 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2702 }
2703
2704 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2705 frame top(sp);
2706 assert(top.pc() == ContinuationHelper::return_address_at(sp - frame::sender_sp_ret_address_offset()), "");
2707 DEBUG_ONLY(verify_frame_kind(top, preempt_kind);)
2708 NOT_PRODUCT(int64_t tid = _thread->monitor_owner_id();)
2709
2710 // Finish the VTMS transition.
2711 assert(_thread->is_in_vthread_transition(), "must be");
2712 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2713 if (is_vthread) {
2714 #if INCLUDE_JVMTI
2715 if (MountUnmountDisabler::notify_jvmti_events()) {
2716 jvmti_mount_end(_thread, _cont, top, preempt_kind);
2717 } else
2718 #endif
2719 { // Faster version of MountUnmountDisabler::end_transition() to avoid
2720 // unnecessary extra instructions from jvmti_mount_end().
2721 java_lang_Thread::set_is_in_vthread_transition(_thread->vthread(), false);
2722 _thread->set_is_in_vthread_transition(false);
2723 }
2724 }
2725
2726 if (fast_case) {
2727 // If we thawed in the slow path the runtime stub/native wrapper frame already
2728 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2729 // we copied the fp patched during freeze, which will now have to be fixed.
2730 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2731 int fsize = top.cb()->frame_size();
2732 patch_pd(top, sp + fsize);
2733 }
2734
2735 if (preempt_kind == Continuation::object_wait) {
2736 // Check now if we need to throw IE exception.
2737 bool throw_ie = _thread->pending_interrupted_exception();
2738 if (throw_ie) {
2739 throw_interrupted_exception(_thread, top);
2740 _thread->set_pending_interrupted_exception(false);
2741 }
2742 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT" after preemption on Object.wait%s", tid, throw_ie ? "(throwing IE)" : "");
2743 } else if (preempt_kind == Continuation::monitorenter) {
2744 if (top.is_runtime_frame()) {
2745 // The continuation might now run on a different platform thread than the previous time so
2746 // we need to adjust the current thread saved in the stub frame before restoring registers.
2747 JavaThread** thread_addr = frame::saved_thread_address(top);
2748 if (thread_addr != nullptr) *thread_addr = _thread;
2749 }
2750 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT " after preemption on monitorenter", tid);
2751 } else {
2752 // We need to redo the original call into the VM. First though, we need
2753 // to exit the monitor we just acquired (except on preemption cancelled
2754 // case where it was already released).
2755 assert(preempt_kind == Continuation::object_locker, "");
2756 if (_init_lock != nullptr) _init_lock->exit(_thread);
2757 sp = redo_vmcall(_thread, top);
2758 }
2759 return sp;
2760 }
2761
2762 intptr_t* ThawBase::redo_vmcall(JavaThread* current, frame& top) {
2763 assert(!current->preempting(), "");
2764 NOT_PRODUCT(int64_t tid = current->monitor_owner_id();)
2765 intptr_t* sp = top.sp();
2766
2767 {
2768 HandleMarkCleaner hmc(current); // Cleanup all handles (including so._conth) before returning to Java.
2769 ContinuationWrapper::SafepointOp so(current, _cont);
2770 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2771
2772 Method* m = top.interpreter_frame_method();
2773 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
2774 Bytecodes::Code code = current_bytecode.code();
2775 log_develop_trace(continuations, preempt)("Redoing InterpreterRuntime::%s for " INT64_FORMAT, code == Bytecodes::Code::_new ? "_new" : "resolve_from_cache", tid);
2776
2777 // These InterpreterRuntime entry points use JRT_ENTRY which uses a HandleMarkCleaner.
2778 // Create a HandeMark to avoid destroying so._conth.
2779 HandleMark hm(current);
2780 DEBUG_ONLY(JavaThread::AtRedoVMCall apvmc(current);)
2781 if (code == Bytecodes::Code::_new) {
2782 InterpreterRuntime::_new(current, m->constants(), current_bytecode.get_index_u2(code));
2783 } else {
2784 InterpreterRuntime::resolve_from_cache(current, code);
2785 }
2786 }
2787
2788 if (current->preempting()) {
2789 // Preempted again so we just arrange to return to preempt stub to unmount.
