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/continuationGCSupport.inline.hpp"
32 #include "gc/shared/gc_globals.hpp"
33 #include "gc/shared/barrierSet.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/oopsHierarchy.hpp"
46 #include "oops/objArrayOop.inline.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/smallRegisterMap.inline.hpp"
64 #include "runtime/sharedRuntime.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 void adjust_interpreted_frame_unextended_sp(frame& f);
476 static inline void prepare_freeze_interpreted_top_frame(frame& f);
477 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
478
479 protected:
480 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
481 bool freeze_fast_new_chunk(stackChunkOop chunk);
482 };
483
484 template <typename ConfigT>
485 class Freeze : public FreezeBase {
486 private:
487 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
488
489 public:
490 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
491 : FreezeBase(thread, cont, frame_sp, preempt) {}
492
493 freeze_result try_freeze_fast();
494
495 protected:
496 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
497 };
498
499 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
500 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
501 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
502
503 assert(_thread != nullptr, "");
504 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
505
506 DEBUG_ONLY(_cont.entry()->verify_cookie();)
507
508 assert(!Interpreter::contains(_cont.entryPC()), "");
509
510 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
511 #ifdef _LP64
512 if (((intptr_t)_bottom_address & 0xf) != 0) {
513 _bottom_address--;
514 }
515 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
516 #endif
517
518 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
519 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
520 assert(_bottom_address != nullptr, "");
521 assert(_bottom_address <= _cont.entrySP(), "");
522 DEBUG_ONLY(_last_write = nullptr;)
523
524 assert(_cont.chunk_invariant(), "");
525 assert(!Interpreter::contains(_cont.entryPC()), "");
526 #if !defined(PPC64) || defined(ZERO)
527 static const int do_yield_frame_size = frame::metadata_words;
528 #else
529 static const int do_yield_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
530 #endif
531 // With preemption doYield() might not have been resolved yet
532 assert(_preempt || ContinuationEntry::do_yield_nmethod()->frame_size() == do_yield_frame_size, "");
533
534 if (preempt) {
535 _last_frame = _thread->last_frame();
536 }
537
538 // properties of the continuation on the stack; all sizes are in words
539 _cont_stack_top = frame_sp + (!preempt ? do_yield_frame_size : 0); // we don't freeze the doYield stub frame
540 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
541 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
542
543 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
544 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
545 assert(cont_size() > 0, "");
546
547 if (LockingMode != LM_LIGHTWEIGHT) {
548 _monitors_in_lockstack = 0;
549 } else {
550 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
551 }
552 }
553
554 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
555 _freeze_size = 0;
556 _total_align_size = 0;
557 NOT_PRODUCT(_frames = 0;)
558 }
559
560 void FreezeBase::freeze_lockstack(stackChunkOop chunk) {
561 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
562
563 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
564 chunk->set_lockstack_size(checked_cast<uint8_t>(_monitors_in_lockstack));
565 chunk->set_has_lockstack(true);
566 }
567
568 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
569 stackChunkOop chunk = _cont.tail();
570 chunk->copy_from_stack_to_chunk(from, to, size);
571 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
572
573 #ifdef ASSERT
574 if (_last_write != nullptr) {
575 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
576 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
577 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
578 _last_write = to;
579 }
580 #endif
581 }
582
583 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
584 #ifdef ASSERT
585 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
586 if (watermark == nullptr) {
587 return;
588 }
589 ContinuationEntry* ce = thread->last_continuation();
590 RegisterMap map(thread,
591 RegisterMap::UpdateMap::include,
592 RegisterMap::ProcessFrames::skip,
593 RegisterMap::WalkContinuation::skip);
594 map.set_include_argument_oops(false);
595 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
596 watermark->assert_is_frame_safe(f);
597 }
598 #endif // ASSERT
599 }
600
601 #ifdef ASSERT
602 static bool monitors_on_stack(JavaThread* thread) {
603 assert_frames_in_continuation_are_safe(thread);
604 ContinuationEntry* ce = thread->last_continuation();
605 RegisterMap map(thread,
606 RegisterMap::UpdateMap::include,
607 RegisterMap::ProcessFrames::skip,
608 RegisterMap::WalkContinuation::skip);
609 map.set_include_argument_oops(false);
610 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
611 if ((f.is_interpreted_frame() && ContinuationHelper::InterpretedFrame::is_owning_locks(f)) ||
612 (f.is_compiled_frame() && ContinuationHelper::CompiledFrame::is_owning_locks(map.thread(), &map, f)) ||
613 (f.is_native_frame() && ContinuationHelper::NativeFrame::is_owning_locks(map.thread(), f))) {
614 return true;
615 }
616 }
617 return false;
618 }
619 #endif // ASSERT
620
621 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
622 void FreezeBase::unwind_frames() {
623 ContinuationEntry* entry = _cont.entry();
624 entry->flush_stack_processing(_thread);
625 assert_frames_in_continuation_are_safe(_thread);
626 JFR_ONLY(Jfr::check_and_process_sample_request(_thread);)
627 assert(LockingMode != LM_LEGACY || !monitors_on_stack(_thread), "unexpected monitors on stack");
628 set_anchor_to_entry(_thread, entry);
629 }
630
631 template <typename ConfigT>
632 freeze_result Freeze<ConfigT>::try_freeze_fast() {
633 assert(_thread->thread_state() == _thread_in_vm, "");
634 assert(_thread->cont_fastpath(), "");
635
636 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
637 assert(_fast_freeze_size == 0, "");
638
639 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
640 if (freeze_fast_new_chunk(chunk)) {
641 return freeze_ok;
642 }
643 if (_thread->has_pending_exception()) {
644 return freeze_exception;
645 }
646
647 // TODO R REMOVE when deopt change is fixed
648 assert(!_thread->cont_fastpath() || _barriers, "");
649 log_develop_trace(continuations)("-- RETRYING SLOW --");
650 return freeze_slow();
651 }
652
653 // Returns size needed if the continuation fits, otherwise 0.
654 int FreezeBase::size_if_fast_freeze_available() {
655 stackChunkOop chunk = _cont.tail();
656 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
657 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
658 return 0;
659 }
660
661 int total_size_needed = cont_size();
662 const int chunk_sp = chunk->sp();
663
664 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
665 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
666 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
667 // although that would require changing stackChunkOopDesc::is_empty
668 if (!chunk->is_empty()) {
669 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
670 }
671
672 total_size_needed += _monitors_in_lockstack;
673
674 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
675 bool available = chunk_free_room >= total_size_needed;
676 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
677 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
678 return available ? total_size_needed : 0;
679 }
680
681 void FreezeBase::freeze_fast_existing_chunk() {
682 stackChunkOop chunk = _cont.tail();
683
684 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
685 assert(_fast_freeze_size > 0, "");
686
687 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
688 DEBUG_ONLY(_empty = false;)
689 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
690 #ifdef ASSERT
691 {
692 intptr_t* retaddr_slot = (chunk->sp_address()
693 - frame::sender_sp_ret_address_offset());
694 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
695 "unexpected saved return address");
696 }
697 #endif
698
699 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
700 // we overlap; we'll overwrite the chunk's top frame's callee arguments
701 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
702 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
703
704 // increase max_size by what we're freezing minus the overlap
705 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
706
707 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
708 assert(bottom_sp == _bottom_address, "");
709 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
710 // should have a return barrier (installed back when we thawed it).
711 #ifdef ASSERT
712 {
713 intptr_t* retaddr_slot = (bottom_sp
714 - frame::sender_sp_ret_address_offset());
715 assert(ContinuationHelper::return_address_at(retaddr_slot)
716 == StubRoutines::cont_returnBarrier(),
717 "should be the continuation return barrier");
718 }
719 #endif
720 // We copy the fp from the chunk back to the stack because it contains some caller data,
721 // including, possibly, an oop that might have gone stale since we thawed.
722 patch_stack_pd(bottom_sp, chunk->sp_address());
723 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
724
725 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
726 } else { // the chunk is empty
727 const int chunk_start_sp = chunk->stack_size();
728
729 DEBUG_ONLY(_empty = true;)
730 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
731
732 chunk->set_max_thawing_size(cont_size());
733 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
734 chunk->set_sp(chunk->bottom());
735
736 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
737 }
738 }
739
740 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
741 DEBUG_ONLY(_empty = true;)
742
743 // Install new chunk
744 _cont.set_tail(chunk);
745
746 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
747 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
748 return false;
749 }
750
751 chunk->set_max_thawing_size(cont_size());
752
753 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
754 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
755 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
756 assert(chunk_start_sp == chunk->stack_size(), "");
757
758 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
759
760 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
761
762 return true;
763 }
764
765 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
766 assert(chunk != nullptr, "");
767 assert(!chunk->has_mixed_frames(), "");
768 assert(!chunk->is_gc_mode(), "");
769 assert(!chunk->has_bitmap(), "");
770 assert(!chunk->requires_barriers(), "");
771 assert(chunk == _cont.tail(), "");
772
773 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
774 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
775 // will either see no continuation on the stack, or a consistent chunk.
776 unwind_frames();
777
778 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
779 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
780 assert(chunk_start_sp <= chunk->stack_size(), "");
781 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
782
783 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
784 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
785
786 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
787 #ifdef ASSERT
788 if (!_empty) {
789 intptr_t* retaddr_slot = (_orig_chunk_sp
790 - frame::sender_sp_ret_address_offset());
791 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
792 "unexpected saved return address");
793 }
794 #endif
795
796 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
797 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
798 intptr_t* from = _cont_stack_top - frame::metadata_words_at_bottom;
799 intptr_t* to = chunk_top - frame::metadata_words_at_bottom;
800 copy_to_chunk(from, to, cont_size() + frame::metadata_words_at_bottom);
801 // Because we're not patched yet, the chunk is now in a bad state
802
803 // patch return pc of the bottom-most frozen frame (now in the chunk)
804 // with the actual caller's return address
805 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
806 - _cont.argsize()
807 - frame::metadata_words_at_top
808 - frame::sender_sp_ret_address_offset());
809 #ifdef ASSERT
810 if (!_empty) {
811 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
812 == StubRoutines::cont_returnBarrier(),
813 "should be the continuation return barrier");
814 }
815 #endif
816 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
817 chunk->pc());
818
819 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
820 chunk->set_sp(chunk_new_sp);
821
822 // set chunk->pc to the return address of the topmost frame in the chunk
823 if (_preempt) {
824 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
825 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
826 // be the pc we used when creating the oopmap. Get the top's frame last pc from
827 // the anchor instead.
