1 /*
2 * Copyright (c) 2018, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "classfile/javaClasses.inline.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "code/codeCache.inline.hpp"
28 #include "code/nmethod.inline.hpp"
29 #include "code/vmreg.inline.hpp"
30 #include "compiler/oopMap.inline.hpp"
31 #include "gc/shared/continuationGCSupport.inline.hpp"
32 #include "gc/shared/gc_globals.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/memAllocator.hpp"
35 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
36 #include "interpreter/bytecodeStream.hpp"
37 #include "interpreter/interpreter.hpp"
38 #include "interpreter/interpreterRuntime.hpp"
39 #include "jfr/jfrEvents.hpp"
40 #include "logging/log.hpp"
41 #include "logging/logStream.hpp"
42 #include "oops/access.inline.hpp"
43 #include "oops/constantPool.inline.hpp"
44 #include "oops/method.inline.hpp"
45 #include "oops/oopsHierarchy.hpp"
46 #include "oops/objArrayOop.inline.hpp"
47 #include "oops/stackChunkOop.inline.hpp"
48 #include "prims/jvmtiThreadState.hpp"
49 #include "runtime/arguments.hpp"
50 #include "runtime/continuation.hpp"
51 #include "runtime/continuationEntry.inline.hpp"
52 #include "runtime/continuationHelper.inline.hpp"
53 #include "runtime/continuationJavaClasses.inline.hpp"
54 #include "runtime/continuationWrapper.inline.hpp"
55 #include "runtime/frame.inline.hpp"
56 #include "runtime/interfaceSupport.inline.hpp"
57 #include "runtime/javaThread.inline.hpp"
58 #include "runtime/jniHandles.inline.hpp"
59 #include "runtime/keepStackGCProcessed.hpp"
60 #include "runtime/objectMonitor.inline.hpp"
61 #include "runtime/orderAccess.hpp"
62 #include "runtime/prefetch.inline.hpp"
63 #include "runtime/smallRegisterMap.inline.hpp"
64 #include "runtime/sharedRuntime.hpp"
65 #include "runtime/stackChunkFrameStream.inline.hpp"
66 #include "runtime/stackFrameStream.inline.hpp"
67 #include "runtime/stackOverflow.hpp"
68 #include "runtime/stackWatermarkSet.inline.hpp"
69 #include "runtime/vframe.inline.hpp"
70 #include "runtime/vframe_hp.hpp"
71 #include "utilities/debug.hpp"
72 #include "utilities/exceptions.hpp"
73 #include "utilities/macros.hpp"
74 #include "utilities/vmError.hpp"
75 #if INCLUDE_ZGC
76 #include "gc/z/zStackChunkGCData.inline.hpp"
77 #endif
78 #ifdef COMPILER1
79 #include "c1/c1_Runtime1.hpp"
80 #endif
81 #ifdef COMPILER2
82 #include "opto/runtime.hpp"
83 #endif
84
85 #include <type_traits>
86
87 /*
88 * This file contains the implementation of continuation freezing (yield) and thawing (run).
89 *
90 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
91 * would likely call these operations many thousands of times per second second, on every core.
92 *
93 * Freeze might be called every time the application performs any I/O operation, every time it
94 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
95 * multiple times in each of those cases, as it is called by the return barrier, which may be
96 * invoked on method return.
97 *
98 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
99 * example, every effort is made to avoid Java-VM transitions as much as possible.
100 *
101 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
102 * and so frames simply copied, and the bottom-most one is patched.
103 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
104 * and absolute pointers, and barriers invoked.
105 */
106
107 /************************************************
108
109 Thread-stack layout on freeze/thaw.
110 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
111
112 +----------------------------+
113 | . |
114 | . |
115 | . |
116 | carrier frames |
117 | |
118 |----------------------------|
119 | |
120 | Continuation.run |
121 | |
122 |============================|
123 | enterSpecial frame |
124 | pc |
125 | rbp |
126 | ----- |
127 ^ | int argsize | = ContinuationEntry
128 | | oopDesc* cont |
129 | | oopDesc* chunk |
130 | | ContinuationEntry* parent |
131 | | ... |
132 | |============================| <------ JavaThread::_cont_entry = entry->sp()
133 | | ? alignment word ? |
134 | |----------------------------| <--\
135 | | | |
136 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
137 Address | | | | Caller is still in the chunk.
138 | |----------------------------| |
139 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
140 | | rbp | | isn't the last one in the continuation.
141 | | | |
142 | | frame | |
143 | | | |
144 +----------------------------| \__ Continuation frames to be frozen/thawed
145 | | /
146 | frame | |
147 | | |
148 |----------------------------| |
149 | | |
150 | frame | |
151 | | |
152 |----------------------------| <--/
153 | |
154 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
155 | | instead of a doYield stub
156 |============================| <- the sp passed to freeze
157 | |
158 | Native freeze/thaw frames |
159 | . |
160 | . |
161 | . |
162 +----------------------------+
163
164 ************************************************/
165
166 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing
167
168 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
169 #if CONT_JFR
170 #define CONT_JFR_ONLY(code) code
171 #else
172 #define CONT_JFR_ONLY(code)
173 #endif
174
175 // TODO: See AbstractAssembler::generate_stack_overflow_check,
176 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
177 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
178
179 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
180
181 // Used to just annotatate cold/hot branches
182 #define LIKELY(condition) (condition)
183 #define UNLIKELY(condition) (condition)
184
185 // debugging functions
186 #ifdef ASSERT
187 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
188
189 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
190
191 static void do_deopt_after_thaw(JavaThread* thread);
192 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
193 static void log_frames(JavaThread* thread, bool dolog = false);
194 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp);
195 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
196 static void verify_frame_kind(const frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr = nullptr, const char** code_name_ptr = nullptr, int* bci_ptr = nullptr);
197
198 #define assert_pfl(p, ...) \
199 do { \
200 if (!(p)) { \
201 JavaThread* t = JavaThread::active(); \
202 if (t->has_last_Java_frame()) { \
203 tty->print_cr("assert(" #p ") failed:"); \
204 t->print_frame_layout(); \
205 } \
206 } \
207 vmassert(p, __VA_ARGS__); \
208 } while(0)
209
210 #else
211 static void verify_continuation(oop continuation) { }
212 #define assert_pfl(p, ...)
213 #endif
214
215 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
216 template<typename ConfigT, bool preempt> static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp);
217
218 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
219 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
220
221
222 // Entry point to freeze. Transitions are handled manually
223 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
224 template<typename ConfigT>
225 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
226 assert(sp == current->frame_anchor()->last_Java_sp(), "");
227
228 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
229 current->set_cont_fastpath(nullptr);
230 }
231
232 return checked_cast<int>(ConfigT::freeze(current, sp));
233 JRT_END
234
235 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
236 return prepare_thaw_internal(thread, return_barrier);
237 JRT_END
238
239 template<typename ConfigT>
240 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
241 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
242 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
243 // Also the preemption case with JVMTI events enabled might safepoint so
244 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
245 // JRT_ENTRY instead?
246 ResetNoHandleMark rnhm;
247 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
248
249 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
250 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
251 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
252 JRT_END
253
254 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
255 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
256 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
257 }
258 JVM_END
259
260 ///////////
261
262 enum class oop_kind { NARROW, WIDE };
263 template <oop_kind oops, typename BarrierSetT>
264 class Config {
265 public:
266 typedef Config<oops, BarrierSetT> SelfT;
267 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
268
269 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
270 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
271 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
272 return res;
273 }
274
275 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
276 return freeze_internal<SelfT, true>(thread, sp);
277 }
278
279 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
280 return thaw_internal<SelfT>(thread, kind);
281 }
282 };
283
284 #ifdef _WINDOWS
285 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
286 address new_sp = sp - size;
287 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
288
289 if (new_sp < watermark) {
290 size_t page_size = os::vm_page_size();
291 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
292 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
293 while (pages_to_touch-- > 0) {
294 last_touched_page -= page_size;
295 *last_touched_page = 0;
296 }
297 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
298 }
299 }
300 #endif
301
302 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
303 const size_t page_size = os::vm_page_size();
304 if (size > page_size) {
305 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
306 return false;
307 }
308 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
309 }
310 return true;
311 }
312
313 #ifdef ASSERT
314 static oop get_continuation(JavaThread* thread) {
315 assert(thread != nullptr, "");
316 assert(thread->threadObj() != nullptr, "");
317 return java_lang_Thread::continuation(thread->threadObj());
318 }
319 #endif // ASSERT
320
321 inline void clear_anchor(JavaThread* thread) {
322 thread->frame_anchor()->clear();
323 }
324
325 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
326 assert(pc != nullptr, "");
327
328 JavaFrameAnchor* anchor = thread->frame_anchor();
329 anchor->set_last_Java_sp(sp);
330 anchor->set_last_Java_pc(pc);
331 ContinuationHelper::set_anchor_pd(anchor, sp);
332
333 assert(thread->has_last_Java_frame(), "");
334 assert(thread->last_frame().cb() != nullptr, "");
335 }
336
337 static void set_anchor(JavaThread* thread, intptr_t* sp) {
338 address pc = ContinuationHelper::return_address_at(
339 sp - frame::sender_sp_ret_address_offset());
340 set_anchor(thread, sp, pc);
341 }
342
343 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
344 JavaFrameAnchor* anchor = thread->frame_anchor();
345 anchor->set_last_Java_sp(entry->entry_sp());
346 anchor->set_last_Java_pc(entry->entry_pc());
347 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
348
349 assert(thread->has_last_Java_frame(), "");
350 assert(thread->last_frame().cb() != nullptr, "");
351 }
352
353 #if CONT_JFR
354 class FreezeThawJfrInfo : public StackObj {
355 short _e_size;
356 short _e_num_interpreted_frames;
357 public:
358
359 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
360 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
361 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
362 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
363 };
364
365 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
366 if (e->should_commit()) {
367 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
368 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
369 e->set_continuationClass(continuation->klass());
370 e->set_interpretedFrames(_e_num_interpreted_frames);
371 e->set_size(_e_size);
372 e->commit();
373 }
374 }
375 #endif // CONT_JFR
376
377 /////////////// FREEZE ////
378
379 class FreezeBase : public StackObj {
380 protected:
381 JavaThread* const _thread;
382 ContinuationWrapper& _cont;
383 bool _barriers; // only set when we allocate a chunk
384
385 intptr_t* _bottom_address;
386
387 // Used for preemption only
388 const bool _preempt;
389 frame _last_frame;
390
391 // Used to support freezing with held monitors
392 int _monitors_in_lockstack;
393
394 int _freeze_size; // total size of all frames plus metadata in words.
