1 /*
2 * Copyright (c) 2018, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "classfile/javaClasses.inline.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "code/codeCache.inline.hpp"
28 #include "code/nmethod.inline.hpp"
29 #include "code/vmreg.inline.hpp"
30 #include "compiler/oopMap.inline.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/continuationGCSupport.inline.hpp"
33 #include "gc/shared/gc_globals.hpp"
34 #include "gc/shared/memAllocator.hpp"
35 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
36 #include "interpreter/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/objArrayOop.inline.hpp"
43 #include "oops/oopsHierarchy.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/sharedRuntime.hpp"
61 #include "runtime/smallRegisterMap.inline.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 #if INCLUDE_JFR
74 #include "jfr/jfr.inline.hpp"
75 #endif
76
77 #include <type_traits>
78
79 /*
80 * This file contains the implementation of continuation freezing (yield) and thawing (run).
81 *
82 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
83 * would likely call these operations many thousands of times per second second, on every core.
84 *
85 * Freeze might be called every time the application performs any I/O operation, every time it
86 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
87 * multiple times in each of those cases, as it is called by the return barrier, which may be
88 * invoked on method return.
89 *
90 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
91 * example, every effort is made to avoid Java-VM transitions as much as possible.
92 *
93 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
94 * and so frames simply copied, and the bottom-most one is patched.
95 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
96 * and absolute pointers, and barriers invoked.
97 */
98
99 /************************************************
100
101 Thread-stack layout on freeze/thaw.
102 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
103
104 +----------------------------+
105 | . |
106 | . |
107 | . |
108 | carrier frames |
109 | |
110 |----------------------------|
111 | |
112 | Continuation.run |
113 | |
114 |============================|
115 | enterSpecial frame |
116 | pc |
117 | rbp |
118 | ----- |
119 ^ | int argsize | = ContinuationEntry
120 | | oopDesc* cont |
121 | | oopDesc* chunk |
122 | | ContinuationEntry* parent |
123 | | ... |
124 | |============================| <------ JavaThread::_cont_entry = entry->sp()
125 | | ? alignment word ? |
126 | |----------------------------| <--\
127 | | | |
128 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
129 Address | | | | Caller is still in the chunk.
130 | |----------------------------| |
131 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
132 | | rbp | | isn't the last one in the continuation.
133 | | | |
134 | | frame | |
135 | | | |
136 +----------------------------| \__ Continuation frames to be frozen/thawed
137 | | /
138 | frame | |
139 | | |
140 |----------------------------| |
141 | | |
142 | frame | |
143 | | |
144 |----------------------------| <--/
145 | |
146 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
147 | | instead of a doYield stub
148 |============================| <- the sp passed to freeze
149 | |
150 | Native freeze/thaw frames |
151 | . |
152 | . |
153 | . |
154 +----------------------------+
155
156 ************************************************/
157
158 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing
159
160 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
161 #if CONT_JFR
162 #define CONT_JFR_ONLY(code) code
163 #else
164 #define CONT_JFR_ONLY(code)
165 #endif
166
167 // TODO: See AbstractAssembler::generate_stack_overflow_check,
168 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
169 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
170
171 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
172
173 // Used to just annotatate cold/hot branches
174 #define LIKELY(condition) (condition)
175 #define UNLIKELY(condition) (condition)
176
177 // debugging functions
178 #ifdef ASSERT
179 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
180
181 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
182
183 static void do_deopt_after_thaw(JavaThread* thread);
184 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
185 static void log_frames(JavaThread* thread);
186 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempted);
187 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
188
189 #define assert_pfl(p, ...) \
190 do { \
191 if (!(p)) { \
192 JavaThread* t = JavaThread::active(); \
193 if (t->has_last_Java_frame()) { \
194 tty->print_cr("assert(" #p ") failed:"); \
195 t->print_frame_layout(); \
196 } \
197 } \
198 vmassert(p, __VA_ARGS__); \
199 } while(0)
200
201 #else
202 static void verify_continuation(oop continuation) { }
203 #define assert_pfl(p, ...)
204 #endif
205
206 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
207 template<typename ConfigT, bool preempt> static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp);
208
209 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
210 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
211
212
213 // Entry point to freeze. Transitions are handled manually
214 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
215 template<typename ConfigT>
216 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
217 assert(sp == current->frame_anchor()->last_Java_sp(), "");
218
219 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
220 current->set_cont_fastpath(nullptr);
221 }
222
223 return checked_cast<int>(ConfigT::freeze(current, sp));
224 JRT_END
225
226 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
227 return prepare_thaw_internal(thread, return_barrier);
228 JRT_END
229
230 template<typename ConfigT>
231 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
232 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
233 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
234 // Also the preemption case with JVMTI events enabled might safepoint so
235 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
236 // JRT_ENTRY instead?
237 ResetNoHandleMark rnhm;
238 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
239
240 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
241 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
242 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
243 JRT_END
244
245 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
246 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
247 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
248 }
249 JVM_END
250
251 ///////////
252
253 enum class oop_kind { NARROW, WIDE };
254 template <oop_kind oops, typename BarrierSetT>
255 class Config {
256 public:
257 typedef Config<oops, BarrierSetT> SelfT;
258 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
259
260 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
261 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
262 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
263 return res;
264 }
265
266 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
267 return freeze_internal<SelfT, true>(thread, sp);
268 }
269
270 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
271 return thaw_internal<SelfT>(thread, kind);
272 }
273 };
274
275 #ifdef _WINDOWS
276 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
277 address new_sp = sp - size;
278 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
279
280 if (new_sp < watermark) {
281 size_t page_size = os::vm_page_size();
282 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
283 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
284 while (pages_to_touch-- > 0) {
285 last_touched_page -= page_size;
286 *last_touched_page = 0;
287 }
288 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
289 }
290 }
291 #endif
292
293 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
294 const size_t page_size = os::vm_page_size();
295 if (size > page_size) {
296 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
297 return false;
298 }
299 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
300 }
301 return true;
302 }
303
304 #ifdef ASSERT
305 static oop get_continuation(JavaThread* thread) {
306 assert(thread != nullptr, "");
307 assert(thread->threadObj() != nullptr, "");
308 return java_lang_Thread::continuation(thread->threadObj());
309 }
310 #endif // ASSERT
311
312 inline void clear_anchor(JavaThread* thread) {
313 thread->frame_anchor()->clear();
314 }
315
316 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
317 assert(pc != nullptr, "");
318
319 JavaFrameAnchor* anchor = thread->frame_anchor();
320 anchor->set_last_Java_sp(sp);
321 anchor->set_last_Java_pc(pc);
322 ContinuationHelper::set_anchor_pd(anchor, sp);
323
324 assert(thread->has_last_Java_frame(), "");
325 assert(thread->last_frame().cb() != nullptr, "");
326 }
327
328 static void set_anchor(JavaThread* thread, intptr_t* sp) {
329 address pc = ContinuationHelper::return_address_at(
330 sp - frame::sender_sp_ret_address_offset());
331 set_anchor(thread, sp, pc);
332 }
333
334 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
335 JavaFrameAnchor* anchor = thread->frame_anchor();
336 anchor->set_last_Java_sp(entry->entry_sp());
337 anchor->set_last_Java_pc(entry->entry_pc());
338 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
339
340 assert(thread->has_last_Java_frame(), "");
341 assert(thread->last_frame().cb() != nullptr, "");
342 }
343
344 #if CONT_JFR
345 class FreezeThawJfrInfo : public StackObj {
346 short _e_size;
347 short _e_num_interpreted_frames;
348 public:
349
350 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
351 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
352 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
353 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
354 };
355
356 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
357 if (e->should_commit()) {
358 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
359 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
360 e->set_continuationClass(continuation->klass());
361 e->set_interpretedFrames(_e_num_interpreted_frames);
362 e->set_size(_e_size);
363 e->commit();
364 }
365 }
366 #endif // CONT_JFR
367
368 /////////////// FREEZE ////
369
370 class FreezeBase : public StackObj {
371 protected:
372 JavaThread* const _thread;
373 ContinuationWrapper& _cont;
374 bool _barriers; // only set when we allocate a chunk
375
376 intptr_t* _bottom_address;
377
378 // Used for preemption only
379 const bool _preempt;
380 frame _last_frame;
381
382 // Used to support freezing with held monitors
383 int _monitors_in_lockstack;
384
385 int _freeze_size; // total size of all frames plus metadata in words.