2790 sp = push_preempt_adapter();
2791 current->set_preempt_alternate_return(nullptr);
2792 bool cancelled = current->preemption_cancelled();
2793 if (cancelled) {
2794 // Since preemption was cancelled, the thread will call thaw again from the preempt
2795 // stub. These retries could happen several times due to contention on the init_lock,
2796 // so just let the vthread umount to give a chance for other vthreads to run.
2797 current->set_preemption_cancelled(false);
2798 oop vthread = current->vthread();
2799 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
2800 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::YIELDING);
2801 #if INCLUDE_JVMTI
2802 if (current->contended_entered_monitor() != nullptr) {
2803 current->set_contended_entered_monitor(nullptr);
2804 }
2805 #endif
2806 }
2807 log_develop_trace(continuations, preempt)("Preempted " INT64_FORMAT " again%s", tid, cancelled ? "(preemption cancelled, setting state to YIELDING)" : "");
2808 } else {
2809 log_develop_trace(continuations, preempt)("Call succesful, resuming " INT64_FORMAT, tid);
2810 }
2811 return sp;
2812 }
2813
2814 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2815 HandleMarkCleaner hm(current); // Cleanup all handles (including so._conth) before returning to Java.
2816 ContinuationWrapper::SafepointOp so(current, _cont);
2817 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2818 JRT_BLOCK
2819 THROW(vmSymbols::java_lang_InterruptedException());
2820 JRT_BLOCK_END
2821 }
2822
2823 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top) {
2824 assert(hf.is_interpreted_frame(), "");
2825
2826 if (UNLIKELY(seen_by_gc())) {
2827 if (is_top && _process_args_at_top) {
2828 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_with_args());
2829 } else {
2830 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2831 }
2832 }
2833
2834 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2835
2836 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2837
2838 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2839
2840 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2841
2842 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2843 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2844 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2845 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2846
2847 assert(hf.is_heap_frame(), "should be");
2848 assert(!f.is_heap_frame(), "should not be");
2849
2850 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2851 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2852
2853 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2854 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2855 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2856
2857 // Make sure the relativized locals is already set.
2858 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2859
2860 derelativize_interpreted_frame_metadata(hf, f);
2861 patch(f, caller, is_bottom_frame);
2862
2863 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2864 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2865
2866 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2867
2868 maybe_set_fastpath(f.sp());
2869
2870 Method* m = hf.interpreter_frame_method();
2871 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2872 const int locals = m->max_locals();
2873
2874 if (!is_bottom_frame) {
2875 // can only fix caller once this frame is thawed (due to callee saved regs)
2876 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2877 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2878 assert(hf.is_heap_frame(), "should be");
2879 address start = (address)(heap_frame_bottom - locals);
2880 address end = (address)heap_frame_bottom;
2881 clear_bitmap_bits(start, end);
2882 }
2883
2884 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2885 DEBUG_ONLY(address return_pc = ContinuationHelper::InterpretedFrame::return_pc(f);)
2886 assert(return_pc == _caller_raw_pc || (is_bottom_frame && return_pc == StubRoutines::cont_returnBarrier()), "wrong return pc");
2887 assert(f.pc() == f.raw_pc(), "");
2888 DEBUG_ONLY(_caller_raw_pc = f.pc();)
2889 caller = f;
2890 }
2891
2892 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2893 assert(hf.is_compiled_frame(), "");
2894 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2895
2896 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2897 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2898 }
2899
2900 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2901
2902 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2903
2904 assert(caller.sp() == caller.unextended_sp(), "");
2905
2906 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2907 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2908 }
2909
2910 int fsize = 0;
2911 int added_argsize = 0;
2912 bool augmented = hf.was_augmented_on_entry(fsize);
2913 if (!augmented) {
2914 added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2915 fsize += added_argsize;
2916 }
2917 assert(!is_bottom_frame || !augmented, "");
2918
2919 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2920 // yet laid out in the stack, and so the original_pc is not stored in it.