395 int _total_align_size;
396
397 intptr_t* _cont_stack_top;
398 intptr_t* _cont_stack_bottom;
399
400 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
401
402 #ifdef ASSERT
403 intptr_t* _orig_chunk_sp;
404 int _fast_freeze_size;
405 bool _empty;
406 #endif
407
408 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
409
410 NOT_PRODUCT(int _frames;)
411 DEBUG_ONLY(intptr_t* _last_write;)
412
413 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
414
415 public:
416 NOINLINE freeze_result freeze_slow();
417 void freeze_fast_existing_chunk();
418
419 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
420 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
421
422 inline int size_if_fast_freeze_available();
423
424 inline frame& last_frame() { return _last_frame; }
425
426 #ifdef ASSERT
427 bool check_valid_fast_path();
428 #endif
429
430 protected:
431 inline void init_rest();
432 void throw_stack_overflow_on_humongous_chunk();
433
434 // fast path
435 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
436 inline void unwind_frames();
437 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
438
439 // slow path
440 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
441
442 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
443
444 private:
445 // slow path
446 frame freeze_start_frame();
447 frame freeze_start_frame_on_preempt();
448 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
449 inline frame freeze_start_frame_yield_stub();
450 template<typename FKind>
451 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
452 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
453 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
454 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
455 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
456 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
457 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
458 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
459 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
460 NOINLINE void finish_freeze(const frame& f, const frame& top);
461
462 void freeze_lockstack(stackChunkOop chunk);
463
464 inline bool stack_overflow();
465
466 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
467 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
468 template<typename FKind> static inline frame sender(const frame& f);
469 template<typename FKind> frame new_heap_frame(frame& f, frame& caller);
470 inline void set_top_frame_metadata_pd(const frame& hf);
471 inline void patch_pd(frame& callee, const frame& caller);
472 void adjust_interpreted_frame_unextended_sp(frame& f);
473 static inline void prepare_freeze_interpreted_top_frame(frame& f);
474 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
475
476 protected:
477 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
478 bool freeze_fast_new_chunk(stackChunkOop chunk);
479 };
480
481 template <typename ConfigT>
482 class Freeze : public FreezeBase {
483 private:
484 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
485
486 public:
487 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
488 : FreezeBase(thread, cont, frame_sp, preempt) {}
489
490 freeze_result try_freeze_fast();
491
492 protected:
493 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
494 };
495
496 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
497 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
498 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
499
500 assert(_thread != nullptr, "");
501 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
502
503 DEBUG_ONLY(_cont.entry()->verify_cookie();)
504
505 assert(!Interpreter::contains(_cont.entryPC()), "");
506
507 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
508 #ifdef _LP64
509 if (((intptr_t)_bottom_address & 0xf) != 0) {
510 _bottom_address--;
511 }
512 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
513 #endif
514
515 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
516 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
517 assert(_bottom_address != nullptr, "");
518 assert(_bottom_address <= _cont.entrySP(), "");
519 DEBUG_ONLY(_last_write = nullptr;)
520
521 assert(_cont.chunk_invariant(), "");
522 assert(!Interpreter::contains(_cont.entryPC()), "");
523 #if !defined(PPC64) || defined(ZERO)
524 static const int do_yield_frame_size = frame::metadata_words;
525 #else
526 static const int do_yield_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
527 #endif
528 // With preemption doYield() might not have been resolved yet
529 assert(_preempt || ContinuationEntry::do_yield_nmethod()->frame_size() == do_yield_frame_size, "");
530
531 if (preempt) {
532 _last_frame = _thread->last_frame();
533 }
534
535 // properties of the continuation on the stack; all sizes are in words
536 _cont_stack_top = frame_sp + (!preempt ? do_yield_frame_size : 0); // we don't freeze the doYield stub frame
537 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
538 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
539
540 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
541 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
542 assert(cont_size() > 0, "");
543
544 if (LockingMode != LM_LIGHTWEIGHT) {
545 _monitors_in_lockstack = 0;
546 } else {
547 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
548 }
549 }
550
551 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
552 _freeze_size = 0;
553 _total_align_size = 0;
554 NOT_PRODUCT(_frames = 0;)
555 }
556
557 void FreezeBase::freeze_lockstack(stackChunkOop chunk) {
558 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
559
560 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
561 chunk->set_lockstack_size(checked_cast<uint8_t>(_monitors_in_lockstack));
562 chunk->set_has_lockstack(true);
563 }
564
565 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
566 stackChunkOop chunk = _cont.tail();
567 chunk->copy_from_stack_to_chunk(from, to, size);
568 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
569
570 #ifdef ASSERT
571 if (_last_write != nullptr) {
572 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
573 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
574 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
575 _last_write = to;
576 }
577 #endif
578 }
579
580 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
581 void FreezeBase::unwind_frames() {
582 ContinuationEntry* entry = _cont.entry();
583 entry->flush_stack_processing(_thread);
584 set_anchor_to_entry(_thread, entry);
585 }
586
587 template <typename ConfigT>
588 freeze_result Freeze<ConfigT>::try_freeze_fast() {
589 assert(_thread->thread_state() == _thread_in_vm, "");
590 assert(_thread->cont_fastpath(), "");
591
592 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
593 assert(_fast_freeze_size == 0, "");
594
595 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
596 if (freeze_fast_new_chunk(chunk)) {
597 return freeze_ok;
598 }
599 if (_thread->has_pending_exception()) {
600 return freeze_exception;
601 }
602
603 // TODO R REMOVE when deopt change is fixed
604 assert(!_thread->cont_fastpath() || _barriers, "");
605 log_develop_trace(continuations)("-- RETRYING SLOW --");
606 return freeze_slow();
607 }
608
609 // Returns size needed if the continuation fits, otherwise 0.
610 int FreezeBase::size_if_fast_freeze_available() {
611 stackChunkOop chunk = _cont.tail();
612 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
613 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
614 return 0;
615 }
616
617 int total_size_needed = cont_size();
618 const int chunk_sp = chunk->sp();
619
620 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
621 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
622 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
623 // although that would require changing stackChunkOopDesc::is_empty
624 if (!chunk->is_empty()) {
625 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
626 }
627
628 total_size_needed += _monitors_in_lockstack;
629
630 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
631 bool available = chunk_free_room >= total_size_needed;
632 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
633 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
634 return available ? total_size_needed : 0;
635 }
636
637 void FreezeBase::freeze_fast_existing_chunk() {
638 stackChunkOop chunk = _cont.tail();
639
640 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
641 assert(_fast_freeze_size > 0, "");
642
643 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
644 DEBUG_ONLY(_empty = false;)
645 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
646 #ifdef ASSERT
647 {
648 intptr_t* retaddr_slot = (chunk->sp_address()
649 - frame::sender_sp_ret_address_offset());
650 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
651 "unexpected saved return address");
652 }
653 #endif
654
655 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
656 // we overlap; we'll overwrite the chunk's top frame's callee arguments
657 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
658 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
659
660 // increase max_size by what we're freezing minus the overlap
661 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
662
663 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
664 assert(bottom_sp == _bottom_address, "");
665 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
666 // should have a return barrier (installed back when we thawed it).
667 #ifdef ASSERT
668 {
669 intptr_t* retaddr_slot = (bottom_sp
670 - frame::sender_sp_ret_address_offset());
671 assert(ContinuationHelper::return_address_at(retaddr_slot)
672 == StubRoutines::cont_returnBarrier(),
673 "should be the continuation return barrier");
674 }
675 #endif
676 // We copy the fp from the chunk back to the stack because it contains some caller data,
677 // including, possibly, an oop that might have gone stale since we thawed.
678 patch_stack_pd(bottom_sp, chunk->sp_address());
679 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
680
681 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
682 } else { // the chunk is empty
683 const int chunk_start_sp = chunk->stack_size();
684
685 DEBUG_ONLY(_empty = true;)
686 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
687
688 chunk->set_max_thawing_size(cont_size());
689 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
690 chunk->set_sp(chunk->bottom());
691
692 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
693 }
694 }
695
696 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
697 DEBUG_ONLY(_empty = true;)
698
699 // Install new chunk
700 _cont.set_tail(chunk);
701
702 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
703 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
704 return false;
705 }
706
707 chunk->set_max_thawing_size(cont_size());
708
709 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
710 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
711 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
712 assert(chunk_start_sp == chunk->stack_size(), "");
713
714 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
715
716 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
717
718 return true;
719 }
720
721 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
722 assert(chunk != nullptr, "");
723 assert(!chunk->has_mixed_frames(), "");
724 assert(!chunk->is_gc_mode(), "");
725 assert(!chunk->has_bitmap(), "");
726 assert(!chunk->requires_barriers(), "");
727 assert(chunk == _cont.tail(), "");
728
729 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
730 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
731 // will either see no continuation on the stack, or a consistent chunk.
732 unwind_frames();
733
734 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
735 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
736 assert(chunk_start_sp <= chunk->stack_size(), "");
737 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
738
739 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
740 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
741
742 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
743 #ifdef ASSERT
744 if (!_empty) {
745 intptr_t* retaddr_slot = (_orig_chunk_sp
746 - frame::sender_sp_ret_address_offset());
747 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
748 "unexpected saved return address");
749 }
750 #endif
751
752 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
753 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
754 intptr_t* from = _cont_stack_top - frame::metadata_words_at_bottom;
755 intptr_t* to = chunk_top - frame::metadata_words_at_bottom;
756 copy_to_chunk(from, to, cont_size() + frame::metadata_words_at_bottom);
757 // Because we're not patched yet, the chunk is now in a bad state
758
759 // patch return pc of the bottom-most frozen frame (now in the chunk)
760 // with the actual caller's return address
761 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
762 - _cont.argsize()
763 - frame::metadata_words_at_top
764 - frame::sender_sp_ret_address_offset());
765 #ifdef ASSERT
766 if (!_empty) {
767 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
768 == StubRoutines::cont_returnBarrier(),
769 "should be the continuation return barrier");
770 }
771 #endif
772 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
773 chunk->pc());
774
775 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
776 chunk->set_sp(chunk_new_sp);
777
778 // set chunk->pc to the return address of the topmost frame in the chunk
779 if (_preempt) {
780 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
781 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
782 // be the pc we used when creating the oopmap. Get the top's frame last pc from
783 // the anchor instead.
784 address last_pc = _last_frame.pc();
785 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
786 chunk->set_pc(last_pc);
787 } else {
788 chunk->set_pc(ContinuationHelper::return_address_at(
789 _cont_stack_top - frame::sender_sp_ret_address_offset()));
790 }
791
792 if (_monitors_in_lockstack > 0) {
793 freeze_lockstack(chunk);
794 }
795
796 _cont.write();
797
798 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
799 LogTarget(Trace, continuations) lt;
800 if (lt.develop_is_enabled()) {
801 LogStream ls(lt);
802 chunk->print_on(true, &ls);
803 }
804
805 // Verification
806 assert(_cont.chunk_invariant(), "");
807 chunk->verify();
808
809 #if CONT_JFR
810 EventContinuationFreezeFast e;
811 if (e.should_commit()) {
812 e.set_id(cast_from_oop<u8>(chunk));
813 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
814 e.set_size(cont_size() << LogBytesPerWord);
815 e.commit();
816 }
817 #endif
818 }
819
820 NOINLINE freeze_result FreezeBase::freeze_slow() {
821 #ifdef ASSERT
822 ResourceMark rm;
823 #endif
824
825 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
826 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
827
828 #if CONT_JFR
829 EventContinuationFreezeSlow e;
830 if (e.should_commit()) {
831 e.set_id(cast_from_oop<u8>(_cont.continuation()));
832 e.commit();
833 }
834 #endif
835
836 init_rest();
837
838 HandleMark hm(Thread::current());
839
840 frame f = freeze_start_frame();
841
842 LogTarget(Debug, continuations) lt;
843 if (lt.develop_is_enabled()) {
844 LogStream ls(lt);
845 f.print_on(&ls);
846 }
847
848 frame caller; // the frozen caller in the chunk
849 freeze_result res = recurse_freeze(f, caller, 0, false, true);
850
851 if (res == freeze_ok) {
852 finish_freeze(f, caller);
853 _cont.write();
854 }
855
856 return res;
857 }
858
859 frame FreezeBase::freeze_start_frame() {
860 if (LIKELY(!_preempt)) {
861 return freeze_start_frame_yield_stub();
862 } else {
863 return freeze_start_frame_on_preempt();
864 }
865 }
866
867 frame FreezeBase::freeze_start_frame_yield_stub() {
868 frame f = _thread->last_frame();
869 assert(ContinuationEntry::do_yield_nmethod()->contains(f.pc()), "must be");
870 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
871 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
872 return f;
873 }
874
875 frame FreezeBase::freeze_start_frame_on_preempt() {
876 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
877 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
878 return _last_frame;
879 }
880
881 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
882 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
883 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
884 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
885 || ((top && _preempt) == f.is_native_frame()), "");
886
887 if (stack_overflow()) {
888 return freeze_exception;
889 }
890
891 if (f.is_compiled_frame()) {
892 if (UNLIKELY(f.oop_map() == nullptr)) {
893 // special native frame
894 return freeze_pinned_native;
895 }
896 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
897 } else if (f.is_interpreted_frame()) {
898 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
899 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
900 } else if (top && _preempt) {
901 assert(f.is_native_frame() || f.is_runtime_frame(), "");
902 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
903 } else {
904 // Frame can't be frozen. Most likely the call_stub or upcall_stub
905 // which indicates there are further natives frames up the stack.