386 int _total_align_size;
387
388 intptr_t* _cont_stack_top;
389 intptr_t* _cont_stack_bottom;
390
391 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
392
393 #ifdef ASSERT
394 intptr_t* _orig_chunk_sp;
395 int _fast_freeze_size;
396 bool _empty;
397 #endif
398
399 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
400
401 NOT_PRODUCT(int _frames;)
402 DEBUG_ONLY(intptr_t* _last_write;)
403
404 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
405
406 public:
407 NOINLINE freeze_result freeze_slow();
408 void freeze_fast_existing_chunk();
409
410 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
411 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
412
413 inline int size_if_fast_freeze_available();
414
415 inline frame& last_frame() { return _last_frame; }
416
417 #ifdef ASSERT
418 bool check_valid_fast_path();
419 #endif
420
421 protected:
422 inline void init_rest();
423 void throw_stack_overflow_on_humongous_chunk();
424
425 // fast path
426 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
427 inline void unwind_frames();
428 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
429
430 // slow path
431 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
432
433 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
434
435 private:
436 // slow path
437 frame freeze_start_frame();
438 frame freeze_start_frame_on_preempt();
439 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
440 inline frame freeze_start_frame_yield_stub();
441 template<typename FKind>
442 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
443 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
444 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
445 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
446 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
447 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
448 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
449 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
450 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
451 NOINLINE void finish_freeze(const frame& f, const frame& top);
452
453 void freeze_lockstack(stackChunkOop chunk);
454
455 inline bool stack_overflow();
456
457 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
458 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
459 template<typename FKind> static inline frame sender(const frame& f);
460 template<typename FKind> frame new_heap_frame(frame& f, frame& caller);
461 inline void set_top_frame_metadata_pd(const frame& hf);
462 inline void patch_pd(frame& callee, const frame& caller);
463 inline void patch_pd_unused(intptr_t* sp);
464 void adjust_interpreted_frame_unextended_sp(frame& f);
465 static inline void prepare_freeze_interpreted_top_frame(frame& f);
466 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
467
468 protected:
469 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
470 bool freeze_fast_new_chunk(stackChunkOop chunk);
471 };
472
473 template <typename ConfigT>
474 class Freeze : public FreezeBase {
475 private:
476 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
477
478 public:
479 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
480 : FreezeBase(thread, cont, frame_sp, preempt) {}
481
482 freeze_result try_freeze_fast();
483
484 protected:
485 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
486 };
487
488 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
489 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
490 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
491
492 assert(_thread != nullptr, "");
493 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
494
495 DEBUG_ONLY(_cont.entry()->verify_cookie();)
496
497 assert(!Interpreter::contains(_cont.entryPC()), "");
498
499 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
500 #ifdef _LP64
501 if (((intptr_t)_bottom_address & 0xf) != 0) {
502 _bottom_address--;
503 }
504 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
505 #endif
506
507 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
508 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
509 assert(_bottom_address != nullptr, "");
510 assert(_bottom_address <= _cont.entrySP(), "");
511 DEBUG_ONLY(_last_write = nullptr;)
512
513 assert(_cont.chunk_invariant(), "");
514 assert(!Interpreter::contains(_cont.entryPC()), "");
515 #if !defined(PPC64) || defined(ZERO)
516 static const int doYield_stub_frame_size = frame::metadata_words;
517 #else
518 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
519 #endif
520 // With preemption doYield() might not have been resolved yet
521 assert(_preempt || SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size, "");
522
523 if (preempt) {
524 _last_frame = _thread->last_frame();
525 }
526
527 // properties of the continuation on the stack; all sizes are in words
528 _cont_stack_top = frame_sp + (!preempt ? doYield_stub_frame_size : 0); // we don't freeze the doYield stub frame
529 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
530 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
531
532 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
533 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
534 assert(cont_size() > 0, "");
535
536 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
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 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
569 #ifdef ASSERT
570 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
571 if (watermark == nullptr) {
572 return;
573 }
574 ContinuationEntry* ce = thread->last_continuation();
575 RegisterMap map(thread,
576 RegisterMap::UpdateMap::include,
577 RegisterMap::ProcessFrames::skip,
578 RegisterMap::WalkContinuation::skip);
579 map.set_include_argument_oops(false);
580 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
581 watermark->assert_is_frame_safe(f);
582 }
583 #endif // ASSERT
584 }
585
586 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
587 void FreezeBase::unwind_frames() {
588 ContinuationEntry* entry = _cont.entry();
589 entry->flush_stack_processing(_thread);
590 assert_frames_in_continuation_are_safe(_thread);
591 JFR_ONLY(Jfr::check_and_process_sample_request(_thread);)
592 set_anchor_to_entry(_thread, entry);
593 }
594
595 template <typename ConfigT>
596 freeze_result Freeze<ConfigT>::try_freeze_fast() {
597 assert(_thread->thread_state() == _thread_in_vm, "");
598 assert(_thread->cont_fastpath(), "");
599
600 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
601 assert(_fast_freeze_size == 0, "");
602
603 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
604 if (freeze_fast_new_chunk(chunk)) {
605 return freeze_ok;
606 }
607 if (_thread->has_pending_exception()) {
608 return freeze_exception;
609 }
610
611 // TODO R REMOVE when deopt change is fixed
612 assert(!_thread->cont_fastpath() || _barriers, "");
613 log_develop_trace(continuations)("-- RETRYING SLOW --");
614 return freeze_slow();
615 }
616
617 // Returns size needed if the continuation fits, otherwise 0.
618 int FreezeBase::size_if_fast_freeze_available() {
619 stackChunkOop chunk = _cont.tail();
620 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
621 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
622 return 0;
623 }
624
625 int total_size_needed = cont_size();
626 const int chunk_sp = chunk->sp();
627
628 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
629 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
630 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
631 // although that would require changing stackChunkOopDesc::is_empty
632 if (!chunk->is_empty()) {
633 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
634 }
635
636 total_size_needed += _monitors_in_lockstack;
637
638 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
639 bool available = chunk_free_room >= total_size_needed;
640 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
641 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
642 return available ? total_size_needed : 0;
643 }
644
645 void FreezeBase::freeze_fast_existing_chunk() {
646 stackChunkOop chunk = _cont.tail();
647
648 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
649 assert(_fast_freeze_size > 0, "");
650
651 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
652 DEBUG_ONLY(_empty = false;)
653 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
654 #ifdef ASSERT
655 {
656 intptr_t* retaddr_slot = (chunk->sp_address()
657 - frame::sender_sp_ret_address_offset());
658 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
659 "unexpected saved return address");
660 }
661 #endif
662
663 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
664 // we overlap; we'll overwrite the chunk's top frame's callee arguments
665 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
666 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
667
668 // increase max_size by what we're freezing minus the overlap
669 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
670
671 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
672 assert(bottom_sp == _bottom_address, "");
673 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
674 // should have a return barrier (installed back when we thawed it).
675 #ifdef ASSERT
676 {
677 intptr_t* retaddr_slot = (bottom_sp
678 - frame::sender_sp_ret_address_offset());
679 assert(ContinuationHelper::return_address_at(retaddr_slot)
680 == StubRoutines::cont_returnBarrier(),
681 "should be the continuation return barrier");
682 }
683 #endif
684 // We copy the fp from the chunk back to the stack because it contains some caller data,
685 // including, possibly, an oop that might have gone stale since we thawed.
686 patch_stack_pd(bottom_sp, chunk->sp_address());
687 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
688
689 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
690 } else { // the chunk is empty
691 const int chunk_start_sp = chunk->stack_size();
692
693 DEBUG_ONLY(_empty = true;)
694 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
695
696 chunk->set_max_thawing_size(cont_size());
697 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
698 chunk->set_sp(chunk->bottom());
699
700 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
701 }
702 }
703
704 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
705 DEBUG_ONLY(_empty = true;)
706
707 // Install new chunk
708 _cont.set_tail(chunk);
709
710 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
711 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
712 return false;
713 }
714
715 chunk->set_max_thawing_size(cont_size());
716
717 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
718 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
719 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
720 assert(chunk_start_sp == chunk->stack_size(), "");
721
722 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
723
724 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
725
726 return true;
727 }
728
729 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
730 assert(chunk != nullptr, "");
731 assert(!chunk->has_mixed_frames(), "");
732 assert(!chunk->is_gc_mode(), "");
733 assert(!chunk->has_bitmap(), "");
734 assert(!chunk->requires_barriers(), "");
735 assert(chunk == _cont.tail(), "");
736
737 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
738 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
739 // will either see no continuation on the stack, or a consistent chunk.
740 unwind_frames();
741
742 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
743 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
744 assert(chunk_start_sp <= chunk->stack_size(), "");
745 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
746
747 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
748 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
749
750 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
751 #ifdef ASSERT
752 if (!_empty) {
753 intptr_t* retaddr_slot = (_orig_chunk_sp
754 - frame::sender_sp_ret_address_offset());
755 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
756 "unexpected saved return address");
757 }
758 #endif
759
760 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
761 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
762
763 int adjust = frame::metadata_words_at_bottom;
764 #if INCLUDE_ASAN && defined(AARCH64)
765 // Reading at offset frame::metadata_words_at_bottom from _cont_stack_top
766 // will accesss memory at the callee frame, which on preemption cases will
767 // be the VM native method being called. The Arm 64-bit ABI doesn't specify
768 // a location where the frame record (returnpc+fp) has to be stored within
769 // a stack frame, and GCC currently chooses to save it at the top of the
770 // frame (lowest address). ASan treats this memory access in the callee as
771 // an overflow access to one of the locals stored in that frame. For these
772 // preemption cases we don't need to read these words anyways so we avoid it.
773 if (_preempt) {
774 adjust = 0;
775 }
776 #endif
777 intptr_t* from = _cont_stack_top - adjust;
778 intptr_t* to = chunk_top - adjust;
779 copy_to_chunk(from, to, cont_size() + adjust);
780 // Because we're not patched yet, the chunk is now in a bad state
781
782 // patch return pc of the bottom-most frozen frame (now in the chunk)
783 // with the actual caller's return address
784 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
785 - _cont.argsize()
786 - frame::metadata_words_at_top
787 - frame::sender_sp_ret_address_offset());
788 #ifdef ASSERT
789 if (!_empty) {
790 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
791 == StubRoutines::cont_returnBarrier(),
792 "should be the continuation return barrier");
793 }
794 #endif
795 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
796 chunk->pc());
797
798 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
799 chunk->set_sp(chunk_new_sp);
800
801 // set chunk->pc to the return address of the topmost frame in the chunk
802 if (_preempt) {
803 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
804 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
805 // be the pc we used when creating the oopmap. Get the top's frame last pc from
806 // the anchor instead.
807 address last_pc = _last_frame.pc();
808 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
809 chunk->set_pc(last_pc);
810 // For stub/native frames the fp is not used while frozen, and will be constructed
811 // again when thawing the frame (see ThawBase::handle_preempted_continuation). We
812 // patch it with a special bad address to help with debugging, particularly when
813 // inspecting frames and identifying invalid accesses.