828 address last_pc = _last_frame.pc();
829 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
830 chunk->set_pc(last_pc);
831 } else {
832 chunk->set_pc(ContinuationHelper::return_address_at(
833 _cont_stack_top - frame::sender_sp_ret_address_offset()));
834 }
835
836 if (_monitors_in_lockstack > 0) {
837 freeze_lockstack(chunk);
838 }
839
840 _cont.write();
841
842 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
843 LogTarget(Trace, continuations) lt;
844 if (lt.develop_is_enabled()) {
845 LogStream ls(lt);
846 chunk->print_on(true, &ls);
847 }
848
849 // Verification
850 assert(_cont.chunk_invariant(), "");
851 chunk->verify();
852
853 #if CONT_JFR
854 EventContinuationFreezeFast e;
855 if (e.should_commit()) {
856 e.set_id(cast_from_oop<u8>(chunk));
857 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
858 e.set_size(cont_size() << LogBytesPerWord);
859 e.commit();
860 }
861 #endif
862 }
863
864 NOINLINE freeze_result FreezeBase::freeze_slow() {
865 #ifdef ASSERT
866 ResourceMark rm;
867 #endif
868
869 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
870 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
871
872 #if CONT_JFR
873 EventContinuationFreezeSlow e;
874 if (e.should_commit()) {
875 e.set_id(cast_from_oop<u8>(_cont.continuation()));
876 e.commit();
877 }
878 #endif
879
880 init_rest();
881
882 HandleMark hm(Thread::current());
883
884 frame f = freeze_start_frame();
885
886 LogTarget(Debug, continuations) lt;
887 if (lt.develop_is_enabled()) {
888 LogStream ls(lt);
889 f.print_on(&ls);
890 }
891
892 frame caller; // the frozen caller in the chunk
893 freeze_result res = recurse_freeze(f, caller, 0, false, true);
894
895 if (res == freeze_ok) {
896 finish_freeze(f, caller);
897 _cont.write();
898 }
899
900 return res;
901 }
902
903 frame FreezeBase::freeze_start_frame() {
904 if (LIKELY(!_preempt)) {
905 return freeze_start_frame_yield_stub();
906 } else {
907 return freeze_start_frame_on_preempt();
908 }
909 }
910
911 frame FreezeBase::freeze_start_frame_yield_stub() {
912 frame f = _thread->last_frame();
913 assert(ContinuationEntry::do_yield_nmethod()->contains(f.pc()), "must be");
914 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
915 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
916 return f;
917 }
918
919 frame FreezeBase::freeze_start_frame_on_preempt() {
920 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
921 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
922 return _last_frame;
923 }
924
925 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
926 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
927 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
928 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
929 || ((top && _preempt) == f.is_native_frame()), "");
930
931 if (stack_overflow()) {
932 return freeze_exception;
933 }
934
935 if (f.is_compiled_frame()) {
936 if (UNLIKELY(f.oop_map() == nullptr)) {
937 // special native frame
938 return freeze_pinned_native;
939 }
940 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
941 } else if (f.is_interpreted_frame()) {
942 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
943 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
944 } else if (top && _preempt) {
945 assert(f.is_native_frame() || f.is_runtime_frame(), "");
946 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
947 } else {
948 // Frame can't be frozen. Most likely the call_stub or upcall_stub
949 // which indicates there are further natives frames up the stack.
950 return freeze_pinned_native;
951 }
952 }
953
954 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
955 // See also StackChunkFrameStream<frame_kind>::frame_size()
956 template<typename FKind>
957 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
958 assert(FKind::is_instance(f), "");
959
960 assert(fsize > 0, "");
961 assert(argsize >= 0, "");
962 _freeze_size += fsize;
963 NOT_PRODUCT(_frames++;)
964
965 assert(FKind::frame_bottom(f) <= _bottom_address, "");
966
967 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
968 // enterSpecial and an interpreted frame
969 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
970 return finalize_freeze(f, caller, argsize); // recursion end
971 } else {
972 frame senderf = sender<FKind>(f);
973 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
974 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
975 return result;
976 }
977 }
978
979 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
980 LogTarget(Trace, continuations) lt;
981 if (lt.develop_is_enabled()) {
982 LogStream ls(lt);
983 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
984 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
985 f.print_value_on(&ls);
986 }
987 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
988 }
989
990 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
991 LogTarget(Trace, continuations) lt;
992 if (lt.develop_is_enabled()) {
993 LogStream ls(lt);
994 DEBUG_ONLY(hf.print_value_on(&ls);)
995 assert(hf.is_heap_frame(), "should be");
996 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
997 if (is_bottom_frame) {
998 ls.print_cr("bottom h-frame:");
999 hf.print_on(&ls);
1000 }
1001 }
1002 }
1003
1004 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
1005 // See also StackChunkFrameStream<frame_kind>::frame_size()
1006 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
1007 int argsize = argsize_md - frame::metadata_words_at_top;
1008 assert(callee.is_interpreted_frame()
1009 || ContinuationHelper::Frame::is_stub(callee.cb())
1010 || callee.cb()->as_nmethod()->is_osr_method()
1011 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
1012 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
1013 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
1014
1015 LogTarget(Trace, continuations) lt;
1016
1017 #ifdef ASSERT
1018 bool empty = _cont.is_empty();
1019 log_develop_trace(continuations)("empty: %d", empty);
1020 #endif
1021
1022 stackChunkOop chunk = _cont.tail();
1023
1024 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1025
1026 _freeze_size += frame::metadata_words; // for top frame's metadata
1027
1028 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1029 int unextended_sp = -1;
1030 if (chunk != nullptr) {
1031 if (!chunk->is_empty()) {
1032 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1033 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1034 bool top_interpreted = Interpreter::contains(chunk->pc());
1035 if (callee.is_interpreted_frame() == top_interpreted) {
1036 overlap = argsize_md;
1037 }
1038 } else {
1039 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1040 }
1041 }
1042
1043 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1044
1045 _freeze_size -= overlap;
1046 assert(_freeze_size >= 0, "");
1047
1048 assert(chunk == nullptr || chunk->is_empty()
1049 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1050 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1051
1052 _freeze_size += _monitors_in_lockstack;
1053
1054 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1055 bool allocated_old_in_freeze_fast = _barriers;
1056 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1057 "Chunk allocated in freeze_fast is of insufficient size "
1058 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1059 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1060
1061 DEBUG_ONLY(bool empty_chunk = true);
1062 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1063 // ALLOCATE NEW CHUNK
1064
1065 if (lt.develop_is_enabled()) {
1066 LogStream ls(lt);
1067 if (chunk == nullptr) {
1068 ls.print_cr("no chunk");
1069 } else {
1070 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1071 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1072 chunk->print_on(&ls);
1073 }
1074 }
1075
1076 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1077 // overlap = 0;
1078
1079 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1080 if (chunk == nullptr) {
1081 return freeze_exception;
1082 }
1083
1084 // Install new chunk
1085 _cont.set_tail(chunk);
1086 assert(chunk->is_empty(), "");
1087 } else {
1088 // REUSE EXISTING CHUNK
1089 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1090 if (chunk->is_empty()) {
1091 int sp = chunk->stack_size() - argsize_md;
1092 chunk->set_sp(sp);
1093 chunk->set_bottom(sp);
1094 _freeze_size += overlap;
1095 assert(chunk->max_thawing_size() == 0, "");
1096 } DEBUG_ONLY(else empty_chunk = false;)
1097 }
1098 assert(!chunk->is_gc_mode(), "");
1099 assert(!chunk->has_bitmap(), "");
1100 chunk->set_has_mixed_frames(true);
1101
1102 assert(chunk->requires_barriers() == _barriers, "");
1103 assert(!_barriers || chunk->is_empty(), "");
1104
1105 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1106 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1107
1108 if (_preempt) {
1109 frame top_frame = _thread->last_frame();
1110 if (top_frame.is_interpreted_frame()) {
1111 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1112 // We need it so that on resume we can restore the sp to the right place, since
1113 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1114 // We do it now that we know freezing will be successful.
1115 prepare_freeze_interpreted_top_frame(top_frame);
1116 }
1117
1118 // Do this now so should_process_args_at_top() is set before calling finish_freeze
1119 // in case we might need to apply GC barriers to frames in this stackChunk.
1120 if (_thread->at_preemptable_init()) {
1121 assert(top_frame.is_interpreted_frame(), "only InterpreterRuntime::_new/resolve_from_cache allowed");
1122 chunk->set_at_klass_init(true);
1123 Method* m = top_frame.interpreter_frame_method();
1124 Bytecode current_bytecode = Bytecode(m, top_frame.interpreter_frame_bcp());
1125 Bytecodes::Code code = current_bytecode.code();
1126 int exp_size = top_frame.interpreter_frame_expression_stack_size();
1127 if (code == Bytecodes::Code::_invokestatic && exp_size > 0) {
1128 chunk->set_has_args_at_top(true);
1129 }
1130 }
1131 }
1132
1133 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1134 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1135 // will either see no continuation or a consistent chunk.
1136 unwind_frames();
1137
1138 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1139
1140 if (lt.develop_is_enabled()) {
1141 LogStream ls(lt);
1142 ls.print_cr("top chunk:");
1143 chunk->print_on(&ls);
1144 }
1145
1146 if (_monitors_in_lockstack > 0) {
1147 freeze_lockstack(chunk);
1148 }
1149
1150 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1151 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1152
1153 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1154
1155 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1156 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1157 #ifdef ASSERT
1158 if (lt.develop_is_enabled()) {
1159 LogStream ls(lt);
1160 ls.print_cr("top hframe before (freeze):");
1161 assert(caller.is_heap_frame(), "should be");
1162 caller.print_on(&ls);
1163 }
1164
1165 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1166
1167 frame entry = sender(callee);
1168
1169 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1170 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1171 #endif
1172
1173 return freeze_ok_bottom;
1174 }
1175
1176 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1177 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1178 if (is_bottom_frame) {
1179 // If we're the bottom frame, we need to replace the return barrier with the real
1180 // caller's pc.