906 return freeze_pinned_native;
907 }
908 }
909
910 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
911 // See also StackChunkFrameStream<frame_kind>::frame_size()
912 template<typename FKind>
913 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
914 assert(FKind::is_instance(f), "");
915
916 assert(fsize > 0, "");
917 assert(argsize >= 0, "");
918 _freeze_size += fsize;
919 NOT_PRODUCT(_frames++;)
920
921 assert(FKind::frame_bottom(f) <= _bottom_address, "");
922
923 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
924 // enterSpecial and an interpreted frame
925 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
926 return finalize_freeze(f, caller, argsize); // recursion end
927 } else {
928 frame senderf = sender<FKind>(f);
929 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
930 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
931 return result;
932 }
933 }
934
935 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
936 LogTarget(Trace, continuations) lt;
937 if (lt.develop_is_enabled()) {
938 LogStream ls(lt);
939 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
940 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
941 f.print_value_on(&ls);
942 }
943 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
944 }
945
946 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
947 LogTarget(Trace, continuations) lt;
948 if (lt.develop_is_enabled()) {
949 LogStream ls(lt);
950 DEBUG_ONLY(hf.print_value_on(&ls);)
951 assert(hf.is_heap_frame(), "should be");
952 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
953 if (is_bottom_frame) {
954 ls.print_cr("bottom h-frame:");
955 hf.print_on(&ls);
956 }
957 }
958 }
959
960 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
961 // See also StackChunkFrameStream<frame_kind>::frame_size()
962 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
963 int argsize = argsize_md - frame::metadata_words_at_top;
964 assert(callee.is_interpreted_frame()
965 || ContinuationHelper::Frame::is_stub(callee.cb())
966 || callee.cb()->as_nmethod()->is_osr_method()
967 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
968 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
969 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
970
971 LogTarget(Trace, continuations) lt;
972
973 #ifdef ASSERT
974 bool empty = _cont.is_empty();
975 log_develop_trace(continuations)("empty: %d", empty);
976 #endif
977
978 stackChunkOop chunk = _cont.tail();
979
980 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
981
982 _freeze_size += frame::metadata_words; // for top frame's metadata
983
984 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
985 int unextended_sp = -1;
986 if (chunk != nullptr) {
987 if (!chunk->is_empty()) {
988 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
989 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
990 bool top_interpreted = Interpreter::contains(chunk->pc());
991 if (callee.is_interpreted_frame() == top_interpreted) {
992 overlap = argsize_md;
993 }
994 } else {
995 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
996 }
997 }
998
999 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1000
1001 _freeze_size -= overlap;
1002 assert(_freeze_size >= 0, "");
1003
1004 assert(chunk == nullptr || chunk->is_empty()
1005 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1006 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1007
1008 _freeze_size += _monitors_in_lockstack;
1009
1010 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1011 bool allocated_old_in_freeze_fast = _barriers;
1012 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1013 "Chunk allocated in freeze_fast is of insufficient size "
1014 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1015 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1016
1017 DEBUG_ONLY(bool empty_chunk = true);
1018 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1019 // ALLOCATE NEW CHUNK
1020
1021 if (lt.develop_is_enabled()) {
1022 LogStream ls(lt);
1023 if (chunk == nullptr) {
1024 ls.print_cr("no chunk");
1025 } else {
1026 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1027 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1028 chunk->print_on(&ls);
1029 }
1030 }
1031
1032 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1033 // overlap = 0;
1034
1035 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1036 if (chunk == nullptr) {
1037 return freeze_exception;
1038 }
1039
1040 // Install new chunk
1041 _cont.set_tail(chunk);
1042 assert(chunk->is_empty(), "");
1043 } else {
1044 // REUSE EXISTING CHUNK
1045 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1046 if (chunk->is_empty()) {
1047 int sp = chunk->stack_size() - argsize_md;
1048 chunk->set_sp(sp);
1049 chunk->set_bottom(sp);
1050 _freeze_size += overlap;
1051 assert(chunk->max_thawing_size() == 0, "");
1052 } DEBUG_ONLY(else empty_chunk = false;)
1053 }
1054 assert(!chunk->is_gc_mode(), "");
1055 assert(!chunk->has_bitmap(), "");
1056 chunk->set_has_mixed_frames(true);
1057
1058 assert(chunk->requires_barriers() == _barriers, "");
1059 assert(!_barriers || chunk->is_empty(), "");
1060
1061 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1062 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1063
1064 if (_preempt) {
1065 frame top_frame = _thread->last_frame();
1066 if (top_frame.is_interpreted_frame()) {
1067 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1068 // We need it so that on resume we can restore the sp to the right place, since
1069 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1070 // We do it now that we know freezing will be successful.
1071 prepare_freeze_interpreted_top_frame(top_frame);
1072 }
1073
1074 // Do this now so should_process_args_at_top() is set before calling finish_freeze
1075 // in case we might need to apply GC barriers to frames in this stackChunk.
1076 if (_thread->at_preemptable_init()) {
1077 assert(top_frame.is_interpreted_frame(), "only InterpreterRuntime::_new/resolve_from_cache allowed");
1078 chunk->set_at_klass_init(true);
1079 Method* m = top_frame.interpreter_frame_method();
1080 Bytecode current_bytecode = Bytecode(m, top_frame.interpreter_frame_bcp());
1081 Bytecodes::Code code = current_bytecode.code();
1082 int exp_size = top_frame.interpreter_frame_expression_stack_size();
1083 if (code == Bytecodes::Code::_invokestatic && exp_size > 0) {
1084 chunk->set_has_args_at_top(true);
1085 }
1086 }
1087 }
1088
1089 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1090 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1091 // will either see no continuation or a consistent chunk.
1092 unwind_frames();
1093
1094 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1095
1096 if (lt.develop_is_enabled()) {
1097 LogStream ls(lt);
1098 ls.print_cr("top chunk:");
1099 chunk->print_on(&ls);
1100 }
1101
1102 if (_monitors_in_lockstack > 0) {
1103 freeze_lockstack(chunk);
1104 }
1105
1106 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1107 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1108
1109 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1110
1111 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1112 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1113 #ifdef ASSERT
1114 if (lt.develop_is_enabled()) {
1115 LogStream ls(lt);
1116 ls.print_cr("top hframe before (freeze):");
1117 assert(caller.is_heap_frame(), "should be");
1118 caller.print_on(&ls);
1119 }
1120
1121 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1122
1123 frame entry = sender(callee);
1124
1125 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1126 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1127 #endif
1128
1129 return freeze_ok_bottom;
1130 }
1131
1132 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1133 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1134 if (is_bottom_frame) {
1135 // If we're the bottom frame, we need to replace the return barrier with the real
1136 // caller's pc.
1137 address last_pc = caller.pc();
1138 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1139 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1140 } else {
1141 assert(!caller.is_empty(), "");
1142 }
1143
1144 patch_pd(hf, caller);
1145
1146 if (f.is_interpreted_frame()) {
1147 assert(hf.is_heap_frame(), "should be");
1148 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1149 }
1150
1151 #ifdef ASSERT
1152 if (hf.is_compiled_frame()) {
1153 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1154 log_develop_trace(continuations)("Freezing deoptimized frame");
1155 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1156 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1157 }
1158 }
1159 #endif
1160 }
1161
1162 #ifdef ASSERT
1163 static void verify_frame_top(const frame& f, intptr_t* top) {
1164 ResourceMark rm;
1165 InterpreterOopMap mask;
1166 f.interpreted_frame_oop_map(&mask);
1167 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1168 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1169 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1170 }
1171 #endif // ASSERT
1172
1173 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1174 // See also StackChunkFrameStream<frame_kind>::frame_size()
1175 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1176 int callee_argsize /* incl. metadata */,
1177 bool callee_interpreted) {
1178 adjust_interpreted_frame_unextended_sp(f);
1179
1180 // The frame's top never includes the stack arguments to the callee
1181 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1182 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1183 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1184
1185 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1186
1187 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1188 // including metadata between f and its args
1189 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1190
1191 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1192 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1193 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1194 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1195
1196 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1197 if (UNLIKELY(result > freeze_ok_bottom)) {
1198 return result;
1199 }
1200
1201 bool is_bottom_frame = result == freeze_ok_bottom;
1202 assert(!caller.is_empty() || is_bottom_frame, "");
1203
1204 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1205
1206 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1207 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1208
1209 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1210 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1211 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1212
1213 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1214 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1215 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1216 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1217
1218 relativize_interpreted_frame_metadata(f, hf);
1219
1220 patch(f, hf, caller, is_bottom_frame);
1221
1222 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1223 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1224 caller = hf;
1225
1226 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1227 frame_method->record_gc_epoch();
1228
1229 return freeze_ok;
1230 }
1231
1232 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1233 // See also StackChunkFrameStream<frame_kind>::frame_size()
1234 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1235 int callee_argsize /* incl. metadata */,
1236 bool callee_interpreted) {
1237 // The frame's top never includes the stack arguments to the callee
1238 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1239 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1240 // including metadata between f and its stackargs
1241 const int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1242 const int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1243
1244 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d",
1245 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1246 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1247 _freeze_size, fsize, argsize);
1248 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1249 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1250
1251 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1252 if (UNLIKELY(result > freeze_ok_bottom)) {
1253 return result;
1254 }
1255
1256 bool is_bottom_frame = result == freeze_ok_bottom;
1257 assert(!caller.is_empty() || is_bottom_frame, "");
1258
1259 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1260
1261 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller);
1262
1263 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1264
1265 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1266 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1267
1268 if (caller.is_interpreted_frame()) {
1269 // When thawing the frame we might need to add alignment (see Thaw::align)
1270 _total_align_size += frame::align_wiggle;
1271 }
1272
1273 patch(f, hf, caller, is_bottom_frame);
1274
1275 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1276
1277 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1278 caller = hf;
1279 return freeze_ok;
1280 }
1281
1282 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1283 DEBUG_ONLY(frame fsender = sender(f);)
1284 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1285
1286 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1287 const int fsize = f.cb()->frame_size();
1288
1289 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1290 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1291
1292 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1293 if (UNLIKELY(result > freeze_ok_bottom)) {
1294 return result;
1295 }
1296
1297 assert(result == freeze_ok, "should have caller");
1298 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1299
1300 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1301 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1302
1303 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1304
1305 patch(f, hf, caller, false /*is_bottom_frame*/);
1306
1307 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1308
1309 caller = hf;
1310 return freeze_ok;
1311 }
1312
1313 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1314 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1315 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1316 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1317 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1318 // caller into the native convention. This is so that we have a valid oopMap in case of
1319 // having to block in the slow path. But that would require freezing those registers too
1320 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1321 // given that this would be a rare case anyways just pin the vthread to the carrier.
1322 return freeze_pinned_native;
1323 }
1324
1325 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1326 // There are no stackargs but argsize must include the metadata
1327 const int argsize = frame::metadata_words_at_top;
1328 const int fsize = f.cb()->frame_size() + argsize;
1329
1330 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1331 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1332
1333 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1334 if (UNLIKELY(result > freeze_ok_bottom)) {
1335 return result;
1336 }
1337
1338 assert(result == freeze_ok, "should have caller frame");
1339 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1340
1341 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1342 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1343
1344 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1345
1346 if (caller.is_interpreted_frame()) {
1347 // When thawing the frame we might need to add alignment (see Thaw::align)
1348 _total_align_size += frame::align_wiggle;
1349 }
1350
1351 patch(f, hf, caller, false /* is_bottom_frame */);
1352
1353 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1354
1355 caller = hf;
1356 return freeze_ok;
1357 }
1358
1359 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1360 stackChunkOop chunk = _cont.tail();
1361
1362 LogTarget(Trace, continuations) lt;
1363 if (lt.develop_is_enabled()) {
1364 LogStream ls(lt);
1365 assert(top.is_heap_frame(), "should be");
1366 top.print_on(&ls);
1367 }
1368
1369 set_top_frame_metadata_pd(top);
1370
1371 chunk->set_sp(chunk->to_offset(top.sp()));
1372 chunk->set_pc(top.pc());
1373
1374 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1375
1376 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1377
1378 // At this point the chunk is consistent
1379
1380 if (UNLIKELY(_barriers)) {
1381 log_develop_trace(continuations)("do barriers on old chunk");
1382 // Serial and Parallel GC can allocate objects directly into the old generation.