814 patch_pd_unused(chunk_top);
815 } else {
816 chunk->set_pc(ContinuationHelper::return_address_at(
817 _cont_stack_top - frame::sender_sp_ret_address_offset()));
818 }
819
820 if (_monitors_in_lockstack > 0) {
821 freeze_lockstack(chunk);
822 }
823
824 _cont.write();
825
826 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
827 LogTarget(Trace, continuations) lt;
828 if (lt.develop_is_enabled()) {
829 LogStream ls(lt);
830 chunk->print_on(true, &ls);
831 }
832
833 // Verification
834 assert(_cont.chunk_invariant(), "");
835 chunk->verify();
836
837 #if CONT_JFR
838 EventContinuationFreezeFast e;
839 if (e.should_commit()) {
840 e.set_id(cast_from_oop<u8>(chunk));
841 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
842 e.set_size(cont_size() << LogBytesPerWord);
843 e.commit();
844 }
845 #endif
846 }
847
848 NOINLINE freeze_result FreezeBase::freeze_slow() {
849 #ifdef ASSERT
850 ResourceMark rm;
851 #endif
852
853 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
854 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
855
856 #if CONT_JFR
857 EventContinuationFreezeSlow e;
858 if (e.should_commit()) {
859 e.set_id(cast_from_oop<u8>(_cont.continuation()));
860 e.commit();
861 }
862 #endif
863
864 init_rest();
865
866 HandleMark hm(Thread::current());
867
868 frame f = freeze_start_frame();
869
870 LogTarget(Debug, continuations) lt;
871 if (lt.develop_is_enabled()) {
872 LogStream ls(lt);
873 f.print_on(&ls);
874 }
875
876 frame caller; // the frozen caller in the chunk
877 freeze_result res = recurse_freeze(f, caller, 0, false, true);
878
879 if (res == freeze_ok) {
880 finish_freeze(f, caller);
881 _cont.write();
882 }
883
884 return res;
885 }
886
887 frame FreezeBase::freeze_start_frame() {
888 if (LIKELY(!_preempt)) {
889 return freeze_start_frame_yield_stub();
890 } else {
891 return freeze_start_frame_on_preempt();
892 }
893 }
894
895 frame FreezeBase::freeze_start_frame_yield_stub() {
896 frame f = _thread->last_frame();
897 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be");
898 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
899 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
900 return f;
901 }
902
903 frame FreezeBase::freeze_start_frame_on_preempt() {
904 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
905 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
906 return _last_frame;
907 }
908
909 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
910 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
911 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
912 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
913 || ((top && _preempt) == f.is_native_frame()), "");
914
915 if (stack_overflow()) {
916 return freeze_exception;
917 }
918
919 if (f.is_compiled_frame()) {
920 if (UNLIKELY(f.oop_map() == nullptr)) {
921 // special native frame
922 return freeze_pinned_native;
923 }
924 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
925 } else if (f.is_interpreted_frame()) {
926 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
927 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
928 } else if (top && _preempt) {
929 assert(f.is_native_frame() || f.is_runtime_frame(), "");
930 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
931 } else {
932 // Frame can't be frozen. Most likely the call_stub or upcall_stub
933 // which indicates there are further natives frames up the stack.
934 return freeze_pinned_native;
935 }
936 }
937
938 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
939 // See also StackChunkFrameStream<frame_kind>::frame_size()
940 template<typename FKind>
941 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
942 assert(FKind::is_instance(f), "");
943
944 assert(fsize > 0, "");
945 assert(argsize >= 0, "");
946 _freeze_size += fsize;
947 NOT_PRODUCT(_frames++;)
948
949 assert(FKind::frame_bottom(f) <= _bottom_address, "");
950
951 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
952 // enterSpecial and an interpreted frame
953 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
954 return finalize_freeze(f, caller, argsize); // recursion end
955 } else {
956 frame senderf = sender<FKind>(f);
957 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
958 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
959 return result;
960 }
961 }
962
963 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
964 LogTarget(Trace, continuations) lt;
965 if (lt.develop_is_enabled()) {
966 LogStream ls(lt);
967 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
968 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
969 f.print_value_on(&ls);
970 }
971 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
972 }
973
974 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
975 LogTarget(Trace, continuations) lt;
976 if (lt.develop_is_enabled()) {
977 LogStream ls(lt);
978 DEBUG_ONLY(hf.print_value_on(&ls);)
979 assert(hf.is_heap_frame(), "should be");
980 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
981 if (is_bottom_frame) {
982 ls.print_cr("bottom h-frame:");
983 hf.print_on(&ls);
984 }
985 }
986 }
987
988 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
989 // See also StackChunkFrameStream<frame_kind>::frame_size()
990 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
991 int argsize = argsize_md - frame::metadata_words_at_top;
992 assert(callee.is_interpreted_frame()
993 || ContinuationHelper::Frame::is_stub(callee.cb())
994 || callee.cb()->as_nmethod()->is_osr_method()
995 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
996 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
997 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
998
999 LogTarget(Trace, continuations) lt;
1000
1001 #ifdef ASSERT
1002 bool empty = _cont.is_empty();
1003 log_develop_trace(continuations)("empty: %d", empty);
1004 #endif
1005
1006 stackChunkOop chunk = _cont.tail();
1007
1008 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1009
1010 _freeze_size += frame::metadata_words; // for top frame's metadata
1011
1012 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1013 int unextended_sp = -1;
1014 if (chunk != nullptr) {
1015 if (!chunk->is_empty()) {
1016 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1017 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1018 bool top_interpreted = Interpreter::contains(chunk->pc());
1019 if (callee.is_interpreted_frame() == top_interpreted) {
1020 overlap = argsize_md;
1021 }
1022 } else {
1023 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1024 }
1025 }
1026
1027 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1028
1029 _freeze_size -= overlap;
1030 assert(_freeze_size >= 0, "");
1031
1032 assert(chunk == nullptr || chunk->is_empty()
1033 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1034 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1035
1036 _freeze_size += _monitors_in_lockstack;
1037
1038 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1039 bool allocated_old_in_freeze_fast = _barriers;
1040 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1041 "Chunk allocated in freeze_fast is of insufficient size "
1042 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1043 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1044
1045 DEBUG_ONLY(bool empty_chunk = true);
1046 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1047 // ALLOCATE NEW CHUNK
1048
1049 if (lt.develop_is_enabled()) {
1050 LogStream ls(lt);
1051 if (chunk == nullptr) {
1052 ls.print_cr("no chunk");
1053 } else {
1054 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1055 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1056 chunk->print_on(&ls);
1057 }
1058 }
1059
1060 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1061 // overlap = 0;
1062
1063 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1064 if (chunk == nullptr) {
1065 return freeze_exception;
1066 }
1067
1068 // Install new chunk
1069 _cont.set_tail(chunk);
1070 assert(chunk->is_empty(), "");
1071 } else {
1072 // REUSE EXISTING CHUNK
1073 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1074 if (chunk->is_empty()) {
1075 int sp = chunk->stack_size() - argsize_md;
1076 chunk->set_sp(sp);
1077 chunk->set_bottom(sp);
1078 _freeze_size += overlap;
1079 assert(chunk->max_thawing_size() == 0, "");
1080 } DEBUG_ONLY(else empty_chunk = false;)
1081 }
1082 assert(!chunk->is_gc_mode(), "");
1083 assert(!chunk->has_bitmap(), "");
1084 chunk->set_has_mixed_frames(true);
1085
1086 assert(chunk->requires_barriers() == _barriers, "");
1087 assert(!_barriers || chunk->is_empty(), "");
1088
1089 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1090 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1091
1092 if (_preempt) {
1093 frame f = _thread->last_frame();
1094 if (f.is_interpreted_frame()) {
1095 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1096 // We need it so that on resume we can restore the sp to the right place, since
1097 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1098 // We do it now that we know freezing will be successful.
1099 prepare_freeze_interpreted_top_frame(f);
1100 }
1101 }
1102
1103 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1104 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1105 // will either see no continuation or a consistent chunk.
1106 unwind_frames();
1107
1108 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1109
1110 if (lt.develop_is_enabled()) {
1111 LogStream ls(lt);
1112 ls.print_cr("top chunk:");
1113 chunk->print_on(&ls);
1114 }
1115
1116 if (_monitors_in_lockstack > 0) {
1117 freeze_lockstack(chunk);
1118 }
1119
1120 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1121 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1122
1123 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1124
1125 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1126 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1127 #ifdef ASSERT
1128 if (lt.develop_is_enabled()) {
1129 LogStream ls(lt);
1130 ls.print_cr("top hframe before (freeze):");
1131 assert(caller.is_heap_frame(), "should be");
1132 caller.print_on(&ls);
1133 }
1134
1135 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1136
1137 frame entry = sender(callee);
1138
1139 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1140 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1141 #endif
1142
1143 return freeze_ok_bottom;
1144 }
1145
1146 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1147 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1148 if (is_bottom_frame) {
1149 // If we're the bottom frame, we need to replace the return barrier with the real
1150 // caller's pc.
1151 address last_pc = caller.pc();
1152 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1153 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1154 } else {
1155 assert(!caller.is_empty(), "");
1156 }
1157
1158 patch_pd(hf, caller);
1159
1160 if (f.is_interpreted_frame()) {
1161 assert(hf.is_heap_frame(), "should be");
1162 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1163 }
1164
1165 #ifdef ASSERT
1166 if (hf.is_compiled_frame()) {
1167 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1168 log_develop_trace(continuations)("Freezing deoptimized frame");
1169 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1170 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1171 }
1172 }
1173 #endif
1174 }
1175
1176 #ifdef ASSERT
1177 static void verify_frame_top(const frame& f, intptr_t* top) {
1178 ResourceMark rm;
1179 InterpreterOopMap mask;
1180 f.interpreted_frame_oop_map(&mask);
1181 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1182 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1183 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1184 }
1185 #endif // ASSERT
1186
1187 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1188 // See also StackChunkFrameStream<frame_kind>::frame_size()
1189 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1190 int callee_argsize /* incl. metadata */,
1191 bool callee_interpreted) {
1192 adjust_interpreted_frame_unextended_sp(f);
1193
1194 // The frame's top never includes the stack arguments to the callee
1195 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1196 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1197 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1198
1199 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1200
1201 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1202 // including metadata between f and its args
1203 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1204
1205 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1206 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1207 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1208 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1209
1210 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1211 if (UNLIKELY(result > freeze_ok_bottom)) {
1212 return result;
1213 }
1214
1215 bool is_bottom_frame = result == freeze_ok_bottom;
1216 assert(!caller.is_empty() || is_bottom_frame, "");
1217
1218 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1219
1220 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1221 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1222
1223 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1224 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1225 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1226
1227 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1228 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1229 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1230 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1231
1232 relativize_interpreted_frame_metadata(f, hf);
1233
1234 patch(f, hf, caller, is_bottom_frame);
1235
1236 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1237 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1238 caller = hf;
1239
1240 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1241 frame_method->record_gc_epoch();
1242
1243 return freeze_ok;
1244 }
1245
1246 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1247 // See also StackChunkFrameStream<frame_kind>::frame_size()
1248 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1249 int callee_argsize /* incl. metadata */,
1250 bool callee_interpreted) {
1251 // The frame's top never includes the stack arguments to the callee
1252 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1253 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1254 // including metadata between f and its stackargs
1255 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1256 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1257
1258 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d",
1259 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1260 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1261 _freeze_size, fsize, argsize);
1262 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1263 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1264
1265 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1266 if (UNLIKELY(result > freeze_ok_bottom)) {
1267 return result;
1268 }
1269
1270 bool is_bottom_frame = result == freeze_ok_bottom;
1271 assert(!caller.is_empty() || is_bottom_frame, "");
1272
1273 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1274
1275 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller);
1276
1277 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1278
1279 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1280 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1281
1282 if (caller.is_interpreted_frame()) {
1283 // When thawing the frame we might need to add alignment (see Thaw::align)
1284 _total_align_size += frame::align_wiggle;
1285 }
1286
1287 patch(f, hf, caller, is_bottom_frame);
1288
1289 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1290
1291 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1292 caller = hf;
1293 return freeze_ok;
1294 }
1295
1296 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1297 DEBUG_ONLY(frame fsender = sender(f);)
1298 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1299
1300 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1301 const int fsize = f.cb()->frame_size();
1302
1303 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1304 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1305
1306 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1307 if (UNLIKELY(result > freeze_ok_bottom)) {
1308 return result;
1309 }
1310
1311 assert(result == freeze_ok, "should have caller");
1312 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1313
1314 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1315 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1316
1317 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1318
1319 patch(f, hf, caller, false /*is_bottom_frame*/);
1320
1321 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1322
1323 caller = hf;
1324 return freeze_ok;
1325 }
1326
1327 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1328 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1329 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1330 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1331 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1332 // caller into the native convention. This is so that we have a valid oopMap in case of
1333 // having to block in the slow path. But that would require freezing those registers too
1334 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1335 // given that this would be a rare case anyways just pin the vthread to the carrier.
