1 /*
2 * Copyright (c) 2018, 2025, 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 "gc/shared/barrierSet.hpp"
32 #include "gc/shared/continuationGCSupport.inline.hpp"
33 #include "gc/shared/gc_globals.hpp"
34 #include "gc/shared/memAllocator.hpp"
35 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
36 #include "interpreter/bytecodeStream.hpp"
37 #include "interpreter/interpreter.hpp"
38 #include "interpreter/interpreterRuntime.hpp"
39 #include "jfr/jfrEvents.hpp"
40 #include "logging/log.hpp"
41 #include "logging/logStream.hpp"
42 #include "oops/access.inline.hpp"
43 #include "oops/constantPool.inline.hpp"
44 #include "oops/method.inline.hpp"
45 #include "oops/objArrayOop.inline.hpp"
46 #include "oops/oopsHierarchy.hpp"
47 #include "oops/stackChunkOop.inline.hpp"
48 #include "prims/jvmtiThreadState.hpp"
49 #include "runtime/arguments.hpp"
50 #include "runtime/continuation.hpp"
51 #include "runtime/continuationEntry.inline.hpp"
52 #include "runtime/continuationHelper.inline.hpp"
53 #include "runtime/continuationJavaClasses.inline.hpp"
54 #include "runtime/continuationWrapper.inline.hpp"
55 #include "runtime/frame.inline.hpp"
56 #include "runtime/interfaceSupport.inline.hpp"
57 #include "runtime/javaThread.inline.hpp"
58 #include "runtime/jniHandles.inline.hpp"
59 #include "runtime/keepStackGCProcessed.hpp"
60 #include "runtime/objectMonitor.inline.hpp"
61 #include "runtime/orderAccess.hpp"
62 #include "runtime/prefetch.inline.hpp"
63 #include "runtime/sharedRuntime.hpp"
64 #include "runtime/smallRegisterMap.inline.hpp"
65 #include "runtime/stackChunkFrameStream.inline.hpp"
66 #include "runtime/stackFrameStream.inline.hpp"
67 #include "runtime/stackOverflow.hpp"
68 #include "runtime/stackWatermarkSet.inline.hpp"
69 #include "runtime/vframe.inline.hpp"
70 #include "runtime/vframe_hp.hpp"
71 #include "utilities/debug.hpp"
72 #include "utilities/exceptions.hpp"
73 #include "utilities/macros.hpp"
74 #include "utilities/vmError.hpp"
75 #if INCLUDE_ZGC
76 #include "gc/z/zStackChunkGCData.inline.hpp"
77 #endif
78 #if INCLUDE_JFR
79 #include "jfr/jfr.inline.hpp"
80 #endif
81 #ifdef COMPILER1
82 #include "c1/c1_Runtime1.hpp"
83 #endif
84 #ifdef COMPILER2
85 #include "opto/runtime.hpp"
86 #endif
87
88 #include <type_traits>
89
90 /*
91 * This file contains the implementation of continuation freezing (yield) and thawing (run).
92 *
93 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
94 * would likely call these operations many thousands of times per second second, on every core.
95 *
96 * Freeze might be called every time the application performs any I/O operation, every time it
97 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
98 * multiple times in each of those cases, as it is called by the return barrier, which may be
99 * invoked on method return.
100 *
101 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
102 * example, every effort is made to avoid Java-VM transitions as much as possible.
103 *
104 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
105 * and so frames simply copied, and the bottom-most one is patched.
106 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
107 * and absolute pointers, and barriers invoked.
108 */
109
110 /************************************************
111
112 Thread-stack layout on freeze/thaw.
113 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
114
115 +----------------------------+
116 | . |
117 | . |
118 | . |
119 | carrier frames |
120 | |
121 |----------------------------|
122 | |
123 | Continuation.run |
124 | |
125 |============================|
126 | enterSpecial frame |
127 | pc |
128 | rbp |
129 | ----- |
130 ^ | int argsize | = ContinuationEntry
131 | | oopDesc* cont |
132 | | oopDesc* chunk |
133 | | ContinuationEntry* parent |
134 | | ... |
135 | |============================| <------ JavaThread::_cont_entry = entry->sp()
136 | | ? alignment word ? |
137 | |----------------------------| <--\
138 | | | |
139 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
140 Address | | | | Caller is still in the chunk.
141 | |----------------------------| |
142 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
143 | | rbp | | isn't the last one in the continuation.
144 | | | |
145 | | frame | |
146 | | | |
147 +----------------------------| \__ Continuation frames to be frozen/thawed
148 | | /
149 | frame | |
150 | | |
151 |----------------------------| |
152 | | |
153 | frame | |
154 | | |
155 |----------------------------| <--/
156 | |
157 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
158 | | instead of a doYield stub
159 |============================| <- the sp passed to freeze
160 | |
161 | Native freeze/thaw frames |
162 | . |
163 | . |
164 | . |
165 +----------------------------+
166
167 ************************************************/
168
169 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing
170
171 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
172 #if CONT_JFR
173 #define CONT_JFR_ONLY(code) code
174 #else
175 #define CONT_JFR_ONLY(code)
176 #endif
177
178 // TODO: See AbstractAssembler::generate_stack_overflow_check,
179 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
180 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
181
182 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
183
184 // Used to just annotatate cold/hot branches
185 #define LIKELY(condition) (condition)
186 #define UNLIKELY(condition) (condition)
187
188 // debugging functions
189 #ifdef ASSERT
190 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
191
192 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
193
194 static void do_deopt_after_thaw(JavaThread* thread);
195 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
196 static void log_frames(JavaThread* thread, bool dolog = false);
197 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp);
198 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
199 static void verify_frame_kind(const frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr = nullptr, const char** code_name_ptr = nullptr, int* bci_ptr = nullptr);
200
201 #define assert_pfl(p, ...) \
202 do { \
203 if (!(p)) { \
204 JavaThread* t = JavaThread::active(); \
205 if (t->has_last_Java_frame()) { \
206 tty->print_cr("assert(" #p ") failed:"); \
207 t->print_frame_layout(); \
208 } \
209 } \
210 vmassert(p, __VA_ARGS__); \
211 } while(0)
212
213 #else
214 static void verify_continuation(oop continuation) { }
215 #define assert_pfl(p, ...)
216 #endif
217
218 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
219 template<typename ConfigT, bool preempt> static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp);
220
221 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
222 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
223
224
225 // Entry point to freeze. Transitions are handled manually
226 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
227 template<typename ConfigT>
228 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
229 assert(sp == current->frame_anchor()->last_Java_sp(), "");
230
231 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
232 current->set_cont_fastpath(nullptr);
233 }
234
235 return checked_cast<int>(ConfigT::freeze(current, sp));
236 JRT_END
237
238 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
239 return prepare_thaw_internal(thread, return_barrier);
240 JRT_END
241
242 template<typename ConfigT>
243 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
244 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
245 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
246 // Also the preemption case with JVMTI events enabled might safepoint so
247 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
248 // JRT_ENTRY instead?
249 ResetNoHandleMark rnhm;
250 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
251
252 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
253 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
254 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
255 JRT_END
256
257 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
258 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
259 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
260 }
261 JVM_END
262
263 ///////////
264
265 enum class oop_kind { NARROW, WIDE };
266 template <oop_kind oops, typename BarrierSetT>
267 class Config {
268 public:
269 typedef Config<oops, BarrierSetT> SelfT;
270 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
271
272 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
273 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
274 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
275 return res;
276 }
277
278 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
279 return freeze_internal<SelfT, true>(thread, sp);
280 }
281
282 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
283 return thaw_internal<SelfT>(thread, kind);
284 }
285 };
286
287 #ifdef _WINDOWS
288 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
289 address new_sp = sp - size;
290 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
291
292 if (new_sp < watermark) {
293 size_t page_size = os::vm_page_size();
294 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
295 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
296 while (pages_to_touch-- > 0) {
297 last_touched_page -= page_size;
298 *last_touched_page = 0;
299 }
300 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
301 }
302 }
303 #endif
304
305 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
306 const size_t page_size = os::vm_page_size();
307 if (size > page_size) {
308 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
309 return false;
310 }
311 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
312 }
313 return true;
314 }
315
316 #ifdef ASSERT
317 static oop get_continuation(JavaThread* thread) {
318 assert(thread != nullptr, "");
319 assert(thread->threadObj() != nullptr, "");
320 return java_lang_Thread::continuation(thread->threadObj());
321 }
322 #endif // ASSERT
323
324 inline void clear_anchor(JavaThread* thread) {
325 thread->frame_anchor()->clear();
326 }
327
328 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
329 assert(pc != nullptr, "");
330
331 JavaFrameAnchor* anchor = thread->frame_anchor();
332 anchor->set_last_Java_sp(sp);
333 anchor->set_last_Java_pc(pc);
334 ContinuationHelper::set_anchor_pd(anchor, sp);
335
336 assert(thread->has_last_Java_frame(), "");
337 assert(thread->last_frame().cb() != nullptr, "");
338 }
339
340 static void set_anchor(JavaThread* thread, intptr_t* sp) {
341 address pc = ContinuationHelper::return_address_at(
342 sp - frame::sender_sp_ret_address_offset());
343 set_anchor(thread, sp, pc);
344 }
345
346 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
347 JavaFrameAnchor* anchor = thread->frame_anchor();
348 anchor->set_last_Java_sp(entry->entry_sp());
349 anchor->set_last_Java_pc(entry->entry_pc());
350 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
351
352 assert(thread->has_last_Java_frame(), "");
353 assert(thread->last_frame().cb() != nullptr, "");
354 }
355
356 #if CONT_JFR
357 class FreezeThawJfrInfo : public StackObj {
358 short _e_size;
359 short _e_num_interpreted_frames;
360 public:
361
362 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
363 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
364 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
365 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
366 };
367
368 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
369 if (e->should_commit()) {
370 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
371 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
372 e->set_continuationClass(continuation->klass());
373 e->set_interpretedFrames(_e_num_interpreted_frames);
374 e->set_size(_e_size);
375 e->commit();
376 }
377 }
378 #endif // CONT_JFR
379
380 /////////////// FREEZE ////
381
382 class FreezeBase : public StackObj {
383 protected:
384 JavaThread* const _thread;
385 ContinuationWrapper& _cont;
386 bool _barriers; // only set when we allocate a chunk
387
388 intptr_t* _bottom_address;
389
390 // Used for preemption only
391 const bool _preempt;
392 frame _last_frame;
393
394 // Used to support freezing with held monitors
395 int _monitors_in_lockstack;
396
397 int _freeze_size; // total size of all frames plus metadata in words.