1383 // Then we want to relativize the derived pointers eagerly so that
1384 // old chunks are all in GC mode.
1385 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1386 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1387 if (UseShenandoahGC) {
1388 _cont.tail()->relativize_derived_pointers_concurrently();
1389 } else {
1390 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1391 }
1392 // For objects in the old generation we must maintain the remembered set
1393 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1394 }
1395
1396 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1397 if (lt.develop_is_enabled()) {
1398 LogStream ls(lt);
1399 chunk->print_on(true, &ls);
1400 }
1401
1402 if (lt.develop_is_enabled()) {
1403 LogStream ls(lt);
1404 ls.print_cr("top hframe after (freeze):");
1405 assert(_cont.last_frame().is_heap_frame(), "should be");
1406 _cont.last_frame().print_on(&ls);
1407 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1408 }
1409
1410 assert(_cont.chunk_invariant(), "");
1411 }
1412
1413 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1414 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1415 assert(t == JavaThread::current(), "");
1416 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1417 if (!_preempt) {
1418 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1419 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1420 }
1421 return true;
1422 }
1423 return false;
1424 }
1425
1426 class StackChunkAllocator : public MemAllocator {
1427 const size_t _stack_size;
1428 int _argsize_md;
1429 ContinuationWrapper& _continuation_wrapper;
1430 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1431 mutable bool _took_slow_path;
1432
1433 // Does the minimal amount of initialization needed for a TLAB allocation.
1434 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1435 virtual oop initialize(HeapWord* mem) const override {
1436 assert(_stack_size > 0, "");
1437 assert(_stack_size <= max_jint, "");
1438 assert(_word_size > _stack_size, "");
1439
1440 // zero out fields (but not the stack)
1441 const size_t hs = oopDesc::header_size();
1442 if (oopDesc::has_klass_gap()) {
1443 oopDesc::set_klass_gap(mem, 0);
1444 }
1445 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1446
1447 int bottom = (int)_stack_size - _argsize_md;
1448
1449 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1450 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1451 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1452
1453 return finish(mem);
1454 }
1455
1456 stackChunkOop allocate_fast() const {
1457 if (!UseTLAB) {
1458 return nullptr;
1459 }
1460
1461 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1462 if (mem == nullptr) {
1463 return nullptr;
1464 }
1465
1466 oop obj = initialize(mem);
1467 return stackChunkOopDesc::cast(obj);
1468 }
1469
1470 public:
1471 StackChunkAllocator(Klass* klass,
1472 size_t word_size,
1473 Thread* thread,
1474 size_t stack_size,
1475 int argsize_md,
1476 ContinuationWrapper& continuation_wrapper,
1477 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1478 : MemAllocator(klass, word_size, thread),
1479 _stack_size(stack_size),
1480 _argsize_md(argsize_md),
1481 _continuation_wrapper(continuation_wrapper),
1482 _jvmti_event_collector(jvmti_event_collector),
1483 _took_slow_path(false) {}
1484
1485 // Provides it's own, specialized allocation which skips instrumentation
1486 // if the memory can be allocated without going to a slow-path.
1487 stackChunkOop allocate() const {
1488 // First try to allocate without any slow-paths or instrumentation.
1489 stackChunkOop obj = allocate_fast();
1490 if (obj != nullptr) {
1491 return obj;
1492 }
1493
1494 // Now try full-blown allocation with all expensive operations,
1495 // including potentially safepoint operations.
1496 _took_slow_path = true;
1497
1498 // Protect unhandled Loom oops
1499 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1500
1501 // Can safepoint
1502 _jvmti_event_collector->start();
1503
1504 // Can safepoint
1505 return stackChunkOopDesc::cast(MemAllocator::allocate());
1506 }
1507
1508 bool took_slow_path() const {
1509 return _took_slow_path;
1510 }
1511 };
1512
1513 template <typename ConfigT>
1514 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1515 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1516
1517 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1518 size_t size_in_words = klass->instance_size(stack_size);
1519
1520 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1521 if (!_preempt) {
1522 throw_stack_overflow_on_humongous_chunk();
1523 }
1524 return nullptr;
1525 }
1526
1527 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1528 assert(current == JavaThread::current(), "should be current");
1529
1530 // Allocate the chunk.
1531 //
1532 // This might safepoint while allocating, but all safepointing due to
1533 // instrumentation have been deferred. This property is important for
1534 // some GCs, as this ensures that the allocated object is in the young
1535 // generation / newly allocated memory.
1536 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1537 stackChunkOop chunk = allocator.allocate();
1538
1539 if (chunk == nullptr) {
1540 return nullptr; // OOME
1541 }
1542
1543 // assert that chunk is properly initialized
1544 assert(chunk->stack_size() == (int)stack_size, "");
1545 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1546 assert(chunk->sp() == chunk->bottom(), "");
1547 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1548 assert(chunk->max_thawing_size() == 0, "");
1549 assert(chunk->pc() == nullptr, "");
1550 assert(chunk->is_empty(), "");
1551 assert(chunk->flags() == 0, "");
1552 assert(chunk->is_gc_mode() == false, "");
1553 assert(chunk->lockstack_size() == 0, "");
1554
1555 // fields are uninitialized
1556 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1557 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1558
1559 #if INCLUDE_ZGC
1560 if (UseZGC) {
1561 ZStackChunkGCData::initialize(chunk);
1562 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1563 _barriers = false;
1564 } else
1565 #endif
1566 #if INCLUDE_SHENANDOAHGC
1567 if (UseShenandoahGC) {
1568 _barriers = chunk->requires_barriers();
1569 } else
1570 #endif
1571 {
1572 if (!allocator.took_slow_path()) {
1573 // Guaranteed to be in young gen / newly allocated memory
1574 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1575 _barriers = false;
1576 } else {
1577 // Some GCs could put direct allocations in old gen for slow-path
1578 // allocations; need to explicitly check if that was the case.
1579 _barriers = chunk->requires_barriers();
1580 }
1581 }
1582
1583 if (_barriers) {
1584 log_develop_trace(continuations)("allocation requires barriers");
1585 }
1586
1587 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1588
1589 return chunk;
1590 }
1591
1592 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1593 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1594 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1595 }
1596
1597 class AnchorMark : public StackObj {
1598 JavaThread* _current;
1599 frame& _top_frame;
1600 intptr_t* _last_sp_from_frame;
1601 bool _is_interpreted;
1602
1603 public:
1604 AnchorMark(JavaThread* current, frame& f) : _current(current), _top_frame(f), _is_interpreted(false) {
1605 intptr_t* sp = anchor_mark_set_pd();
1606 set_anchor(_current, sp);
1607 }
1608 ~AnchorMark() {
1609 clear_anchor(_current);
1610 anchor_mark_clear_pd();
1611 }
1612 inline intptr_t* anchor_mark_set_pd();
1613 inline void anchor_mark_clear_pd();
1614 };
1615
1616 #if INCLUDE_JVMTI
1617 static int num_java_frames(ContinuationWrapper& cont) {
1618 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1619 int count = 0;
1620 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1621 count += chunk->num_java_frames();
1622 }
1623 return count;
1624 }
1625
1626 static void invalidate_jvmti_stack(JavaThread* thread) {
1627 if (thread->is_interp_only_mode()) {
1628 JvmtiThreadState *state = thread->jvmti_thread_state();
1629 if (state != nullptr)
1630 state->invalidate_cur_stack_depth();
1631 }
1632 }
1633
1634 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1635 if (JvmtiExport::can_post_frame_pop()) {
1636 int num_frames = num_java_frames(cont);
1637
1638 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1639 JvmtiExport::continuation_yield_cleanup(JavaThread::current(), num_frames);
1640 }
1641 invalidate_jvmti_stack(thread);
1642 }
1643
1644 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, Continuation::preempt_kind pk) {
1645 assert(current->vthread() != nullptr, "must be");
1646
1647 HandleMarkCleaner hm(current); // Cleanup vth and so._conth Handles
1648 Handle vth(current, current->vthread());
1649 ContinuationWrapper::SafepointOp so(current, cont);
1650
1651 AnchorMark am(current, top); // Set anchor so that the stack is walkable.
1652
1653 JRT_BLOCK
1654 JvmtiVTMSTransitionDisabler::VTMS_vthread_mount((jthread)vth.raw_value(), false);
1655
1656 if (current->pending_contended_entered_event()) {
1657 // No monitor JVMTI events for ObjectLocker case.
1658 if (pk != Continuation::object_locker) {
1659 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1660 }
1661 current->set_contended_entered_monitor(nullptr);
1662 }
1663 JRT_BLOCK_END
1664 }
1665 #endif // INCLUDE_JVMTI
1666
1667 #ifdef ASSERT
1668 static bool monitors_on_stack(JavaThread* thread) {
1669 ContinuationEntry* ce = thread->last_continuation();
1670 RegisterMap map(thread,
1671 RegisterMap::UpdateMap::include,
1672 RegisterMap::ProcessFrames::include,
1673 RegisterMap::WalkContinuation::skip);
1674 map.set_include_argument_oops(false);
1675 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
1676 if ((f.is_interpreted_frame() && ContinuationHelper::InterpretedFrame::is_owning_locks(f)) ||
1677 (f.is_compiled_frame() && ContinuationHelper::CompiledFrame::is_owning_locks(map.thread(), &map, f)) ||
1678 (f.is_native_frame() && ContinuationHelper::NativeFrame::is_owning_locks(map.thread(), f))) {
1679 return true;
1680 }
1681 }
1682 return false;
1683 }
1684
1685 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1686 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1687 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1688 bool FreezeBase::check_valid_fast_path() {
1689 ContinuationEntry* ce = _thread->last_continuation();
1690 RegisterMap map(_thread,
1691 RegisterMap::UpdateMap::skip,
1692 RegisterMap::ProcessFrames::skip,
1693 RegisterMap::WalkContinuation::skip);
1694 map.set_include_argument_oops(false);
1695 bool is_top_frame = true;
1696 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1697 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1698 return false;
1699 }
1700 }
1701 return true;
1702 }
1703
1704 static void verify_frame_kind(const frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr, const char** code_name_ptr, int* bci_ptr) {
1705 JavaThread* current = JavaThread::current();
1706 ResourceMark rm(current);
1707
1708 Method* m;
1709 const char* code_name;
1710 int bci;
1711 if (preempt_kind == Continuation::monitorenter) {
1712 assert(top.is_interpreted_frame() || top.is_runtime_frame(), "");
1713 bool at_sync_method;
1714 if (top.is_interpreted_frame()) {
1715 m = top.interpreter_frame_method();
1716 assert(!m->is_native() || m->is_synchronized(), "invalid method %s", m->external_name());
1717 address bcp = top.interpreter_frame_bcp();
1718 assert(bcp != 0 || m->is_native(), "");
1719 at_sync_method = m->is_synchronized() && (bcp == 0 || bcp == m->code_base());
1720 // bcp is advanced on monitorenter before making the VM call, adjust for that.