1336 return freeze_pinned_native;
1337 }
1338
1339 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1340 // There are no stackargs but argsize must include the metadata
1341 const int argsize = frame::metadata_words_at_top;
1342 const int fsize = f.cb()->frame_size() + argsize;
1343
1344 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1345 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1346
1347 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1348 if (UNLIKELY(result > freeze_ok_bottom)) {
1349 return result;
1350 }
1351
1352 assert(result == freeze_ok, "should have caller frame");
1353 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1354
1355 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1356 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1357
1358 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1359
1360 if (caller.is_interpreted_frame()) {
1361 // When thawing the frame we might need to add alignment (see Thaw::align)
1362 _total_align_size += frame::align_wiggle;
1363 }
1364
1365 patch(f, hf, caller, false /* is_bottom_frame */);
1366
1367 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1368
1369 caller = hf;
1370 return freeze_ok;
1371 }
1372
1373 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1374 stackChunkOop chunk = _cont.tail();
1375
1376 LogTarget(Trace, continuations) lt;
1377 if (lt.develop_is_enabled()) {
1378 LogStream ls(lt);
1379 assert(top.is_heap_frame(), "should be");
1380 top.print_on(&ls);
1381 }
1382
1383 set_top_frame_metadata_pd(top);
1384
1385 chunk->set_sp(chunk->to_offset(top.sp()));
1386 chunk->set_pc(top.pc());
1387
1388 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1389
1390 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1391
1392 // At this point the chunk is consistent
1393
1394 if (UNLIKELY(_barriers)) {
1395 log_develop_trace(continuations)("do barriers on old chunk");
1396 // Serial and Parallel GC can allocate objects directly into the old generation.
1397 // Then we want to relativize the derived pointers eagerly so that
1398 // old chunks are all in GC mode.
1399 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1400 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1401 if (UseShenandoahGC) {
1402 _cont.tail()->relativize_derived_pointers_concurrently();
1403 } else {
1404 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1405 }
1406 // For objects in the old generation we must maintain the remembered set
1407 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1408 }
1409
1410 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1411 if (lt.develop_is_enabled()) {
1412 LogStream ls(lt);
1413 chunk->print_on(true, &ls);
1414 }
1415
1416 if (lt.develop_is_enabled()) {
1417 LogStream ls(lt);
1418 ls.print_cr("top hframe after (freeze):");
1419 assert(_cont.last_frame().is_heap_frame(), "should be");
1420 _cont.last_frame().print_on(&ls);
1421 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1422 }
1423
1424 assert(_cont.chunk_invariant(), "");
1425 }
1426
1427 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1428 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1429 assert(t == JavaThread::current(), "");
1430 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1431 if (!_preempt) {
1432 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1433 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1434 }
1435 return true;
1436 }
1437 return false;
1438 }
1439
1440 class StackChunkAllocator : public MemAllocator {
1441 const size_t _stack_size;
1442 int _argsize_md;
1443 ContinuationWrapper& _continuation_wrapper;
1444 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1445 mutable bool _took_slow_path;
1446
1447 // Does the minimal amount of initialization needed for a TLAB allocation.
1448 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1449 virtual oop initialize(HeapWord* mem) const override {
1450 assert(_stack_size > 0, "");
1451 assert(_stack_size <= max_jint, "");
1452 assert(_word_size > _stack_size, "");
1453
1454 // zero out fields (but not the stack)
1455 const size_t hs = oopDesc::header_size();
1456 if (oopDesc::has_klass_gap()) {
1457 oopDesc::set_klass_gap(mem, 0);
1458 }
1459 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1460
1461 int bottom = (int)_stack_size - _argsize_md;
1462
1463 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1464 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1465 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1466
1467 return finish(mem);
1468 }
1469
1470 stackChunkOop allocate_fast() const {
1471 if (!UseTLAB) {
1472 return nullptr;
1473 }
1474
1475 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1476 if (mem == nullptr) {
1477 return nullptr;
1478 }
1479
1480 oop obj = initialize(mem);
1481 return stackChunkOopDesc::cast(obj);
1482 }
1483
1484 public:
1485 StackChunkAllocator(Klass* klass,
1486 size_t word_size,
1487 Thread* thread,
1488 size_t stack_size,
1489 int argsize_md,
1490 ContinuationWrapper& continuation_wrapper,
1491 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1492 : MemAllocator(klass, word_size, thread),
1493 _stack_size(stack_size),
1494 _argsize_md(argsize_md),
1495 _continuation_wrapper(continuation_wrapper),
1496 _jvmti_event_collector(jvmti_event_collector),
1497 _took_slow_path(false) {}
1498
1499 // Provides it's own, specialized allocation which skips instrumentation
1500 // if the memory can be allocated without going to a slow-path.
1501 stackChunkOop allocate() const {
1502 // First try to allocate without any slow-paths or instrumentation.
1503 stackChunkOop obj = allocate_fast();
1504 if (obj != nullptr) {
1505 return obj;
1506 }
1507
1508 // Now try full-blown allocation with all expensive operations,
1509 // including potentially safepoint operations.
1510 _took_slow_path = true;
1511
1512 // Protect unhandled Loom oops
1513 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1514
1515 // Can safepoint
1516 _jvmti_event_collector->start();
1517
1518 // Can safepoint
1519 return stackChunkOopDesc::cast(MemAllocator::allocate());
1520 }
1521
1522 bool took_slow_path() const {
1523 return _took_slow_path;
1524 }
1525 };
1526
1527 template <typename ConfigT>
1528 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1529 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1530
1531 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1532 size_t size_in_words = klass->instance_size(stack_size);
1533
1534 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1535 if (!_preempt) {
1536 throw_stack_overflow_on_humongous_chunk();
1537 }
1538 return nullptr;
1539 }
1540
1541 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1542 assert(current == JavaThread::current(), "should be current");
1543
1544 // Allocate the chunk.
1545 //
1546 // This might safepoint while allocating, but all safepointing due to
1547 // instrumentation have been deferred. This property is important for
1548 // some GCs, as this ensures that the allocated object is in the young
1549 // generation / newly allocated memory.
1550 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1551 stackChunkOop chunk = allocator.allocate();
1552
1553 if (chunk == nullptr) {
1554 return nullptr; // OOME
1555 }
1556
1557 // assert that chunk is properly initialized
1558 assert(chunk->stack_size() == (int)stack_size, "");
1559 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1560 assert(chunk->sp() == chunk->bottom(), "");
1561 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1562 assert(chunk->max_thawing_size() == 0, "");
1563 assert(chunk->pc() == nullptr, "");
1564 assert(chunk->is_empty(), "");
1565 assert(chunk->flags() == 0, "");
1566 assert(chunk->is_gc_mode() == false, "");
1567 assert(chunk->lockstack_size() == 0, "");
1568
1569 // fields are uninitialized
1570 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1571 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1572
1573 #if INCLUDE_ZGC
1574 if (UseZGC) {
1575 ZStackChunkGCData::initialize(chunk);
1576 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1577 _barriers = false;
1578 } else
1579 #endif
1580 #if INCLUDE_SHENANDOAHGC
1581 if (UseShenandoahGC) {
1582 _barriers = chunk->requires_barriers();
1583 } else
1584 #endif
1585 {
1586 if (!allocator.took_slow_path()) {
1587 // Guaranteed to be in young gen / newly allocated memory
1588 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1589 _barriers = false;
1590 } else {
1591 // Some GCs could put direct allocations in old gen for slow-path
1592 // allocations; need to explicitly check if that was the case.
1593 _barriers = chunk->requires_barriers();
1594 }
1595 }
1596
1597 if (_barriers) {
1598 log_develop_trace(continuations)("allocation requires barriers");
1599 }
1600
1601 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1602
1603 return chunk;
1604 }
1605
1606 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1607 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1608 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1609 }
1610
1611 #if INCLUDE_JVMTI
1612 static int num_java_frames(ContinuationWrapper& cont) {
1613 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1614 int count = 0;
1615 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1616 count += chunk->num_java_frames();
1617 }
1618 return count;
1619 }
1620
1621 static void invalidate_jvmti_stack(JavaThread* thread) {
1622 JvmtiThreadState *state = thread->jvmti_thread_state();
1623 if (state != nullptr) {
1624 state->invalidate_cur_stack_depth();
1625 }
1626 }
1627
1628 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1629 if (!cont.entry()->is_virtual_thread() && JvmtiExport::has_frame_pops(thread)) {
1630 int num_frames = num_java_frames(cont);
1631
1632 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1633 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames);
1634 }
1635 invalidate_jvmti_stack(thread);
1636 }
1637
1638 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top) {
1639 assert(current->vthread() != nullptr, "must be");
1640
1641 HandleMarkCleaner hm(current);
1642 Handle vth(current, current->vthread());
1643
1644 ContinuationWrapper::SafepointOp so(current, cont);
1645
1646 // Since we might safepoint set the anchor so that the stack can be walked.