1181 address last_pc = caller.pc();
1182 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1183 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1184 } else {
1185 assert(!caller.is_empty(), "");
1186 }
1187
1188 patch_pd(hf, caller);
1189
1190 if (f.is_interpreted_frame()) {
1191 assert(hf.is_heap_frame(), "should be");
1192 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1193 }
1194
1195 #ifdef ASSERT
1196 if (hf.is_compiled_frame()) {
1197 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1198 log_develop_trace(continuations)("Freezing deoptimized frame");
1199 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1200 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1201 }
1202 }
1203 #endif
1204 }
1205
1206 #ifdef ASSERT
1207 static void verify_frame_top(const frame& f, intptr_t* top) {
1208 ResourceMark rm;
1209 InterpreterOopMap mask;
1210 f.interpreted_frame_oop_map(&mask);
1211 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1212 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1213 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1214 }
1215 #endif // ASSERT
1216
1217 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1218 // See also StackChunkFrameStream<frame_kind>::frame_size()
1219 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1220 int callee_argsize /* incl. metadata */,
1221 bool callee_interpreted) {
1222 adjust_interpreted_frame_unextended_sp(f);
1223
1224 // The frame's top never includes the stack arguments to the callee
1225 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1226 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1227 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1228
1229 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1230
1231 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1232 // including metadata between f and its args
1233 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1234
1235 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1236 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1237 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1238 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1239
1240 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1241 if (UNLIKELY(result > freeze_ok_bottom)) {
1242 return result;
1243 }
1244
1245 bool is_bottom_frame = result == freeze_ok_bottom;
1246 assert(!caller.is_empty() || is_bottom_frame, "");
1247
1248 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1249
1250 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1251 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1252
1253 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1254 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1255 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1256
1257 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1258 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1259 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1260 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1261
1262 relativize_interpreted_frame_metadata(f, hf);
1263
1264 patch(f, hf, caller, is_bottom_frame);
1265
1266 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1267 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1268 caller = hf;
1269
1270 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1271 frame_method->record_gc_epoch();
1272
1273 return freeze_ok;
1274 }
1275
1276 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1277 // See also StackChunkFrameStream<frame_kind>::frame_size()
1278 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1279 int callee_argsize /* incl. metadata */,
1280 bool callee_interpreted) {
1281 // The frame's top never includes the stack arguments to the callee
1282 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1283 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1284 // including metadata between f and its stackargs
1285 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1286 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1287
1288 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d",
1289 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1290 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1291 _freeze_size, fsize, argsize);
1292 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1293 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1294
1295 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1296 if (UNLIKELY(result > freeze_ok_bottom)) {
1297 return result;
1298 }
1299
1300 bool is_bottom_frame = result == freeze_ok_bottom;
1301 assert(!caller.is_empty() || is_bottom_frame, "");
1302
1303 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1304
1305 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller);
1306
1307 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1308
1309 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1310 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1311
1312 if (caller.is_interpreted_frame()) {
1313 // When thawing the frame we might need to add alignment (see Thaw::align)
1314 _total_align_size += frame::align_wiggle;
1315 }
1316
1317 patch(f, hf, caller, is_bottom_frame);
1318
1319 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1320
1321 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1322 caller = hf;
1323 return freeze_ok;
1324 }
1325
1326 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1327 DEBUG_ONLY(frame fsender = sender(f);)
1328 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1329
1330 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1331 const int fsize = f.cb()->frame_size();
1332
1333 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1334 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1335
1336 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1337 if (UNLIKELY(result > freeze_ok_bottom)) {
1338 return result;
1339 }
1340
1341 assert(result == freeze_ok, "should have caller");
1342 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1343
1344 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1345 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1346
1347 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1348
1349 patch(f, hf, caller, false /*is_bottom_frame*/);
1350
1351 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1352
1353 caller = hf;
1354 return freeze_ok;
1355 }
1356
1357 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1358 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1359 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1360 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1361 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1362 // caller into the native convention. This is so that we have a valid oopMap in case of
1363 // having to block in the slow path. But that would require freezing those registers too
1364 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1365 // given that this would be a rare case anyways just pin the vthread to the carrier.
1366 return freeze_pinned_native;
1367 }
1368
1369 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1370 // There are no stackargs but argsize must include the metadata
1371 const int argsize = frame::metadata_words_at_top;
1372 const int fsize = f.cb()->frame_size() + argsize;
1373
1374 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1375 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1376
1377 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1378 if (UNLIKELY(result > freeze_ok_bottom)) {
1379 return result;
1380 }
1381
1382 assert(result == freeze_ok, "should have caller frame");
1383 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1384
1385 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1386 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1387
1388 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1389
1390 if (caller.is_interpreted_frame()) {
1391 // When thawing the frame we might need to add alignment (see Thaw::align)
1392 _total_align_size += frame::align_wiggle;
1393 }
1394
1395 patch(f, hf, caller, false /* is_bottom_frame */);
1396
1397 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1398
1399 caller = hf;
1400 return freeze_ok;
1401 }
1402
1403 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1404 stackChunkOop chunk = _cont.tail();
1405
1406 LogTarget(Trace, continuations) lt;
1407 if (lt.develop_is_enabled()) {
1408 LogStream ls(lt);
1409 assert(top.is_heap_frame(), "should be");
1410 top.print_on(&ls);
1411 }
1412
1413 set_top_frame_metadata_pd(top);
1414
1415 chunk->set_sp(chunk->to_offset(top.sp()));
1416 chunk->set_pc(top.pc());
1417
1418 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1419
1420 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1421
1422 // At this point the chunk is consistent
1423
1424 if (UNLIKELY(_barriers)) {
1425 log_develop_trace(continuations)("do barriers on old chunk");
1426 // Serial and Parallel GC can allocate objects directly into the old generation.
1427 // Then we want to relativize the derived pointers eagerly so that
1428 // old chunks are all in GC mode.
1429 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1430 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1431 if (UseShenandoahGC) {
1432 _cont.tail()->relativize_derived_pointers_concurrently();
1433 } else {
1434 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1435 }
1436 // For objects in the old generation we must maintain the remembered set
1437 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1438 }
1439
1440 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1441 if (lt.develop_is_enabled()) {
1442 LogStream ls(lt);
1443 chunk->print_on(true, &ls);
1444 }
1445
1446 if (lt.develop_is_enabled()) {
1447 LogStream ls(lt);
1448 ls.print_cr("top hframe after (freeze):");
1449 assert(_cont.last_frame().is_heap_frame(), "should be");
1450 _cont.last_frame().print_on(&ls);
1451 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1452 }
1453
1454 assert(_cont.chunk_invariant(), "");
1455 }
1456
1457 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1458 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1459 assert(t == JavaThread::current(), "");
1460 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1461 if (!_preempt) {
1462 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1463 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1464 }
1465 return true;
1466 }
1467 return false;
1468 }
1469
1470 class StackChunkAllocator : public MemAllocator {
1471 const size_t _stack_size;
1472 int _argsize_md;
1473 ContinuationWrapper& _continuation_wrapper;
1474 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1475 mutable bool _took_slow_path;
1476
1477 // Does the minimal amount of initialization needed for a TLAB allocation.
1478 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1479 virtual oop initialize(HeapWord* mem) const override {
1480 assert(_stack_size > 0, "");
1481 assert(_stack_size <= max_jint, "");
1482 assert(_word_size > _stack_size, "");
1483
1484 // zero out fields (but not the stack)
1485 const size_t hs = oopDesc::header_size();
1486 if (oopDesc::has_klass_gap()) {
1487 oopDesc::set_klass_gap(mem, 0);
1488 }
1489 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1490
1491 int bottom = (int)_stack_size - _argsize_md;
1492
1493 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1494 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1495 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1496
1497 return finish(mem);
1498 }
1499
1500 stackChunkOop allocate_fast() const {
1501 if (!UseTLAB) {
1502 return nullptr;
1503 }
1504
1505 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1506 if (mem == nullptr) {
1507 return nullptr;
1508 }
1509
1510 oop obj = initialize(mem);
1511 return stackChunkOopDesc::cast(obj);
1512 }
1513
1514 public:
1515 StackChunkAllocator(Klass* klass,
1516 size_t word_size,
1517 Thread* thread,
1518 size_t stack_size,
1519 int argsize_md,
1520 ContinuationWrapper& continuation_wrapper,
1521 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1522 : MemAllocator(klass, word_size, thread),
1523 _stack_size(stack_size),
1524 _argsize_md(argsize_md),
1525 _continuation_wrapper(continuation_wrapper),
1526 _jvmti_event_collector(jvmti_event_collector),
1527 _took_slow_path(false) {}
1528
1529 // Provides it's own, specialized allocation which skips instrumentation
1530 // if the memory can be allocated without going to a slow-path.
1531 stackChunkOop allocate() const {
1532 // First try to allocate without any slow-paths or instrumentation.
1533 stackChunkOop obj = allocate_fast();
1534 if (obj != nullptr) {
1535 return obj;
1536 }
1537
1538 // Now try full-blown allocation with all expensive operations,
1539 // including potentially safepoint operations.
1540 _took_slow_path = true;
1541
1542 // Protect unhandled Loom oops
1543 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1544
1545 // Can safepoint
1546 _jvmti_event_collector->start();
1547
1548 // Can safepoint
1549 return stackChunkOopDesc::cast(MemAllocator::allocate());
1550 }
1551
1552 bool took_slow_path() const {
1553 return _took_slow_path;
1554 }
1555 };
1556
1557 template <typename ConfigT>
1558 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1559 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1560
1561 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1562 size_t size_in_words = klass->instance_size(stack_size);
1563
1564 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1565 if (!_preempt) {
1566 throw_stack_overflow_on_humongous_chunk();
1567 }
1568 return nullptr;
1569 }
1570
1571 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1572 assert(current == JavaThread::current(), "should be current");
1573
1574 // Allocate the chunk.
1575 //
1576 // This might safepoint while allocating, but all safepointing due to
1577 // instrumentation have been deferred. This property is important for
1578 // some GCs, as this ensures that the allocated object is in the young
1579 // generation / newly allocated memory.
1580 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1581 stackChunkOop chunk = allocator.allocate();
1582
1583 if (chunk == nullptr) {
1584 return nullptr; // OOME
1585 }
1586
1587 // assert that chunk is properly initialized
1588 assert(chunk->stack_size() == (int)stack_size, "");
1589 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1590 assert(chunk->sp() == chunk->bottom(), "");
1591 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1592 assert(chunk->max_thawing_size() == 0, "");
1593 assert(chunk->pc() == nullptr, "");
1594 assert(chunk->is_empty(), "");
1595 assert(chunk->flags() == 0, "");
1596 assert(chunk->is_gc_mode() == false, "");
1597 assert(chunk->lockstack_size() == 0, "");
1598
1599 // fields are uninitialized
1600 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1601 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1602
1603 #if INCLUDE_ZGC
1604 if (UseZGC) {
1605 ZStackChunkGCData::initialize(chunk);
1606 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1607 _barriers = false;
1608 } else
1609 #endif
1610 #if INCLUDE_SHENANDOAHGC
1611 if (UseShenandoahGC) {
1612 _barriers = chunk->requires_barriers();
1613 } else
1614 #endif
1615 {
1616 if (!allocator.took_slow_path()) {
1617 // Guaranteed to be in young gen / newly allocated memory
1618 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1619 _barriers = false;
1620 } else {
1621 // Some GCs could put direct allocations in old gen for slow-path
1622 // allocations; need to explicitly check if that was the case.