398 int _total_align_size;
399
400 intptr_t* _cont_stack_top;
401 intptr_t* _cont_stack_bottom;
402
403 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
404
405 #ifdef ASSERT
406 intptr_t* _orig_chunk_sp;
407 int _fast_freeze_size;
408 bool _empty;
409 #endif
410
411 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
412
413 NOT_PRODUCT(int _frames;)
414 DEBUG_ONLY(intptr_t* _last_write;)
415
416 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
417
418 public:
419 NOINLINE freeze_result freeze_slow();
420 void freeze_fast_existing_chunk();
421
422 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
423 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
424
425 inline int size_if_fast_freeze_available();
426
427 inline frame& last_frame() { return _last_frame; }
428
429 #ifdef ASSERT
430 bool check_valid_fast_path();
431 #endif
432
433 protected:
434 inline void init_rest();
435 void throw_stack_overflow_on_humongous_chunk();
436
437 // fast path
438 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
439 inline void unwind_frames();
440 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
441
442 // slow path
443 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
444
445 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
446
447 private:
448 // slow path
449 frame freeze_start_frame();
450 frame freeze_start_frame_on_preempt();
451 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
452 inline frame freeze_start_frame_yield_stub();
453 template<typename FKind>
454 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
455 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
456 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
457 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
458 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
459 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
460 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
461 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
462 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
463 NOINLINE void finish_freeze(const frame& f, const frame& top);
464
465 void freeze_lockstack(stackChunkOop chunk);
466
467 inline bool stack_overflow();
468
469 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
470 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
471 template<typename FKind> static inline frame sender(const frame& f);
472 template<typename FKind> frame new_heap_frame(frame& f, frame& caller);
473 inline void set_top_frame_metadata_pd(const frame& hf);
474 inline void patch_pd(frame& callee, const frame& caller);
475 inline void patch_pd_unused(intptr_t* sp);
476 void adjust_interpreted_frame_unextended_sp(frame& f);
477 static inline void prepare_freeze_interpreted_top_frame(frame& f);
478 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
479
480 protected:
481 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
482 bool freeze_fast_new_chunk(stackChunkOop chunk);
483 };
484
485 template <typename ConfigT>
486 class Freeze : public FreezeBase {
487 private:
488 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
489
490 public:
491 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
492 : FreezeBase(thread, cont, frame_sp, preempt) {}
493
494 freeze_result try_freeze_fast();
495
496 protected:
497 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
498 };
499
500 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
501 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
502 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
503
504 assert(_thread != nullptr, "");
505 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
506
507 DEBUG_ONLY(_cont.entry()->verify_cookie();)
508
509 assert(!Interpreter::contains(_cont.entryPC()), "");
510
511 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
512 #ifdef _LP64
513 if (((intptr_t)_bottom_address & 0xf) != 0) {
514 _bottom_address--;
515 }
516 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
517 #endif
518
519 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
520 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
521 assert(_bottom_address != nullptr, "");
522 assert(_bottom_address <= _cont.entrySP(), "");
523 DEBUG_ONLY(_last_write = nullptr;)
524
525 assert(_cont.chunk_invariant(), "");
526 assert(!Interpreter::contains(_cont.entryPC()), "");
527 #if !defined(PPC64) || defined(ZERO)
528 static const int do_yield_frame_size = frame::metadata_words;
529 #else
530 static const int do_yield_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
531 #endif
532 // With preemption doYield() might not have been resolved yet
533 assert(_preempt || ContinuationEntry::do_yield_nmethod()->frame_size() == do_yield_frame_size, "");
534
535 if (preempt) {
536 _last_frame = _thread->last_frame();
537 }
538
539 // properties of the continuation on the stack; all sizes are in words
540 _cont_stack_top = frame_sp + (!preempt ? do_yield_frame_size : 0); // we don't freeze the doYield stub frame
541 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
542 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
543
544 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
545 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
546 assert(cont_size() > 0, "");
547
548 if (LockingMode != LM_LIGHTWEIGHT) {
549 _monitors_in_lockstack = 0;
550 } else {
551 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
552 }
553 }
554
555 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
556 _freeze_size = 0;
557 _total_align_size = 0;
558 NOT_PRODUCT(_frames = 0;)
559 }
560
561 void FreezeBase::freeze_lockstack(stackChunkOop chunk) {
562 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
563
564 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
565 chunk->set_lockstack_size(checked_cast<uint8_t>(_monitors_in_lockstack));
566 chunk->set_has_lockstack(true);
567 }
568
569 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
570 stackChunkOop chunk = _cont.tail();
571 chunk->copy_from_stack_to_chunk(from, to, size);
572 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
573
574 #ifdef ASSERT
575 if (_last_write != nullptr) {
576 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
577 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
578 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
579 _last_write = to;
580 }
581 #endif
582 }
583
584 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
585 #ifdef ASSERT
586 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
587 if (watermark == nullptr) {
588 return;
589 }
590 ContinuationEntry* ce = thread->last_continuation();
591 RegisterMap map(thread,
592 RegisterMap::UpdateMap::include,
593 RegisterMap::ProcessFrames::skip,
594 RegisterMap::WalkContinuation::skip);
595 map.set_include_argument_oops(false);
596 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
597 watermark->assert_is_frame_safe(f);
598 }
599 #endif // ASSERT
600 }
601
602 #ifdef ASSERT
603 static bool monitors_on_stack(JavaThread* thread) {
604 assert_frames_in_continuation_are_safe(thread);
605 ContinuationEntry* ce = thread->last_continuation();
606 RegisterMap map(thread,
607 RegisterMap::UpdateMap::include,
608 RegisterMap::ProcessFrames::skip,
609 RegisterMap::WalkContinuation::skip);
610 map.set_include_argument_oops(false);
611 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
612 if ((f.is_interpreted_frame() && ContinuationHelper::InterpretedFrame::is_owning_locks(f)) ||
613 (f.is_compiled_frame() && ContinuationHelper::CompiledFrame::is_owning_locks(map.thread(), &map, f)) ||
614 (f.is_native_frame() && ContinuationHelper::NativeFrame::is_owning_locks(map.thread(), f))) {
615 return true;
616 }
617 }
618 return false;
619 }
620 #endif // ASSERT
621
622 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
623 void FreezeBase::unwind_frames() {
624 ContinuationEntry* entry = _cont.entry();
625 entry->flush_stack_processing(_thread);
626 assert_frames_in_continuation_are_safe(_thread);
627 JFR_ONLY(Jfr::check_and_process_sample_request(_thread);)
628 assert(LockingMode != LM_LEGACY || !monitors_on_stack(_thread), "unexpected monitors on stack");
629 set_anchor_to_entry(_thread, entry);
630 }
631
632 template <typename ConfigT>
633 freeze_result Freeze<ConfigT>::try_freeze_fast() {
634 assert(_thread->thread_state() == _thread_in_vm, "");
635 assert(_thread->cont_fastpath(), "");
636
637 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
638 assert(_fast_freeze_size == 0, "");
639
640 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
641 if (freeze_fast_new_chunk(chunk)) {
642 return freeze_ok;
643 }
644 if (_thread->has_pending_exception()) {
645 return freeze_exception;
646 }
647
648 // TODO R REMOVE when deopt change is fixed
649 assert(!_thread->cont_fastpath() || _barriers, "");
650 log_develop_trace(continuations)("-- RETRYING SLOW --");
651 return freeze_slow();
652 }
653
654 // Returns size needed if the continuation fits, otherwise 0.
655 int FreezeBase::size_if_fast_freeze_available() {
656 stackChunkOop chunk = _cont.tail();
657 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
658 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
659 return 0;
660 }
661
662 int total_size_needed = cont_size();
663 const int chunk_sp = chunk->sp();
664
665 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
666 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
667 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
668 // although that would require changing stackChunkOopDesc::is_empty
669 if (!chunk->is_empty()) {
670 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
671 }
672
673 total_size_needed += _monitors_in_lockstack;
674
675 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
676 bool available = chunk_free_room >= total_size_needed;
677 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
678 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
679 return available ? total_size_needed : 0;
680 }
681
682 void FreezeBase::freeze_fast_existing_chunk() {
683 stackChunkOop chunk = _cont.tail();
684
685 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
686 assert(_fast_freeze_size > 0, "");
687
688 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
689 DEBUG_ONLY(_empty = false;)
690 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
691 #ifdef ASSERT
692 {
693 intptr_t* retaddr_slot = (chunk->sp_address()
694 - frame::sender_sp_ret_address_offset());
695 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
696 "unexpected saved return address");
697 }
698 #endif
699
700 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
701 // we overlap; we'll overwrite the chunk's top frame's callee arguments
702 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
703 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
704
705 // increase max_size by what we're freezing minus the overlap
706 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
707
708 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
709 assert(bottom_sp == _bottom_address, "");
710 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
711 // should have a return barrier (installed back when we thawed it).
712 #ifdef ASSERT
713 {
714 intptr_t* retaddr_slot = (bottom_sp
715 - frame::sender_sp_ret_address_offset());
716 assert(ContinuationHelper::return_address_at(retaddr_slot)
717 == StubRoutines::cont_returnBarrier(),
718 "should be the continuation return barrier");
719 }
720 #endif
721 // We copy the fp from the chunk back to the stack because it contains some caller data,
722 // including, possibly, an oop that might have gone stale since we thawed.
723 patch_stack_pd(bottom_sp, chunk->sp_address());
724 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
725
726 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
727 } else { // the chunk is empty
728 const int chunk_start_sp = chunk->stack_size();
729
730 DEBUG_ONLY(_empty = true;)
731 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
732
733 chunk->set_max_thawing_size(cont_size());
734 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
735 chunk->set_sp(chunk->bottom());
736
737 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
738 }
739 }
740
741 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
742 DEBUG_ONLY(_empty = true;)
743
744 // Install new chunk
745 _cont.set_tail(chunk);
746
747 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
748 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
749 return false;
750 }
751
752 chunk->set_max_thawing_size(cont_size());
753
754 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
755 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
756 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
757 assert(chunk_start_sp == chunk->stack_size(), "");
758
759 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
760
761 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
762
763 return true;
764 }
765
766 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
767 assert(chunk != nullptr, "");
768 assert(!chunk->has_mixed_frames(), "");
769 assert(!chunk->is_gc_mode(), "");
770 assert(!chunk->has_bitmap(), "");
771 assert(!chunk->requires_barriers(), "");
772 assert(chunk == _cont.tail(), "");
773
774 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
775 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
776 // will either see no continuation on the stack, or a consistent chunk.
777 unwind_frames();
778
779 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
780 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
781 assert(chunk_start_sp <= chunk->stack_size(), "");
782 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
783
784 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
785 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
786
787 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
788 #ifdef ASSERT
789 if (!_empty) {
790 intptr_t* retaddr_slot = (_orig_chunk_sp
791 - frame::sender_sp_ret_address_offset());
792 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
793 "unexpected saved return address");
794 }
795 #endif
796
797 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
798 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
799
800 int adjust = frame::metadata_words_at_bottom;
801 #if INCLUDE_ASAN && defined(AARCH64)
802 // Reading at offset frame::metadata_words_at_bottom from _cont_stack_top
803 // will accesss memory at the callee frame, which on preemption cases will
804 // be the VM native method being called. The Arm 64-bit ABI doesn't specify
805 // a location where the frame record (returnpc+fp) has to be stored within
806 // a stack frame, and GCC currently chooses to save it at the top of the
807 // frame (lowest address). ASan treats this memory access in the callee as
808 // an overflow access to one of the locals stored in that frame. For these
809 // preemption cases we don't need to read these words anyways so we avoid it.
810 if (_preempt) {
811 adjust = 0;
812 }
813 #endif
814 intptr_t* from = _cont_stack_top - adjust;
815 intptr_t* to = chunk_top - adjust;
816 copy_to_chunk(from, to, cont_size() + adjust);
817 // Because we're not patched yet, the chunk is now in a bad state
818
819 // patch return pc of the bottom-most frozen frame (now in the chunk)
820 // with the actual caller's return address
821 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
822 - _cont.argsize()
823 - frame::metadata_words_at_top
824 - frame::sender_sp_ret_address_offset());
825 #ifdef ASSERT
826 if (!_empty) {
827 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
828 == StubRoutines::cont_returnBarrier(),
829 "should be the continuation return barrier");
830 }
831 #endif
832 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
833 chunk->pc());
834
835 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
836 chunk->set_sp(chunk_new_sp);
837
838 // set chunk->pc to the return address of the topmost frame in the chunk
839 if (_preempt) {
840 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
841 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
842 // be the pc we used when creating the oopmap. Get the top's frame last pc from
843 // the anchor instead.
844 address last_pc = _last_frame.pc();
845 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
846 chunk->set_pc(last_pc);
847 // For stub/native frames the fp is not used while frozen, and will be constructed
848 // again when thawing the frame (see ThawBase::handle_preempted_continuation). We
849 // patch it with a special bad address to help with debugging, particularly when
850 // inspecting frames and identifying invalid accesses.