1721 bool at_sync_bytecode = bcp > m->code_base() && Bytecode(m, bcp - 1).code() == Bytecodes::Code::_monitorenter;
1722 assert(at_sync_method || at_sync_bytecode, "");
1723 bci = at_sync_method ? -1 : top.interpreter_frame_bci();
1724 } else {
1725 CodeBlob* cb = top.cb();
1726 RegisterMap reg_map(current,
1727 RegisterMap::UpdateMap::skip,
1728 RegisterMap::ProcessFrames::skip,
1729 RegisterMap::WalkContinuation::skip);
1730 frame fr = top.sender(®_map);
1731 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1732 compiledVFrame* cvf = compiledVFrame::cast(vf);
1733 m = cvf->method();
1734 bci = cvf->scope()->bci();
1735 at_sync_method = bci == SynchronizationEntryBCI;
1736 assert(!at_sync_method || m->is_synchronized(), "bci is %d but method %s is not synchronized", bci, m->external_name());
1737 bool is_c1_monitorenter = false, is_c2_monitorenter = false;
1738 COMPILER1_PRESENT(is_c1_monitorenter = cb == Runtime1::blob_for(C1StubId::monitorenter_id) ||
1739 cb == Runtime1::blob_for(C1StubId::monitorenter_nofpu_id);)
1740 COMPILER2_PRESENT(is_c2_monitorenter = cb == CodeCache::find_blob(OptoRuntime::complete_monitor_locking_Java());)
1741 assert(is_c1_monitorenter || is_c2_monitorenter, "wrong runtime stub frame");
1742 }
1743 code_name = at_sync_method ? "synchronized method" : "monitorenter";
1744 } else if (preempt_kind == Continuation::object_wait) {
1745 assert(top.is_interpreted_frame() || top.is_native_frame(), "");
1746 m = top.is_interpreted_frame() ? top.interpreter_frame_method() : top.cb()->as_nmethod()->method();
1747 assert(m->is_object_wait0(), "");
1748 bci = 0;
1749 code_name = "";
1750 } else {
1751 assert(preempt_kind == Continuation::object_locker, "invalid preempt kind");
1752 assert(top.is_interpreted_frame(), "");
1753 m = top.interpreter_frame_method();
1754 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
1755 Bytecodes::Code code = current_bytecode.code();
1756 assert(code == Bytecodes::Code::_new || code == Bytecodes::Code::_invokestatic ||
1757 (code == Bytecodes::Code::_getstatic || code == Bytecodes::Code::_putstatic), "invalid bytecode");
1758 bci = top.interpreter_frame_bci();
1759 code_name = Bytecodes::name(current_bytecode.code());
1760 }
1761 assert(bci >= 0 || m->is_synchronized(), "invalid bci:%d at method %s", bci, m->external_name());
1762
1763 if (m_ptr != nullptr) {
1764 *m_ptr = m;
1765 *code_name_ptr = code_name;
1766 *bci_ptr = bci;
1767 }
1768 }
1769
1770 static void log_preempt_after_freeze(ContinuationWrapper& cont) {
1771 JavaThread* current = cont.thread();
1772 StackChunkFrameStream<ChunkFrames::Mixed> sfs(cont.tail());
1773 frame top_frame = sfs.to_frame();
1774 bool at_init = current->at_preemptable_init();
1775 bool at_enter = current->current_pending_monitor() != nullptr;
1776 bool at_wait = current->current_waiting_monitor() != nullptr;
1777 assert((at_enter && !at_wait) || (!at_enter && at_wait), "");
1778 Continuation::preempt_kind pk = at_init ? Continuation::object_locker : at_enter ? Continuation::monitorenter : Continuation::object_wait;
1779
1780 Method* m = nullptr;
1781 const char* code_name = nullptr;
1782 int bci = InvalidFrameStateBci;
1783 verify_frame_kind(top_frame, pk, &m, &code_name, &bci);
1784 assert(m != nullptr && code_name != nullptr && bci != InvalidFrameStateBci, "should be set");
1785
1786 ResourceMark rm(current);
1787 if (bci < 0) {
1788 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " while synchronizing on %smethod %s", current->monitor_owner_id(), m->is_native() ? "native " : "", m->external_name());
1789 } else if (m->is_object_wait0()) {
1790 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at native method %s", current->monitor_owner_id(), m->external_name());
1791 } else {
1792 Klass* k = current->preempt_init_klass();
1793 assert(k != nullptr || !at_init, "");
1794 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at %s(bci:%d) in method %s %s%s", current->monitor_owner_id(),
1795 code_name, bci, m->external_name(), at_init ? "trying to initialize klass " : "", at_init ? k->external_name() : "");
1796 }
1797 }
1798 #endif // ASSERT
1799
1800 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1801 verify_continuation(cont.continuation());
1802 assert(!cont.is_empty(), "");
1803
1804 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1805 return freeze_ok;
1806 }
1807
1808 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1809 if (UNLIKELY(res != freeze_ok)) {
1810 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1811 verify_continuation(cont.continuation());
1812 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1813 return res;
1814 }
1815
1816 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1817 return freeze_epilog(cont);
1818 }
1819
1820 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1821 if (UNLIKELY(res != freeze_ok)) {
1822 verify_continuation(cont.continuation());
1823 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1824 return res;
1825 }
1826
1827 // Set up things so that on return to Java we jump to preempt stub.
1828 patch_return_pc_with_preempt_stub(old_last_frame);
1829 cont.tail()->set_preempted(true);
1830 DEBUG_ONLY(log_preempt_after_freeze(cont);)
1831 return freeze_epilog(cont);
1832 }
1833
1834 template<typename ConfigT, bool preempt>
1835 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1836 assert(!current->has_pending_exception(), "");
1837
1838 #ifdef ASSERT
1839 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1840 log_frames(current, false);
1841 #endif
1842
1843 CONT_JFR_ONLY(EventContinuationFreeze event;)
1844
1845 ContinuationEntry* entry = current->last_continuation();
1846
1847 oop oopCont = entry->cont_oop(current);
1848 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1849 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1850
1851 verify_continuation(oopCont);
1852 ContinuationWrapper cont(current, oopCont);
1853 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1854
1855 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1856
1857 assert(LockingMode != LM_LEGACY || (monitors_on_stack(current) == ((current->held_monitor_count() - current->jni_monitor_count()) > 0)),
1858 "Held monitor count and locks on stack invariant: " INT64_FORMAT " JNI: " INT64_FORMAT, (int64_t)current->held_monitor_count(), (int64_t)current->jni_monitor_count());
1859 assert(LockingMode == LM_LEGACY || (current->held_monitor_count() == 0 && current->jni_monitor_count() == 0),
1860 "Held monitor count should only be used for LM_LEGACY: " INT64_FORMAT " JNI: " INT64_FORMAT, (int64_t)current->held_monitor_count(), (int64_t)current->jni_monitor_count());
1861
1862 if (entry->is_pinned() || current->held_monitor_count() > 0) {
1863 log_develop_debug(continuations)("PINNED due to critical section/hold monitor");
1864 verify_continuation(cont.continuation());
1865 freeze_result res = entry->is_pinned() ? freeze_pinned_cs : freeze_pinned_monitor;
1866 if (!preempt) {
1867 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1868 }
1869 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1870 // Avoid Thread.yield() loops without safepoint polls.
1871 if (SafepointMechanism::should_process(current) && !preempt) {
1872 cont.done(); // allow safepoint
1873 ThreadInVMfromJava tivmfj(current);
1874 }
1875 return res;
1876 }
1877
1878 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1879
1880 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1881 bool fast = UseContinuationFastPath && current->cont_fastpath();
1882 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1883 freeze.freeze_fast_existing_chunk();
1884 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1885 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1886 }
1887
1888 if (preempt) {
1889 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1890 freeze.set_jvmti_event_collector(&jsoaec);
1891
1892 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1893
1894 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1895 preempt_epilog(cont, res, freeze.last_frame());
1896 return res;
1897 }
1898
1899 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1900 assert(current == JavaThread::current(), "must be current thread");
1901 JRT_BLOCK
1902 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1903 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1904 freeze.set_jvmti_event_collector(&jsoaec);
1905
1906 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1907
1908 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1909 freeze_epilog(current, cont, res);
1910 cont.done(); // allow safepoint in the transition back to Java
1911 return res;
1912 JRT_BLOCK_END
1913 }
1914
1915 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1916 ContinuationEntry* entry = thread->last_continuation();
1917 if (entry == nullptr) {
1918 return freeze_ok;
1919 }
1920 if (entry->is_pinned()) {
1921 return freeze_pinned_cs;
1922 } else if (thread->held_monitor_count() > 0) {
1923 return freeze_pinned_monitor;
1924 }
1925
1926 RegisterMap map(thread,
1927 RegisterMap::UpdateMap::include,
1928 RegisterMap::ProcessFrames::skip,
1929 RegisterMap::WalkContinuation::skip);
1930 map.set_include_argument_oops(false);
1931 frame f = thread->last_frame();
1932
1933 if (!safepoint) {
1934 f = f.sender(&map); // this is the yield frame
1935 } else { // safepoint yield
1936 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1937 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1938 #else
1939 Unimplemented();
1940 #endif
1941 if (!Interpreter::contains(f.pc())) {
1942 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1943 assert(f.oop_map() != nullptr, "must be");
1944 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1945 }
1946 }
1947
1948 while (true) {
1949 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1950 return freeze_pinned_native;
1951 }
1952
1953 f = f.sender(&map);
1954 if (!Continuation::is_frame_in_continuation(entry, f)) {
1955 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
1956 if (scope == cont_scope) {
1957 break;
1958 }
1959 intx monitor_count = entry->parent_held_monitor_count();
1960 entry = entry->parent();
1961 if (entry == nullptr) {
1962 break;
1963 }
1964 if (entry->is_pinned()) {
1965 return freeze_pinned_cs;
1966 } else if (monitor_count > 0) {
1967 return freeze_pinned_monitor;
1968 }
1969 }
1970 }
1971 return freeze_ok;
1972 }
1973
1974 /////////////// THAW ////
1975
1976 static int thaw_size(stackChunkOop chunk) {
1977 int size = chunk->max_thawing_size();
1978 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
1979 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
1980 return size;
1981 }
1982
1983 // make room on the stack for thaw
1984 // returns the size in bytes, or 0 on failure
1985 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
1986 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
1987
1988 assert(thread == JavaThread::current(), "");
1989
1990 ContinuationEntry* ce = thread->last_continuation();
1991 assert(ce != nullptr, "");
1992 oop continuation = ce->cont_oop(thread);
1993 assert(continuation == get_continuation(thread), "");
1994 verify_continuation(continuation);
1995
1996 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
1997 assert(chunk != nullptr, "");
1998
1999 // The tail can be empty because it might still be available for another freeze.
2000 // However, here we want to thaw, so we get rid of it (it will be GCed).
2001 if (UNLIKELY(chunk->is_empty())) {
2002 chunk = chunk->parent();
2003 assert(chunk != nullptr, "");
2004 assert(!chunk->is_empty(), "");
2005 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
2006 }
2007
2008 // Verification
2009 chunk->verify();
2010 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
2011
2012 // Only make space for the last chunk because we only thaw from the last chunk
2013 int size = thaw_size(chunk) << LogBytesPerWord;
2014
2015 const address bottom = (address)thread->last_continuation()->entry_sp();
2016 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
2017 // for the Java frames in the check below.
2018 if (!stack_overflow_check(thread, size + 300, bottom)) {
2019 return 0;
2020 }
2021
2022 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
2023 p2i(bottom), p2i(bottom - size), size);
2024 return size;
2025 }
2026
2027 class ThawBase : public StackObj {
2028 protected:
2029 JavaThread* _thread;
2030 ContinuationWrapper& _cont;
2031 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
2032
2033 intptr_t* _fastpath;
2034 bool _barriers;
2035 bool _preempted_case;
2036 bool _process_args_at_top;
2037 intptr_t* _top_unextended_sp_before_thaw;
2038 int _align_size;
2039 DEBUG_ONLY(intptr_t* _top_stack_address);
2040
2041 // Only used for some preemption cases.