1623 _barriers = chunk->requires_barriers();
1624 }
1625 }
1626
1627 if (_barriers) {
1628 log_develop_trace(continuations)("allocation requires barriers");
1629 }
1630
1631 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1632
1633 return chunk;
1634 }
1635
1636 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1637 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1638 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1639 }
1640
1641 class AnchorMark : public StackObj {
1642 JavaThread* _current;
1643 frame& _top_frame;
1644 intptr_t* _last_sp_from_frame;
1645 bool _is_interpreted;
1646
1647 public:
1648 AnchorMark(JavaThread* current, frame& f) : _current(current), _top_frame(f), _is_interpreted(false) {
1649 intptr_t* sp = anchor_mark_set_pd();
1650 set_anchor(_current, sp);
1651 }
1652 ~AnchorMark() {
1653 clear_anchor(_current);
1654 anchor_mark_clear_pd();
1655 }
1656 inline intptr_t* anchor_mark_set_pd();
1657 inline void anchor_mark_clear_pd();
1658 };
1659
1660 #if INCLUDE_JVMTI
1661 static int num_java_frames(ContinuationWrapper& cont) {
1662 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1663 int count = 0;
1664 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1665 count += chunk->num_java_frames();
1666 }
1667 return count;
1668 }
1669
1670 static void invalidate_jvmti_stack(JavaThread* thread) {
1671 if (thread->is_interp_only_mode()) {
1672 JvmtiThreadState *state = thread->jvmti_thread_state();
1673 if (state != nullptr)
1674 state->invalidate_cur_stack_depth();
1675 }
1676 }
1677
1678 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1679 if (JvmtiExport::can_post_frame_pop()) {
1680 int num_frames = num_java_frames(cont);
1681
1682 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1683 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames);
1684 }
1685 invalidate_jvmti_stack(thread);
1686 }
1687
1688 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, Continuation::preempt_kind pk) {
1689 assert(current->vthread() != nullptr, "must be");
1690
1691 HandleMarkCleaner hm(current); // Cleanup vth and so._conth Handles
1692 Handle vth(current, current->vthread());
1693 ContinuationWrapper::SafepointOp so(current, cont);
1694
1695 AnchorMark am(current, top); // Set anchor so that the stack is walkable.
1696
1697 JRT_BLOCK
1698 JvmtiVTMSTransitionDisabler::VTMS_vthread_mount((jthread)vth.raw_value(), false);
1699
1700 if (current->pending_contended_entered_event()) {
1701 // No monitor JVMTI events for ObjectLocker case.
1702 if (pk != Continuation::object_locker) {
1703 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1704 }
1705 current->set_contended_entered_monitor(nullptr);
1706 }
1707 JRT_BLOCK_END
1708 }
1709 #endif // INCLUDE_JVMTI
1710
1711 #ifdef ASSERT
1712 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1713 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1714 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1715 bool FreezeBase::check_valid_fast_path() {
1716 ContinuationEntry* ce = _thread->last_continuation();
1717 RegisterMap map(_thread,
1718 RegisterMap::UpdateMap::skip,
1719 RegisterMap::ProcessFrames::skip,
1720 RegisterMap::WalkContinuation::skip);
1721 map.set_include_argument_oops(false);
1722 bool is_top_frame = true;
1723 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1724 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1725 return false;
1726 }
1727 }
1728 return true;
1729 }
1730
1731 static void verify_frame_kind(const frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr, const char** code_name_ptr, int* bci_ptr) {
1732 JavaThread* current = JavaThread::current();
1733 ResourceMark rm(current);
1734
1735 Method* m;
1736 const char* code_name;
1737 int bci;
1738 if (preempt_kind == Continuation::monitorenter) {
1739 assert(top.is_interpreted_frame() || top.is_runtime_frame(), "");
1740 bool at_sync_method;
1741 if (top.is_interpreted_frame()) {
1742 m = top.interpreter_frame_method();
1743 assert(!m->is_native() || m->is_synchronized(), "invalid method %s", m->external_name());
1744 address bcp = top.interpreter_frame_bcp();
1745 assert(bcp != 0 || m->is_native(), "");
1746 at_sync_method = m->is_synchronized() && (bcp == 0 || bcp == m->code_base());
1747 // bcp is advanced on monitorenter before making the VM call, adjust for that.
1748 bool at_sync_bytecode = bcp > m->code_base() && Bytecode(m, bcp - 1).code() == Bytecodes::Code::_monitorenter;
1749 assert(at_sync_method || at_sync_bytecode, "");
1750 bci = at_sync_method ? -1 : top.interpreter_frame_bci();
1751 } else {
1752 CodeBlob* cb = top.cb();
1753 RegisterMap reg_map(current,
1754 RegisterMap::UpdateMap::skip,
1755 RegisterMap::ProcessFrames::skip,
1756 RegisterMap::WalkContinuation::skip);
1757 frame fr = top.sender(®_map);
1758 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1759 compiledVFrame* cvf = compiledVFrame::cast(vf);
1760 m = cvf->method();
1761 bci = cvf->scope()->bci();
1762 at_sync_method = bci == SynchronizationEntryBCI;
1763 assert(!at_sync_method || m->is_synchronized(), "bci is %d but method %s is not synchronized", bci, m->external_name());
1764 bool is_c1_monitorenter = false, is_c2_monitorenter = false;
1765 COMPILER1_PRESENT(is_c1_monitorenter = cb == Runtime1::blob_for(C1StubId::monitorenter_id) ||
1766 cb == Runtime1::blob_for(C1StubId::monitorenter_nofpu_id);)
1767 COMPILER2_PRESENT(is_c2_monitorenter = cb == CodeCache::find_blob(OptoRuntime::complete_monitor_locking_Java());)
1768 assert(is_c1_monitorenter || is_c2_monitorenter, "wrong runtime stub frame");
1769 }
1770 code_name = at_sync_method ? "synchronized method" : "monitorenter";
1771 } else if (preempt_kind == Continuation::object_wait) {
1772 assert(top.is_interpreted_frame() || top.is_native_frame(), "");
1773 m = top.is_interpreted_frame() ? top.interpreter_frame_method() : top.cb()->as_nmethod()->method();
1774 assert(m->is_object_wait0(), "");
1775 bci = 0;
1776 code_name = "";
1777 } else {
1778 assert(preempt_kind == Continuation::object_locker, "invalid preempt kind");
1779 assert(top.is_interpreted_frame(), "");
1780 m = top.interpreter_frame_method();
1781 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
1782 Bytecodes::Code code = current_bytecode.code();
1783 assert(code == Bytecodes::Code::_new || code == Bytecodes::Code::_invokestatic ||
1784 (code == Bytecodes::Code::_getstatic || code == Bytecodes::Code::_putstatic), "invalid bytecode");
1785 bci = top.interpreter_frame_bci();
1786 code_name = Bytecodes::name(current_bytecode.code());
1787 }
1788 assert(bci >= 0 || m->is_synchronized(), "invalid bci:%d at method %s", bci, m->external_name());
1789
1790 if (m_ptr != nullptr) {
1791 *m_ptr = m;
1792 *code_name_ptr = code_name;
1793 *bci_ptr = bci;
1794 }
1795 }
1796
1797 static void log_preempt_after_freeze(ContinuationWrapper& cont) {
1798 JavaThread* current = cont.thread();
1799 StackChunkFrameStream<ChunkFrames::Mixed> sfs(cont.tail());
1800 frame top_frame = sfs.to_frame();
1801 bool at_init = current->at_preemptable_init();
1802 bool at_enter = current->current_pending_monitor() != nullptr;
1803 bool at_wait = current->current_waiting_monitor() != nullptr;
1804 assert((at_enter && !at_wait) || (!at_enter && at_wait), "");
1805 Continuation::preempt_kind pk = at_init ? Continuation::object_locker : at_enter ? Continuation::monitorenter : Continuation::object_wait;
1806
1807 Method* m = nullptr;
1808 const char* code_name = nullptr;
1809 int bci = InvalidFrameStateBci;
1810 verify_frame_kind(top_frame, pk, &m, &code_name, &bci);
1811 assert(m != nullptr && code_name != nullptr && bci != InvalidFrameStateBci, "should be set");
1812
1813 ResourceMark rm(current);
1814 if (bci < 0) {
1815 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " while synchronizing on %smethod %s", current->monitor_owner_id(), m->is_native() ? "native " : "", m->external_name());
1816 } else if (m->is_object_wait0()) {
1817 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at native method %s", current->monitor_owner_id(), m->external_name());
1818 } else {
1819 Klass* k = current->preempt_init_klass();
1820 assert(k != nullptr || !at_init, "");
1821 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at %s(bci:%d) in method %s %s%s", current->monitor_owner_id(),
1822 code_name, bci, m->external_name(), at_init ? "trying to initialize klass " : "", at_init ? k->external_name() : "");
1823 }
1824 }
1825 #endif // ASSERT
1826
1827 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1828 verify_continuation(cont.continuation());
1829 assert(!cont.is_empty(), "");
1830
1831 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1832 return freeze_ok;
1833 }
1834
1835 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1836 if (UNLIKELY(res != freeze_ok)) {
1837 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1838 verify_continuation(cont.continuation());
1839 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1840 return res;
1841 }
1842
1843 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1844 return freeze_epilog(cont);
1845 }
1846
1847 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1848 if (UNLIKELY(res != freeze_ok)) {
1849 verify_continuation(cont.continuation());
1850 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1851 return res;
1852 }
1853
1854 // Set up things so that on return to Java we jump to preempt stub.
1855 patch_return_pc_with_preempt_stub(old_last_frame);
1856 cont.tail()->set_preempted(true);
1857 DEBUG_ONLY(log_preempt_after_freeze(cont);)
1858 return freeze_epilog(cont);
1859 }
1860
1861 template<typename ConfigT, bool preempt>
1862 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1863 assert(!current->has_pending_exception(), "");
1864
1865 #ifdef ASSERT
1866 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1867 log_frames(current, false);
1868 #endif
1869
1870 CONT_JFR_ONLY(EventContinuationFreeze event;)
1871
1872 ContinuationEntry* entry = current->last_continuation();
1873
1874 oop oopCont = entry->cont_oop(current);
1875 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1876 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1877
1878 verify_continuation(oopCont);
1879 ContinuationWrapper cont(current, oopCont);
1880 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1881
1882 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1883
1884 assert(LockingMode == LM_LEGACY || (current->held_monitor_count() == 0 && current->jni_monitor_count() == 0),
1885 "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());
1886
1887 if (entry->is_pinned() || current->held_monitor_count() > 0) {
1888 log_develop_debug(continuations)("PINNED due to critical section/hold monitor");
1889 verify_continuation(cont.continuation());
1890 freeze_result res = entry->is_pinned() ? freeze_pinned_cs : freeze_pinned_monitor;
1891 if (!preempt) {
1892 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1893 }
1894 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1895 // Avoid Thread.yield() loops without safepoint polls.