2921 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2922 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame, augmented ? fsize - hf.cb()->frame_size() : 0);
2923 assert((int)(caller.sp() - f.sp()) == (augmented ? fsize : f.cb()->frame_size()), "");
2924
2925 intptr_t* const stack_frame_top = f.sp();
2926 intptr_t* const heap_frame_top = hf.unextended_sp();
2927 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2928 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2929 // copy metadata, except the metadata at the top of the (unextended) entry frame
2930 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2931
2932 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2933 // (we might have one padding word for alignment)
2934 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2935 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2936
2937 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2938
2939 patch(f, caller, is_bottom_frame);
2940
2941 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2942 assert(!f.is_deoptimized_frame(), "");
2943 if (hf.is_deoptimized_frame()) {
2944 maybe_set_fastpath(f.sp());
2945 f.set_deoptimized();
2946 } else if (_thread->is_interp_only_mode()
2947 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2948 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2949 // cannot rely on nmethod patching for deopt.
2950 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2951
2952 log_develop_trace(continuations)("Deoptimizing thawed frame");
2953 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2954
2955 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2956 assert(f.is_deoptimized_frame(), "");
2957 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2958 maybe_set_fastpath(f.sp());
2959 assert(!_should_patch_caller_pc, "");
2960 _should_patch_caller_pc = true;
2961 }
2962
2963 if (!is_bottom_frame) {
2964 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2965 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2966 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2967 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2968 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2969 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2970 clear_bitmap_bits(start, start + argsize_in_bytes);
2971 }
2972
2973 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2974 DEBUG_ONLY(address return_pc = ContinuationHelper::CompiledFrame::return_pc(f);)
2975 assert(return_pc == _caller_raw_pc || (is_bottom_frame && return_pc == StubRoutines::cont_returnBarrier()), "wrong return pc");
2976 DEBUG_ONLY(_caller_raw_pc = f.raw_pc();)
2977 caller = f;
2978 }
2979
2980 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2981 DEBUG_ONLY(_frames++;)
2982
2983 if (UNLIKELY(seen_by_gc())) {
2984 // Process the stub's caller here since we might need the full map.
2985 RegisterMap map(nullptr,
2986 RegisterMap::UpdateMap::include,
2987 RegisterMap::ProcessFrames::skip,
2988 RegisterMap::WalkContinuation::skip);
2989 map.set_include_argument_oops(false);
2990 _stream.next(&map);
2991 assert(!_stream.is_done(), "");
2992 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2993 } else {
2994 _stream.next(SmallRegisterMap::instance_no_args());
2995 assert(!_stream.is_done(), "");
2996 }
2997
2998 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2999
3000 assert(caller.is_compiled_frame(), "");
3001 assert(caller.sp() == caller.unextended_sp(), "");
3002
3003 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
3004
3005 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
3006 intptr_t* stack_frame_top = f.sp();
3007 intptr_t* heap_frame_top = hf.sp();
3008 int fsize = ContinuationHelper::StubFrame::size(hf);
3009
3010 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
3011 fsize + frame::metadata_words);
3012
3013 patch(f, caller, false /*is_bottom_frame*/);
3014
3015 // can only fix caller once this frame is thawed (due to callee saved regs)
3016 RegisterMap map(nullptr,
3017 RegisterMap::UpdateMap::include,
3018 RegisterMap::ProcessFrames::skip,
3019 RegisterMap::WalkContinuation::skip);
3020 map.set_include_argument_oops(false);
3021 f.oop_map()->update_register_map(&f, &map);
3022 ContinuationHelper::update_register_map_with_callee(caller, &map);
3023 _cont.tail()->fix_thawed_frame(caller, &map);
3024
3025 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
3026 assert(ContinuationHelper::StubFrame::return_pc(f) == _caller_raw_pc, "wrong return pc");
3027 caller = f;
3028 }
3029
3030 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
3031 assert(hf.is_native_frame(), "");
3032 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
3033
3034 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
3035 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
3036 }
3037
3038 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
3039 assert(!is_bottom_frame, "");
3040
3041 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
3042
3043 assert(caller.sp() == caller.unextended_sp(), "");
3044
3045 if (caller.is_interpreted_frame()) {
3046 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
3047 }
3048
3049 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
3050 // yet laid out in the stack, and so the original_pc is not stored in it.