2042 ObjectMonitor* _monitor;
2043
2044 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
2045
2046 NOT_PRODUCT(int _frames;)
2047
2048 protected:
2049 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2050 _thread(thread), _cont(cont),
2051 _fastpath(nullptr) {
2052 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2053 assert (cont.tail() != nullptr, "no last chunk");
2054 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2055 }
2056
2057 void clear_chunk(stackChunkOop chunk);
2058 template<bool check_stub>
2059 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2060 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2061
2062 void thaw_lockstack(stackChunkOop chunk);
2063
2064 // fast path
2065 inline void prefetch_chunk_pd(void* start, int size_words);
2066 void patch_return(intptr_t* sp, bool is_last);
2067
2068 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
2069 inline intptr_t* push_cleanup_continuation();
2070 inline intptr_t* push_preempt_adapter();
2071 intptr_t* redo_vmcall(JavaThread* current, frame& top);
2072 void throw_interrupted_exception(JavaThread* current, frame& top);
2073
2074 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
2075 void finish_thaw(frame& f);
2076
2077 private:
2078 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2079 void finalize_thaw(frame& entry, int argsize);
2080
2081 inline bool seen_by_gc();
2082
2083 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2084 inline void after_thaw_java_frame(const frame& f, bool bottom);
2085 inline void patch(frame& f, const frame& caller, bool bottom);
2086 void clear_bitmap_bits(address start, address end);
2087
2088 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top);
2089 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2090 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2091 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
2092
2093 void push_return_frame(frame& f);
2094 inline frame new_entry_frame();
2095 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom);
2096 inline void patch_pd(frame& f, const frame& sender);
2097 inline void patch_pd(frame& f, intptr_t* caller_sp);
2098 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2099
2100 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2101
2102 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2103
2104 public:
2105 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2106 };
2107
2108 template <typename ConfigT>
2109 class Thaw : public ThawBase {
2110 public:
2111 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2112
2113 inline bool can_thaw_fast(stackChunkOop chunk) {
2114 return !_barriers
2115 && _thread->cont_fastpath_thread_state()
2116 && !chunk->has_thaw_slowpath_condition()
2117 && !PreserveFramePointer;
2118 }
2119
2120 inline intptr_t* thaw(Continuation::thaw_kind kind);
2121 template<bool check_stub = false>
2122 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2123 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2124 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2125 };
2126
2127 template <typename ConfigT>
2128 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2129 verify_continuation(_cont.continuation());
2130 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2131 assert(!_cont.is_empty(), "");
2132
2133 stackChunkOop chunk = _cont.tail();
2134 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2135 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2136
2137 _barriers = chunk->requires_barriers();
2138 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2139 : thaw_slow(chunk, kind);
2140 }
2141
2142 class ReconstructedStack : public StackObj {
2143 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2144 int _thaw_size;
2145 int _argsize;
2146 public:
2147 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2148 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2149 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2150 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2151 // We're at most one alignment word away from entrySP
2152 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2153 }
2154
2155 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2156
2157 // top and bottom stack pointers
2158 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2159 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2160
2161 // several operations operate on the totality of the stack being reconstructed,
2162 // including the metadata words
2163 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2164 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2165 };
2166
2167 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2168 chunk->set_sp(chunk->bottom());
2169 chunk->set_max_thawing_size(0);
2170 }
2171
2172 template<bool check_stub>
2173 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2174 bool empty = false;
2175 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2176 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2177 assert(chunk_sp == f.sp(), "");
2178 assert(chunk_sp == f.unextended_sp(), "");
2179
2180 int frame_size = f.cb()->frame_size();
2181 argsize = f.stack_argsize();
2182
2183 assert(!f.is_stub() || check_stub, "");
2184 if (check_stub && f.is_stub()) {
2185 // If we don't thaw the top compiled frame too, after restoring the saved
2186 // registers back in Java, we would hit the return barrier to thaw one more
2187 // frame effectively overwriting the restored registers during that call.
2188 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2189 assert(!f.is_done(), "");
2190
2191 f.get_cb();
2192 assert(f.is_compiled(), "");
2193 frame_size += f.cb()->frame_size();
2194 argsize = f.stack_argsize();
2195
2196 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2197 // The caller of the runtime stub when the continuation is preempted is not at a
2198 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2199 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2200 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2201 }
2202 }
2203
2204 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2205 empty = f.is_done();
2206 assert(!empty || argsize == chunk->argsize(), "");
2207
2208 if (empty) {
2209 clear_chunk(chunk);
2210 } else {
2211 chunk->set_sp(chunk->sp() + frame_size);
2212 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2213 // We set chunk->pc to the return pc into the next frame
2214 chunk->set_pc(f.pc());
2215 #ifdef ASSERT
2216 {
2217 intptr_t* retaddr_slot = (chunk_sp
2218 + frame_size
2219 - frame::sender_sp_ret_address_offset());
2220 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2221 "unexpected pc");
2222 }
2223 #endif
2224 }
2225 assert(empty == chunk->is_empty(), "");
2226 // returns the size required to store the frame on stack, and because it is a
2227 // compiled frame, it must include a copy of the arguments passed by the caller
2228 return frame_size + argsize + frame::metadata_words_at_top;
2229 }
2230
2231 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2232 int lockStackSize = chunk->lockstack_size();
2233 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2234
2235 oop tmp_lockstack[LockStack::CAPACITY];
2236 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2237 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2238
2239 chunk->set_lockstack_size(0);
2240 chunk->set_has_lockstack(false);
2241 }
2242
2243 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2244 assert(to >= _top_stack_address, "overwrote past thawing space"
2245 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2246 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2247 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2248 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2249 }
2250
2251 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2252 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2253
2254 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2255 ContinuationHelper::patch_return_address_at(
2256 sp - frame::sender_sp_ret_address_offset(),
2257 pc);
2258 }
2259
2260 template <typename ConfigT>
2261 template<bool check_stub>
2262 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2263 assert(chunk == _cont.tail(), "");
2264 assert(!chunk->has_mixed_frames(), "");
2265 assert(!chunk->requires_barriers(), "");
2266 assert(!chunk->has_bitmap(), "");
2267 assert(!_thread->is_interp_only_mode(), "");
2268
2269 LogTarget(Trace, continuations) lt;
2270 if (lt.develop_is_enabled()) {
2271 LogStream ls(lt);
2272 ls.print_cr("thaw_fast");
2273 chunk->print_on(true, &ls);
2274 }
2275
2276 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2277 static const int threshold = 500; // words
2278
2279 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2280 int argsize, thaw_size;
2281
2282 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2283
2284 bool partial, empty;
2285 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2286 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2287
2288 partial = false;
2289 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2290 clear_chunk(chunk);
2291 thaw_size = full_chunk_size;
2292 empty = true;
2293 } else { // thaw a single frame
2294 partial = true;
2295 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2296 empty = chunk->is_empty();
2297 }
2298
2299 // Are we thawing the last frame(s) in the continuation
2300 const bool is_last = empty && chunk->parent() == nullptr;
2301 assert(!is_last || argsize == 0, "");
2302
2303 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2304 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2305
2306 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2307
2308 // also copy metadata words at frame bottom
2309 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2310
2311 // update the ContinuationEntry
2312 _cont.set_argsize(argsize);
2313 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2314 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2315
2316 // install the return barrier if not last frame, or the entry's pc if last
2317 patch_return(rs.bottom_sp(), is_last);
2318
2319 // insert the back links from callee to caller frames
2320 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2321
2322 assert(is_last == _cont.is_empty(), "");
2323 assert(_cont.chunk_invariant(), "");
2324
2325 #if CONT_JFR
2326 EventContinuationThawFast e;
2327 if (e.should_commit()) {
2328 e.set_id(cast_from_oop<u8>(chunk));
2329 e.set_size(thaw_size << LogBytesPerWord);
2330 e.set_full(!partial);
2331 e.commit();
2332 }
2333 #endif
2334
2335 #ifdef ASSERT
2336 set_anchor(_thread, rs.sp());
2337 log_frames(_thread);
2338 if (LoomDeoptAfterThaw) {
2339 do_deopt_after_thaw(_thread);
2340 }
2341 clear_anchor(_thread);
2342 #endif
2343
2344 return rs.sp();
2345 }
2346
2347 inline bool ThawBase::seen_by_gc() {
2348 return _barriers || _cont.tail()->is_gc_mode();
2349 }
2350
2351 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2352 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2353 if (UseZGC || UseShenandoahGC) {
2354 chunk->relativize_derived_pointers_concurrently();
2355 }
2356 #endif
2357 }
2358
2359 template <typename ConfigT>
2360 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2361 Continuation::preempt_kind preempt_kind;
2362 bool retry_fast_path = false;
2363
2364 _process_args_at_top = false;
2365 _preempted_case = chunk->preempted();
2366 if (_preempted_case) {
2367 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2368 if (waiter != nullptr) {
2369 // Mounted again after preemption. Resume the pending monitor operation,
2370 // which will be either a monitorenter or Object.wait() call.
2371 ObjectMonitor* mon = waiter->monitor();
2372 preempt_kind = waiter->is_wait() ? Continuation::object_wait : Continuation::monitorenter;
2373
2374 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2375 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2376 if (!mon_acquired) {
2377 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2378 log_trace(continuations, tracking)("Failed to acquire monitor, unmounting again");
2379 return push_cleanup_continuation();
2380 }
2381 _monitor = mon; // remember monitor since we might need it on handle_preempted_continuation()
2382 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2383 } else {
2384 // Preemption cancelled in moniterenter case. We actually acquired
2385 // the monitor after freezing all frames so nothing to do. In case
2386 // of preemption on ObjectLocker during klass init, we released the
2387 // monitor already at ~ObjectLocker so here we just set _monitor to
2388 // nullptr so we know there is no need to release it later.
2389 preempt_kind = Continuation::monitorenter;
2390 _monitor = nullptr;
2391 }
2392
2393 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2394 relativize_chunk_concurrently(chunk);
2395
2396 if (chunk->at_klass_init()) {
2397 preempt_kind = Continuation::object_locker;
2398 chunk->set_at_klass_init(false);
2399 _process_args_at_top = chunk->has_args_at_top();
2400 if (_process_args_at_top) chunk->set_has_args_at_top(false);
2401 }
2402 chunk->set_preempted(false);
2403 retry_fast_path = true;
2404 } else {
2405 relativize_chunk_concurrently(chunk);
2406 }
2407
2408 // On first thaw after freeze restore oops to the lockstack if any.
2409 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2410 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2411 thaw_lockstack(chunk);
2412 retry_fast_path = true;
2413 }
2414
2415 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2416 // and FLAG_PREEMPTED flags from the stackChunk.
2417 if (retry_fast_path && can_thaw_fast(chunk)) {
2418 intptr_t* sp = thaw_fast<true>(chunk);
2419 if (_preempted_case) {
2420 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2421 }
2422 return sp;
2423 }
2424
2425 LogTarget(Trace, continuations) lt;
2426 if (lt.develop_is_enabled()) {
2427 LogStream ls(lt);
2428 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2429 chunk->print_on(true, &ls);
2430 }
2431
2432 #if CONT_JFR
2433 EventContinuationThawSlow e;
2434 if (e.should_commit()) {
2435 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2436 e.commit();
2437 }
2438 #endif
2439
2440 DEBUG_ONLY(_frames = 0;)
2441 _align_size = 0;
2442 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2443
2444 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2445 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2446
2447 frame heap_frame = _stream.to_frame();
2448 if (lt.develop_is_enabled()) {
2449 LogStream ls(lt);
2450 ls.print_cr("top hframe before (thaw):");
2451 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2452 heap_frame.print_value_on(&ls);
2453 }
2454
2455 frame caller; // the thawed caller on the stack
2456 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2457 finish_thaw(caller); // caller is now the topmost thawed frame
2458 _cont.write();
2459
2460 assert(_cont.chunk_invariant(), "");
2461
2462 JVMTI_ONLY(invalidate_jvmti_stack(_thread));
2463
2464 _thread->set_cont_fastpath(_fastpath);
2465
2466 intptr_t* sp = caller.sp();
2467
2468 if (_preempted_case) {
2469 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2470 }
2471 return sp;
2472 }
2473
2474 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2475 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2476 assert(!_cont.is_empty(), "no more frames");
2477 assert(num_frames > 0, "");
2478 assert(!heap_frame.is_empty(), "");
2479
2480 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2481 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2482 } else if (!heap_frame.is_interpreted_frame()) {
2483 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2484 } else {
2485 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames, top_on_preempt_case);
2486 }
2487 }
2488
2489 template<typename FKind>
2490 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2491 assert(num_frames > 0, "");
2492
2493 DEBUG_ONLY(_frames++;)
2494
2495 int argsize = _stream.stack_argsize();
2496
2497 _stream.next(SmallRegisterMap::instance_no_args());
2498 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2499
2500 // we never leave a compiled caller of an interpreted frame as the top frame in the chunk
2501 // as it makes detecting that situation and adjusting unextended_sp tricky
2502 if (num_frames == 1 && !_stream.is_done() && FKind::interpreted && _stream.is_compiled()) {
2503 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2504 num_frames++;
2505 }
2506
2507 if (num_frames == 1 || _stream.is_done()) { // end recursion
2508 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2509 return true; // bottom
2510 } else { // recurse
2511 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2512 return false;
2513 }
2514 }
2515
2516 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2517 stackChunkOop chunk = _cont.tail();
2518
2519 if (!_stream.is_done()) {
2520 assert(_stream.sp() >= chunk->sp_address(), "");
2521 chunk->set_sp(chunk->to_offset(_stream.sp()));
2522 chunk->set_pc(_stream.pc());
2523 } else {
2524 chunk->set_sp(chunk->bottom());
2525 chunk->set_pc(nullptr);
2526 }
2527 assert(_stream.is_done() == chunk->is_empty(), "");
2528
2529 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2530 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2531
2532 _cont.set_argsize(argsize);
2533 entry = new_entry_frame();
2534
2535 assert(entry.sp() == _cont.entrySP(), "");
2536 assert(Continuation::is_continuation_enterSpecial(entry), "");
2537 assert(_cont.is_entry_frame(entry), "");
2538 }
2539
2540 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2541 LogTarget(Trace, continuations) lt;
2542 if (lt.develop_is_enabled()) {
2543 LogStream ls(lt);
2544 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2545 assert(hf.is_heap_frame(), "should be");
2546 hf.print_value_on(&ls);
2547 }
2548 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2549 }
2550
2551 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2552 #ifdef ASSERT
2553 LogTarget(Trace, continuations) lt;
2554 if (lt.develop_is_enabled()) {
2555 LogStream ls(lt);
2556 ls.print_cr("thawed frame:");
2557 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2558 }
2559 #endif
2560 }
2561
2562 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom) {
2563 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2564 if (bottom) {
2565 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2566 : StubRoutines::cont_returnBarrier());
2567 } else {
2568 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2569 // If the caller is not deoptimized, pc is unchanged.