1896 if (SafepointMechanism::should_process(current) && !preempt) {
1897 cont.done(); // allow safepoint
1898 ThreadInVMfromJava tivmfj(current);
1899 }
1900 return res;
1901 }
1902
1903 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1904
1905 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1906 bool fast = UseContinuationFastPath && current->cont_fastpath();
1907 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1908 freeze.freeze_fast_existing_chunk();
1909 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1910 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1911 }
1912
1913 if (preempt) {
1914 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1915 freeze.set_jvmti_event_collector(&jsoaec);
1916
1917 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1918
1919 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1920 preempt_epilog(cont, res, freeze.last_frame());
1921 return res;
1922 }
1923
1924 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1925 assert(current == JavaThread::current(), "must be current thread");
1926 JRT_BLOCK
1927 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1928 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1929 freeze.set_jvmti_event_collector(&jsoaec);
1930
1931 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1932
1933 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1934 freeze_epilog(current, cont, res);
1935 cont.done(); // allow safepoint in the transition back to Java
1936 return res;
1937 JRT_BLOCK_END
1938 }
1939
1940 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1941 ContinuationEntry* entry = thread->last_continuation();
1942 if (entry == nullptr) {
1943 return freeze_ok;
1944 }
1945 if (entry->is_pinned()) {
1946 return freeze_pinned_cs;
1947 } else if (thread->held_monitor_count() > 0) {
1948 return freeze_pinned_monitor;
1949 }
1950
1951 RegisterMap map(thread,
1952 RegisterMap::UpdateMap::include,
1953 RegisterMap::ProcessFrames::skip,
1954 RegisterMap::WalkContinuation::skip);
1955 map.set_include_argument_oops(false);
1956 frame f = thread->last_frame();
1957
1958 if (!safepoint) {
1959 f = f.sender(&map); // this is the yield frame
1960 } else { // safepoint yield
1961 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1962 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1963 #else
1964 Unimplemented();
1965 #endif
1966 if (!Interpreter::contains(f.pc())) {
1967 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1968 assert(f.oop_map() != nullptr, "must be");
1969 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1970 }
1971 }
1972
1973 while (true) {
1974 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1975 return freeze_pinned_native;
1976 }
1977
1978 f = f.sender(&map);
1979 if (!Continuation::is_frame_in_continuation(entry, f)) {
1980 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
1981 if (scope == cont_scope) {
1982 break;
1983 }
1984 intx monitor_count = entry->parent_held_monitor_count();
1985 entry = entry->parent();
1986 if (entry == nullptr) {
1987 break;
1988 }
1989 if (entry->is_pinned()) {
1990 return freeze_pinned_cs;
1991 } else if (monitor_count > 0) {
1992 return freeze_pinned_monitor;
1993 }
1994 }
1995 }
1996 return freeze_ok;
1997 }
1998
1999 /////////////// THAW ////
2000
2001 static int thaw_size(stackChunkOop chunk) {
2002 int size = chunk->max_thawing_size();
2003 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
2004 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
2005 return size;
2006 }
2007
2008 // make room on the stack for thaw
2009 // returns the size in bytes, or 0 on failure
2010 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
2011 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
2012
2013 assert(thread == JavaThread::current(), "");
2014
2015 ContinuationEntry* ce = thread->last_continuation();
2016 assert(ce != nullptr, "");
2017 oop continuation = ce->cont_oop(thread);
2018 assert(continuation == get_continuation(thread), "");
2019 verify_continuation(continuation);
2020
2021 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
2022 assert(chunk != nullptr, "");
2023
2024 // The tail can be empty because it might still be available for another freeze.
2025 // However, here we want to thaw, so we get rid of it (it will be GCed).
2026 if (UNLIKELY(chunk->is_empty())) {
2027 chunk = chunk->parent();
2028 assert(chunk != nullptr, "");
2029 assert(!chunk->is_empty(), "");
2030 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
2031 }
2032
2033 // Verification
2034 chunk->verify();
2035 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
2036
2037 // Only make space for the last chunk because we only thaw from the last chunk
2038 int size = thaw_size(chunk) << LogBytesPerWord;
2039
2040 const address bottom = (address)thread->last_continuation()->entry_sp();
2041 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
2042 // for the Java frames in the check below.
2043 if (!stack_overflow_check(thread, size + 300, bottom)) {
2044 return 0;
2045 }
2046
2047 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
2048 p2i(bottom), p2i(bottom - size), size);
2049 return size;
2050 }
2051
2052 class ThawBase : public StackObj {
2053 protected:
2054 JavaThread* _thread;
2055 ContinuationWrapper& _cont;
2056 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
2057
2058 intptr_t* _fastpath;
2059 bool _barriers;
2060 bool _preempted_case;
2061 bool _process_args_at_top;
2062 intptr_t* _top_unextended_sp_before_thaw;
2063 int _align_size;
2064 DEBUG_ONLY(intptr_t* _top_stack_address);
2065
2066 // Only used for some preemption cases.
2067 ObjectMonitor* _monitor;
2068
2069 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
2070
2071 NOT_PRODUCT(int _frames;)
2072
2073 protected:
2074 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2075 _thread(thread), _cont(cont),
2076 _fastpath(nullptr) {
2077 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2078 assert (cont.tail() != nullptr, "no last chunk");
2079 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2080 }
2081
2082 void clear_chunk(stackChunkOop chunk);
2083 template<bool check_stub>
2084 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2085 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2086
2087 void thaw_lockstack(stackChunkOop chunk);
2088
2089 // fast path
2090 inline void prefetch_chunk_pd(void* start, int size_words);
2091 void patch_return(intptr_t* sp, bool is_last);
2092
2093 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
2094 inline intptr_t* push_cleanup_continuation();
2095 inline intptr_t* push_preempt_adapter();
2096 intptr_t* redo_vmcall(JavaThread* current, frame& top);
2097 void throw_interrupted_exception(JavaThread* current, frame& top);
2098
2099 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
2100 void finish_thaw(frame& f);
2101
2102 private:
2103 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2104 void finalize_thaw(frame& entry, int argsize);
2105
2106 inline bool seen_by_gc();
2107
2108 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2109 inline void after_thaw_java_frame(const frame& f, bool bottom);
2110 inline void patch(frame& f, const frame& caller, bool bottom);
2111 void clear_bitmap_bits(address start, address end);
2112
2113 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top);
2114 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2115 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2116 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
2117
2118 void push_return_frame(frame& f);
2119 inline frame new_entry_frame();
2120 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom);
2121 inline void patch_pd(frame& f, const frame& sender);
2122 inline void patch_pd(frame& f, intptr_t* caller_sp);
2123 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2124
2125 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2126
2127 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2128
2129 public:
2130 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2131 };
2132
2133 template <typename ConfigT>
2134 class Thaw : public ThawBase {
2135 public:
2136 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2137
2138 inline bool can_thaw_fast(stackChunkOop chunk) {
2139 return !_barriers
2140 && _thread->cont_fastpath_thread_state()
2141 && !chunk->has_thaw_slowpath_condition()
2142 && !PreserveFramePointer;
2143 }
2144
2145 inline intptr_t* thaw(Continuation::thaw_kind kind);
2146 template<bool check_stub = false>
2147 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2148 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2149 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2150 };
2151
2152 template <typename ConfigT>
2153 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2154 verify_continuation(_cont.continuation());
2155 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2156 assert(!_cont.is_empty(), "");
2157
2158 stackChunkOop chunk = _cont.tail();
2159 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2160 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2161
2162 _barriers = chunk->requires_barriers();
2163 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2164 : thaw_slow(chunk, kind);
2165 }
2166
2167 class ReconstructedStack : public StackObj {
2168 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2169 int _thaw_size;
2170 int _argsize;
2171 public:
2172 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2173 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2174 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2175 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2176 // We're at most one alignment word away from entrySP
2177 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2178 }
2179
2180 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2181
2182 // top and bottom stack pointers
2183 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2184 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2185
2186 // several operations operate on the totality of the stack being reconstructed,
2187 // including the metadata words
2188 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2189 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2190 };
2191
2192 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2193 chunk->set_sp(chunk->bottom());
2194 chunk->set_max_thawing_size(0);
2195 }
2196
2197 template<bool check_stub>
2198 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2199 bool empty = false;
2200 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2201 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2202 assert(chunk_sp == f.sp(), "");
2203 assert(chunk_sp == f.unextended_sp(), "");
2204
2205 int frame_size = f.cb()->frame_size();
2206 argsize = f.stack_argsize();
2207
2208 assert(!f.is_stub() || check_stub, "");
2209 if (check_stub && f.is_stub()) {
2210 // If we don't thaw the top compiled frame too, after restoring the saved
2211 // registers back in Java, we would hit the return barrier to thaw one more
2212 // frame effectively overwriting the restored registers during that call.