1647 set_anchor(current, top.sp());
1648
1649 JRT_BLOCK
1650 JvmtiVTMSTransitionDisabler::VTMS_vthread_mount((jthread)vth.raw_value(), false);
1651
1652 if (current->pending_contended_entered_event()) {
1653 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1654 current->set_contended_entered_monitor(nullptr);
1655 }
1656 JRT_BLOCK_END
1657
1658 clear_anchor(current);
1659 }
1660 #endif // INCLUDE_JVMTI
1661
1662 #ifdef ASSERT
1663 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1664 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1665 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1666 bool FreezeBase::check_valid_fast_path() {
1667 ContinuationEntry* ce = _thread->last_continuation();
1668 RegisterMap map(_thread,
1669 RegisterMap::UpdateMap::skip,
1670 RegisterMap::ProcessFrames::skip,
1671 RegisterMap::WalkContinuation::skip);
1672 map.set_include_argument_oops(false);
1673 bool is_top_frame = true;
1674 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1675 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1676 return false;
1677 }
1678 }
1679 return true;
1680 }
1681 #endif // ASSERT
1682
1683 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1684 verify_continuation(cont.continuation());
1685 assert(!cont.is_empty(), "");
1686
1687 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1688 return freeze_ok;
1689 }
1690
1691 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1692 if (UNLIKELY(res != freeze_ok)) {
1693 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1694 verify_continuation(cont.continuation());
1695 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1696 return res;
1697 }
1698
1699 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1700 return freeze_epilog(cont);
1701 }
1702
1703 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1704 if (UNLIKELY(res != freeze_ok)) {
1705 verify_continuation(cont.continuation());
1706 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1707 return res;
1708 }
1709
1710 patch_return_pc_with_preempt_stub(old_last_frame);
1711 cont.tail()->set_preempted(true);
1712
1713 return freeze_epilog(cont);
1714 }
1715
1716 template<typename ConfigT, bool preempt>
1717 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1718 assert(!current->has_pending_exception(), "");
1719
1720 #ifdef ASSERT
1721 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1722 log_frames(current);
1723 #endif
1724
1725 CONT_JFR_ONLY(EventContinuationFreeze event;)
1726
1727 ContinuationEntry* entry = current->last_continuation();
1728
1729 oop oopCont = entry->cont_oop(current);
1730 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1731 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1732
1733 verify_continuation(oopCont);
1734 ContinuationWrapper cont(current, oopCont);
1735 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1736
1737 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1738
1739 if (entry->is_pinned()) {
1740 log_develop_debug(continuations)("PINNED due to critical section");
1741 verify_continuation(cont.continuation());
1742 const freeze_result res = freeze_pinned_cs;
1743 if (!preempt) {
1744 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1745 }
1746 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1747 // Avoid Thread.yield() loops without safepoint polls.
1748 if (SafepointMechanism::should_process(current) && !preempt) {
1749 cont.done(); // allow safepoint
1750 ThreadInVMfromJava tivmfj(current);
1751 }
1752 return res;
1753 }
1754
1755 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1756
1757 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1758 bool fast = UseContinuationFastPath && current->cont_fastpath();
1759 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1760 freeze.freeze_fast_existing_chunk();
1761 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1762 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1763 }
1764
1765 if (preempt) {
1766 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1767 freeze.set_jvmti_event_collector(&jsoaec);
1768
1769 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1770
1771 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1772 preempt_epilog(cont, res, freeze.last_frame());
1773 return res;
1774 }
1775
1776 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1777 assert(current == JavaThread::current(), "must be current thread");
1778 JRT_BLOCK
1779 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1780 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1781 freeze.set_jvmti_event_collector(&jsoaec);
1782
1783 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1784
1785 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1786 freeze_epilog(current, cont, res);
1787 cont.done(); // allow safepoint in the transition back to Java
1788 return res;
1789 JRT_BLOCK_END
1790 }
1791
1792 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1793 ContinuationEntry* entry = thread->last_continuation();
1794 if (entry == nullptr) {
1795 return freeze_ok;
1796 }
1797 if (entry->is_pinned()) {
1798 return freeze_pinned_cs;
1799 }
1800
1801 RegisterMap map(thread,
1802 RegisterMap::UpdateMap::include,
1803 RegisterMap::ProcessFrames::skip,
1804 RegisterMap::WalkContinuation::skip);
1805 map.set_include_argument_oops(false);
1806 frame f = thread->last_frame();
1807
1808 if (!safepoint) {
1809 f = f.sender(&map); // this is the yield frame
1810 } else { // safepoint yield
1811 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1812 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1813 #else
1814 Unimplemented();
1815 #endif
1816 if (!Interpreter::contains(f.pc())) {
1817 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1818 assert(f.oop_map() != nullptr, "must be");
1819 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1820 }
1821 }
1822
1823 while (true) {
1824 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1825 return freeze_pinned_native;
1826 }
1827
1828 f = f.sender(&map);
1829 if (!Continuation::is_frame_in_continuation(entry, f)) {
1830 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
1831 if (scope == cont_scope) {
1832 break;
1833 }
1834 entry = entry->parent();
1835 if (entry == nullptr) {
1836 break;
1837 }
1838 if (entry->is_pinned()) {
1839 return freeze_pinned_cs;
1840 }
1841 }
1842 }
1843 return freeze_ok;
1844 }
1845
1846 /////////////// THAW ////
1847
1848 static int thaw_size(stackChunkOop chunk) {
1849 int size = chunk->max_thawing_size();
1850 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
1851 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
1852 return size;
1853 }
1854
1855 // make room on the stack for thaw
1856 // returns the size in bytes, or 0 on failure
1857 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
1858 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
1859
1860 assert(thread == JavaThread::current(), "");
1861
1862 ContinuationEntry* ce = thread->last_continuation();
1863 assert(ce != nullptr, "");
1864 oop continuation = ce->cont_oop(thread);
1865 assert(continuation == get_continuation(thread), "");
1866 verify_continuation(continuation);
1867
1868 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
1869 assert(chunk != nullptr, "");
1870
1871 // The tail can be empty because it might still be available for another freeze.
1872 // However, here we want to thaw, so we get rid of it (it will be GCed).
1873 if (UNLIKELY(chunk->is_empty())) {
1874 chunk = chunk->parent();
1875 assert(chunk != nullptr, "");
1876 assert(!chunk->is_empty(), "");
1877 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
1878 }
1879
1880 // Verification
1881 chunk->verify();
1882 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
1883
1884 // Only make space for the last chunk because we only thaw from the last chunk
1885 int size = thaw_size(chunk) << LogBytesPerWord;
1886
1887 const address bottom = (address)thread->last_continuation()->entry_sp();
1888 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
1889 // for the Java frames in the check below.
1890 if (!stack_overflow_check(thread, size + 300, bottom)) {
1891 return 0;
1892 }
1893
1894 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
1895 p2i(bottom), p2i(bottom - size), size);
1896 return size;
1897 }
1898
1899 class ThawBase : public StackObj {
1900 protected:
1901 JavaThread* _thread;
1902 ContinuationWrapper& _cont;
1903 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
1904
1905 intptr_t* _fastpath;
1906 bool _barriers;
1907 bool _preempted_case;
1908 intptr_t* _top_unextended_sp_before_thaw;
1909 int _align_size;
1910 DEBUG_ONLY(intptr_t* _top_stack_address);
1911
1912 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
1913
1914 NOT_PRODUCT(int _frames;)
1915
1916 protected:
1917 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
1918 _thread(thread), _cont(cont),
1919 _fastpath(nullptr) {
1920 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
1921 assert (cont.tail() != nullptr, "no last chunk");
1922 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
1923 }
1924
1925 void clear_chunk(stackChunkOop chunk);
1926 template<bool check_stub>
1927 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
1928 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
1929
1930 void thaw_lockstack(stackChunkOop chunk);
1931
1932 // fast path
1933 inline void prefetch_chunk_pd(void* start, int size_words);
1934 void patch_return(intptr_t* sp, bool is_last);
1935
1936 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
1937 inline intptr_t* push_cleanup_continuation();
1938 void throw_interrupted_exception(JavaThread* current, frame& top);
1939
1940 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
1941 void finish_thaw(frame& f);
1942
1943 private:
1944 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
1945 void finalize_thaw(frame& entry, int argsize);
1946
1947 inline bool seen_by_gc();
1948
1949 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
1950 inline void after_thaw_java_frame(const frame& f, bool bottom);
1951 inline void patch(frame& f, const frame& caller, bool bottom);
1952 void clear_bitmap_bits(address start, address end);
1953
1954 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames);
1955 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
1956 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
1957 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
1958
1959 void push_return_frame(frame& f);
1960 inline frame new_entry_frame();
1961 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom);
1962 inline void patch_pd(frame& f, const frame& sender);
1963 inline void patch_pd(frame& f, intptr_t* caller_sp);
1964 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
1965
1966 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
1967
1968 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
1969
1970 public:
1971 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
1972 };
1973
1974 template <typename ConfigT>
1975 class Thaw : public ThawBase {
1976 public:
1977 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
1978
1979 inline bool can_thaw_fast(stackChunkOop chunk) {
1980 return !_barriers
1981 && _thread->cont_fastpath_thread_state()
1982 && !chunk->has_thaw_slowpath_condition()
1983 && !PreserveFramePointer;
1984 }
1985
1986 inline intptr_t* thaw(Continuation::thaw_kind kind);
1987 template<bool check_stub = false>
1988 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
1989 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
1990 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
1991 };
1992
1993 template <typename ConfigT>
1994 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
1995 verify_continuation(_cont.continuation());
1996 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
1997 assert(!_cont.is_empty(), "");
1998
1999 stackChunkOop chunk = _cont.tail();
2000 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2001 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2002
2003 _barriers = chunk->requires_barriers();
2004 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2005 : thaw_slow(chunk, kind);
2006 }
2007
2008 class ReconstructedStack : public StackObj {
2009 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2010 int _thaw_size;
2011 int _argsize;
2012 public:
2013 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2014 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2015 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2016 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2017 // We're at most one alignment word away from entrySP
2018 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2019 }
2020
2021 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2022
2023 // top and bottom stack pointers
2024 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2025 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2026
2027 // several operations operate on the totality of the stack being reconstructed,
2028 // including the metadata words
2029 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2030 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2031 };
2032
2033 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2034 chunk->set_sp(chunk->bottom());
2035 chunk->set_max_thawing_size(0);
2036 }
2037
2038 template<bool check_stub>
2039 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2040 bool empty = false;
2041 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2042 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2043 assert(chunk_sp == f.sp(), "");
2044 assert(chunk_sp == f.unextended_sp(), "");
2045
2046 int frame_size = f.cb()->frame_size();
2047 argsize = f.stack_argsize();
2048
2049 assert(!f.is_stub() || check_stub, "");
2050 if (check_stub && f.is_stub()) {
2051 // If we don't thaw the top compiled frame too, after restoring the saved
2052 // registers back in Java, we would hit the return barrier to thaw one more
2053 // frame effectively overwriting the restored registers during that call.