2570 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc());
2571 }
2572
2573 patch_pd(f, caller);
2574
2575 if (f.is_interpreted_frame()) {
2576 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2577 }
2578
2579 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2580 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2581 }
2582
2583 void ThawBase::clear_bitmap_bits(address start, address end) {
2584 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2585 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2586
2587 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2588 // or they will keep objects that are otherwise unreachable alive.
2589
2590 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2591 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2592 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2593 // anyways and we assert that before returning.
2594 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2595 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2596 stackChunkOop chunk = _cont.tail();
2597 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2598 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2599 }
2600
2601 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2602 frame top(sp);
2603 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2604 DEBUG_ONLY(verify_frame_kind(top, preempt_kind);)
2605 NOT_PRODUCT(int64_t tid = _thread->monitor_owner_id();)
2606
2607 #if INCLUDE_JVMTI
2608 // Finish the VTMS transition.
2609 assert(_thread->is_in_VTMS_transition(), "must be");
2610 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2611 if (is_vthread) {
2612 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2613 jvmti_mount_end(_thread, _cont, top, preempt_kind);
2614 } else {
2615 _thread->set_is_in_VTMS_transition(false);
2616 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2617 }
2618 }
2619 #endif
2620
2621 if (fast_case) {
2622 // If we thawed in the slow path the runtime stub/native wrapper frame already
2623 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2624 // we copied the original fp at the time of freeze which now will have to be fixed.
2625 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2626 int fsize = top.cb()->frame_size();
2627 patch_pd(top, sp + fsize);
2628 }
2629
2630 if (preempt_kind == Continuation::object_wait) {
2631 // Check now if we need to throw IE exception.
2632 bool throw_ie = _thread->pending_interrupted_exception();
2633 if (throw_ie) {
2634 throw_interrupted_exception(_thread, top);
2635 _thread->set_pending_interrupted_exception(false);
2636 }
2637 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT" after preemption on Object.wait%s", tid, throw_ie ? "(throwing IE)" : "");
2638 } else if (preempt_kind == Continuation::monitorenter) {
2639 if (top.is_runtime_frame()) {
2640 // The continuation might now run on a different platform thread than the previous time so
2641 // we need to adjust the current thread saved in the stub frame before restoring registers.
2642 JavaThread** thread_addr = frame::saved_thread_address(top);
2643 if (thread_addr != nullptr) *thread_addr = _thread;
2644 }
2645 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT " after preemption on monitorenter", tid);
2646 } else {
2647 // We need to redo the original call into the VM. First though, we need
2648 // to exit the monitor we just acquired (except on preemption cancelled
2649 // case where it was already released).
2650 assert(preempt_kind == Continuation::object_locker, "");
2651 if (_monitor != nullptr) _monitor->exit(_thread);
2652 sp = redo_vmcall(_thread, top);
2653 }
2654 return sp;
2655 }
2656
2657 intptr_t* ThawBase::redo_vmcall(JavaThread* current, frame& top) {
2658 assert(!current->preempting(), "");
2659 NOT_PRODUCT(int64_t tid = current->monitor_owner_id();)
2660 intptr_t* sp = top.sp();
2661
2662 {
2663 HandleMarkCleaner hmc(current); // Cleanup so._conth Handle
2664 ContinuationWrapper::SafepointOp so(current, _cont);
2665 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2666
2667 Method* m = top.interpreter_frame_method();
2668 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
2669 Bytecodes::Code code = current_bytecode.code();
2670 log_develop_trace(continuations, preempt)("Redoing InterpreterRuntime::%s for " INT64_FORMAT, code == Bytecodes::Code::_new ? "_new" : "resolve_from_cache", tid);
2671
2672 // These InterpreterRuntime entry points use JRT_ENTRY which uses a HandleMarkCleaner.
2673 // Create a HandeMark to avoid destroying so._conth.
2674 HandleMark hm(current);
2675 if (code == Bytecodes::Code::_new) {
2676 InterpreterRuntime::_new(current, m->constants(), current_bytecode.get_index_u2(code));
2677 } else {
2678 InterpreterRuntime::resolve_from_cache(current, code);
2679 }
2680 }
2681
2682 if (current->preempting()) {
2683 // Preempted again so we just arrange to return to preempt stub to unmount.
2684 sp = push_preempt_adapter();
2685 current->set_preempt_alternate_return(nullptr);
2686 bool cancelled = current->preemption_cancelled();
2687 if (cancelled) {
2688 // Instead of calling thaw again from the preempt stub just unmount anyways with
2689 // state of YIELDING. This will give a chance for other vthreads to run while
2690 // minimizing repeated loops of "thaw->redo_vmcall->try_preempt->preemption_cancelled->thaw..."
2691 // in case of multiple vthreads contending for the same init_lock().
2692 current->set_preemption_cancelled(false);
2693 oop vthread = current->vthread();
2694 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
2695 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::YIELDING);
2696 }
2697 log_develop_trace(continuations, preempt)("Preempted " INT64_FORMAT " again%s", tid, cancelled ? "(preemption cancelled, setting state to YIELDING)" : "");
2698 } else {
2699 log_develop_trace(continuations, preempt)("Call succesful, resuming " INT64_FORMAT, tid);
2700 }
2701 return sp;
2702 }
2703
2704 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2705 HandleMarkCleaner hm(current); // Cleanup so._conth Handle
2706 ContinuationWrapper::SafepointOp so(current, _cont);
2707 // Since we might safepoint set the anchor so that the stack can be walked.
2708 set_anchor(current, top.sp());
2709 JRT_BLOCK
2710 THROW(vmSymbols::java_lang_InterruptedException());
2711 JRT_BLOCK_END
2712 clear_anchor(current);
2713 }
2714
2715 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top) {
2716 assert(hf.is_interpreted_frame(), "");
2717
2718 if (UNLIKELY(seen_by_gc())) {
2719 if (is_top && _process_args_at_top) {
2720 log_trace(continuations, tracking)("Processing arguments in recurse_thaw_interpreted_frame");
2721 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_with_args());
2722 } else {
2723 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2724 }
2725 }
2726
2727 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2728
2729 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2730
2731 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2732
2733 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2734
2735 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2736 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2737 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2738 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2739
2740 assert(hf.is_heap_frame(), "should be");
2741 assert(!f.is_heap_frame(), "should not be");
2742
2743 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2744 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2745
2746 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2747 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2748 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2749
2750 // Make sure the relativized locals is already set.
2751 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2752
2753 derelativize_interpreted_frame_metadata(hf, f);
2754 patch(f, caller, is_bottom_frame);
2755
2756 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2757 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2758
2759 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2760
2761 maybe_set_fastpath(f.sp());
2762
2763 Method* m = hf.interpreter_frame_method();
2764 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2765 const int locals = m->max_locals();
2766
2767 if (!is_bottom_frame) {
2768 // can only fix caller once this frame is thawed (due to callee saved regs)
2769 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2770 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2771 assert(hf.is_heap_frame(), "should be");
2772 address start = (address)(heap_frame_bottom - locals);
2773 address end = (address)heap_frame_bottom;
2774 clear_bitmap_bits(start, end);
2775 }
2776
2777 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2778 caller = f;
2779 }
2780
2781 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2782 assert(hf.is_compiled_frame(), "");
2783 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2784
2785 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2786 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2787 }
2788
2789 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2790
2791 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2792
2793 assert(caller.sp() == caller.unextended_sp(), "");
2794
2795 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2796 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2797 }
2798
2799 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2800 // yet laid out in the stack, and so the original_pc is not stored in it.
2801 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2802 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame);
2803 intptr_t* const stack_frame_top = f.sp();
2804 intptr_t* const heap_frame_top = hf.unextended_sp();
2805
2806 const int added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2807 int fsize = ContinuationHelper::CompiledFrame::size(hf) + added_argsize;
2808 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2809
2810 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2811 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2812 // copy metadata, except the metadata at the top of the (unextended) entry frame
2813 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2814
2815 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2816 // (we might have one padding word for alignment)
2817 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2818 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2819
2820 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2821
2822 patch(f, caller, is_bottom_frame);
2823
2824 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2825 assert(!f.is_deoptimized_frame(), "");
2826 if (hf.is_deoptimized_frame()) {
2827 maybe_set_fastpath(f.sp());
2828 } else if (_thread->is_interp_only_mode()
2829 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2830 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2831 // cannot rely on nmethod patching for deopt.
2832 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2833
2834 log_develop_trace(continuations)("Deoptimizing thawed frame");
2835 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2836
2837 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2838 assert(f.is_deoptimized_frame(), "");
2839 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2840 maybe_set_fastpath(f.sp());
2841 }
2842
2843 if (!is_bottom_frame) {
2844 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2845 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2846 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2847 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2848 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2849 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2850 clear_bitmap_bits(start, start + argsize_in_bytes);
2851 }
2852
2853 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2854 caller = f;
2855 }
2856
2857 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2858 DEBUG_ONLY(_frames++;)
2859
2860 if (UNLIKELY(seen_by_gc())) {
2861 // Process the stub's caller here since we might need the full map.
2862 RegisterMap map(nullptr,
2863 RegisterMap::UpdateMap::include,
2864 RegisterMap::ProcessFrames::skip,
2865 RegisterMap::WalkContinuation::skip);
2866 map.set_include_argument_oops(false);
2867 _stream.next(&map);
2868 assert(!_stream.is_done(), "");
2869 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2870 } else {
2871 _stream.next(SmallRegisterMap::instance_no_args());
2872 assert(!_stream.is_done(), "");
2873 }
2874
2875 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2876
2877 assert(caller.is_compiled_frame(), "");
2878 assert(caller.sp() == caller.unextended_sp(), "");
2879
2880 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2881
2882 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2883 intptr_t* stack_frame_top = f.sp();
2884 intptr_t* heap_frame_top = hf.sp();
2885 int fsize = ContinuationHelper::StubFrame::size(hf);
2886
2887 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2888 fsize + frame::metadata_words);
2889
2890 patch(f, caller, false /*is_bottom_frame*/);
2891
2892 // can only fix caller once this frame is thawed (due to callee saved regs)
2893 RegisterMap map(nullptr,
2894 RegisterMap::UpdateMap::include,
2895 RegisterMap::ProcessFrames::skip,
2896 RegisterMap::WalkContinuation::skip);
2897 map.set_include_argument_oops(false);
2898 f.oop_map()->update_register_map(&f, &map);
2899 ContinuationHelper::update_register_map_with_callee(caller, &map);
2900 _cont.tail()->fix_thawed_frame(caller, &map);
2901
2902 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2903 caller = f;
2904 }
2905
2906 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2907 assert(hf.is_native_frame(), "");
2908 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2909
2910 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2911 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2912 }
2913
2914 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
2915 assert(!is_bottom_frame, "");
2916
2917 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2918
2919 assert(caller.sp() == caller.unextended_sp(), "");
2920
2921 if (caller.is_interpreted_frame()) {
2922 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
2923 }
2924
2925 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2926 // yet laid out in the stack, and so the original_pc is not stored in it.