851 patch_pd_unused(chunk_top);
852 } else {
853 chunk->set_pc(ContinuationHelper::return_address_at(
854 _cont_stack_top - frame::sender_sp_ret_address_offset()));
855 }
856
857 if (_monitors_in_lockstack > 0) {
858 freeze_lockstack(chunk);
859 }
860
861 _cont.write();
862
863 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
864 LogTarget(Trace, continuations) lt;
865 if (lt.develop_is_enabled()) {
866 LogStream ls(lt);
867 chunk->print_on(true, &ls);
868 }
869
870 // Verification
871 assert(_cont.chunk_invariant(), "");
872 chunk->verify();
873
874 #if CONT_JFR
875 EventContinuationFreezeFast e;
876 if (e.should_commit()) {
877 e.set_id(cast_from_oop<u8>(chunk));
878 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
879 e.set_size(cont_size() << LogBytesPerWord);
880 e.commit();
881 }
882 #endif
883 }
884
885 NOINLINE freeze_result FreezeBase::freeze_slow() {
886 #ifdef ASSERT
887 ResourceMark rm;
888 #endif
889
890 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
891 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
892
893 #if CONT_JFR
894 EventContinuationFreezeSlow e;
895 if (e.should_commit()) {
896 e.set_id(cast_from_oop<u8>(_cont.continuation()));
897 e.commit();
898 }
899 #endif
900
901 init_rest();
902
903 HandleMark hm(Thread::current());
904
905 frame f = freeze_start_frame();
906
907 LogTarget(Debug, continuations) lt;
908 if (lt.develop_is_enabled()) {
909 LogStream ls(lt);
910 f.print_on(&ls);
911 }
912
913 frame caller; // the frozen caller in the chunk
914 freeze_result res = recurse_freeze(f, caller, 0, false, true);
915
916 if (res == freeze_ok) {
917 finish_freeze(f, caller);
918 _cont.write();
919 }
920
921 return res;
922 }
923
924 frame FreezeBase::freeze_start_frame() {
925 if (LIKELY(!_preempt)) {
926 return freeze_start_frame_yield_stub();
927 } else {
928 return freeze_start_frame_on_preempt();
929 }
930 }
931
932 frame FreezeBase::freeze_start_frame_yield_stub() {
933 frame f = _thread->last_frame();
934 assert(ContinuationEntry::do_yield_nmethod()->contains(f.pc()), "must be");
935 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
936 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
937 return f;
938 }
939
940 frame FreezeBase::freeze_start_frame_on_preempt() {
941 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
942 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
943 return _last_frame;
944 }
945
946 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
947 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
948 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
949 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
950 || ((top && _preempt) == f.is_native_frame()), "");
951
952 if (stack_overflow()) {
953 return freeze_exception;
954 }
955
956 if (f.is_compiled_frame()) {
957 if (UNLIKELY(f.oop_map() == nullptr)) {
958 // special native frame
959 return freeze_pinned_native;
960 }
961 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
962 } else if (f.is_interpreted_frame()) {
963 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
964 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
965 } else if (top && _preempt) {
966 assert(f.is_native_frame() || f.is_runtime_frame(), "");
967 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
968 } else {
969 // Frame can't be frozen. Most likely the call_stub or upcall_stub
970 // which indicates there are further natives frames up the stack.
971 return freeze_pinned_native;
972 }
973 }
974
975 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
976 // See also StackChunkFrameStream<frame_kind>::frame_size()
977 template<typename FKind>
978 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
979 assert(FKind::is_instance(f), "");
980
981 assert(fsize > 0, "");
982 assert(argsize >= 0, "");
983 _freeze_size += fsize;
984 NOT_PRODUCT(_frames++;)
985
986 assert(FKind::frame_bottom(f) <= _bottom_address, "");
987
988 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
989 // enterSpecial and an interpreted frame
990 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
991 return finalize_freeze(f, caller, argsize); // recursion end
992 } else {
993 frame senderf = sender<FKind>(f);
994 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
995 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
996 return result;
997 }
998 }
999
1000 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
1001 LogTarget(Trace, continuations) lt;
1002 if (lt.develop_is_enabled()) {
1003 LogStream ls(lt);
1004 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
1005 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
1006 f.print_value_on(&ls);
1007 }
1008 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
1009 }
1010
1011 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
1012 LogTarget(Trace, continuations) lt;
1013 if (lt.develop_is_enabled()) {
1014 LogStream ls(lt);
1015 DEBUG_ONLY(hf.print_value_on(&ls);)
1016 assert(hf.is_heap_frame(), "should be");
1017 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
1018 if (is_bottom_frame) {
1019 ls.print_cr("bottom h-frame:");
1020 hf.print_on(&ls);
1021 }
1022 }
1023 }
1024
1025 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
1026 // See also StackChunkFrameStream<frame_kind>::frame_size()
1027 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
1028 int argsize = argsize_md - frame::metadata_words_at_top;
1029 assert(callee.is_interpreted_frame()
1030 || ContinuationHelper::Frame::is_stub(callee.cb())
1031 || callee.cb()->as_nmethod()->is_osr_method()
1032 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
1033 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
1034 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
1035
1036 LogTarget(Trace, continuations) lt;
1037
1038 #ifdef ASSERT
1039 bool empty = _cont.is_empty();
1040 log_develop_trace(continuations)("empty: %d", empty);
1041 #endif
1042
1043 stackChunkOop chunk = _cont.tail();
1044
1045 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1046
1047 _freeze_size += frame::metadata_words; // for top frame's metadata
1048
1049 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1050 int unextended_sp = -1;
1051 if (chunk != nullptr) {
1052 if (!chunk->is_empty()) {
1053 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1054 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1055 bool top_interpreted = Interpreter::contains(chunk->pc());
1056 if (callee.is_interpreted_frame() == top_interpreted) {
1057 overlap = argsize_md;
1058 }
1059 } else {
1060 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1061 }
1062 }
1063
1064 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1065
1066 _freeze_size -= overlap;
1067 assert(_freeze_size >= 0, "");
1068
1069 assert(chunk == nullptr || chunk->is_empty()
1070 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1071 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1072
1073 _freeze_size += _monitors_in_lockstack;
1074
1075 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1076 bool allocated_old_in_freeze_fast = _barriers;
1077 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1078 "Chunk allocated in freeze_fast is of insufficient size "
1079 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1080 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1081
1082 DEBUG_ONLY(bool empty_chunk = true);
1083 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1084 // ALLOCATE NEW CHUNK
1085
1086 if (lt.develop_is_enabled()) {
1087 LogStream ls(lt);
1088 if (chunk == nullptr) {
1089 ls.print_cr("no chunk");
1090 } else {
1091 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1092 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1093 chunk->print_on(&ls);
1094 }
1095 }
1096
1097 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1098 // overlap = 0;
1099
1100 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1101 if (chunk == nullptr) {
1102 return freeze_exception;
1103 }
1104
1105 // Install new chunk
1106 _cont.set_tail(chunk);
1107 assert(chunk->is_empty(), "");
1108 } else {
1109 // REUSE EXISTING CHUNK
1110 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1111 if (chunk->is_empty()) {
1112 int sp = chunk->stack_size() - argsize_md;
1113 chunk->set_sp(sp);
1114 chunk->set_bottom(sp);
1115 _freeze_size += overlap;
1116 assert(chunk->max_thawing_size() == 0, "");
1117 } DEBUG_ONLY(else empty_chunk = false;)
1118 }
1119 assert(!chunk->is_gc_mode(), "");
1120 assert(!chunk->has_bitmap(), "");
1121 chunk->set_has_mixed_frames(true);
1122
1123 assert(chunk->requires_barriers() == _barriers, "");
1124 assert(!_barriers || chunk->is_empty(), "");
1125
1126 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1127 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1128
1129 if (_preempt) {
1130 frame top_frame = _thread->last_frame();
1131 if (top_frame.is_interpreted_frame()) {
1132 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1133 // We need it so that on resume we can restore the sp to the right place, since
1134 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1135 // We do it now that we know freezing will be successful.
1136 prepare_freeze_interpreted_top_frame(top_frame);
1137 }
1138
1139 // Do this now so should_process_args_at_top() is set before calling finish_freeze
1140 // in case we might need to apply GC barriers to frames in this stackChunk.
1141 if (_thread->at_preemptable_init()) {
1142 assert(top_frame.is_interpreted_frame(), "only InterpreterRuntime::_new/resolve_from_cache allowed");
1143 chunk->set_at_klass_init(true);
1144 Method* m = top_frame.interpreter_frame_method();
1145 Bytecode current_bytecode = Bytecode(m, top_frame.interpreter_frame_bcp());
1146 Bytecodes::Code code = current_bytecode.code();
1147 int exp_size = top_frame.interpreter_frame_expression_stack_size();
1148 if (code == Bytecodes::Code::_invokestatic && exp_size > 0) {
1149 chunk->set_has_args_at_top(true);
1150 }
1151 }
1152 }
1153
1154 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1155 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1156 // will either see no continuation or a consistent chunk.
1157 unwind_frames();
1158
1159 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1160
1161 if (lt.develop_is_enabled()) {
1162 LogStream ls(lt);
1163 ls.print_cr("top chunk:");
1164 chunk->print_on(&ls);
1165 }
1166
1167 if (_monitors_in_lockstack > 0) {
1168 freeze_lockstack(chunk);
1169 }
1170
1171 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1172 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1173
1174 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1175
1176 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1177 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1178 #ifdef ASSERT
1179 if (lt.develop_is_enabled()) {
1180 LogStream ls(lt);
1181 ls.print_cr("top hframe before (freeze):");
1182 assert(caller.is_heap_frame(), "should be");
1183 caller.print_on(&ls);
1184 }
1185
1186 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1187
1188 frame entry = sender(callee);
1189
1190 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1191 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1192 #endif
1193
1194 return freeze_ok_bottom;
1195 }
1196
1197 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1198 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1199 if (is_bottom_frame) {
1200 // If we're the bottom frame, we need to replace the return barrier with the real
1201 // caller's pc.
1202 address last_pc = caller.pc();
1203 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1204 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1205 } else {
1206 assert(!caller.is_empty(), "");
1207 }
1208
1209 patch_pd(hf, caller);
1210
1211 if (f.is_interpreted_frame()) {
1212 assert(hf.is_heap_frame(), "should be");
1213 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1214 }
1215
1216 #ifdef ASSERT
1217 if (hf.is_compiled_frame()) {
1218 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1219 log_develop_trace(continuations)("Freezing deoptimized frame");
1220 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1221 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1222 }
1223 }
1224 #endif
1225 }
1226
1227 #ifdef ASSERT
1228 static void verify_frame_top(const frame& f, intptr_t* top) {
1229 ResourceMark rm;
1230 InterpreterOopMap mask;
1231 f.interpreted_frame_oop_map(&mask);
1232 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1233 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1234 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1235 }
1236 #endif // ASSERT
1237
1238 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1239 // See also StackChunkFrameStream<frame_kind>::frame_size()
1240 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1241 int callee_argsize /* incl. metadata */,
1242 bool callee_interpreted) {
1243 adjust_interpreted_frame_unextended_sp(f);
1244
1245 // The frame's top never includes the stack arguments to the callee
1246 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1247 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1248 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1249
1250 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1251
1252 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1253 // including metadata between f and its args
1254 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1255
1256 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1257 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1258 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1259 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1260
1261 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1262 if (UNLIKELY(result > freeze_ok_bottom)) {
1263 return result;
1264 }
1265
1266 bool is_bottom_frame = result == freeze_ok_bottom;
1267 assert(!caller.is_empty() || is_bottom_frame, "");
1268
1269 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1270
1271 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1272 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1273
1274 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1275 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1276 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1277
1278 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1279 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1280 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1281 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1282
1283 relativize_interpreted_frame_metadata(f, hf);
1284
1285 patch(f, hf, caller, is_bottom_frame);
1286
1287 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1288 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1289 caller = hf;
1290
1291 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1292 frame_method->record_gc_epoch();
1293
1294 return freeze_ok;
1295 }
1296
1297 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1298 // See also StackChunkFrameStream<frame_kind>::frame_size()
1299 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1300 int callee_argsize /* incl. metadata */,
1301 bool callee_interpreted) {
1302 // The frame's top never includes the stack arguments to the callee
1303 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1304 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1305 // including metadata between f and its stackargs
1306 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1307 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1308
1309 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d",
1310 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1311 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1312 _freeze_size, fsize, argsize);
1313 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1314 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1315
1316 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1317 if (UNLIKELY(result > freeze_ok_bottom)) {
1318 return result;
1319 }
1320
1321 bool is_bottom_frame = result == freeze_ok_bottom;
1322 assert(!caller.is_empty() || is_bottom_frame, "");
1323
1324 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1325
1326 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller);
1327
1328 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1329
1330 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1331 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1332
1333 if (caller.is_interpreted_frame()) {
1334 // When thawing the frame we might need to add alignment (see Thaw::align)
1335 _total_align_size += frame::align_wiggle;
1336 }
1337
1338 patch(f, hf, caller, is_bottom_frame);
1339
1340 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1341
1342 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1343 caller = hf;
1344 return freeze_ok;
1345 }
1346
1347 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1348 DEBUG_ONLY(frame fsender = sender(f);)
1349 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1350
1351 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1352 const int fsize = f.cb()->frame_size();
1353
1354 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1355 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1356
1357 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1358 if (UNLIKELY(result > freeze_ok_bottom)) {
1359 return result;
1360 }
1361
1362 assert(result == freeze_ok, "should have caller");
1363 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1364
1365 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1366 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1367
1368 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1369
1370 patch(f, hf, caller, false /*is_bottom_frame*/);
1371
1372 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1373
1374 caller = hf;
1375 return freeze_ok;
1376 }
1377
1378 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1379 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1380 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1381 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1382 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1383 // caller into the native convention. This is so that we have a valid oopMap in case of
1384 // having to block in the slow path. But that would require freezing those registers too
1385 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1386 // given that this would be a rare case anyways just pin the vthread to the carrier.
