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 "cppstdlib/type_traits.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/continuationGCSupport.inline.hpp"
34 #include "gc/shared/gc_globals.hpp"
35 #include "gc/shared/memAllocator.hpp"
36 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
37 #include "interpreter/bytecodeStream.hpp"
38 #include "interpreter/interpreter.hpp"
39 #include "interpreter/interpreterRuntime.hpp"
40 #include "jfr/jfrEvents.hpp"
41 #include "logging/log.hpp"
42 #include "logging/logStream.hpp"
43 #include "oops/access.inline.hpp"
44 #include "oops/constantPool.inline.hpp"
45 #include "oops/method.inline.hpp"
46 #include "oops/objArrayOop.inline.hpp"
47 #include "oops/oopsHierarchy.hpp"
48 #include "oops/stackChunkOop.inline.hpp"
49 #include "prims/jvmtiThreadState.hpp"
50 #include "runtime/arguments.hpp"
51 #include "runtime/continuation.hpp"
52 #include "runtime/continuationEntry.inline.hpp"
53 #include "runtime/continuationHelper.inline.hpp"
54 #include "runtime/continuationJavaClasses.inline.hpp"
55 #include "runtime/continuationWrapper.inline.hpp"
56 #include "runtime/frame.inline.hpp"
57 #include "runtime/interfaceSupport.inline.hpp"
58 #include "runtime/javaThread.inline.hpp"
59 #include "runtime/jniHandles.inline.hpp"
60 #include "runtime/keepStackGCProcessed.hpp"
61 #include "runtime/mountUnmountDisabler.hpp"
62 #include "runtime/objectMonitor.inline.hpp"
63 #include "runtime/orderAccess.hpp"
64 #include "runtime/prefetch.inline.hpp"
65 #include "runtime/sharedRuntime.hpp"
66 #include "runtime/smallRegisterMap.inline.hpp"
67 #include "runtime/stackChunkFrameStream.inline.hpp"
68 #include "runtime/stackFrameStream.inline.hpp"
69 #include "runtime/stackOverflow.hpp"
70 #include "runtime/stackWatermarkSet.inline.hpp"
71 #include "runtime/vframe.inline.hpp"
72 #include "runtime/vframe_hp.hpp"
73 #include "utilities/debug.hpp"
74 #include "utilities/exceptions.hpp"
75 #include "utilities/macros.hpp"
76 #include "utilities/vmError.hpp"
77 #if INCLUDE_ZGC
78 #include "gc/z/zStackChunkGCData.inline.hpp"
79 #endif
80 #if INCLUDE_JFR
81 #include "jfr/jfr.inline.hpp"
82 #endif
83 #ifdef COMPILER1
84 #include "c1/c1_Runtime1.hpp"
85 #endif
86 #ifdef COMPILER2
87 #include "opto/runtime.hpp"
88 #endif
89
90 /*
91 * This file contains the implementation of continuation freezing (yield) and thawing (run).
92 *
93 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
94 * would likely call these operations many thousands of times per second second, on every core.
95 *
96 * Freeze might be called every time the application performs any I/O operation, every time it
97 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
98 * multiple times in each of those cases, as it is called by the return barrier, which may be
99 * invoked on method return.
100 *
101 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
102 * example, every effort is made to avoid Java-VM transitions as much as possible.
103 *
104 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
105 * and so frames simply copied, and the bottom-most one is patched.
106 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
107 * and absolute pointers, and barriers invoked.
108 */
109
110 /************************************************
111
112 Thread-stack layout on freeze/thaw.
113 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
114
115 +----------------------------+
116 | . |
117 | . |
118 | . |
119 | carrier frames |
120 | |
121 |----------------------------|
122 | |
123 | Continuation.run |
124 | |
125 |============================|
126 | enterSpecial frame |
127 | pc |
128 | rbp |
129 | ----- |
130 ^ | int argsize | = ContinuationEntry
131 | | oopDesc* cont |
132 | | oopDesc* chunk |
133 | | ContinuationEntry* parent |
134 | | ... |
135 | |============================| <------ JavaThread::_cont_entry = entry->sp()
136 | | ? alignment word ? |
137 | |----------------------------| <--\
138 | | | |
139 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
140 Address | | | | Caller is still in the chunk.
141 | |----------------------------| |
142 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
143 | | rbp | | isn't the last one in the continuation.