2927 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2928 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
2929 intptr_t* const stack_frame_top = f.sp();
2930 intptr_t* const heap_frame_top = hf.unextended_sp();
2931
2932 int fsize = ContinuationHelper::NativeFrame::size(hf);
2933 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
2934
2935 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2936 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2937 int sz = fsize + frame::metadata_words_at_bottom;
2938
2939 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2940
2941 patch(f, caller, false /* bottom */);
2942
2943 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2944 assert(!f.is_deoptimized_frame(), "");
2945 assert(!hf.is_deoptimized_frame(), "");
2946 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
2947
2948 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2949 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2950
2951 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
2952 caller = f;
2953 }
2954
2955 void ThawBase::finish_thaw(frame& f) {
2956 stackChunkOop chunk = _cont.tail();
2957
2958 if (chunk->is_empty()) {
2959 // Only remove chunk from list if it can't be reused for another freeze
2960 if (seen_by_gc()) {
2961 _cont.set_tail(chunk->parent());
2962 } else {
2963 chunk->set_has_mixed_frames(false);
2964 }
2965 chunk->set_max_thawing_size(0);
2966 } else {
2967 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
2968 }
2969 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
2970
2971 if (!is_aligned(f.sp(), frame::frame_alignment)) {
2972 assert(f.is_interpreted_frame(), "");
2973 f.set_sp(align_down(f.sp(), frame::frame_alignment));
2974 }
2975 push_return_frame(f);
2976 // can only fix caller after push_return_frame (due to callee saved regs)
2977 if (_process_args_at_top) {
2978 log_trace(continuations, tracking)("Processing arguments in finish_thaw");
2979 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_with_args());
2980 } else {
2981 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_no_args());
2982 }
2983
2984 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
2985
2986 log_develop_trace(continuations)("thawed %d frames", _frames);
2987
2988 LogTarget(Trace, continuations) lt;
2989 if (lt.develop_is_enabled()) {
2990 LogStream ls(lt);
2991 ls.print_cr("top hframe after (thaw):");
2992 _cont.last_frame().print_value_on(&ls);
2993 }
2994 }
2995
2996 void ThawBase::push_return_frame(frame& f) { // see generate_cont_thaw
2997 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
2998 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
2999
3000 LogTarget(Trace, continuations) lt;
3001 if (lt.develop_is_enabled()) {
3002 LogStream ls(lt);
3003 ls.print_cr("push_return_frame");
3004 f.print_value_on(&ls);
3005 }
3006
3007 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
3008 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
3009 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
3010 ContinuationHelper::push_pd(f);
3011
3012 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
3013 }
3014
3015 // returns new top sp
3016 // called after preparations (stack overflow check and making room)
3017 template<typename ConfigT>
3018 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
3019 assert(thread == JavaThread::current(), "Must be current thread");
3020
3021 CONT_JFR_ONLY(EventContinuationThaw event;)
3022
3023 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
3024
3025 ContinuationEntry* entry = thread->last_continuation();
3026 assert(entry != nullptr, "");
3027 oop oopCont = entry->cont_oop(thread);
3028
3029 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
3030 assert(oopCont == get_continuation(thread), "");
3031 verify_continuation(oopCont);
3032
3033 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
3034
3035 ContinuationWrapper cont(thread, oopCont);
3036 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
3037
3038 #ifdef ASSERT
3039 set_anchor_to_entry(thread, cont.entry());
3040 log_frames(thread);
3041 clear_anchor(thread);
3042 #endif
3043
3044 Thaw<ConfigT> thw(thread, cont);
3045 intptr_t* const sp = thw.thaw(kind);
3046 assert(is_aligned(sp, frame::frame_alignment), "");
3047 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp);)
3048
3049 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
3050
3051 verify_continuation(cont.continuation());
3052 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
3053
3054 return sp;
3055 }
3056
3057 #ifdef ASSERT
3058 static void do_deopt_after_thaw(JavaThread* thread) {
3059 int i = 0;
3060 StackFrameStream fst(thread, true, false);
3061 fst.register_map()->set_include_argument_oops(false);
3062 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3063 for (; !fst.is_done(); fst.next()) {
3064 if (fst.current()->cb()->is_nmethod()) {
3065 nmethod* nm = fst.current()->cb()->as_nmethod();
3066 if (!nm->method()->is_continuation_native_intrinsic()) {
3067 nm->make_deoptimized();
3068 }
3069 }
3070 }
3071 }
3072
3073 class ThawVerifyOopsClosure: public OopClosure {
3074 intptr_t* _p;
3075 outputStream* _st;
3076 bool is_good_oop(oop o) {
3077 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
3078 }
3079 public:
3080 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
3081 intptr_t* p() { return _p; }
3082 void reset() { _p = nullptr; }
3083
3084 virtual void do_oop(oop* p) {
3085 oop o = *p;
3086 if (o == nullptr || is_good_oop(o)) {
3087 return;
3088 }
3089 _p = (intptr_t*)p;
3090 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
3091 }
3092 virtual void do_oop(narrowOop* p) {
3093 oop o = RawAccess<>::oop_load(p);
3094 if (o == nullptr || is_good_oop(o)) {
3095 return;
3096 }
3097 _p = (intptr_t*)p;
3098 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
3099 }
3100 };
3101
3102 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
3103 assert(thread->has_last_Java_frame(), "");
3104
3105 ResourceMark rm;
3106 ThawVerifyOopsClosure cl(st);
3107 NMethodToOopClosure cf(&cl, false);
3108
3109 StackFrameStream fst(thread, true, false);
3110 fst.register_map()->set_include_argument_oops(false);
3111 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3112 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
3113 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
3114 st->print_cr(">>> do_verify_after_thaw deopt");
3115 fst.current()->deoptimize(nullptr);
3116 fst.current()->print_on(st);
3117 }
3118
3119 fst.current()->oops_do(&cl, &cf, fst.register_map());
3120 if (cl.p() != nullptr) {
3121 frame fr = *fst.current();
3122 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
3123 fr.print_on(st);
3124 if (!fr.is_interpreted_frame()) {
3125 st->print_cr("size: %d argsize: %d",
3126 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
3127 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
3128 }
3129 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
3130 if (reg != nullptr) {
3131 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
3132 }
3133 cl.reset();
3134 DEBUG_ONLY(thread->print_frame_layout();)
3135 if (chunk != nullptr) {
3136 chunk->print_on(true, st);
3137 }
3138 return false;
3139 }
3140 }
3141 return true;
3142 }
3143
3144 static void log_frames(JavaThread* thread, bool dolog) {
3145 const static int show_entry_callers = 3;
3146 LogTarget(Trace, continuations, tracking) lt;
3147 if (!lt.develop_is_enabled() || !dolog) {
3148 return;
3149 }
3150 LogStream ls(lt);
3151
3152 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3153 if (!thread->has_last_Java_frame()) {
3154 ls.print_cr("NO ANCHOR!");
3155 }
3156
3157 RegisterMap map(thread,
3158 RegisterMap::UpdateMap::include,
3159 RegisterMap::ProcessFrames::include,
3160 RegisterMap::WalkContinuation::skip);
3161 map.set_include_argument_oops(false);
3162
3163 if (false) {
3164 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3165 f.print_on(&ls);
3166 }
3167 } else {
3168 map.set_skip_missing(true);
3169 ResetNoHandleMark rnhm;
3170 ResourceMark rm;
3171 HandleMark hm(Thread::current());
3172 FrameValues values;
3173
3174 int i = 0;
3175 int post_entry = -1;
3176 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3177 f.describe(values, i, &map, i == 0);
3178 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3179 post_entry++;
3180 if (post_entry >= show_entry_callers)
3181 break;
3182 }
3183 values.print_on(thread, &ls);
3184 }
3185
3186 ls.print_cr("======= end frames =========");
3187 }
3188
3189 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp) {
3190 intptr_t* sp0 = sp;
3191 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
3192
3193 bool preempted = false;
3194 stackChunkOop tail = cont.tail();
3195 if (tail != nullptr && tail->preempted()) {
3196 // Still preempted (monitor not acquired) so no frames were thawed.
3197 set_anchor(thread, cont.entrySP(), cont.entryPC());
3198 preempted = true;
3199 } else {
3200 set_anchor(thread, sp0);
3201 }
3202
3203 log_frames(thread);
3204 if (LoomVerifyAfterThaw) {
3205 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
3206 }
3207 assert(ContinuationEntry::assert_entry_frame_laid_out(thread, preempted), "");
3208 clear_anchor(thread);
3209
3210 LogTarget(Trace, continuations) lt;
3211 if (lt.develop_is_enabled()) {
3212 LogStream ls(lt);
3213 ls.print_cr("Jumping to frame (thaw):");
3214 frame(sp).print_value_on(&ls);
3215 }
3216 }
3217 #endif // ASSERT
3218
3219 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3220
3221 #ifdef ASSERT
3222 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3223 ResourceMark rm;
3224 FrameValues values;
3225 assert(f.get_cb() != nullptr, "");
3226 RegisterMap map(f.is_heap_frame() ?
3227 nullptr :
3228 JavaThread::current(),
3229 RegisterMap::UpdateMap::include,
3230 RegisterMap::ProcessFrames::skip,
3231 RegisterMap::WalkContinuation::skip);
3232 map.set_include_argument_oops(false);
3233 map.set_skip_missing(true);
3234 if (callee_complete) {
3235 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3236 }
3237 const_cast<frame&>(f).describe(values, 0, &map, true);
3238 values.print_on(static_cast<JavaThread*>(nullptr), st);
3239 }
3240 #endif
3241
3242 static address thaw_entry = nullptr;
3243 static address freeze_entry = nullptr;
3244 static address freeze_preempt_entry = nullptr;
3245
3246 address Continuation::thaw_entry() {
3247 return ::thaw_entry;
3248 }
3249
3250 address Continuation::freeze_entry() {
3251 return ::freeze_entry;
3252 }
3253
3254 address Continuation::freeze_preempt_entry() {
3255 return ::freeze_preempt_entry;
3256 }
3257
3258 class ConfigResolve {
3259 public:
3260 static void resolve() { resolve_compressed(); }
3261
3262 static void resolve_compressed() {
3263 UseCompressedOops ? resolve_gc<true>()
3264 : resolve_gc<false>();
3265 }
3266
3267 private:
3268 template <bool use_compressed>
3269 static void resolve_gc() {
3270 BarrierSet* bs = BarrierSet::barrier_set();
3271 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3272 switch (bs->kind()) {
3273 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3274 case BarrierSet::bs_name: { \
3275 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3276 } \
3277 break;
3278 FOR_EACH_CONCRETE_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3279 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3280
3281 default:
3282 fatal("BarrierSet resolving not implemented");
3283 };
3284 }
3285
3286 template <bool use_compressed, typename BarrierSetT>
3287 static void resolve() {
3288 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3289
3290 freeze_entry = (address)freeze<SelectedConfigT>;
3291 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3292
3293 // If we wanted, we could templatize by kind and have three different thaw entries
3294 thaw_entry = (address)thaw<SelectedConfigT>;
3295 }
3296 };
3297
3298 void Continuation::init() {
3299 ConfigResolve::resolve();
3300 }