1387 return freeze_pinned_native;
1388 }
1389
1390 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1391 // There are no stackargs but argsize must include the metadata
1392 const int argsize = frame::metadata_words_at_top;
1393 const int fsize = f.cb()->frame_size() + argsize;
1394
1395 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1396 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1397
1398 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1399 if (UNLIKELY(result > freeze_ok_bottom)) {
1400 return result;
1401 }
1402
1403 assert(result == freeze_ok, "should have caller frame");
1404 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1405
1406 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1407 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1408
1409 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1410
1411 if (caller.is_interpreted_frame()) {
1412 // When thawing the frame we might need to add alignment (see Thaw::align)
1413 _total_align_size += frame::align_wiggle;
1414 }
1415
1416 patch(f, hf, caller, false /* is_bottom_frame */);
1417
1418 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1419
1420 caller = hf;
1421 return freeze_ok;
1422 }
1423
1424 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1425 stackChunkOop chunk = _cont.tail();
1426
1427 LogTarget(Trace, continuations) lt;
1428 if (lt.develop_is_enabled()) {
1429 LogStream ls(lt);
1430 assert(top.is_heap_frame(), "should be");
1431 top.print_on(&ls);
1432 }
1433
1434 set_top_frame_metadata_pd(top);
1435
1436 chunk->set_sp(chunk->to_offset(top.sp()));
1437 chunk->set_pc(top.pc());
1438
1439 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1440
1441 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1442
1443 // At this point the chunk is consistent
1444
1445 if (UNLIKELY(_barriers)) {
1446 log_develop_trace(continuations)("do barriers on old chunk");
1447 // Serial and Parallel GC can allocate objects directly into the old generation.
1448 // Then we want to relativize the derived pointers eagerly so that
1449 // old chunks are all in GC mode.
1450 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1451 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1452 if (UseShenandoahGC) {
1453 _cont.tail()->relativize_derived_pointers_concurrently();
1454 } else {
1455 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1456 }
1457 // For objects in the old generation we must maintain the remembered set
1458 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1459 }
1460
1461 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1462 if (lt.develop_is_enabled()) {
1463 LogStream ls(lt);
1464 chunk->print_on(true, &ls);
1465 }
1466
1467 if (lt.develop_is_enabled()) {
1468 LogStream ls(lt);
1469 ls.print_cr("top hframe after (freeze):");
1470 assert(_cont.last_frame().is_heap_frame(), "should be");
1471 _cont.last_frame().print_on(&ls);
1472 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1473 }
1474
1475 assert(_cont.chunk_invariant(), "");
1476 }
1477
1478 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1479 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1480 assert(t == JavaThread::current(), "");
1481 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1482 if (!_preempt) {
1483 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1484 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1485 }
1486 return true;
1487 }
1488 return false;
1489 }
1490
1491 class StackChunkAllocator : public MemAllocator {
1492 const size_t _stack_size;
1493 int _argsize_md;
1494 ContinuationWrapper& _continuation_wrapper;
1495 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1496 mutable bool _took_slow_path;
1497
1498 // Does the minimal amount of initialization needed for a TLAB allocation.
1499 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1500 virtual oop initialize(HeapWord* mem) const override {
1501 assert(_stack_size > 0, "");
1502 assert(_stack_size <= max_jint, "");
1503 assert(_word_size > _stack_size, "");
1504
1505 // zero out fields (but not the stack)
1506 const size_t hs = oopDesc::header_size();
1507 if (oopDesc::has_klass_gap()) {
1508 oopDesc::set_klass_gap(mem, 0);
1509 }
1510 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1511
1512 int bottom = (int)_stack_size - _argsize_md;
1513
1514 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1515 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1516 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1517
1518 return finish(mem);
1519 }
1520
1521 stackChunkOop allocate_fast() const {
1522 if (!UseTLAB) {
1523 return nullptr;
1524 }
1525
1526 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1527 if (mem == nullptr) {
1528 return nullptr;
1529 }
1530
1531 oop obj = initialize(mem);
1532 return stackChunkOopDesc::cast(obj);
1533 }
1534
1535 public:
1536 StackChunkAllocator(Klass* klass,
1537 size_t word_size,
1538 Thread* thread,
1539 size_t stack_size,
1540 int argsize_md,
1541 ContinuationWrapper& continuation_wrapper,
1542 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1543 : MemAllocator(klass, word_size, thread),
1544 _stack_size(stack_size),
1545 _argsize_md(argsize_md),
1546 _continuation_wrapper(continuation_wrapper),
1547 _jvmti_event_collector(jvmti_event_collector),
1548 _took_slow_path(false) {}
1549
1550 // Provides it's own, specialized allocation which skips instrumentation
1551 // if the memory can be allocated without going to a slow-path.
1552 stackChunkOop allocate() const {
1553 // First try to allocate without any slow-paths or instrumentation.
1554 stackChunkOop obj = allocate_fast();
1555 if (obj != nullptr) {
1556 return obj;
1557 }
1558
1559 // Now try full-blown allocation with all expensive operations,
1560 // including potentially safepoint operations.
1561 _took_slow_path = true;
1562
1563 // Protect unhandled Loom oops
1564 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1565
1566 // Can safepoint
1567 _jvmti_event_collector->start();
1568
1569 // Can safepoint
1570 return stackChunkOopDesc::cast(MemAllocator::allocate());
1571 }
1572
1573 bool took_slow_path() const {
1574 return _took_slow_path;
1575 }
1576 };
1577
1578 template <typename ConfigT>
1579 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1580 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1581
1582 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1583 size_t size_in_words = klass->instance_size(stack_size);
1584
1585 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1586 if (!_preempt) {
1587 throw_stack_overflow_on_humongous_chunk();
1588 }
1589 return nullptr;
1590 }
1591
1592 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1593 assert(current == JavaThread::current(), "should be current");
1594
1595 // Allocate the chunk.
1596 //
1597 // This might safepoint while allocating, but all safepointing due to
1598 // instrumentation have been deferred. This property is important for
1599 // some GCs, as this ensures that the allocated object is in the young
1600 // generation / newly allocated memory.
1601 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1602 stackChunkOop chunk = allocator.allocate();
1603
1604 if (chunk == nullptr) {
1605 return nullptr; // OOME
1606 }
1607
1608 // assert that chunk is properly initialized
1609 assert(chunk->stack_size() == (int)stack_size, "");
1610 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1611 assert(chunk->sp() == chunk->bottom(), "");
1612 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1613 assert(chunk->max_thawing_size() == 0, "");
1614 assert(chunk->pc() == nullptr, "");
1615 assert(chunk->is_empty(), "");
1616 assert(chunk->flags() == 0, "");
1617 assert(chunk->is_gc_mode() == false, "");
1618 assert(chunk->lockstack_size() == 0, "");
1619
1620 // fields are uninitialized
1621 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1622 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1623
1624 #if INCLUDE_ZGC
1625 if (UseZGC) {
1626 ZStackChunkGCData::initialize(chunk);
1627 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1628 _barriers = false;
1629 } else
1630 #endif
1631 #if INCLUDE_SHENANDOAHGC
1632 if (UseShenandoahGC) {
1633 _barriers = chunk->requires_barriers();
1634 } else
1635 #endif
1636 {
1637 if (!allocator.took_slow_path()) {
1638 // Guaranteed to be in young gen / newly allocated memory
1639 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1640 _barriers = false;
1641 } else {
1642 // Some GCs could put direct allocations in old gen for slow-path
1643 // allocations; need to explicitly check if that was the case.
1644 _barriers = chunk->requires_barriers();
1645 }
1646 }
1647
1648 if (_barriers) {
1649 log_develop_trace(continuations)("allocation requires barriers");
1650 }
1651
1652 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1653
1654 return chunk;
1655 }
1656
1657 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1658 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1659 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1660 }
1661
1662 class AnchorMark : public StackObj {
1663 JavaThread* _current;
1664 frame& _top_frame;
1665 intptr_t* _last_sp_from_frame;
1666 bool _is_interpreted;
1667
1668 public:
1669 AnchorMark(JavaThread* current, frame& f) : _current(current), _top_frame(f), _is_interpreted(false) {
1670 intptr_t* sp = anchor_mark_set_pd();
1671 set_anchor(_current, sp);
1672 }
1673 ~AnchorMark() {
1674 clear_anchor(_current);
1675 anchor_mark_clear_pd();
1676 }
1677 inline intptr_t* anchor_mark_set_pd();
1678 inline void anchor_mark_clear_pd();
1679 };
1680
1681 #if INCLUDE_JVMTI
1682 static int num_java_frames(ContinuationWrapper& cont) {
1683 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1684 int count = 0;
1685 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1686 count += chunk->num_java_frames();
1687 }
1688 return count;
1689 }
1690
1691 static void invalidate_jvmti_stack(JavaThread* thread) {
1692 if (thread->is_interp_only_mode()) {
1693 JvmtiThreadState *state = thread->jvmti_thread_state();
1694 if (state != nullptr)
1695 state->invalidate_cur_stack_depth();
1696 }
1697 }
1698
1699 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1700 if (JvmtiExport::can_post_frame_pop()) {
1701 int num_frames = num_java_frames(cont);
1702
1703 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1704 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames);
1705 }
1706 invalidate_jvmti_stack(thread);
1707 }
1708
1709 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, Continuation::preempt_kind pk) {
1710 assert(current->vthread() != nullptr, "must be");
1711
1712 HandleMarkCleaner hm(current); // Cleanup vth and so._conth Handles
1713 Handle vth(current, current->vthread());
1714 ContinuationWrapper::SafepointOp so(current, cont);
1715
1716 AnchorMark am(current, top); // Set anchor so that the stack is walkable.
1717
1718 JRT_BLOCK
1719 JvmtiVTMSTransitionDisabler::VTMS_vthread_mount((jthread)vth.raw_value(), false);
1720
1721 if (current->pending_contended_entered_event()) {
1722 // No monitor JVMTI events for ObjectLocker case.
1723 if (pk != Continuation::object_locker) {
1724 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1725 }
1726 current->set_contended_entered_monitor(nullptr);
1727 }
1728 JRT_BLOCK_END
1729 }
1730 #endif // INCLUDE_JVMTI
1731
1732 #ifdef ASSERT
1733 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1734 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1735 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1736 bool FreezeBase::check_valid_fast_path() {
1737 ContinuationEntry* ce = _thread->last_continuation();
1738 RegisterMap map(_thread,
1739 RegisterMap::UpdateMap::skip,
1740 RegisterMap::ProcessFrames::skip,
1741 RegisterMap::WalkContinuation::skip);
1742 map.set_include_argument_oops(false);
1743 bool is_top_frame = true;
1744 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1745 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1746 return false;
1747 }
1748 }
1749 return true;
1750 }
1751
1752 static void verify_frame_kind(const frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr, const char** code_name_ptr, int* bci_ptr) {
1753 JavaThread* current = JavaThread::current();
1754 ResourceMark rm(current);
1755
1756 Method* m;
1757 const char* code_name;
1758 int bci;
1759 if (preempt_kind == Continuation::monitorenter) {
1760 assert(top.is_interpreted_frame() || top.is_runtime_frame(), "");
1761 bool at_sync_method;
1762 if (top.is_interpreted_frame()) {
1763 m = top.interpreter_frame_method();
1764 assert(!m->is_native() || m->is_synchronized(), "invalid method %s", m->external_name());
1765 address bcp = top.interpreter_frame_bcp();
1766 assert(bcp != 0 || m->is_native(), "");
1767 at_sync_method = m->is_synchronized() && (bcp == 0 || bcp == m->code_base());
1768 // bcp is advanced on monitorenter before making the VM call, adjust for that.