3051 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
3052 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
3053 intptr_t* const stack_frame_top = f.sp();
3054 intptr_t* const heap_frame_top = hf.unextended_sp();
3055
3056 int fsize = ContinuationHelper::NativeFrame::size(hf);
3057 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
3058
3059 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
3060 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
3061 int sz = fsize + frame::metadata_words_at_bottom;
3062
3063 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
3064
3065 patch(f, caller, false /* bottom */);
3066
3067 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
3068 assert(!f.is_deoptimized_frame(), "");
3069 assert(!hf.is_deoptimized_frame(), "");
3070 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
3071
3072 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
3073 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
3074
3075 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
3076 assert(ContinuationHelper::NativeFrame::return_pc(f) == _caller_raw_pc, "wrong return pc");
3077 caller = f;
3078 }
3079
3080 void ThawBase::finish_thaw(frame& f) {
3081 stackChunkOop chunk = _cont.tail();
3082
3083 if (chunk->is_empty()) {
3084 // Only remove chunk from list if it can't be reused for another freeze
3085 if (seen_by_gc()) {
3086 _cont.set_tail(chunk->parent());
3087 } else {
3088 chunk->set_has_mixed_frames(false);
3089 }
3090 chunk->set_max_thawing_size(0);
3091 } else {
3092 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
3093 }
3094 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
3095
3096 if (!is_aligned(f.sp(), frame::frame_alignment)) {
3097 assert(f.is_interpreted_frame(), "");
3098 f.set_sp(align_down(f.sp(), frame::frame_alignment));
3099 }
3100 push_return_frame(f);
3101 // can only fix caller after push_return_frame (due to callee saved regs)
3102 if (_process_args_at_top) {
3103 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_with_args());
3104 } else {
3105 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_no_args());
3106 }
3107
3108 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
3109
3110 log_develop_trace(continuations)("thawed %d frames", _frames);
3111
3112 LogTarget(Trace, continuations) lt;
3113 if (lt.develop_is_enabled()) {
3114 LogStream ls(lt);
3115 ls.print_cr("top hframe after (thaw):");
3116 _cont.last_frame().print_value_on(&ls);
3117 }
3118 }
3119
3120 void ThawBase::push_return_frame(const frame& f) { // see generate_cont_thaw
3121 assert(!f.is_compiled_frame() || verify_deopt_state(f), "");
3122
3123 LogTarget(Trace, continuations) lt;
3124 if (lt.develop_is_enabled()) {
3125 LogStream ls(lt);
3126 ls.print_cr("push_return_frame");
3127 f.print_value_on(&ls);
3128 }
3129
3130 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
3131 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
3132 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
3133 ContinuationHelper::push_pd(f);
3134
3135 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
3136 }
3137
3138 // returns new top sp
3139 // called after preparations (stack overflow check and making room)
3140 template<typename ConfigT>
3141 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
3142 assert(thread == JavaThread::current(), "Must be current thread");
3143
3144 CONT_JFR_ONLY(EventContinuationThaw event;)
3145
3146 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
3147
3148 ContinuationEntry* entry = thread->last_continuation();
3149 assert(entry != nullptr, "");
3150 oop oopCont = entry->cont_oop(thread);
3151
3152 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
3153 assert(oopCont == get_continuation(thread), "");
3154 verify_continuation(oopCont);
3155
3156 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
3157
3158 ContinuationWrapper cont(thread, oopCont);
3159 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
3160
3161 #ifdef ASSERT
3162 set_anchor_to_entry(thread, cont.entry());
3163 log_frames(thread);
3164 clear_anchor(thread);
3165 #endif
3166
3167 Thaw<ConfigT> thw(thread, cont);
3168 intptr_t* const sp = thw.thaw(kind);
3169 assert(is_aligned(sp, frame::frame_alignment), "");
3170 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp);)
3171
3172 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
3173
3174 verify_continuation(cont.continuation());
3175 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
3176
3177 return sp;
3178 }
3179
3180 #ifdef ASSERT