2213 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2214 assert(!f.is_done(), "");
2215
2216 f.get_cb();
2217 assert(f.is_compiled(), "");
2218 frame_size += f.cb()->frame_size();
2219 argsize = f.stack_argsize();
2220
2221 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2222 // The caller of the runtime stub when the continuation is preempted is not at a
2223 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2224 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2225 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2226 }
2227 }
2228
2229 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2230 empty = f.is_done();
2231 assert(!empty || argsize == chunk->argsize(), "");
2232
2233 if (empty) {
2234 clear_chunk(chunk);
2235 } else {
2236 chunk->set_sp(chunk->sp() + frame_size);
2237 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2238 // We set chunk->pc to the return pc into the next frame
2239 chunk->set_pc(f.pc());
2240 #ifdef ASSERT
2241 {
2242 intptr_t* retaddr_slot = (chunk_sp
2243 + frame_size
2244 - frame::sender_sp_ret_address_offset());
2245 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2246 "unexpected pc");
2247 }
2248 #endif
2249 }
2250 assert(empty == chunk->is_empty(), "");
2251 // returns the size required to store the frame on stack, and because it is a
2252 // compiled frame, it must include a copy of the arguments passed by the caller
2253 return frame_size + argsize + frame::metadata_words_at_top;
2254 }
2255
2256 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2257 int lockStackSize = chunk->lockstack_size();
2258 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2259
2260 oop tmp_lockstack[LockStack::CAPACITY];
2261 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2262 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2263
2264 chunk->set_lockstack_size(0);
2265 chunk->set_has_lockstack(false);
2266 }
2267
2268 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2269 assert(to >= _top_stack_address, "overwrote past thawing space"
2270 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2271 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2272 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2273 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2274 }
2275
2276 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2277 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2278
2279 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2280 ContinuationHelper::patch_return_address_at(
2281 sp - frame::sender_sp_ret_address_offset(),
2282 pc);
2283 }
2284
2285 template <typename ConfigT>
2286 template<bool check_stub>
2287 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2288 assert(chunk == _cont.tail(), "");
2289 assert(!chunk->has_mixed_frames(), "");
2290 assert(!chunk->requires_barriers(), "");
2291 assert(!chunk->has_bitmap(), "");
2292 assert(!_thread->is_interp_only_mode(), "");
2293
2294 LogTarget(Trace, continuations) lt;
2295 if (lt.develop_is_enabled()) {
2296 LogStream ls(lt);
2297 ls.print_cr("thaw_fast");
2298 chunk->print_on(true, &ls);
2299 }
2300
2301 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2302 static const int threshold = 500; // words
2303
2304 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2305 int argsize, thaw_size;
2306
2307 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2308
2309 bool partial, empty;
2310 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2311 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2312
2313 partial = false;
2314 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2315 clear_chunk(chunk);
2316 thaw_size = full_chunk_size;
2317 empty = true;
2318 } else { // thaw a single frame
2319 partial = true;
2320 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2321 empty = chunk->is_empty();
2322 }
2323
2324 // Are we thawing the last frame(s) in the continuation
2325 const bool is_last = empty && chunk->parent() == nullptr;
2326 assert(!is_last || argsize == 0, "");
2327
2328 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2329 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2330
2331 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2332
2333 // also copy metadata words at frame bottom
2334 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2335
2336 // update the ContinuationEntry
2337 _cont.set_argsize(argsize);
2338 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2339 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2340
2341 // install the return barrier if not last frame, or the entry's pc if last
2342 patch_return(rs.bottom_sp(), is_last);
2343
2344 // insert the back links from callee to caller frames
2345 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2346
2347 assert(is_last == _cont.is_empty(), "");
2348 assert(_cont.chunk_invariant(), "");
2349
2350 #if CONT_JFR
2351 EventContinuationThawFast e;
2352 if (e.should_commit()) {
2353 e.set_id(cast_from_oop<u8>(chunk));
2354 e.set_size(thaw_size << LogBytesPerWord);
2355 e.set_full(!partial);
2356 e.commit();
2357 }
2358 #endif
2359
2360 #ifdef ASSERT
2361 set_anchor(_thread, rs.sp());
2362 log_frames(_thread);
2363 if (LoomDeoptAfterThaw) {
2364 do_deopt_after_thaw(_thread);
2365 }
2366 clear_anchor(_thread);
2367 #endif
2368
2369 return rs.sp();
2370 }
2371
2372 inline bool ThawBase::seen_by_gc() {
2373 return _barriers || _cont.tail()->is_gc_mode();
2374 }
2375
2376 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2377 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2378 if (UseZGC || UseShenandoahGC) {
2379 chunk->relativize_derived_pointers_concurrently();
2380 }
2381 #endif
2382 }
2383
2384 template <typename ConfigT>
2385 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2386 Continuation::preempt_kind preempt_kind;
2387 bool retry_fast_path = false;
2388
2389 _process_args_at_top = false;
2390 _preempted_case = chunk->preempted();
2391 if (_preempted_case) {
2392 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2393 if (waiter != nullptr) {
2394 // Mounted again after preemption. Resume the pending monitor operation,
2395 // which will be either a monitorenter or Object.wait() call.
2396 ObjectMonitor* mon = waiter->monitor();
2397 preempt_kind = waiter->is_wait() ? Continuation::object_wait : Continuation::monitorenter;
2398
2399 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2400 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2401 if (!mon_acquired) {
2402 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2403 log_trace(continuations, tracking)("Failed to acquire monitor, unmounting again");
2404 return push_cleanup_continuation();
2405 }
2406 _monitor = mon; // remember monitor since we might need it on handle_preempted_continuation()
2407 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2408 } else {
2409 // Preemption cancelled in moniterenter case. We actually acquired
2410 // the monitor after freezing all frames so nothing to do. In case
2411 // of preemption on ObjectLocker during klass init, we released the
2412 // monitor already at ~ObjectLocker so here we just set _monitor to
2413 // nullptr so we know there is no need to release it later.
2414 preempt_kind = Continuation::monitorenter;
2415 _monitor = nullptr;
2416 }
2417
2418 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2419 relativize_chunk_concurrently(chunk);
2420
2421 if (chunk->at_klass_init()) {
2422 preempt_kind = Continuation::object_locker;
2423 chunk->set_at_klass_init(false);
2424 _process_args_at_top = chunk->has_args_at_top();
2425 if (_process_args_at_top) chunk->set_has_args_at_top(false);
2426 }
2427 chunk->set_preempted(false);
2428 retry_fast_path = true;
2429 } else {
2430 relativize_chunk_concurrently(chunk);
2431 }
2432
2433 // On first thaw after freeze restore oops to the lockstack if any.
2434 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2435 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2436 thaw_lockstack(chunk);
2437 retry_fast_path = true;
2438 }
2439
2440 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2441 // and FLAG_PREEMPTED flags from the stackChunk.
2442 if (retry_fast_path && can_thaw_fast(chunk)) {
2443 intptr_t* sp = thaw_fast<true>(chunk);
2444 if (_preempted_case) {
2445 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2446 }
2447 return sp;
2448 }
2449
2450 LogTarget(Trace, continuations) lt;
2451 if (lt.develop_is_enabled()) {
2452 LogStream ls(lt);
2453 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2454 chunk->print_on(true, &ls);
2455 }
2456
2457 #if CONT_JFR
2458 EventContinuationThawSlow e;
2459 if (e.should_commit()) {
2460 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2461 e.commit();
2462 }
2463 #endif
2464
2465 DEBUG_ONLY(_frames = 0;)
2466 _align_size = 0;
2467 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2468
2469 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2470 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2471
2472 frame heap_frame = _stream.to_frame();
2473 if (lt.develop_is_enabled()) {
2474 LogStream ls(lt);
2475 ls.print_cr("top hframe before (thaw):");
2476 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2477 heap_frame.print_value_on(&ls);
2478 }
2479
2480 frame caller; // the thawed caller on the stack
2481 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2482 finish_thaw(caller); // caller is now the topmost thawed frame
2483 _cont.write();
2484
2485 assert(_cont.chunk_invariant(), "");
2486
2487 JVMTI_ONLY(invalidate_jvmti_stack(_thread));
2488
2489 _thread->set_cont_fastpath(_fastpath);
2490
2491 intptr_t* sp = caller.sp();
2492
2493 if (_preempted_case) {
2494 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2495 }
2496 return sp;
2497 }
2498
2499 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2500 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2501 assert(!_cont.is_empty(), "no more frames");
2502 assert(num_frames > 0, "");
2503 assert(!heap_frame.is_empty(), "");
2504
2505 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2506 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2507 } else if (!heap_frame.is_interpreted_frame()) {
2508 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2509 } else {
2510 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames, top_on_preempt_case);
2511 }
2512 }
2513
2514 template<typename FKind>
2515 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2516 assert(num_frames > 0, "");
2517
2518 DEBUG_ONLY(_frames++;)
2519
2520 int argsize = _stream.stack_argsize();
2521
2522 _stream.next(SmallRegisterMap::instance_no_args());
2523 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2524
2525 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk
2526 // as it makes detecting that situation and adjusting unextended_sp tricky
2527 if (num_frames == 1 && !_stream.is_done() && FKind::interpreted && _stream.is_compiled()) {
2528 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2529 num_frames++;
2530 }
2531
2532 if (num_frames == 1 || _stream.is_done()) { // end recursion
2533 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2534 return true; // bottom
2535 } else { // recurse
2536 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2537 return false;
2538 }
2539 }
2540
2541 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2542 stackChunkOop chunk = _cont.tail();
2543
2544 if (!_stream.is_done()) {
2545 assert(_stream.sp() >= chunk->sp_address(), "");
2546 chunk->set_sp(chunk->to_offset(_stream.sp()));
2547 chunk->set_pc(_stream.pc());
2548 } else {
2549 chunk->set_sp(chunk->bottom());
2550 chunk->set_pc(nullptr);
2551 }
2552 assert(_stream.is_done() == chunk->is_empty(), "");
2553
2554 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2555 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2556
2557 _cont.set_argsize(argsize);
2558 entry = new_entry_frame();
2559
2560 assert(entry.sp() == _cont.entrySP(), "");
2561 assert(Continuation::is_continuation_enterSpecial(entry), "");
2562 assert(_cont.is_entry_frame(entry), "");
2563 }
2564
2565 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2566 LogTarget(Trace, continuations) lt;
2567 if (lt.develop_is_enabled()) {
2568 LogStream ls(lt);
2569 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2570 assert(hf.is_heap_frame(), "should be");
2571 hf.print_value_on(&ls);
2572 }
2573 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2574 }
2575
2576 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2577 #ifdef ASSERT
2578 LogTarget(Trace, continuations) lt;
2579 if (lt.develop_is_enabled()) {
2580 LogStream ls(lt);
2581 ls.print_cr("thawed frame:");
2582 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2583 }
2584 #endif
2585 }
2586
2587 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) {
2588 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2589 if (bottom) {
2590 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2591 : StubRoutines::cont_returnBarrier());
2592 } else {
2593 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2594 // If the caller is not deoptimized, pc is unchanged.
2595 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc());
2596 }
2597
2598 patch_pd(f, caller);
2599
2600 if (f.is_interpreted_frame()) {
2601 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2602 }
2603
2604 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2605 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2606 }
2607
2608 void ThawBase::clear_bitmap_bits(address start, address end) {
2609 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2610 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2611
2612 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2613 // or they will keep objects that are otherwise unreachable alive.
2614
2615 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2616 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2617 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2618 // anyways and we assert that before returning.
2619 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2620 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2621 stackChunkOop chunk = _cont.tail();
2622 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2623 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2624 }
2625
2626 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2627 frame top(sp);
2628 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2629 DEBUG_ONLY(verify_frame_kind(top, preempt_kind);)
2630 NOT_PRODUCT(int64_t tid = _thread->monitor_owner_id();)
2631
2632 #if INCLUDE_JVMTI
2633 // Finish the VTMS transition.