1769 bool at_sync_bytecode = bcp > m->code_base() && Bytecode(m, bcp - 1).code() == Bytecodes::Code::_monitorenter;
1770 assert(at_sync_method || at_sync_bytecode, "");
1771 bci = at_sync_method ? -1 : top.interpreter_frame_bci();
1772 } else {
1773 CodeBlob* cb = top.cb();
1774 RegisterMap reg_map(current,
1775 RegisterMap::UpdateMap::skip,
1776 RegisterMap::ProcessFrames::skip,
1777 RegisterMap::WalkContinuation::skip);
1778 frame fr = top.sender(®_map);
1779 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1780 compiledVFrame* cvf = compiledVFrame::cast(vf);
1781 m = cvf->method();
1782 bci = cvf->scope()->bci();
1783 at_sync_method = bci == SynchronizationEntryBCI;
1784 assert(!at_sync_method || m->is_synchronized(), "bci is %d but method %s is not synchronized", bci, m->external_name());
1785 bool is_c1_monitorenter = false, is_c2_monitorenter = false;
1786 COMPILER1_PRESENT(is_c1_monitorenter = cb == Runtime1::blob_for(StubId::c1_monitorenter_id) ||
1787 cb == Runtime1::blob_for(StubId::c1_monitorenter_nofpu_id);)
1788 COMPILER2_PRESENT(is_c2_monitorenter = cb == CodeCache::find_blob(OptoRuntime::complete_monitor_locking_Java());)
1789 assert(is_c1_monitorenter || is_c2_monitorenter, "wrong runtime stub frame");
1790 }
1791 code_name = at_sync_method ? "synchronized method" : "monitorenter";
1792 } else if (preempt_kind == Continuation::object_wait) {
1793 assert(top.is_interpreted_frame() || top.is_native_frame(), "");
1794 m = top.is_interpreted_frame() ? top.interpreter_frame_method() : top.cb()->as_nmethod()->method();
1795 assert(m->is_object_wait0(), "");
1796 bci = 0;
1797 code_name = "";
1798 } else {
1799 assert(preempt_kind == Continuation::object_locker, "invalid preempt kind");
1800 assert(top.is_interpreted_frame(), "");
1801 m = top.interpreter_frame_method();
1802 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
1803 Bytecodes::Code code = current_bytecode.code();
1804 assert(code == Bytecodes::Code::_new || code == Bytecodes::Code::_invokestatic ||
1805 (code == Bytecodes::Code::_getstatic || code == Bytecodes::Code::_putstatic), "invalid bytecode");
1806 bci = top.interpreter_frame_bci();
1807 code_name = Bytecodes::name(current_bytecode.code());
1808 }
1809 assert(bci >= 0 || m->is_synchronized(), "invalid bci:%d at method %s", bci, m->external_name());
1810
1811 if (m_ptr != nullptr) {
1812 *m_ptr = m;
1813 *code_name_ptr = code_name;
1814 *bci_ptr = bci;
1815 }
1816 }
1817
1818 static void log_preempt_after_freeze(ContinuationWrapper& cont) {
1819 JavaThread* current = cont.thread();
1820 StackChunkFrameStream<ChunkFrames::Mixed> sfs(cont.tail());
1821 frame top_frame = sfs.to_frame();
1822 bool at_init = current->at_preemptable_init();
1823 bool at_enter = current->current_pending_monitor() != nullptr;
1824 bool at_wait = current->current_waiting_monitor() != nullptr;
1825 assert((at_enter && !at_wait) || (!at_enter && at_wait), "");
1826 Continuation::preempt_kind pk = at_init ? Continuation::object_locker : at_enter ? Continuation::monitorenter : Continuation::object_wait;
1827
1828 Method* m = nullptr;
1829 const char* code_name = nullptr;
1830 int bci = InvalidFrameStateBci;
1831 verify_frame_kind(top_frame, pk, &m, &code_name, &bci);
1832 assert(m != nullptr && code_name != nullptr && bci != InvalidFrameStateBci, "should be set");
1833
1834 ResourceMark rm(current);
1835 if (bci < 0) {
1836 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " while synchronizing on %smethod %s", current->monitor_owner_id(), m->is_native() ? "native " : "", m->external_name());
1837 } else if (m->is_object_wait0()) {
1838 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at native method %s", current->monitor_owner_id(), m->external_name());
1839 } else {
1840 Klass* k = current->preempt_init_klass();
1841 assert(k != nullptr || !at_init, "");
1842 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at %s(bci:%d) in method %s %s%s", current->monitor_owner_id(),
1843 code_name, bci, m->external_name(), at_init ? "trying to initialize klass " : "", at_init ? k->external_name() : "");
1844 }
1845 }
1846 #endif // ASSERT
1847
1848 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1849 verify_continuation(cont.continuation());
1850 assert(!cont.is_empty(), "");
1851
1852 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1853 return freeze_ok;
1854 }
1855
1856 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1857 if (UNLIKELY(res != freeze_ok)) {
1858 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1859 verify_continuation(cont.continuation());
1860 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1861 return res;
1862 }
1863
1864 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1865 return freeze_epilog(cont);
1866 }
1867
1868 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1869 if (UNLIKELY(res != freeze_ok)) {
1870 verify_continuation(cont.continuation());
1871 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1872 return res;
1873 }
1874
1875 // Set up things so that on return to Java we jump to preempt stub.
1876 patch_return_pc_with_preempt_stub(old_last_frame);
1877 cont.tail()->set_preempted(true);
1878 DEBUG_ONLY(log_preempt_after_freeze(cont);)
1879 return freeze_epilog(cont);
1880 }
1881
1882 template<typename ConfigT, bool preempt>
1883 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1884 assert(!current->has_pending_exception(), "");
1885
1886 #ifdef ASSERT
1887 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1888 log_frames(current, false);
1889 #endif
1890
1891 CONT_JFR_ONLY(EventContinuationFreeze event;)
1892
1893 ContinuationEntry* entry = current->last_continuation();
1894
1895 oop oopCont = entry->cont_oop(current);
1896 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1897 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1898
1899 verify_continuation(oopCont);
1900 ContinuationWrapper cont(current, oopCont);
1901 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1902
1903 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1904
1905 assert(LockingMode == LM_LEGACY || (current->held_monitor_count() == 0 && current->jni_monitor_count() == 0),
1906 "Held monitor count should only be used for LM_LEGACY: " INT64_FORMAT " JNI: " INT64_FORMAT, (int64_t)current->held_monitor_count(), (int64_t)current->jni_monitor_count());
1907
1908 if (entry->is_pinned() || current->held_monitor_count() > 0) {
1909 log_develop_debug(continuations)("PINNED due to critical section/hold monitor");
1910 verify_continuation(cont.continuation());
1911 freeze_result res = entry->is_pinned() ? freeze_pinned_cs : freeze_pinned_monitor;
1912 if (!preempt) {
1913 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1914 }
1915 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1916 // Avoid Thread.yield() loops without safepoint polls.
1917 if (SafepointMechanism::should_process(current) && !preempt) {
1918 cont.done(); // allow safepoint
1919 ThreadInVMfromJava tivmfj(current);
1920 }
1921 return res;
1922 }
1923
1924 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1925
1926 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1927 bool fast = UseContinuationFastPath && current->cont_fastpath();
1928 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1929 freeze.freeze_fast_existing_chunk();
1930 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1931 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1932 }
1933
1934 if (preempt) {
1935 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1936 freeze.set_jvmti_event_collector(&jsoaec);
1937
1938 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1939
1940 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1941 preempt_epilog(cont, res, freeze.last_frame());
1942 return res;
1943 }
1944
1945 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1946 assert(current == JavaThread::current(), "must be current thread");
1947 JRT_BLOCK
1948 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1949 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1950 freeze.set_jvmti_event_collector(&jsoaec);
1951
1952 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1953
1954 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1955 freeze_epilog(current, cont, res);
1956 cont.done(); // allow safepoint in the transition back to Java
1957 return res;
1958 JRT_BLOCK_END
1959 }
1960
1961 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1962 ContinuationEntry* entry = thread->last_continuation();
1963 if (entry == nullptr) {
1964 return freeze_ok;
1965 }
1966 if (entry->is_pinned()) {
1967 return freeze_pinned_cs;
1968 } else if (thread->held_monitor_count() > 0) {
1969 return freeze_pinned_monitor;
1970 }
1971
1972 RegisterMap map(thread,
1973 RegisterMap::UpdateMap::include,
1974 RegisterMap::ProcessFrames::skip,
1975 RegisterMap::WalkContinuation::skip);
1976 map.set_include_argument_oops(false);
1977 frame f = thread->last_frame();
1978
1979 if (!safepoint) {
1980 f = f.sender(&map); // this is the yield frame
1981 } else { // safepoint yield
1982 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1983 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1984 #else
1985 Unimplemented();
1986 #endif
1987 if (!Interpreter::contains(f.pc())) {
1988 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1989 assert(f.oop_map() != nullptr, "must be");
1990 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1991 }
1992 }
1993
1994 while (true) {
1995 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1996 return freeze_pinned_native;
1997 }
1998
1999 f = f.sender(&map);
2000 if (!Continuation::is_frame_in_continuation(entry, f)) {
2001 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
2002 if (scope == cont_scope) {
2003 break;
2004 }
2005 intx monitor_count = entry->parent_held_monitor_count();
2006 entry = entry->parent();
2007 if (entry == nullptr) {
2008 break;
2009 }
2010 if (entry->is_pinned()) {
2011 return freeze_pinned_cs;
2012 } else if (monitor_count > 0) {
2013 return freeze_pinned_monitor;
2014 }
2015 }
2016 }
2017 return freeze_ok;
2018 }
2019
2020 /////////////// THAW ////
2021
2022 static int thaw_size(stackChunkOop chunk) {
2023 int size = chunk->max_thawing_size();
2024 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
2025 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
2026 return size;
2027 }
2028
2029 // make room on the stack for thaw
2030 // returns the size in bytes, or 0 on failure
2031 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
2032 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
2033
2034 assert(thread == JavaThread::current(), "");
2035
2036 ContinuationEntry* ce = thread->last_continuation();
2037 assert(ce != nullptr, "");
2038 oop continuation = ce->cont_oop(thread);
2039 assert(continuation == get_continuation(thread), "");
2040 verify_continuation(continuation);
2041
2042 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
2043 assert(chunk != nullptr, "");
2044
2045 // The tail can be empty because it might still be available for another freeze.
2046 // However, here we want to thaw, so we get rid of it (it will be GCed).
2047 if (UNLIKELY(chunk->is_empty())) {
2048 chunk = chunk->parent();
2049 assert(chunk != nullptr, "");
2050 assert(!chunk->is_empty(), "");
2051 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
2052 }
2053
2054 // Verification
2055 chunk->verify();
2056 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
2057
2058 // Only make space for the last chunk because we only thaw from the last chunk
2059 int size = thaw_size(chunk) << LogBytesPerWord;
2060
2061 const address bottom = (address)thread->last_continuation()->entry_sp();
2062 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
2063 // for the Java frames in the check below.
2064 if (!stack_overflow_check(thread, size + 300, bottom)) {
2065 return 0;
2066 }
2067
2068 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
2069 p2i(bottom), p2i(bottom - size), size);
2070 return size;
2071 }
2072
2073 class ThawBase : public StackObj {
2074 protected:
2075 JavaThread* _thread;
2076 ContinuationWrapper& _cont;
2077 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
2078
2079 intptr_t* _fastpath;
2080 bool _barriers;
2081 bool _preempted_case;
2082 bool _process_args_at_top;
2083 intptr_t* _top_unextended_sp_before_thaw;
2084 int _align_size;
2085 DEBUG_ONLY(intptr_t* _top_stack_address);
2086
2087 // Only used for some preemption cases.