2054 f.next(SmallRegisterMap::instance(), true /* stop */);
2055 assert(!f.is_done(), "");
2056
2057 f.get_cb();
2058 assert(f.is_compiled(), "");
2059 frame_size += f.cb()->frame_size();
2060 argsize = f.stack_argsize();
2061
2062 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2063 // The caller of the runtime stub when the continuation is preempted is not at a
2064 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2065 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2066 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2067 }
2068 }
2069
2070 f.next(SmallRegisterMap::instance(), true /* stop */);
2071 empty = f.is_done();
2072 assert(!empty || argsize == chunk->argsize(), "");
2073
2074 if (empty) {
2075 clear_chunk(chunk);
2076 } else {
2077 chunk->set_sp(chunk->sp() + frame_size);
2078 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2079 // We set chunk->pc to the return pc into the next frame
2080 chunk->set_pc(f.pc());
2081 #ifdef ASSERT
2082 {
2083 intptr_t* retaddr_slot = (chunk_sp
2084 + frame_size
2085 - frame::sender_sp_ret_address_offset());
2086 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2087 "unexpected pc");
2088 }
2089 #endif
2090 }
2091 assert(empty == chunk->is_empty(), "");
2092 // returns the size required to store the frame on stack, and because it is a
2093 // compiled frame, it must include a copy of the arguments passed by the caller
2094 return frame_size + argsize + frame::metadata_words_at_top;
2095 }
2096
2097 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2098 int lockStackSize = chunk->lockstack_size();
2099 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2100
2101 oop tmp_lockstack[LockStack::CAPACITY];
2102 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2103 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2104
2105 chunk->set_lockstack_size(0);
2106 chunk->set_has_lockstack(false);
2107 }
2108
2109 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2110 assert(to >= _top_stack_address, "overwrote past thawing space"
2111 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2112 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2113 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2114 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2115 }
2116
2117 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2118 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2119
2120 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2121 ContinuationHelper::patch_return_address_at(
2122 sp - frame::sender_sp_ret_address_offset(),
2123 pc);
2124 }
2125
2126 template <typename ConfigT>
2127 template<bool check_stub>
2128 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2129 assert(chunk == _cont.tail(), "");
2130 assert(!chunk->has_mixed_frames(), "");
2131 assert(!chunk->requires_barriers(), "");
2132 assert(!chunk->has_bitmap(), "");
2133 assert(!_thread->is_interp_only_mode(), "");
2134
2135 LogTarget(Trace, continuations) lt;
2136 if (lt.develop_is_enabled()) {
2137 LogStream ls(lt);
2138 ls.print_cr("thaw_fast");
2139 chunk->print_on(true, &ls);
2140 }
2141
2142 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2143 static const int threshold = 500; // words
2144
2145 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2146 int argsize, thaw_size;
2147
2148 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2149
2150 bool partial, empty;
2151 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2152 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2153
2154 partial = false;
2155 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2156 clear_chunk(chunk);
2157 thaw_size = full_chunk_size;
2158 empty = true;
2159 } else { // thaw a single frame
2160 partial = true;
2161 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2162 empty = chunk->is_empty();
2163 }
2164
2165 // Are we thawing the last frame(s) in the continuation
2166 const bool is_last = empty && chunk->parent() == nullptr;
2167 assert(!is_last || argsize == 0, "");
2168
2169 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2170 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2171
2172 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2173
2174 // also copy metadata words at frame bottom
2175 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2176
2177 // update the ContinuationEntry
2178 _cont.set_argsize(argsize);
2179 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2180 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2181
2182 // install the return barrier if not last frame, or the entry's pc if last
2183 patch_return(rs.bottom_sp(), is_last);
2184
2185 // insert the back links from callee to caller frames
2186 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2187
2188 assert(is_last == _cont.is_empty(), "");
2189 assert(_cont.chunk_invariant(), "");
2190
2191 #if CONT_JFR
2192 EventContinuationThawFast e;
2193 if (e.should_commit()) {
2194 e.set_id(cast_from_oop<u8>(chunk));
2195 e.set_size(thaw_size << LogBytesPerWord);
2196 e.set_full(!partial);
2197 e.commit();
2198 }
2199 #endif
2200
2201 #ifdef ASSERT
2202 set_anchor(_thread, rs.sp());
2203 log_frames(_thread);
2204 if (LoomDeoptAfterThaw) {
2205 do_deopt_after_thaw(_thread);
2206 }
2207 clear_anchor(_thread);
2208 #endif
2209
2210 return rs.sp();
2211 }
2212
2213 inline bool ThawBase::seen_by_gc() {
2214 return _barriers || _cont.tail()->is_gc_mode();
2215 }
2216
2217 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2218 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2219 if (UseZGC || UseShenandoahGC) {
2220 chunk->relativize_derived_pointers_concurrently();
2221 }
2222 #endif
2223 }
2224
2225 template <typename ConfigT>
2226 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2227 Continuation::preempt_kind preempt_kind;
2228 bool retry_fast_path = false;
2229
2230 _preempted_case = chunk->preempted();
2231 if (_preempted_case) {
2232 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2233 if (waiter != nullptr) {
2234 // Mounted again after preemption. Resume the pending monitor operation,
2235 // which will be either a monitorenter or Object.wait() call.
2236 ObjectMonitor* mon = waiter->monitor();
2237 preempt_kind = waiter->is_wait() ? Continuation::freeze_on_wait : Continuation::freeze_on_monitorenter;
2238
2239 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2240 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2241 if (!mon_acquired) {
2242 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2243 return push_cleanup_continuation();
2244 }
2245 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2246 } else {
2247 // Preemption cancelled in moniterenter case. We actually acquired
2248 // the monitor after freezing all frames so nothing to do.
2249 preempt_kind = Continuation::freeze_on_monitorenter;
2250 }
2251 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2252 relativize_chunk_concurrently(chunk);
2253 chunk->set_preempted(false);
2254 retry_fast_path = true;
2255 } else {
2256 relativize_chunk_concurrently(chunk);
2257 }
2258
2259 // On first thaw after freeze restore oops to the lockstack if any.
2260 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2261 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2262 thaw_lockstack(chunk);
2263 retry_fast_path = true;
2264 }
2265
2266 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2267 // and FLAG_PREEMPTED flags from the stackChunk.
2268 if (retry_fast_path && can_thaw_fast(chunk)) {
2269 intptr_t* sp = thaw_fast<true>(chunk);
2270 if (_preempted_case) {
2271 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2272 }
2273 return sp;
2274 }
2275
2276 LogTarget(Trace, continuations) lt;
2277 if (lt.develop_is_enabled()) {
2278 LogStream ls(lt);
2279 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2280 chunk->print_on(true, &ls);
2281 }
2282
2283 #if CONT_JFR
2284 EventContinuationThawSlow e;
2285 if (e.should_commit()) {
2286 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2287 e.commit();
2288 }
2289 #endif
2290
2291 DEBUG_ONLY(_frames = 0;)
2292 _align_size = 0;
2293 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2294
2295 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2296 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2297
2298 frame heap_frame = _stream.to_frame();
2299 if (lt.develop_is_enabled()) {
2300 LogStream ls(lt);
2301 ls.print_cr("top hframe before (thaw):");
2302 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2303 heap_frame.print_value_on(&ls);
2304 }
2305
2306 frame caller; // the thawed caller on the stack
2307 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2308 finish_thaw(caller); // caller is now the topmost thawed frame
2309 _cont.write();
2310
2311 assert(_cont.chunk_invariant(), "");
2312
2313 JVMTI_ONLY(invalidate_jvmti_stack(_thread));
2314
2315 _thread->set_cont_fastpath(_fastpath);
2316
2317 intptr_t* sp = caller.sp();
2318
2319 if (_preempted_case) {
2320 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2321 }
2322 return sp;
2323 }
2324
2325 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2326 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2327 assert(!_cont.is_empty(), "no more frames");
2328 assert(num_frames > 0, "");
2329 assert(!heap_frame.is_empty(), "");
2330
2331 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2332 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2333 } else if (!heap_frame.is_interpreted_frame()) {
2334 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2335 } else {
2336 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames);
2337 }
2338 }
2339
2340 template<typename FKind>
2341 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2342 assert(num_frames > 0, "");
2343
2344 DEBUG_ONLY(_frames++;)
2345
2346 int argsize = _stream.stack_argsize();
2347
2348 _stream.next(SmallRegisterMap::instance());
2349 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2350
2351 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk
2352 // as it makes detecting that situation and adjusting unextended_sp tricky
2353 if (num_frames == 1 && !_stream.is_done() && FKind::interpreted && _stream.is_compiled()) {
2354 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2355 num_frames++;
2356 }
2357
2358 if (num_frames == 1 || _stream.is_done()) { // end recursion
2359 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2360 return true; // bottom
2361 } else { // recurse
2362 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2363 return false;
2364 }
2365 }
2366
2367 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2368 stackChunkOop chunk = _cont.tail();
2369
2370 if (!_stream.is_done()) {
2371 assert(_stream.sp() >= chunk->sp_address(), "");
2372 chunk->set_sp(chunk->to_offset(_stream.sp()));
2373 chunk->set_pc(_stream.pc());
2374 } else {
2375 chunk->set_sp(chunk->bottom());
2376 chunk->set_pc(nullptr);
2377 }
2378 assert(_stream.is_done() == chunk->is_empty(), "");
2379
2380 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2381 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2382
2383 _cont.set_argsize(argsize);
2384 entry = new_entry_frame();
2385
2386 assert(entry.sp() == _cont.entrySP(), "");
2387 assert(Continuation::is_continuation_enterSpecial(entry), "");
2388 assert(_cont.is_entry_frame(entry), "");
2389 }
2390
2391 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2392 LogTarget(Trace, continuations) lt;
2393 if (lt.develop_is_enabled()) {
2394 LogStream ls(lt);
2395 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2396 assert(hf.is_heap_frame(), "should be");
2397 hf.print_value_on(&ls);
2398 }
2399 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2400 }
2401
2402 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2403 #ifdef ASSERT
2404 LogTarget(Trace, continuations) lt;
2405 if (lt.develop_is_enabled()) {
2406 LogStream ls(lt);
2407 ls.print_cr("thawed frame:");
2408 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2409 }
2410 #endif
2411 }
2412
2413 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) {
2414 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2415 if (bottom) {
2416 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2417 : StubRoutines::cont_returnBarrier());
2418 } else {
2419 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2420 // If the caller is not deoptimized, pc is unchanged.