2634 assert(_thread->is_in_VTMS_transition(), "must be");
2635 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2636 if (is_vthread) {
2637 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2638 jvmti_mount_end(_thread, _cont, top, preempt_kind);
2639 } else {
2640 _thread->set_is_in_VTMS_transition(false);
2641 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2642 }
2643 }
2644 #endif
2645
2646 if (fast_case) {
2647 // If we thawed in the slow path the runtime stub/native wrapper frame already
2648 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2649 // we copied the original fp at the time of freeze which now will have to be fixed.
2650 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2651 int fsize = top.cb()->frame_size();
2652 patch_pd(top, sp + fsize);
2653 }
2654
2655 if (preempt_kind == Continuation::object_wait) {
2656 // Check now if we need to throw IE exception.
2657 bool throw_ie = _thread->pending_interrupted_exception();
2658 if (throw_ie) {
2659 throw_interrupted_exception(_thread, top);
2660 _thread->set_pending_interrupted_exception(false);
2661 }
2662 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT" after preemption on Object.wait%s", tid, throw_ie ? "(throwing IE)" : "");
2663 } else if (preempt_kind == Continuation::monitorenter) {
2664 if (top.is_runtime_frame()) {
2665 // The continuation might now run on a different platform thread than the previous time so
2666 // we need to adjust the current thread saved in the stub frame before restoring registers.
2667 JavaThread** thread_addr = frame::saved_thread_address(top);
2668 if (thread_addr != nullptr) *thread_addr = _thread;
2669 }
2670 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT " after preemption on monitorenter", tid);
2671 } else {
2672 // We need to redo the original call into the VM. First though, we need
2673 // to exit the monitor we just acquired (except on preemption cancelled
2674 // case where it was already released).
2675 assert(preempt_kind == Continuation::object_locker, "");
2676 if (_monitor != nullptr) _monitor->exit(_thread);
2677 sp = redo_vmcall(_thread, top);
2678 }
2679 return sp;
2680 }
2681
2682 intptr_t* ThawBase::redo_vmcall(JavaThread* current, frame& top) {
2683 assert(!current->preempting(), "");
2684 NOT_PRODUCT(int64_t tid = current->monitor_owner_id();)
2685 intptr_t* sp = top.sp();
2686
2687 {
2688 HandleMarkCleaner hmc(current); // Cleanup so._conth Handle
2689 ContinuationWrapper::SafepointOp so(current, _cont);
2690 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2691
2692 Method* m = top.interpreter_frame_method();
2693 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
2694 Bytecodes::Code code = current_bytecode.code();
2695 log_develop_trace(continuations, preempt)("Redoing InterpreterRuntime::%s for " INT64_FORMAT, code == Bytecodes::Code::_new ? "_new" : "resolve_from_cache", tid);
2696
2697 // These InterpreterRuntime entry points use JRT_ENTRY which uses a HandleMarkCleaner.
2698 // Create a HandeMark to avoid destroying so._conth.
2699 HandleMark hm(current);
2700 if (code == Bytecodes::Code::_new) {
2701 InterpreterRuntime::_new(current, m->constants(), current_bytecode.get_index_u2(code));
2702 } else {
2703 InterpreterRuntime::resolve_from_cache(current, code);
2704 }
2705 }
2706
2707 if (current->preempting()) {
2708 // Preempted again so we just arrange to return to preempt stub to unmount.
2709 sp = push_preempt_adapter();
2710 current->set_preempt_alternate_return(nullptr);
2711 bool cancelled = current->preemption_cancelled();
2712 if (cancelled) {
2713 // Instead of calling thaw again from the preempt stub just unmount anyways with
2714 // state of YIELDING. This will give a chance for other vthreads to run while
2715 // minimizing repeated loops of "thaw->redo_vmcall->try_preempt->preemption_cancelled->thaw..."
2716 // in case of multiple vthreads contending for the same init_lock().
2717 current->set_preemption_cancelled(false);
2718 oop vthread = current->vthread();
2719 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
2720 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::YIELDING);
2721 }
2722 log_develop_trace(continuations, preempt)("Preempted " INT64_FORMAT " again%s", tid, cancelled ? "(preemption cancelled, setting state to YIELDING)" : "");
2723 } else {
2724 log_develop_trace(continuations, preempt)("Call succesful, resuming " INT64_FORMAT, tid);
2725 }
2726 return sp;
2727 }
2728
2729 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2730 HandleMarkCleaner hm(current); // Cleanup so._conth Handle
2731 ContinuationWrapper::SafepointOp so(current, _cont);
2732 // Since we might safepoint set the anchor so that the stack can be walked.
2733 set_anchor(current, top.sp());
2734 JRT_BLOCK
2735 THROW(vmSymbols::java_lang_InterruptedException());
2736 JRT_BLOCK_END
2737 clear_anchor(current);
2738 }
2739
2740 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top) {
2741 assert(hf.is_interpreted_frame(), "");
2742
2743 if (UNLIKELY(seen_by_gc())) {
2744 if (is_top && _process_args_at_top) {
2745 log_trace(continuations, tracking)("Processing arguments in recurse_thaw_interpreted_frame");
2746 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_with_args());
2747 } else {
2748 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2749 }
2750 }
2751
2752 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2753
2754 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2755
2756 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2757
2758 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2759
2760 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2761 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2762 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2763 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2764
2765 assert(hf.is_heap_frame(), "should be");
2766 assert(!f.is_heap_frame(), "should not be");
2767
2768 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2769 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2770
2771 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2772 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2773 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2774
2775 // Make sure the relativized locals is already set.
2776 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2777
2778 derelativize_interpreted_frame_metadata(hf, f);
2779 patch(f, caller, is_bottom_frame);
2780
2781 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2782 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2783
2784 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2785
2786 maybe_set_fastpath(f.sp());
2787
2788 Method* m = hf.interpreter_frame_method();
2789 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2790 const int locals = m->max_locals();
2791
2792 if (!is_bottom_frame) {
2793 // can only fix caller once this frame is thawed (due to callee saved regs)
2794 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2795 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2796 assert(hf.is_heap_frame(), "should be");
2797 address start = (address)(heap_frame_bottom - locals);
2798 address end = (address)heap_frame_bottom;
2799 clear_bitmap_bits(start, end);
2800 }
2801
2802 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2803 caller = f;
2804 }
2805
2806 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2807 assert(hf.is_compiled_frame(), "");
2808 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2809
2810 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2811 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2812 }
2813
2814 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2815
2816 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2817
2818 assert(caller.sp() == caller.unextended_sp(), "");
2819
2820 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2821 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2822 }
2823
2824 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2825 // yet laid out in the stack, and so the original_pc is not stored in it.
2826 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2827 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame);
2828 intptr_t* const stack_frame_top = f.sp();
2829 intptr_t* const heap_frame_top = hf.unextended_sp();
2830
2831 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2832 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize;
2833 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2834
2835 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2836 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2837 // copy metadata, except the metadata at the top of the (unextended) entry frame
2838 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2839
2840 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2841 // (we might have one padding word for alignment)
2842 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2843 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2844
2845 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2846
2847 patch(f, caller, is_bottom_frame);
2848
2849 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2850 assert(!f.is_deoptimized_frame(), "");
2851 if (hf.is_deoptimized_frame()) {
2852 maybe_set_fastpath(f.sp());
2853 } else if (_thread->is_interp_only_mode()
2854 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2855 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2856 // cannot rely on nmethod patching for deopt.
2857 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2858
2859 log_develop_trace(continuations)("Deoptimizing thawed frame");
2860 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2861
2862 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2863 assert(f.is_deoptimized_frame(), "");
2864 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2865 maybe_set_fastpath(f.sp());
2866 }
2867
2868 if (!is_bottom_frame) {
2869 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2870 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2871 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2872 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2873 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2874 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2875 clear_bitmap_bits(start, start + argsize_in_bytes);
2876 }
2877
2878 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2879 caller = f;
2880 }
2881
2882 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2883 DEBUG_ONLY(_frames++;)
2884
2885 if (UNLIKELY(seen_by_gc())) {
2886 // Process the stub's caller here since we might need the full map.
2887 RegisterMap map(nullptr,
2888 RegisterMap::UpdateMap::include,
2889 RegisterMap::ProcessFrames::skip,
2890 RegisterMap::WalkContinuation::skip);
2891 map.set_include_argument_oops(false);
2892 _stream.next(&map);
2893 assert(!_stream.is_done(), "");
2894 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2895 } else {
2896 _stream.next(SmallRegisterMap::instance_no_args());
2897 assert(!_stream.is_done(), "");
2898 }
2899
2900 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2901
2902 assert(caller.is_compiled_frame(), "");
2903 assert(caller.sp() == caller.unextended_sp(), "");
2904
2905 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2906
2907 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2908 intptr_t* stack_frame_top = f.sp();
2909 intptr_t* heap_frame_top = hf.sp();
2910 int fsize = ContinuationHelper::StubFrame::size(hf);
2911
2912 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2913 fsize + frame::metadata_words);
2914
2915 patch(f, caller, false /*is_bottom_frame*/);
2916
2917 // can only fix caller once this frame is thawed (due to callee saved regs)
2918 RegisterMap map(nullptr,
2919 RegisterMap::UpdateMap::include,
2920 RegisterMap::ProcessFrames::skip,
2921 RegisterMap::WalkContinuation::skip);
2922 map.set_include_argument_oops(false);
2923 f.oop_map()->update_register_map(&f, &map);
2924 ContinuationHelper::update_register_map_with_callee(caller, &map);
2925 _cont.tail()->fix_thawed_frame(caller, &map);
2926
2927 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2928 caller = f;
2929 }
2930
2931 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2932 assert(hf.is_native_frame(), "");
2933 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2934
2935 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2936 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2937 }
2938
2939 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
2940 assert(!is_bottom_frame, "");
2941
2942 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2943
2944 assert(caller.sp() == caller.unextended_sp(), "");
2945
2946 if (caller.is_interpreted_frame()) {
2947 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
2948 }
2949
2950 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2951 // yet laid out in the stack, and so the original_pc is not stored in it.