2088 ObjectMonitor* _monitor;
2089
2090 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
2091
2092 NOT_PRODUCT(int _frames;)
2093
2094 protected:
2095 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2096 _thread(thread), _cont(cont),
2097 _fastpath(nullptr) {
2098 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2099 assert (cont.tail() != nullptr, "no last chunk");
2100 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2101 }
2102
2103 void clear_chunk(stackChunkOop chunk);
2104 template<bool check_stub>
2105 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2106 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2107
2108 void thaw_lockstack(stackChunkOop chunk);
2109
2110 // fast path
2111 inline void prefetch_chunk_pd(void* start, int size_words);
2112 void patch_return(intptr_t* sp, bool is_last);
2113
2114 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
2115 inline intptr_t* push_cleanup_continuation();
2116 inline intptr_t* push_preempt_adapter();
2117 intptr_t* redo_vmcall(JavaThread* current, frame& top);
2118 void throw_interrupted_exception(JavaThread* current, frame& top);
2119
2120 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
2121 void finish_thaw(frame& f);
2122
2123 private:
2124 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2125 void finalize_thaw(frame& entry, int argsize);
2126
2127 inline bool seen_by_gc();
2128
2129 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2130 inline void after_thaw_java_frame(const frame& f, bool bottom);
2131 inline void patch(frame& f, const frame& caller, bool bottom);
2132 void clear_bitmap_bits(address start, address end);
2133
2134 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top);
2135 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2136 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2137 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
2138
2139 void push_return_frame(frame& f);
2140 inline frame new_entry_frame();
2141 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom);
2142 inline void patch_pd(frame& f, const frame& sender);
2143 inline void patch_pd(frame& f, intptr_t* caller_sp);
2144 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2145
2146 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2147
2148 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2149
2150 public:
2151 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2152 };
2153
2154 template <typename ConfigT>
2155 class Thaw : public ThawBase {
2156 public:
2157 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2158
2159 inline bool can_thaw_fast(stackChunkOop chunk) {
2160 return !_barriers
2161 && _thread->cont_fastpath_thread_state()
2162 && !chunk->has_thaw_slowpath_condition()
2163 && !PreserveFramePointer;
2164 }
2165
2166 inline intptr_t* thaw(Continuation::thaw_kind kind);
2167 template<bool check_stub = false>
2168 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2169 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2170 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2171 };
2172
2173 template <typename ConfigT>
2174 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2175 verify_continuation(_cont.continuation());
2176 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2177 assert(!_cont.is_empty(), "");
2178
2179 stackChunkOop chunk = _cont.tail();
2180 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2181 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2182
2183 _barriers = chunk->requires_barriers();
2184 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2185 : thaw_slow(chunk, kind);
2186 }
2187
2188 class ReconstructedStack : public StackObj {
2189 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2190 int _thaw_size;
2191 int _argsize;
2192 public:
2193 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2194 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2195 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2196 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2197 // We're at most one alignment word away from entrySP
2198 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2199 }
2200
2201 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2202
2203 // top and bottom stack pointers
2204 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2205 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2206
2207 // several operations operate on the totality of the stack being reconstructed,
2208 // including the metadata words
2209 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2210 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2211 };
2212
2213 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2214 chunk->set_sp(chunk->bottom());
2215 chunk->set_max_thawing_size(0);
2216 }
2217
2218 template<bool check_stub>
2219 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2220 bool empty = false;
2221 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2222 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2223 assert(chunk_sp == f.sp(), "");
2224 assert(chunk_sp == f.unextended_sp(), "");
2225
2226 int frame_size = f.cb()->frame_size();
2227 argsize = f.stack_argsize();
2228
2229 assert(!f.is_stub() || check_stub, "");
2230 if (check_stub && f.is_stub()) {
2231 // If we don't thaw the top compiled frame too, after restoring the saved
2232 // registers back in Java, we would hit the return barrier to thaw one more
2233 // frame effectively overwriting the restored registers during that call.
2234 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2235 assert(!f.is_done(), "");
2236
2237 f.get_cb();
2238 assert(f.is_compiled(), "");
2239 frame_size += f.cb()->frame_size();
2240 argsize = f.stack_argsize();
2241
2242 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2243 // The caller of the runtime stub when the continuation is preempted is not at a
2244 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2245 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2246 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2247 }
2248 }
2249
2250 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2251 empty = f.is_done();
2252 assert(!empty || argsize == chunk->argsize(), "");
2253
2254 if (empty) {
2255 clear_chunk(chunk);
2256 } else {
2257 chunk->set_sp(chunk->sp() + frame_size);
2258 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2259 // We set chunk->pc to the return pc into the next frame
2260 chunk->set_pc(f.pc());
2261 #ifdef ASSERT
2262 {
2263 intptr_t* retaddr_slot = (chunk_sp
2264 + frame_size
2265 - frame::sender_sp_ret_address_offset());
2266 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2267 "unexpected pc");
2268 }
2269 #endif
2270 }
2271 assert(empty == chunk->is_empty(), "");
2272 // returns the size required to store the frame on stack, and because it is a
2273 // compiled frame, it must include a copy of the arguments passed by the caller
2274 return frame_size + argsize + frame::metadata_words_at_top;
2275 }
2276
2277 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2278 int lockStackSize = chunk->lockstack_size();
2279 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2280
2281 oop tmp_lockstack[LockStack::CAPACITY];
2282 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2283 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2284
2285 chunk->set_lockstack_size(0);
2286 chunk->set_has_lockstack(false);
2287 }
2288
2289 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2290 assert(to >= _top_stack_address, "overwrote past thawing space"
2291 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2292 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2293 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2294 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2295 }
2296
2297 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2298 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2299
2300 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2301 ContinuationHelper::patch_return_address_at(
2302 sp - frame::sender_sp_ret_address_offset(),
2303 pc);
2304 }
2305
2306 template <typename ConfigT>
2307 template<bool check_stub>
2308 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2309 assert(chunk == _cont.tail(), "");
2310 assert(!chunk->has_mixed_frames(), "");
2311 assert(!chunk->requires_barriers(), "");
2312 assert(!chunk->has_bitmap(), "");
2313 assert(!_thread->is_interp_only_mode(), "");
2314
2315 LogTarget(Trace, continuations) lt;
2316 if (lt.develop_is_enabled()) {
2317 LogStream ls(lt);
2318 ls.print_cr("thaw_fast");
2319 chunk->print_on(true, &ls);
2320 }
2321
2322 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2323 static const int threshold = 500; // words
2324
2325 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2326 int argsize, thaw_size;
2327
2328 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2329
2330 bool partial, empty;
2331 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2332 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2333
2334 partial = false;
2335 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2336 clear_chunk(chunk);
2337 thaw_size = full_chunk_size;
2338 empty = true;
2339 } else { // thaw a single frame
2340 partial = true;
2341 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2342 empty = chunk->is_empty();
2343 }
2344
2345 // Are we thawing the last frame(s) in the continuation
2346 const bool is_last = empty && chunk->parent() == nullptr;
2347 assert(!is_last || argsize == 0, "");
2348
2349 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2350 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2351
2352 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2353
2354 // also copy metadata words at frame bottom
2355 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2356
2357 // update the ContinuationEntry
2358 _cont.set_argsize(argsize);
2359 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2360 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2361
2362 // install the return barrier if not last frame, or the entry's pc if last
2363 patch_return(rs.bottom_sp(), is_last);
2364
2365 // insert the back links from callee to caller frames
2366 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2367
2368 assert(is_last == _cont.is_empty(), "");
2369 assert(_cont.chunk_invariant(), "");
2370
2371 #if CONT_JFR
2372 EventContinuationThawFast e;
2373 if (e.should_commit()) {
2374 e.set_id(cast_from_oop<u8>(chunk));
2375 e.set_size(thaw_size << LogBytesPerWord);
2376 e.set_full(!partial);
2377 e.commit();
2378 }
2379 #endif
2380
2381 #ifdef ASSERT
2382 set_anchor(_thread, rs.sp());
2383 log_frames(_thread);
2384 if (LoomDeoptAfterThaw) {
2385 do_deopt_after_thaw(_thread);
2386 }
2387 clear_anchor(_thread);
2388 #endif
2389
2390 return rs.sp();
2391 }
2392
2393 inline bool ThawBase::seen_by_gc() {
2394 return _barriers || _cont.tail()->is_gc_mode();
2395 }
2396
2397 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2398 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2399 if (UseZGC || UseShenandoahGC) {
2400 chunk->relativize_derived_pointers_concurrently();
2401 }
2402 #endif
2403 }
2404
2405 template <typename ConfigT>
2406 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2407 Continuation::preempt_kind preempt_kind;
2408 bool retry_fast_path = false;
2409
2410 _process_args_at_top = false;
2411 _preempted_case = chunk->preempted();
2412 if (_preempted_case) {
2413 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2414 if (waiter != nullptr) {
2415 // Mounted again after preemption. Resume the pending monitor operation,
2416 // which will be either a monitorenter or Object.wait() call.
2417 ObjectMonitor* mon = waiter->monitor();
2418 preempt_kind = waiter->is_wait() ? Continuation::object_wait : Continuation::monitorenter;
2419
2420 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2421 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2422 if (!mon_acquired) {
2423 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2424 log_trace(continuations, tracking)("Failed to acquire monitor, unmounting again");
2425 return push_cleanup_continuation();
2426 }
2427 _monitor = mon; // remember monitor since we might need it on handle_preempted_continuation()
2428 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2429 JVMTI_ONLY(assert(_thread->contended_entered_monitor() == nullptr || _thread->contended_entered_monitor() == _monitor, ""));
2430 } else {
2431 // Preemption cancelled on moniterenter or ObjectLocker case. We
2432 // actually acquired the monitor after freezing all frames so no
2433 // need to call resume_operation. If this is the ObjectLocker case
2434 // we released the monitor already at ~ObjectLocker, so here we set
2435 // _monitor to nullptr to indicate there is no need to release it later.
2436 preempt_kind = Continuation::monitorenter;
2437 _monitor = nullptr;
2438 }
2439
2440 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2441 relativize_chunk_concurrently(chunk);
2442
2443 if (chunk->at_klass_init()) {
2444 preempt_kind = Continuation::object_locker;
2445 chunk->set_at_klass_init(false);
2446 _process_args_at_top = chunk->has_args_at_top();
2447 if (_process_args_at_top) chunk->set_has_args_at_top(false);
2448 }
2449 chunk->set_preempted(false);
2450 retry_fast_path = true;
2451 } else {
2452 relativize_chunk_concurrently(chunk);
2453 }
2454
2455 // On first thaw after freeze restore oops to the lockstack if any.
2456 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2457 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2458 thaw_lockstack(chunk);
2459 retry_fast_path = true;
2460 }
2461
2462 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2463 // and FLAG_PREEMPTED flags from the stackChunk.