2421 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc());
2422 }
2423
2424 patch_pd(f, caller);
2425
2426 if (f.is_interpreted_frame()) {
2427 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2428 }
2429
2430 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2431 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2432 }
2433
2434 void ThawBase::clear_bitmap_bits(address start, address end) {
2435 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2436 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2437
2438 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2439 // or they will keep objects that are otherwise unreachable alive.
2440
2441 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2442 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2443 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2444 // anyways and we assert that before returning.
2445 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2446 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2447 stackChunkOop chunk = _cont.tail();
2448 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2449 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2450 }
2451
2452 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2453 assert(preempt_kind == Continuation::freeze_on_wait || preempt_kind == Continuation::freeze_on_monitorenter, "");
2454 frame top(sp);
2455 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2456
2457 #if INCLUDE_JVMTI
2458 // Finish the VTMS transition.
2459 assert(_thread->is_in_VTMS_transition(), "must be");
2460 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2461 if (is_vthread) {
2462 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2463 jvmti_mount_end(_thread, _cont, top);
2464 } else {
2465 _thread->set_is_in_VTMS_transition(false);
2466 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2467 }
2468 }
2469 #endif
2470
2471 if (fast_case) {
2472 // If we thawed in the slow path the runtime stub/native wrapper frame already
2473 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2474 // we copied the fp patched during freeze, which will now have to be fixed.
2475 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2476 int fsize = top.cb()->frame_size();
2477 patch_pd(top, sp + fsize);
2478 }
2479
2480 if (preempt_kind == Continuation::freeze_on_wait) {
2481 // Check now if we need to throw IE exception.
2482 if (_thread->pending_interrupted_exception()) {
2483 throw_interrupted_exception(_thread, top);
2484 _thread->set_pending_interrupted_exception(false);
2485 }
2486 } else if (top.is_runtime_frame()) {
2487 // The continuation might now run on a different platform thread than the previous time so
2488 // we need to adjust the current thread saved in the stub frame before restoring registers.
2489 JavaThread** thread_addr = frame::saved_thread_address(top);
2490 if (thread_addr != nullptr) *thread_addr = _thread;
2491 }
2492 return sp;
2493 }
2494
2495 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2496 ContinuationWrapper::SafepointOp so(current, _cont);
2497 // Since we might safepoint set the anchor so that the stack can be walked.
2498 set_anchor(current, top.sp());
2499 JRT_BLOCK
2500 THROW(vmSymbols::java_lang_InterruptedException());
2501 JRT_BLOCK_END
2502 clear_anchor(current);
2503 }
2504
2505 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames) {
2506 assert(hf.is_interpreted_frame(), "");
2507
2508 if (UNLIKELY(seen_by_gc())) {
2509 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance());
2510 }
2511
2512 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2513
2514 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2515
2516 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2517
2518 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2519
2520 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2521 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2522 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2523 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2524
2525 assert(hf.is_heap_frame(), "should be");
2526 assert(!f.is_heap_frame(), "should not be");
2527
2528 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2529 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2530
2531 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2532 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2533 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2534
2535 // Make sure the relativized locals is already set.
2536 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2537
2538 derelativize_interpreted_frame_metadata(hf, f);
2539 patch(f, caller, is_bottom_frame);
2540
2541 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2542 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2543
2544 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2545
2546 maybe_set_fastpath(f.sp());
2547
2548 Method* m = hf.interpreter_frame_method();
2549 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2550 const int locals = m->max_locals();
2551
2552 if (!is_bottom_frame) {
2553 // can only fix caller once this frame is thawed (due to callee saved regs)
2554 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance());
2555 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2556 assert(hf.is_heap_frame(), "should be");
2557 address start = (address)(heap_frame_bottom - locals);
2558 address end = (address)heap_frame_bottom;
2559 clear_bitmap_bits(start, end);
2560 }
2561
2562 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2563 caller = f;
2564 }
2565
2566 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2567 assert(hf.is_compiled_frame(), "");
2568 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2569
2570 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2571 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance());
2572 }
2573
2574 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2575
2576 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2577
2578 assert(caller.sp() == caller.unextended_sp(), "");
2579
2580 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2581 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2582 }
2583
2584 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2585 // yet laid out in the stack, and so the original_pc is not stored in it.
2586 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2587 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame);
2588 intptr_t* const stack_frame_top = f.sp();
2589 intptr_t* const heap_frame_top = hf.unextended_sp();
2590
2591 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2592 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize;
2593 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2594
2595 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2596 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2597 // copy metadata, except the metadata at the top of the (unextended) entry frame
2598 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2599
2600 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2601 // (we might have one padding word for alignment)
2602 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2603 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2604
2605 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2606
2607 patch(f, caller, is_bottom_frame);
2608
2609 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2610 assert(!f.is_deoptimized_frame(), "");
2611 if (hf.is_deoptimized_frame()) {
2612 maybe_set_fastpath(f.sp());
2613 } else if (_thread->is_interp_only_mode()
2614 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2615 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2616 // cannot rely on nmethod patching for deopt.
2617 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2618
2619 log_develop_trace(continuations)("Deoptimizing thawed frame");
2620 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2621
2622 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2623 assert(f.is_deoptimized_frame(), "");
2624 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2625 maybe_set_fastpath(f.sp());
2626 }
2627
2628 if (!is_bottom_frame) {
2629 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2630 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance());
2631 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2632 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2633 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2634 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2635 clear_bitmap_bits(start, start + argsize_in_bytes);
2636 }
2637
2638 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2639 caller = f;
2640 }
2641
2642 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2643 DEBUG_ONLY(_frames++;)
2644
2645 if (UNLIKELY(seen_by_gc())) {
2646 // Process the stub's caller here since we might need the full map.
2647 RegisterMap map(nullptr,
2648 RegisterMap::UpdateMap::include,
2649 RegisterMap::ProcessFrames::skip,
2650 RegisterMap::WalkContinuation::skip);
2651 map.set_include_argument_oops(false);
2652 _stream.next(&map);
2653 assert(!_stream.is_done(), "");
2654 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2655 } else {
2656 _stream.next(SmallRegisterMap::instance());
2657 assert(!_stream.is_done(), "");
2658 }
2659
2660 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2661
2662 assert(caller.is_compiled_frame(), "");
2663 assert(caller.sp() == caller.unextended_sp(), "");
2664
2665 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2666
2667 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2668 intptr_t* stack_frame_top = f.sp();
2669 intptr_t* heap_frame_top = hf.sp();
2670 int fsize = ContinuationHelper::StubFrame::size(hf);
2671
2672 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2673 fsize + frame::metadata_words);
2674
2675 patch(f, caller, false /*is_bottom_frame*/);
2676
2677 // can only fix caller once this frame is thawed (due to callee saved regs)
2678 RegisterMap map(nullptr,
2679 RegisterMap::UpdateMap::include,
2680 RegisterMap::ProcessFrames::skip,
2681 RegisterMap::WalkContinuation::skip);
2682 map.set_include_argument_oops(false);
2683 f.oop_map()->update_register_map(&f, &map);
2684 ContinuationHelper::update_register_map_with_callee(caller, &map);
2685 _cont.tail()->fix_thawed_frame(caller, &map);
2686
2687 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2688 caller = f;
2689 }
2690
2691 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2692 assert(hf.is_native_frame(), "");
2693 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2694
2695 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2696 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance());
2697 }
2698
2699 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
2700 assert(!is_bottom_frame, "");
2701
2702 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2703
2704 assert(caller.sp() == caller.unextended_sp(), "");
2705
2706 if (caller.is_interpreted_frame()) {
2707 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
2708 }
2709
2710 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2711 // yet laid out in the stack, and so the original_pc is not stored in it.