144 | | | |
145 | | frame | |
146 | | | |
147 +----------------------------| \__ Continuation frames to be frozen/thawed
148 | | /
149 | frame | |
150 | | |
151 |----------------------------| |
152 | | |
153 | frame | |
154 | | |
155 |----------------------------| <--/
156 | |
157 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
158 | | instead of a doYield stub
159 |============================| <- the sp passed to freeze
160 | |
161 | Native freeze/thaw frames |
162 | . |
163 | . |
164 | . |
165 +----------------------------+
166
167 ************************************************/
168
169 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing
170
171 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
172 #if CONT_JFR
173 #define CONT_JFR_ONLY(code) code
174 #else
175 #define CONT_JFR_ONLY(code)
176 #endif
177
178 // TODO: See AbstractAssembler::generate_stack_overflow_check,
179 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
180 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
181
182 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
183
184 // Used to just annotatate cold/hot branches
185 #define LIKELY(condition) (condition)
186 #define UNLIKELY(condition) (condition)
187
188 // debugging functions
189 #ifdef ASSERT
190 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
191
192 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
193
194 static void do_deopt_after_thaw(JavaThread* thread);
195 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
196 static void log_frames(JavaThread* thread);
197 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp);
198 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
199 static void verify_frame_kind(frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr = nullptr, const char** code_name_ptr = nullptr, int* bci_ptr = nullptr, stackChunkOop chunk = nullptr);
200
201 #define assert_pfl(p, ...) \
202 do { \
203 if (!(p)) { \
204 JavaThread* t = JavaThread::active(); \
205 if (t->has_last_Java_frame()) { \
206 tty->print_cr("assert(" #p ") failed:"); \
207 t->print_frame_layout(); \
208 } \
209 } \
210 vmassert(p, __VA_ARGS__); \
211 } while(0)
212
213 #else
214 static void verify_continuation(oop continuation) { }
215 #define assert_pfl(p, ...)
216 #endif
217
218 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
219 template<typename ConfigT, bool preempt> static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp);
220
221 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
222 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
223
224
225 // Entry point to freeze. Transitions are handled manually
226 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
227 template<typename ConfigT>
228 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
229 assert(sp == current->frame_anchor()->last_Java_sp(), "");
230
231 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
232 current->set_cont_fastpath(nullptr);
233 }
234
235 return checked_cast<int>(ConfigT::freeze(current, sp));
236 JRT_END
237
238 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
239 return prepare_thaw_internal(thread, return_barrier);
240 JRT_END
241
242 template<typename ConfigT>
243 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
244 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
245 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
246 // Also the preemption case with JVMTI events enabled might safepoint so
247 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
248 // JRT_ENTRY instead?
249 ResetNoHandleMark rnhm;
250 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
251
252 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
253 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
254 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
255 JRT_END
256
257 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
258 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
259 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
260 }
261 JVM_END
262
263 ///////////
264
265 enum class oop_kind { NARROW, WIDE };
266 template <oop_kind oops, typename BarrierSetT>
267 class Config {
268 public:
269 typedef Config<oops, BarrierSetT> SelfT;
270 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
271
272 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
273 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
274 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
275 return res;
276 }
277
278 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
279 return freeze_internal<SelfT, true>(thread, sp);
280 }
281
282 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
283 return thaw_internal<SelfT>(thread, kind);
284 }
285 };
286
287 #ifdef _WINDOWS
288 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
289 address new_sp = sp - size;
290 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
291
292 if (new_sp < watermark) {
293 size_t page_size = os::vm_page_size();
294 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
295 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
296 while (pages_to_touch-- > 0) {
297 last_touched_page -= page_size;
298 *last_touched_page = 0;
299 }
300 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
301 }
302 }
303 #endif
304
305 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
306 const size_t page_size = os::vm_page_size();
307 if (size > page_size) {
308 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
309 return false;
310 }
311 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
312 }
313 return true;
314 }
315
316 #ifdef ASSERT
317 static oop get_continuation(JavaThread* thread) {
318 assert(thread != nullptr, "");
319 assert(thread->threadObj() != nullptr, "");
320 return java_lang_Thread::continuation(thread->threadObj());
321 }
322 #endif // ASSERT
323
324 inline void clear_anchor(JavaThread* thread) {
325 thread->frame_anchor()->clear();
326 }
327
328 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
329 assert(pc != nullptr, "");
330
331 JavaFrameAnchor* anchor = thread->frame_anchor();
332 anchor->set_last_Java_sp(sp);
333 anchor->set_last_Java_pc(pc);
334 ContinuationHelper::set_anchor_pd(anchor, sp);
335
336 assert(thread->has_last_Java_frame(), "");
337 assert(thread->last_frame().cb() != nullptr, "");
338 }
339
340 static void set_anchor(JavaThread* thread, intptr_t* sp) {
341 address pc = ContinuationHelper::return_address_at(
342 sp - frame::sender_sp_ret_address_offset());
343 set_anchor(thread, sp, pc);
344 }
345
346 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
347 JavaFrameAnchor* anchor = thread->frame_anchor();
348 anchor->set_last_Java_sp(entry->entry_sp());
349 anchor->set_last_Java_pc(entry->entry_pc());
350 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
351
352 assert(thread->has_last_Java_frame(), "");
353 assert(thread->last_frame().cb() != nullptr, "");
354 }
355
356 #if CONT_JFR
357 class FreezeThawJfrInfo : public StackObj {
358 short _e_size;
359 short _e_num_interpreted_frames;
360 public:
361
362 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
363 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
364 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
365 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
366 };
367
368 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
369 if (e->should_commit()) {
370 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
371 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
372 e->set_continuationClass(continuation->klass());
373 e->set_interpretedFrames(_e_num_interpreted_frames);
374 e->set_size(_e_size);
375 e->commit();
376 }
377 }
378 #endif // CONT_JFR
379
380 /////////////// FREEZE ////
381
382 class FreezeBase : public StackObj {
383 protected:
384 JavaThread* const _thread;
385 ContinuationWrapper& _cont;
386 bool _barriers; // only set when we allocate a chunk
387
388 intptr_t* _bottom_address;
389
390 // Used for preemption only
391 const bool _preempt;
392 frame _last_frame;
393
394 // Used to support freezing with held monitors
395 int _monitors_in_lockstack;
396
397 int _freeze_size; // total size of all frames plus metadata in words.