2464 if (retry_fast_path && can_thaw_fast(chunk)) {
2465 intptr_t* sp = thaw_fast<true>(chunk);
2466 if (_preempted_case) {
2467 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2468 }
2469 return sp;
2470 }
2471
2472 LogTarget(Trace, continuations) lt;
2473 if (lt.develop_is_enabled()) {
2474 LogStream ls(lt);
2475 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2476 chunk->print_on(true, &ls);
2477 }
2478
2479 #if CONT_JFR
2480 EventContinuationThawSlow e;
2481 if (e.should_commit()) {
2482 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2483 e.commit();
2484 }
2485 #endif
2486
2487 DEBUG_ONLY(_frames = 0;)
2488 _align_size = 0;
2489 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2490
2491 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2492 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2493
2494 frame heap_frame = _stream.to_frame();
2495 if (lt.develop_is_enabled()) {
2496 LogStream ls(lt);
2497 ls.print_cr("top hframe before (thaw):");
2498 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2499 heap_frame.print_value_on(&ls);
2500 }
2501
2502 frame caller; // the thawed caller on the stack
2503 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2504 finish_thaw(caller); // caller is now the topmost thawed frame
2505 _cont.write();
2506
2507 assert(_cont.chunk_invariant(), "");
2508
2509 JVMTI_ONLY(invalidate_jvmti_stack(_thread));
2510
2511 _thread->set_cont_fastpath(_fastpath);
2512
2513 intptr_t* sp = caller.sp();
2514
2515 if (_preempted_case) {
2516 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2517 }
2518 return sp;
2519 }
2520
2521 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2522 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2523 assert(!_cont.is_empty(), "no more frames");
2524 assert(num_frames > 0, "");
2525 assert(!heap_frame.is_empty(), "");
2526
2527 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2528 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2529 } else if (!heap_frame.is_interpreted_frame()) {
2530 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2531 } else {
2532 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames, top_on_preempt_case);
2533 }
2534 }
2535
2536 template<typename FKind>
2537 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2538 assert(num_frames > 0, "");
2539
2540 DEBUG_ONLY(_frames++;)
2541
2542 int argsize = _stream.stack_argsize();
2543
2544 _stream.next(SmallRegisterMap::instance_no_args());
2545 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2546
2547 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk
2548 // as it makes detecting that situation and adjusting unextended_sp tricky
2549 if (num_frames == 1 && !_stream.is_done() && FKind::interpreted && _stream.is_compiled()) {
2550 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2551 num_frames++;
2552 }
2553
2554 if (num_frames == 1 || _stream.is_done()) { // end recursion
2555 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2556 return true; // bottom
2557 } else { // recurse
2558 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2559 return false;
2560 }
2561 }
2562
2563 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2564 stackChunkOop chunk = _cont.tail();
2565
2566 if (!_stream.is_done()) {
2567 assert(_stream.sp() >= chunk->sp_address(), "");
2568 chunk->set_sp(chunk->to_offset(_stream.sp()));
2569 chunk->set_pc(_stream.pc());
2570 } else {
2571 chunk->set_sp(chunk->bottom());
2572 chunk->set_pc(nullptr);
2573 }
2574 assert(_stream.is_done() == chunk->is_empty(), "");
2575
2576 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2577 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2578
2579 _cont.set_argsize(argsize);
2580 entry = new_entry_frame();
2581
2582 assert(entry.sp() == _cont.entrySP(), "");
2583 assert(Continuation::is_continuation_enterSpecial(entry), "");
2584 assert(_cont.is_entry_frame(entry), "");
2585 }
2586
2587 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2588 LogTarget(Trace, continuations) lt;
2589 if (lt.develop_is_enabled()) {
2590 LogStream ls(lt);
2591 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2592 assert(hf.is_heap_frame(), "should be");
2593 hf.print_value_on(&ls);
2594 }
2595 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2596 }
2597
2598 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2599 #ifdef ASSERT
2600 LogTarget(Trace, continuations) lt;
2601 if (lt.develop_is_enabled()) {
2602 LogStream ls(lt);
2603 ls.print_cr("thawed frame:");
2604 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2605 }
2606 #endif
2607 }
2608
2609 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) {
2610 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2611 if (bottom) {
2612 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2613 : StubRoutines::cont_returnBarrier());
2614 } else {
2615 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2616 // If the caller is not deoptimized, pc is unchanged.
2617 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc());
2618 }
2619
2620 patch_pd(f, caller);
2621
2622 if (f.is_interpreted_frame()) {
2623 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2624 }
2625
2626 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2627 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2628 }
2629
2630 void ThawBase::clear_bitmap_bits(address start, address end) {
2631 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2632 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2633
2634 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2635 // or they will keep objects that are otherwise unreachable alive.
2636
2637 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2638 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2639 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2640 // anyways and we assert that before returning.
2641 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2642 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2643 stackChunkOop chunk = _cont.tail();
2644 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2645 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2646 }
2647
2648 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2649 frame top(sp);
2650 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2651 DEBUG_ONLY(verify_frame_kind(top, preempt_kind);)
2652 NOT_PRODUCT(int64_t tid = _thread->monitor_owner_id();)
2653
2654 #if INCLUDE_JVMTI
2655 // Finish the VTMS transition.
2656 assert(_thread->is_in_VTMS_transition(), "must be");
2657 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2658 if (is_vthread) {
2659 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2660 jvmti_mount_end(_thread, _cont, top, preempt_kind);
2661 } else {
2662 _thread->set_is_in_VTMS_transition(false);
2663 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2664 }
2665 }
2666 #endif
2667
2668 if (fast_case) {
2669 // If we thawed in the slow path the runtime stub/native wrapper frame already
2670 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2671 // we copied the fp patched during freeze, which will now have to be fixed.
2672 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2673 int fsize = top.cb()->frame_size();
2674 patch_pd(top, sp + fsize);
2675 }
2676
2677 if (preempt_kind == Continuation::object_wait) {
2678 // Check now if we need to throw IE exception.
2679 bool throw_ie = _thread->pending_interrupted_exception();
2680 if (throw_ie) {
2681 throw_interrupted_exception(_thread, top);
2682 _thread->set_pending_interrupted_exception(false);
2683 }
2684 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT" after preemption on Object.wait%s", tid, throw_ie ? "(throwing IE)" : "");
2685 } else if (preempt_kind == Continuation::monitorenter) {
2686 if (top.is_runtime_frame()) {
2687 // The continuation might now run on a different platform thread than the previous time so
2688 // we need to adjust the current thread saved in the stub frame before restoring registers.
2689 JavaThread** thread_addr = frame::saved_thread_address(top);
2690 if (thread_addr != nullptr) *thread_addr = _thread;
2691 }
2692 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT " after preemption on monitorenter", tid);
2693 } else {
2694 // We need to redo the original call into the VM. First though, we need
2695 // to exit the monitor we just acquired (except on preemption cancelled
2696 // case where it was already released).
2697 assert(preempt_kind == Continuation::object_locker, "");
2698 if (_monitor != nullptr) _monitor->exit(_thread);
2699 sp = redo_vmcall(_thread, top);
2700 }
2701 return sp;
2702 }
2703
2704 intptr_t* ThawBase::redo_vmcall(JavaThread* current, frame& top) {
2705 assert(!current->preempting(), "");
2706 NOT_PRODUCT(int64_t tid = current->monitor_owner_id();)
2707 intptr_t* sp = top.sp();
2708
2709 {
2710 HandleMarkCleaner hmc(current); // Cleanup so._conth Handle
2711 ContinuationWrapper::SafepointOp so(current, _cont);
2712 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2713
2714 Method* m = top.interpreter_frame_method();
2715 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
2716 Bytecodes::Code code = current_bytecode.code();
2717 log_develop_trace(continuations, preempt)("Redoing InterpreterRuntime::%s for " INT64_FORMAT, code == Bytecodes::Code::_new ? "_new" : "resolve_from_cache", tid);
2718
2719 // These InterpreterRuntime entry points use JRT_ENTRY which uses a HandleMarkCleaner.
2720 // Create a HandeMark to avoid destroying so._conth.
2721 HandleMark hm(current);
2722 if (code == Bytecodes::Code::_new) {
2723 InterpreterRuntime::_new(current, m->constants(), current_bytecode.get_index_u2(code));
2724 } else {
2725 InterpreterRuntime::resolve_from_cache(current, code);
2726 }
2727 }
2728
2729 if (current->preempting()) {
2730 // Preempted again so we just arrange to return to preempt stub to unmount.
2731 sp = push_preempt_adapter();
2732 current->set_preempt_alternate_return(nullptr);
2733 bool cancelled = current->preemption_cancelled();
2734 if (cancelled) {
2735 // Instead of calling thaw again from the preempt stub just unmount anyways with
2736 // state of YIELDING. This will give a chance for other vthreads to run while
2737 // minimizing repeated loops of "thaw->redo_vmcall->try_preempt->preemption_cancelled->thaw..."
2738 // in case of multiple vthreads contending for the same init_lock().
2739 current->set_preemption_cancelled(false);
2740 oop vthread = current->vthread();
2741 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
2742 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::YIELDING);
2743 #if INCLUDE_JVMTI
2744 if (current->contended_entered_monitor() != nullptr) {
2745 current->set_contended_entered_monitor(nullptr);
2746 }
2747 #endif
2748 }
2749 log_develop_trace(continuations, preempt)("Preempted " INT64_FORMAT " again%s", tid, cancelled ? "(preemption cancelled, setting state to YIELDING)" : "");
2750 } else {
2751 log_develop_trace(continuations, preempt)("Call succesful, resuming " INT64_FORMAT, tid);
2752 }
2753 return sp;
2754 }
2755
2756 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2757 HandleMarkCleaner hm(current); // Cleanup so._conth Handle
2758 ContinuationWrapper::SafepointOp so(current, _cont);
2759 // Since we might safepoint set the anchor so that the stack can be walked.
2760 set_anchor(current, top.sp());
2761 JRT_BLOCK
2762 THROW(vmSymbols::java_lang_InterruptedException());
2763 JRT_BLOCK_END
2764 clear_anchor(current);
2765 }
2766
2767 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top) {
2768 assert(hf.is_interpreted_frame(), "");
2769
2770 if (UNLIKELY(seen_by_gc())) {
2771 if (is_top && _process_args_at_top) {
2772 log_trace(continuations, tracking)("Processing arguments in recurse_thaw_interpreted_frame");
2773 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_with_args());
2774 } else {
2775 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2776 }
2777 }
2778
2779 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2780
2781 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2782
2783 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2784
2785 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2786
2787 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2788 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2789 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2790 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2791
2792 assert(hf.is_heap_frame(), "should be");
2793 assert(!f.is_heap_frame(), "should not be");
2794
2795 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2796 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2797
2798 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2799 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2800 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2801
2802 // Make sure the relativized locals is already set.
2803 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2804
2805 derelativize_interpreted_frame_metadata(hf, f);
2806 patch(f, caller, is_bottom_frame);
2807
2808 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2809 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2810
2811 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2812
2813 maybe_set_fastpath(f.sp());
2814
2815 Method* m = hf.interpreter_frame_method();
2816 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2817 const int locals = m->max_locals();
2818
2819 if (!is_bottom_frame) {
2820 // can only fix caller once this frame is thawed (due to callee saved regs)
2821 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2822 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2823 assert(hf.is_heap_frame(), "should be");
2824 address start = (address)(heap_frame_bottom - locals);
2825 address end = (address)heap_frame_bottom;
2826 clear_bitmap_bits(start, end);
2827 }
2828
2829 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2830 caller = f;
2831 }
2832
2833 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2834 assert(hf.is_compiled_frame(), "");
2835 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2836
2837 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2838 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2839 }
2840
2841 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2842
2843 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2844
2845 assert(caller.sp() == caller.unextended_sp(), "");
2846
2847 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2848 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2849 }
2850
2851 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2852 // yet laid out in the stack, and so the original_pc is not stored in it.
2853 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2854 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame);
2855 intptr_t* const stack_frame_top = f.sp();
2856 intptr_t* const heap_frame_top = hf.unextended_sp();
2857
2858 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2859 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize;
2860 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2861
2862 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2863 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2864 // copy metadata, except the metadata at the top of the (unextended) entry frame
2865 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2866
2867 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2868 // (we might have one padding word for alignment)
2869 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2870 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2871
2872 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2873
2874 patch(f, caller, is_bottom_frame);
2875
2876 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2877 assert(!f.is_deoptimized_frame(), "");
2878 if (hf.is_deoptimized_frame()) {
2879 maybe_set_fastpath(f.sp());
2880 } else if (_thread->is_interp_only_mode()
2881 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2882 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2883 // cannot rely on nmethod patching for deopt.
2884 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2885
2886 log_develop_trace(continuations)("Deoptimizing thawed frame");
2887 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2888
2889 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2890 assert(f.is_deoptimized_frame(), "");
2891 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2892 maybe_set_fastpath(f.sp());
2893 }
2894
2895 if (!is_bottom_frame) {
2896 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2897 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2898 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2899 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2900 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2901 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2902 clear_bitmap_bits(start, start + argsize_in_bytes);
2903 }
2904
2905 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2906 caller = f;
2907 }
2908
2909 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2910 DEBUG_ONLY(_frames++;)
2911
2912 if (UNLIKELY(seen_by_gc())) {
2913 // Process the stub's caller here since we might need the full map.