2712 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2713 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
2714 intptr_t* const stack_frame_top = f.sp();
2715 intptr_t* const heap_frame_top = hf.unextended_sp();
2716
2717 int fsize = ContinuationHelper::NativeFrame::size(hf);
2718 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2719
2720 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2721 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2722 int sz = fsize + frame::metadata_words_at_bottom;
2723
2724 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2725
2726 patch(f, caller, false /* bottom */);
2727
2728 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2729 assert(!f.is_deoptimized_frame(), "");
2730 assert(!hf.is_deoptimized_frame(), "");
2731 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
2732
2733 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2734 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance());
2735
2736 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
2737 caller = f;
2738 }
2739
2740 void ThawBase::finish_thaw(frame& f) {
2741 stackChunkOop chunk = _cont.tail();
2742
2743 if (chunk->is_empty()) {
2744 // Only remove chunk from list if it can't be reused for another freeze
2745 if (seen_by_gc()) {
2746 _cont.set_tail(chunk->parent());
2747 } else {
2748 chunk->set_has_mixed_frames(false);
2749 }
2750 chunk->set_max_thawing_size(0);
2751 } else {
2752 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
2753 }
2754 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
2755
2756 if (!is_aligned(f.sp(), frame::frame_alignment)) {
2757 assert(f.is_interpreted_frame(), "");
2758 f.set_sp(align_down(f.sp(), frame::frame_alignment));
2759 }
2760 push_return_frame(f);
2761 chunk->fix_thawed_frame(f, SmallRegisterMap::instance()); // can only fix caller after push_return_frame (due to callee saved regs)
2762
2763 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
2764
2765 log_develop_trace(continuations)("thawed %d frames", _frames);
2766
2767 LogTarget(Trace, continuations) lt;
2768 if (lt.develop_is_enabled()) {
2769 LogStream ls(lt);
2770 ls.print_cr("top hframe after (thaw):");
2771 _cont.last_frame().print_value_on(&ls);
2772 }
2773 }
2774
2775 void ThawBase::push_return_frame(frame& f) { // see generate_cont_thaw
2776 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
2777 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
2778
2779 LogTarget(Trace, continuations) lt;
2780 if (lt.develop_is_enabled()) {
2781 LogStream ls(lt);
2782 ls.print_cr("push_return_frame");
2783 f.print_value_on(&ls);
2784 }
2785
2786 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
2787 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
2788 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
2789 ContinuationHelper::push_pd(f);
2790
2791 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
2792 }
2793
2794 // returns new top sp
2795 // called after preparations (stack overflow check and making room)
2796 template<typename ConfigT>
2797 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
2798 assert(thread == JavaThread::current(), "Must be current thread");
2799
2800 CONT_JFR_ONLY(EventContinuationThaw event;)
2801
2802 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
2803
2804 ContinuationEntry* entry = thread->last_continuation();
2805 assert(entry != nullptr, "");
2806 oop oopCont = entry->cont_oop(thread);
2807
2808 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
2809 assert(oopCont == get_continuation(thread), "");
2810 verify_continuation(oopCont);
2811
2812 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
2813
2814 ContinuationWrapper cont(thread, oopCont);
2815 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
2816
2817 #ifdef ASSERT
2818 set_anchor_to_entry(thread, cont.entry());
2819 log_frames(thread);
2820 clear_anchor(thread);
2821 #endif
2822
2823 DEBUG_ONLY(bool preempted = cont.tail()->preempted();)
2824 Thaw<ConfigT> thw(thread, cont);
2825 intptr_t* const sp = thw.thaw(kind);
2826 assert(is_aligned(sp, frame::frame_alignment), "");
2827 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp, preempted);)
2828
2829 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
2830
2831 verify_continuation(cont.continuation());
2832 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
2833
2834 return sp;
2835 }
2836
2837 #ifdef ASSERT
2838 static void do_deopt_after_thaw(JavaThread* thread) {
2839 int i = 0;
2840 StackFrameStream fst(thread, true, false);
2841 fst.register_map()->set_include_argument_oops(false);
2842 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
2843 for (; !fst.is_done(); fst.next()) {
2844 if (fst.current()->cb()->is_nmethod()) {
2845 nmethod* nm = fst.current()->cb()->as_nmethod();
2846 if (!nm->method()->is_continuation_native_intrinsic()) {
2847 nm->make_deoptimized();
2848 }
2849 }
2850 }
2851 }
2852
2853 class ThawVerifyOopsClosure: public OopClosure {
2854 intptr_t* _p;
2855 outputStream* _st;
2856 bool is_good_oop(oop o) {
2857 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
2858 }
2859 public:
2860 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
2861 intptr_t* p() { return _p; }
2862 void reset() { _p = nullptr; }
2863
2864 virtual void do_oop(oop* p) {
2865 oop o = *p;
2866 if (o == nullptr || is_good_oop(o)) {
2867 return;
2868 }
2869 _p = (intptr_t*)p;
2870 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
2871 }
2872 virtual void do_oop(narrowOop* p) {
2873 oop o = RawAccess<>::oop_load(p);
2874 if (o == nullptr || is_good_oop(o)) {
2875 return;
2876 }
2877 _p = (intptr_t*)p;
2878 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
2879 }
2880 };
2881
2882 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
2883 assert(thread->has_last_Java_frame(), "");
2884
2885 ResourceMark rm;
2886 ThawVerifyOopsClosure cl(st);
2887 NMethodToOopClosure cf(&cl, false);
2888
2889 StackFrameStream fst(thread, true, false);
2890 fst.register_map()->set_include_argument_oops(false);
2891 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
2892 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
2893 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
2894 st->print_cr(">>> do_verify_after_thaw deopt");
2895 fst.current()->deoptimize(nullptr);
2896 fst.current()->print_on(st);
2897 }
2898
2899 fst.current()->oops_do(&cl, &cf, fst.register_map());
2900 if (cl.p() != nullptr) {
2901 frame fr = *fst.current();
2902 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
2903 fr.print_on(st);
2904 if (!fr.is_interpreted_frame()) {
2905 st->print_cr("size: %d argsize: %d",
2906 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
2907 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
2908 }
2909 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
2910 if (reg != nullptr) {
2911 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
2912 }
2913 cl.reset();
2914 DEBUG_ONLY(thread->print_frame_layout();)
2915 if (chunk != nullptr) {
2916 chunk->print_on(true, st);
2917 }
2918 return false;
2919 }
2920 }
2921 return true;
2922 }
2923
2924 static void log_frames(JavaThread* thread) {
2925 const static int show_entry_callers = 3;
2926 LogTarget(Trace, continuations) lt;
2927 if (!lt.develop_is_enabled()) {
2928 return;
2929 }
2930 LogStream ls(lt);
2931
2932 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
2933 if (!thread->has_last_Java_frame()) {
2934 ls.print_cr("NO ANCHOR!");
2935 }
2936
2937 RegisterMap map(thread,
2938 RegisterMap::UpdateMap::include,
2939 RegisterMap::ProcessFrames::include,
2940 RegisterMap::WalkContinuation::skip);
2941 map.set_include_argument_oops(false);
2942
2943 if (false) {
2944 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
2945 f.print_on(&ls);
2946 }
2947 } else {
2948 map.set_skip_missing(true);
2949 ResetNoHandleMark rnhm;
2950 ResourceMark rm;
2951 HandleMark hm(Thread::current());
2952 FrameValues values;
2953
2954 int i = 0;
2955 int post_entry = -1;
2956 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
2957 f.describe(values, i, &map, i == 0);
2958 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
2959 post_entry++;
2960 if (post_entry >= show_entry_callers)
2961 break;
2962 }
2963 values.print_on(thread, &ls);
2964 }
2965
2966 ls.print_cr("======= end frames =========");
2967 }
2968
2969 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempted) {
2970 intptr_t* sp0 = sp;
2971 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
2972
2973 if (preempted && sp0 == cont.entrySP()) {
2974 // Still preempted (monitor not acquired) so no frames were thawed.
2975 assert(cont.tail()->preempted(), "");
2976 set_anchor(thread, cont.entrySP(), cont.entryPC());
2977 } else {
2978 set_anchor(thread, sp0);
2979 }
2980
2981 log_frames(thread);
2982 if (LoomVerifyAfterThaw) {
2983 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
2984 }
2985 assert(ContinuationEntry::assert_entry_frame_laid_out(thread), "");
2986 clear_anchor(thread);
2987
2988 LogTarget(Trace, continuations) lt;
2989 if (lt.develop_is_enabled()) {
2990 LogStream ls(lt);
2991 ls.print_cr("Jumping to frame (thaw):");
2992 frame(sp).print_value_on(&ls);
2993 }
2994 }
2995 #endif // ASSERT
2996
2997 #include CPU_HEADER_INLINE(continuationFreezeThaw)
2998
2999 #ifdef ASSERT
3000 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3001 ResourceMark rm;
3002 FrameValues values;
3003 assert(f.get_cb() != nullptr, "");
3004 RegisterMap map(f.is_heap_frame() ?
3005 nullptr :
3006 JavaThread::current(),
3007 RegisterMap::UpdateMap::include,
3008 RegisterMap::ProcessFrames::skip,
3009 RegisterMap::WalkContinuation::skip);
3010 map.set_include_argument_oops(false);
3011 map.set_skip_missing(true);
3012 if (callee_complete) {
3013 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3014 }
3015 const_cast<frame&>(f).describe(values, 0, &map, true);
3016 values.print_on(static_cast<JavaThread*>(nullptr), st);
3017 }
3018 #endif
3019
3020 static address thaw_entry = nullptr;
3021 static address freeze_entry = nullptr;
3022 static address freeze_preempt_entry = nullptr;
3023
3024 address Continuation::thaw_entry() {
3025 return ::thaw_entry;
3026 }
3027
3028 address Continuation::freeze_entry() {
3029 return ::freeze_entry;
3030 }
3031
3032 address Continuation::freeze_preempt_entry() {
3033 return ::freeze_preempt_entry;
3034 }
3035
3036 class ConfigResolve {
3037 public:
3038 static void resolve() { resolve_compressed(); }
3039
3040 static void resolve_compressed() {
3041 UseCompressedOops ? resolve_gc<true>()
3042 : resolve_gc<false>();
3043 }
3044
3045 private:
3046 template <bool use_compressed>
3047 static void resolve_gc() {
3048 BarrierSet* bs = BarrierSet::barrier_set();
3049 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3050 switch (bs->kind()) {
3051 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3052 case BarrierSet::bs_name: { \
3053 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3054 } \
3055 break;
3056 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3057 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3058
3059 default:
3060 fatal("BarrierSet resolving not implemented");
3061 };
3062 }
3063
3064 template <bool use_compressed, typename BarrierSetT>
3065 static void resolve() {
3066 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3067
3068 freeze_entry = (address)freeze<SelectedConfigT>;
3069 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3070
3071 // If we wanted, we could templatize by kind and have three different thaw entries
3072 thaw_entry = (address)thaw<SelectedConfigT>;
3073 }
3074 };
3075
3076 void Continuation::init() {
3077 ConfigResolve::resolve();
3078 }