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