2914 RegisterMap map(nullptr,
2915 RegisterMap::UpdateMap::include,
2916 RegisterMap::ProcessFrames::skip,
2917 RegisterMap::WalkContinuation::skip);
2918 map.set_include_argument_oops(false);
2919 _stream.next(&map);
2920 assert(!_stream.is_done(), "");
2921 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2922 } else {
2923 _stream.next(SmallRegisterMap::instance_no_args());
2924 assert(!_stream.is_done(), "");
2925 }
2926
2927 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2928
2929 assert(caller.is_compiled_frame(), "");
2930 assert(caller.sp() == caller.unextended_sp(), "");
2931
2932 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2933
2934 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2935 intptr_t* stack_frame_top = f.sp();
2936 intptr_t* heap_frame_top = hf.sp();
2937 int fsize = ContinuationHelper::StubFrame::size(hf);
2938
2939 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2940 fsize + frame::metadata_words);
2941
2942 patch(f, caller, false /*is_bottom_frame*/);
2943
2944 // can only fix caller once this frame is thawed (due to callee saved regs)
2945 RegisterMap map(nullptr,
2946 RegisterMap::UpdateMap::include,
2947 RegisterMap::ProcessFrames::skip,
2948 RegisterMap::WalkContinuation::skip);
2949 map.set_include_argument_oops(false);
2950 f.oop_map()->update_register_map(&f, &map);
2951 ContinuationHelper::update_register_map_with_callee(caller, &map);
2952 _cont.tail()->fix_thawed_frame(caller, &map);
2953
2954 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2955 caller = f;
2956 }
2957
2958 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2959 assert(hf.is_native_frame(), "");
2960 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2961
2962 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2963 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2964 }
2965
2966 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
2967 assert(!is_bottom_frame, "");
2968
2969 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2970
2971 assert(caller.sp() == caller.unextended_sp(), "");
2972
2973 if (caller.is_interpreted_frame()) {
2974 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
2975 }
2976
2977 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2978 // yet laid out in the stack, and so the original_pc is not stored in it.
2979 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2980 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
2981 intptr_t* const stack_frame_top = f.sp();
2982 intptr_t* const heap_frame_top = hf.unextended_sp();
2983
2984 int fsize = ContinuationHelper::NativeFrame::size(hf);
2985 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2986
2987 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2988 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2989 int sz = fsize + frame::metadata_words_at_bottom;
2990
2991 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2992
2993 patch(f, caller, false /* bottom */);
2994
2995 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2996 assert(!f.is_deoptimized_frame(), "");
2997 assert(!hf.is_deoptimized_frame(), "");
2998 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
2999
3000 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
3001 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
3002
3003 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
3004 caller = f;
3005 }
3006
3007 void ThawBase::finish_thaw(frame& f) {
3008 stackChunkOop chunk = _cont.tail();
3009
3010 if (chunk->is_empty()) {
3011 // Only remove chunk from list if it can't be reused for another freeze
3012 if (seen_by_gc()) {
3013 _cont.set_tail(chunk->parent());
3014 } else {
3015 chunk->set_has_mixed_frames(false);
3016 }
3017 chunk->set_max_thawing_size(0);
3018 } else {
3019 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
3020 }
3021 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
3022
3023 if (!is_aligned(f.sp(), frame::frame_alignment)) {
3024 assert(f.is_interpreted_frame(), "");
3025 f.set_sp(align_down(f.sp(), frame::frame_alignment));
3026 }
3027 push_return_frame(f);
3028 // can only fix caller after push_return_frame (due to callee saved regs)
3029 if (_process_args_at_top) {
3030 log_trace(continuations, tracking)("Processing arguments in finish_thaw");
3031 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_with_args());
3032 } else {
3033 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_no_args());
3034 }
3035
3036 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
3037
3038 log_develop_trace(continuations)("thawed %d frames", _frames);
3039
3040 LogTarget(Trace, continuations) lt;
3041 if (lt.develop_is_enabled()) {
3042 LogStream ls(lt);
3043 ls.print_cr("top hframe after (thaw):");
3044 _cont.last_frame().print_value_on(&ls);
3045 }
3046 }
3047
3048 void ThawBase::push_return_frame(frame& f) { // see generate_cont_thaw
3049 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
3050 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
3051
3052 LogTarget(Trace, continuations) lt;
3053 if (lt.develop_is_enabled()) {
3054 LogStream ls(lt);
3055 ls.print_cr("push_return_frame");
3056 f.print_value_on(&ls);
3057 }
3058
3059 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
3060 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
3061 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
3062 ContinuationHelper::push_pd(f);
3063
3064 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
3065 }
3066
3067 // returns new top sp
3068 // called after preparations (stack overflow check and making room)
3069 template<typename ConfigT>
3070 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
3071 assert(thread == JavaThread::current(), "Must be current thread");
3072
3073 CONT_JFR_ONLY(EventContinuationThaw event;)
3074
3075 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
3076
3077 ContinuationEntry* entry = thread->last_continuation();
3078 assert(entry != nullptr, "");
3079 oop oopCont = entry->cont_oop(thread);
3080
3081 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
3082 assert(oopCont == get_continuation(thread), "");
3083 verify_continuation(oopCont);
3084
3085 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
3086
3087 ContinuationWrapper cont(thread, oopCont);
3088 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
3089
3090 #ifdef ASSERT
3091 set_anchor_to_entry(thread, cont.entry());
3092 log_frames(thread);
3093 clear_anchor(thread);
3094 #endif
3095
3096 Thaw<ConfigT> thw(thread, cont);
3097 intptr_t* const sp = thw.thaw(kind);
3098 assert(is_aligned(sp, frame::frame_alignment), "");
3099 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp);)
3100
3101 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
3102
3103 verify_continuation(cont.continuation());
3104 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
3105
3106 return sp;
3107 }
3108
3109 #ifdef ASSERT
3110 static void do_deopt_after_thaw(JavaThread* thread) {
3111 int i = 0;
3112 StackFrameStream fst(thread, true, false);
3113 fst.register_map()->set_include_argument_oops(false);
3114 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3115 for (; !fst.is_done(); fst.next()) {
3116 if (fst.current()->cb()->is_nmethod()) {
3117 nmethod* nm = fst.current()->cb()->as_nmethod();
3118 if (!nm->method()->is_continuation_native_intrinsic()) {
3119 nm->make_deoptimized();
3120 }
3121 }
3122 }
3123 }
3124
3125 class ThawVerifyOopsClosure: public OopClosure {
3126 intptr_t* _p;
3127 outputStream* _st;
3128 bool is_good_oop(oop o) {
3129 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
3130 }
3131 public:
3132 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
3133 intptr_t* p() { return _p; }
3134 void reset() { _p = nullptr; }
3135
3136 virtual void do_oop(oop* p) {
3137 oop o = *p;
3138 if (o == nullptr || is_good_oop(o)) {
3139 return;
3140 }
3141 _p = (intptr_t*)p;
3142 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
3143 }
3144 virtual void do_oop(narrowOop* p) {
3145 oop o = RawAccess<>::oop_load(p);
3146 if (o == nullptr || is_good_oop(o)) {
3147 return;
3148 }
3149 _p = (intptr_t*)p;
3150 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
3151 }
3152 };
3153
3154 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
3155 assert(thread->has_last_Java_frame(), "");
3156
3157 ResourceMark rm;
3158 ThawVerifyOopsClosure cl(st);
3159 NMethodToOopClosure cf(&cl, false);
3160
3161 StackFrameStream fst(thread, true, false);
3162 fst.register_map()->set_include_argument_oops(false);
3163 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3164 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
3165 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
3166 st->print_cr(">>> do_verify_after_thaw deopt");
3167 fst.current()->deoptimize(nullptr);
3168 fst.current()->print_on(st);
3169 }
3170
3171 fst.current()->oops_do(&cl, &cf, fst.register_map());
3172 if (cl.p() != nullptr) {
3173 frame fr = *fst.current();
3174 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
3175 fr.print_on(st);
3176 if (!fr.is_interpreted_frame()) {
3177 st->print_cr("size: %d argsize: %d",
3178 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
3179 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
3180 }
3181 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
3182 if (reg != nullptr) {
3183 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
3184 }
3185 cl.reset();
3186 DEBUG_ONLY(thread->print_frame_layout();)
3187 if (chunk != nullptr) {
3188 chunk->print_on(true, st);
3189 }
3190 return false;
3191 }
3192 }
3193 return true;
3194 }
3195
3196 static void log_frames(JavaThread* thread, bool dolog) {
3197 const static int show_entry_callers = 3;
3198 LogTarget(Trace, continuations, tracking) lt;
3199 if (!lt.develop_is_enabled() || !dolog) {
3200 return;
3201 }
3202 LogStream ls(lt);
3203
3204 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3205 if (!thread->has_last_Java_frame()) {
3206 ls.print_cr("NO ANCHOR!");
3207 }
3208
3209 RegisterMap map(thread,
3210 RegisterMap::UpdateMap::include,
3211 RegisterMap::ProcessFrames::include,
3212 RegisterMap::WalkContinuation::skip);
3213 map.set_include_argument_oops(false);
3214
3215 if (false) {
3216 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3217 f.print_on(&ls);
3218 }
3219 } else {
3220 map.set_skip_missing(true);
3221 ResetNoHandleMark rnhm;
3222 ResourceMark rm;
3223 HandleMark hm(Thread::current());
3224 FrameValues values;
3225
3226 int i = 0;
3227 int post_entry = -1;
3228 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3229 f.describe(values, i, &map, i == 0);
3230 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3231 post_entry++;
3232 if (post_entry >= show_entry_callers)
3233 break;
3234 }
3235 values.print_on(thread, &ls);
3236 }
3237
3238 ls.print_cr("======= end frames =========");
3239 }
3240
3241 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp) {
3242 intptr_t* sp0 = sp;
3243 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
3244
3245 bool preempted = false;
3246 stackChunkOop tail = cont.tail();
3247 if (tail != nullptr && tail->preempted()) {
3248 // Still preempted (monitor not acquired) so no frames were thawed.
3249 set_anchor(thread, cont.entrySP(), cont.entryPC());
3250 preempted = true;
3251 } else {
3252 set_anchor(thread, sp0);
3253 }
3254
3255 log_frames(thread);
3256 if (LoomVerifyAfterThaw) {
3257 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
3258 }
3259 assert(ContinuationEntry::assert_entry_frame_laid_out(thread, preempted), "");
3260 clear_anchor(thread);
3261
3262 LogTarget(Trace, continuations) lt;
3263 if (lt.develop_is_enabled()) {
3264 LogStream ls(lt);
3265 ls.print_cr("Jumping to frame (thaw):");
3266 frame(sp).print_value_on(&ls);
3267 }
3268 }
3269 #endif // ASSERT
3270
3271 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3272
3273 #ifdef ASSERT
3274 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3275 ResourceMark rm;
3276 FrameValues values;
3277 assert(f.get_cb() != nullptr, "");
3278 RegisterMap map(f.is_heap_frame() ?
3279 nullptr :
3280 JavaThread::current(),
3281 RegisterMap::UpdateMap::include,
3282 RegisterMap::ProcessFrames::skip,
3283 RegisterMap::WalkContinuation::skip);
3284 map.set_include_argument_oops(false);
3285 map.set_skip_missing(true);
3286 if (callee_complete) {
3287 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3288 }
3289 const_cast<frame&>(f).describe(values, 0, &map, true);
3290 values.print_on(static_cast<JavaThread*>(nullptr), st);
3291 }
3292 #endif
3293
3294 static address thaw_entry = nullptr;
3295 static address freeze_entry = nullptr;
3296 static address freeze_preempt_entry = nullptr;
3297
3298 address Continuation::thaw_entry() {
3299 return ::thaw_entry;
3300 }
3301
3302 address Continuation::freeze_entry() {
3303 return ::freeze_entry;
3304 }
3305
3306 address Continuation::freeze_preempt_entry() {
3307 return ::freeze_preempt_entry;
3308 }
3309
3310 class ConfigResolve {
3311 public:
3312 static void resolve() { resolve_compressed(); }
3313
3314 static void resolve_compressed() {
3315 UseCompressedOops ? resolve_gc<true>()
3316 : resolve_gc<false>();
3317 }
3318
3319 private:
3320 template <bool use_compressed>
3321 static void resolve_gc() {
3322 BarrierSet* bs = BarrierSet::barrier_set();
3323 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3324 switch (bs->kind()) {
3325 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3326 case BarrierSet::bs_name: { \
3327 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3328 } \
3329 break;
3330 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3331 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3332
3333 default:
3334 fatal("BarrierSet resolving not implemented");
3335 };
3336 }
3337
3338 template <bool use_compressed, typename BarrierSetT>
3339 static void resolve() {
3340 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3341
3342 freeze_entry = (address)freeze<SelectedConfigT>;
3343 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3344
3345 // If we wanted, we could templatize by kind and have three different thaw entries
3346 thaw_entry = (address)thaw<SelectedConfigT>;
3347 }
3348 };
3349
3350 void Continuation::init() {
3351 ConfigResolve::resolve();
3352 }