3181 static void do_deopt_after_thaw(JavaThread* thread) {
3182 int i = 0;
3183 StackFrameStream fst(thread, true, false);
3184 fst.register_map()->set_include_argument_oops(false);
3185 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3186 for (; !fst.is_done(); fst.next()) {
3187 if (fst.current()->cb()->is_nmethod()) {
3188 nmethod* nm = fst.current()->cb()->as_nmethod();
3189 if (!nm->method()->is_continuation_native_intrinsic()) {
3190 nm->make_deoptimized();
3191 }
3192 }
3193 }
3194 }
3195
3196 class ThawVerifyOopsClosure: public OopClosure {
3197 intptr_t* _p;
3198 outputStream* _st;
3199 bool is_good_oop(oop o) {
3200 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
3201 }
3202 public:
3203 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
3204 intptr_t* p() { return _p; }
3205 void reset() { _p = nullptr; }
3206
3207 virtual void do_oop(oop* p) {
3208 oop o = *p;
3209 if (o == nullptr || is_good_oop(o)) {
3210 return;
3211 }
3212 _p = (intptr_t*)p;
3213 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
3214 }
3215 virtual void do_oop(narrowOop* p) {
3216 oop o = RawAccess<>::oop_load(p);
3217 if (o == nullptr || is_good_oop(o)) {
3218 return;
3219 }
3220 _p = (intptr_t*)p;
3221 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
3222 }
3223 };
3224
3225 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
3226 assert(thread->has_last_Java_frame(), "");
3227
3228 ResourceMark rm;
3229 ThawVerifyOopsClosure cl(st);
3230 NMethodToOopClosure cf(&cl, false);
3231
3232 StackFrameStream fst(thread, true, false);
3233 fst.register_map()->set_include_argument_oops(false);
3234 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3235 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
3236 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
3237 st->print_cr(">>> do_verify_after_thaw deopt");
3238 fst.current()->deoptimize(nullptr);
3239 fst.current()->print_on(st);
3240 }
3241
3242 fst.current()->oops_do(&cl, &cf, fst.register_map());
3243 if (cl.p() != nullptr) {
3244 frame fr = *fst.current();
3245 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
3246 fr.print_on(st);
3247 if (!fr.is_interpreted_frame()) {
3248 st->print_cr("size: %d argsize: %d",
3249 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
3250 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
3251 }
3252 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
3253 if (reg != nullptr) {
3254 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
3255 }
3256 cl.reset();
3257 DEBUG_ONLY(thread->print_frame_layout();)
3258 if (chunk != nullptr) {
3259 chunk->print_on(true, st);
3260 }
3261 return false;
3262 }
3263 }
3264 return true;
3265 }
3266
3267 static bool verify_deopt_state(const frame& f) {
3268 nmethod* nm = f.cb()->as_nmethod();
3269 assert(f.is_deoptimized_frame() == nm->is_deopt_pc(f.raw_pc()), "");
3270 assert(f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
3271 assert(f.is_deoptimized_frame() == nm->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
3272 return true;
3273 }
3274
3275 static void log_frames(JavaThread* thread) {
3276 const static int show_entry_callers = 3;
3277 LogTarget(Trace, continuations) lt;
3278 if (!lt.develop_is_enabled()) {
3279 return;
3280 }
3281 LogStream ls(lt);
3282
3283 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3284 if (!thread->has_last_Java_frame()) {
3285 ls.print_cr("NO ANCHOR!");
3286 }
3287
3288 RegisterMap map(thread,
3289 RegisterMap::UpdateMap::include,
3290 RegisterMap::ProcessFrames::include,
3291 RegisterMap::WalkContinuation::skip);
3292 map.set_include_argument_oops(false);
3293
3294 if (false) {
3295 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3296 f.print_on(&ls);
3297 }
3298 } else {
3299 map.set_skip_missing(true);
3300 ResetNoHandleMark rnhm;
3301 ResourceMark rm;
3302 HandleMark hm(Thread::current());
3303 FrameValues values;
3304
3305 int i = 0;
3306 int post_entry = -1;
3307 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3308 f.describe(values, i, &map, i == 0);
3309 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3310 post_entry++;
3311 if (post_entry >= show_entry_callers)
3312 break;
3313 }
3314 values.print_on(thread, &ls);
3315 }
3316
3317 ls.print_cr("======= end frames =========");
3318 }
3319
3320 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp) {
3321 intptr_t* sp0 = sp;
3322 bool preempted = false;
3323 stackChunkOop tail = cont.tail();
3324 if (tail != nullptr && tail->preempted()) {
3325 // Still preempted (monitor not acquired) so no frames were thawed.