2952 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2953 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
2954 intptr_t* const stack_frame_top = f.sp();
2955 intptr_t* const heap_frame_top = hf.unextended_sp();
2956
2957 int fsize = ContinuationHelper::NativeFrame::size(hf);
2958 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2959
2960 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2961 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2962 int sz = fsize + frame::metadata_words_at_bottom;
2963
2964 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2965
2966 patch(f, caller, false /* bottom */);
2967
2968 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2969 assert(!f.is_deoptimized_frame(), "");
2970 assert(!hf.is_deoptimized_frame(), "");
2971 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
2972
2973 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2974 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2975
2976 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
2977 caller = f;
2978 }
2979
2980 void ThawBase::finish_thaw(frame& f) {
2981 stackChunkOop chunk = _cont.tail();
2982
2983 if (chunk->is_empty()) {
2984 // Only remove chunk from list if it can't be reused for another freeze
2985 if (seen_by_gc()) {
2986 _cont.set_tail(chunk->parent());
2987 } else {
2988 chunk->set_has_mixed_frames(false);
2989 }
2990 chunk->set_max_thawing_size(0);
2991 } else {
2992 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
2993 }
2994 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
2995
2996 if (!is_aligned(f.sp(), frame::frame_alignment)) {
2997 assert(f.is_interpreted_frame(), "");
2998 f.set_sp(align_down(f.sp(), frame::frame_alignment));
2999 }
3000 push_return_frame(f);
3001 // can only fix caller after push_return_frame (due to callee saved regs)
3002 if (_process_args_at_top) {
3003 log_trace(continuations, tracking)("Processing arguments in finish_thaw");
3004 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_with_args());
3005 } else {
3006 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_no_args());
3007 }
3008
3009 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
3010
3011 log_develop_trace(continuations)("thawed %d frames", _frames);
3012
3013 LogTarget(Trace, continuations) lt;
3014 if (lt.develop_is_enabled()) {
3015 LogStream ls(lt);
3016 ls.print_cr("top hframe after (thaw):");
3017 _cont.last_frame().print_value_on(&ls);
3018 }
3019 }
3020
3021 void ThawBase::push_return_frame(frame& f) { // see generate_cont_thaw
3022 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
3023 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
3024
3025 LogTarget(Trace, continuations) lt;
3026 if (lt.develop_is_enabled()) {
3027 LogStream ls(lt);
3028 ls.print_cr("push_return_frame");
3029 f.print_value_on(&ls);
3030 }
3031
3032 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
3033 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
3034 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
3035 ContinuationHelper::push_pd(f);
3036
3037 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
3038 }
3039
3040 // returns new top sp
3041 // called after preparations (stack overflow check and making room)
3042 template<typename ConfigT>
3043 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
3044 assert(thread == JavaThread::current(), "Must be current thread");
3045
3046 CONT_JFR_ONLY(EventContinuationThaw event;)
3047
3048 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
3049
3050 ContinuationEntry* entry = thread->last_continuation();
3051 assert(entry != nullptr, "");
3052 oop oopCont = entry->cont_oop(thread);
3053
3054 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
3055 assert(oopCont == get_continuation(thread), "");
3056 verify_continuation(oopCont);
3057
3058 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
3059
3060 ContinuationWrapper cont(thread, oopCont);
3061 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
3062
3063 #ifdef ASSERT
3064 set_anchor_to_entry(thread, cont.entry());
3065 log_frames(thread);
3066 clear_anchor(thread);
3067 #endif
3068
3069 Thaw<ConfigT> thw(thread, cont);
3070 intptr_t* const sp = thw.thaw(kind);
3071 assert(is_aligned(sp, frame::frame_alignment), "");
3072 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp);)
3073
3074 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
3075
3076 verify_continuation(cont.continuation());
3077 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
3078
3079 return sp;
3080 }
3081
3082 #ifdef ASSERT
3083 static void do_deopt_after_thaw(JavaThread* thread) {
3084 int i = 0;
3085 StackFrameStream fst(thread, true, false);
3086 fst.register_map()->set_include_argument_oops(false);
3087 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3088 for (; !fst.is_done(); fst.next()) {
3089 if (fst.current()->cb()->is_nmethod()) {
3090 nmethod* nm = fst.current()->cb()->as_nmethod();
3091 if (!nm->method()->is_continuation_native_intrinsic()) {
3092 nm->make_deoptimized();
3093 }
3094 }
3095 }
3096 }
3097
3098 class ThawVerifyOopsClosure: public OopClosure {
3099 intptr_t* _p;
3100 outputStream* _st;
3101 bool is_good_oop(oop o) {
3102 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
3103 }
3104 public:
3105 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
3106 intptr_t* p() { return _p; }
3107 void reset() { _p = nullptr; }
3108
3109 virtual void do_oop(oop* p) {
3110 oop o = *p;
3111 if (o == nullptr || is_good_oop(o)) {
3112 return;
3113 }
3114 _p = (intptr_t*)p;
3115 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
3116 }
3117 virtual void do_oop(narrowOop* p) {
3118 oop o = RawAccess<>::oop_load(p);
3119 if (o == nullptr || is_good_oop(o)) {
3120 return;
3121 }
3122 _p = (intptr_t*)p;
3123 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
3124 }
3125 };
3126
3127 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
3128 assert(thread->has_last_Java_frame(), "");
3129
3130 ResourceMark rm;
3131 ThawVerifyOopsClosure cl(st);
3132 NMethodToOopClosure cf(&cl, false);
3133
3134 StackFrameStream fst(thread, true, false);
3135 fst.register_map()->set_include_argument_oops(false);
3136 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3137 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
3138 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
3139 st->print_cr(">>> do_verify_after_thaw deopt");
3140 fst.current()->deoptimize(nullptr);
3141 fst.current()->print_on(st);
3142 }
3143
3144 fst.current()->oops_do(&cl, &cf, fst.register_map());
3145 if (cl.p() != nullptr) {
3146 frame fr = *fst.current();
3147 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
3148 fr.print_on(st);
3149 if (!fr.is_interpreted_frame()) {
3150 st->print_cr("size: %d argsize: %d",
3151 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
3152 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
3153 }
3154 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
3155 if (reg != nullptr) {
3156 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
3157 }
3158 cl.reset();
3159 DEBUG_ONLY(thread->print_frame_layout();)
3160 if (chunk != nullptr) {
3161 chunk->print_on(true, st);
3162 }
3163 return false;
3164 }
3165 }
3166 return true;
3167 }
3168
3169 static void log_frames(JavaThread* thread, bool dolog) {
3170 const static int show_entry_callers = 3;
3171 LogTarget(Trace, continuations, tracking) lt;
3172 if (!lt.develop_is_enabled() || !dolog) {
3173 return;
3174 }
3175 LogStream ls(lt);
3176
3177 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3178 if (!thread->has_last_Java_frame()) {
3179 ls.print_cr("NO ANCHOR!");
3180 }
3181
3182 RegisterMap map(thread,
3183 RegisterMap::UpdateMap::include,
3184 RegisterMap::ProcessFrames::include,
3185 RegisterMap::WalkContinuation::skip);
3186 map.set_include_argument_oops(false);
3187
3188 if (false) {
3189 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3190 f.print_on(&ls);
3191 }
3192 } else {
3193 map.set_skip_missing(true);
3194 ResetNoHandleMark rnhm;
3195 ResourceMark rm;
3196 HandleMark hm(Thread::current());
3197 FrameValues values;
3198
3199 int i = 0;
3200 int post_entry = -1;
3201 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3202 f.describe(values, i, &map, i == 0);
3203 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3204 post_entry++;
3205 if (post_entry >= show_entry_callers)
3206 break;
3207 }
3208 values.print_on(thread, &ls);
3209 }
3210
3211 ls.print_cr("======= end frames =========");
3212 }
3213
3214 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp) {
3215 intptr_t* sp0 = sp;
3216 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
3217
3218 bool preempted = false;
3219 stackChunkOop tail = cont.tail();
3220 if (tail != nullptr && tail->preempted()) {
3221 // Still preempted (monitor not acquired) so no frames were thawed.
3222 set_anchor(thread, cont.entrySP(), cont.entryPC());
3223 preempted = true;
3224 } else {
3225 set_anchor(thread, sp0);
3226 }
3227
3228 log_frames(thread);
3229 if (LoomVerifyAfterThaw) {
3230 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
3231 }
3232 assert(ContinuationEntry::assert_entry_frame_laid_out(thread, preempted), "");
3233 clear_anchor(thread);
3234
3235 LogTarget(Trace, continuations) lt;
3236 if (lt.develop_is_enabled()) {
3237 LogStream ls(lt);
3238 ls.print_cr("Jumping to frame (thaw):");
3239 frame(sp).print_value_on(&ls);
3240 }
3241 }
3242 #endif // ASSERT
3243
3244 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3245
3246 #ifdef ASSERT
3247 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3248 ResourceMark rm;
3249 FrameValues values;
3250 assert(f.get_cb() != nullptr, "");
3251 RegisterMap map(f.is_heap_frame() ?
3252 nullptr :
3253 JavaThread::current(),
3254 RegisterMap::UpdateMap::include,
3255 RegisterMap::ProcessFrames::skip,
3256 RegisterMap::WalkContinuation::skip);
3257 map.set_include_argument_oops(false);
3258 map.set_skip_missing(true);
3259 if (callee_complete) {
3260 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3261 }
3262 const_cast<frame&>(f).describe(values, 0, &map, true);
3263 values.print_on(static_cast<JavaThread*>(nullptr), st);
3264 }
3265 #endif
3266
3267 static address thaw_entry = nullptr;
3268 static address freeze_entry = nullptr;
3269 static address freeze_preempt_entry = nullptr;
3270
3271 address Continuation::thaw_entry() {
3272 return ::thaw_entry;
3273 }
3274
3275 address Continuation::freeze_entry() {
3276 return ::freeze_entry;
3277 }
3278
3279 address Continuation::freeze_preempt_entry() {
3280 return ::freeze_preempt_entry;
3281 }
3282
3283 class ConfigResolve {
3284 public:
3285 static void resolve() { resolve_compressed(); }
3286
3287 static void resolve_compressed() {
3288 UseCompressedOops ? resolve_gc<true>()
3289 : resolve_gc<false>();
3290 }
3291
3292 private:
3293 template <bool use_compressed>
3294 static void resolve_gc() {
3295 BarrierSet* bs = BarrierSet::barrier_set();
3296 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3297 switch (bs->kind()) {
3298 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3299 case BarrierSet::bs_name: { \
3300 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3301 } \
3302 break;
3303 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3304 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3305
3306 default:
3307 fatal("BarrierSet resolving not implemented");
3308 };
3309 }
3310
3311 template <bool use_compressed, typename BarrierSetT>
3312 static void resolve() {
3313 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3314
3315 freeze_entry = (address)freeze<SelectedConfigT>;
3316 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3317
3318 // If we wanted, we could templatize by kind and have three different thaw entries
3319 thaw_entry = (address)thaw<SelectedConfigT>;
3320 }
3321 };
3322
3323 void Continuation::init() {
3324 ConfigResolve::resolve();
3325 }