3326 set_anchor(thread, cont.entrySP(), cont.entryPC());
3327 preempted = true;
3328 } else {
3329 set_anchor(thread, sp0);
3330 }
3331
3332 log_frames(thread);
3333 if (LoomVerifyAfterThaw) {
3334 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
3335 }
3336 assert(preempted || ContinuationEntry::assert_entry_frame_laid_out(thread), "");
3337 clear_anchor(thread);
3338
3339 LogTarget(Trace, continuations) lt;
3340 if (lt.develop_is_enabled()) {
3341 LogStream ls(lt);
3342 ls.print_cr("Jumping to frame (thaw):");
3343 frame(sp).print_value_on(&ls);
3344 }
3345 }
3346 #endif // ASSERT
3347
3348 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3349
3350 #ifdef ASSERT
3351 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3352 ResourceMark rm;
3353 FrameValues values;
3354 assert(f.get_cb() != nullptr, "");
3355 RegisterMap map(f.is_heap_frame() ?
3356 nullptr :
3357 JavaThread::current(),
3358 RegisterMap::UpdateMap::include,
3359 RegisterMap::ProcessFrames::skip,
3360 RegisterMap::WalkContinuation::skip);
3361 map.set_include_argument_oops(false);
3362 map.set_skip_missing(true);
3363 if (callee_complete) {
3364 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3365 }
3366 const_cast<frame&>(f).describe(values, 0, &map, true);
3367 values.print_on(static_cast<JavaThread*>(nullptr), st);
3368 }
3369 #endif
3370
3371 static address thaw_entry = nullptr;
3372 static address freeze_entry = nullptr;
3373 static address freeze_preempt_entry = nullptr;
3374
3375 address Continuation::thaw_entry() {
3376 return ::thaw_entry;
3377 }
3378
3379 address Continuation::freeze_entry() {
3380 return ::freeze_entry;
3381 }
3382
3383 address Continuation::freeze_preempt_entry() {
3384 return ::freeze_preempt_entry;
3385 }
3386
3387 class ConfigResolve {
3388 public:
3389 static void resolve() { resolve_compressed(); }
3390
3391 static void resolve_compressed() {
3392 UseCompressedOops ? resolve_gc<true>()
3393 : resolve_gc<false>();
3394 }
3395
3396 private:
3397 template <bool use_compressed>
3398 static void resolve_gc() {
3399 BarrierSet* bs = BarrierSet::barrier_set();
3400 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3401 switch (bs->kind()) {
3402 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3403 case BarrierSet::bs_name: { \
3404 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3405 } \
3406 break;
3407 FOR_EACH_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3408 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3409
3410 default:
3411 fatal("BarrierSet resolving not implemented");
3412 };
3413 }
3414
3415 template <bool use_compressed, typename BarrierSetT>
3416 static void resolve() {
3417 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3418
3419 freeze_entry = (address)freeze<SelectedConfigT>;
3420 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3421
3422 // If we wanted, we could templatize by kind and have three different thaw entries
3423 thaw_entry = (address)thaw<SelectedConfigT>;
3424 }
3425 };
3426
3427 void Continuation::init() {
3428 ConfigResolve::resolve();
3429 }