398 int _total_align_size;
399
400 intptr_t* _cont_stack_top;
401 intptr_t* _cont_stack_bottom;
402
403 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
404
405 #ifdef ASSERT
406 intptr_t* _orig_chunk_sp;
407 int _fast_freeze_size;
408 bool _empty;
409 #endif
410
411 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
412
413 NOT_PRODUCT(int _frames;)
414 DEBUG_ONLY(intptr_t* _last_write;)
415
416 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
417
418 public:
419 NOINLINE freeze_result freeze_slow();
420 void freeze_fast_existing_chunk();
421
422 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
423 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
424
425 inline int size_if_fast_freeze_available();
426
427 inline frame& last_frame() { return _last_frame; }
428
429 #ifdef ASSERT
430 bool check_valid_fast_path();
431 #endif
432
433 protected:
434 inline void init_rest();
435 void throw_stack_overflow_on_humongous_chunk();
436
437 // fast path
438 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
439 inline void unwind_frames();
440 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
441
442 // slow path
443 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
444
445 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
446
447 private:
448 // slow path
449 frame freeze_start_frame();
450 frame freeze_start_frame_on_preempt();
451 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
452 inline frame freeze_start_frame_yield_stub();
453 template<typename FKind>
454 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
455 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
456 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
457 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
458 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
459 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
460 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
461 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
462 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
463 NOINLINE void finish_freeze(const frame& f, const frame& top);
464
465 void freeze_lockstack(stackChunkOop chunk);
466
467 inline bool stack_overflow();
468
469 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
470 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
471 template<typename FKind> static inline frame sender(const frame& f);
472 template<typename FKind> frame new_heap_frame(frame& f, frame& caller, int size_adjust = 0);
473 inline void set_top_frame_metadata_pd(const frame& hf);
474 inline void patch_pd(frame& callee, const frame& caller, bool is_bottom_frame);
475 inline void patch_pd_unused(intptr_t* sp);
476 void adjust_interpreted_frame_unextended_sp(frame& f);
477 inline void prepare_freeze_interpreted_top_frame(frame& f);
478 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
479
480 protected:
481 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
482 bool freeze_fast_new_chunk(stackChunkOop chunk);
483 };
484
485 template <typename ConfigT>
486 class Freeze : public FreezeBase {
487 private:
488 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
489
490 public:
491 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
492 : FreezeBase(thread, cont, frame_sp, preempt) {}
493
494 freeze_result try_freeze_fast();
495
496 protected:
497 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
498 };
499
500 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
501 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
502 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
503
504 assert(_thread != nullptr, "");
505 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
506
507 DEBUG_ONLY(_cont.entry()->verify_cookie();)
508
509 assert(!Interpreter::contains(_cont.entryPC()), "");
510
511 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
512 #ifdef _LP64
513 if (((intptr_t)_bottom_address & 0xf) != 0) {
514 _bottom_address--;
515 }
516 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
517 #endif
518
519 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
520 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
521 assert(_bottom_address != nullptr, "");
522 assert(_bottom_address <= _cont.entrySP(), "");
523 DEBUG_ONLY(_last_write = nullptr;)
524
525 assert(_cont.chunk_invariant(), "");
526 assert(!Interpreter::contains(_cont.entryPC()), "");
527 #if !defined(PPC64) || defined(ZERO)
528 static const int doYield_stub_frame_size = frame::metadata_words;
529 #else
530 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
531 #endif
532 // With preemption doYield() might not have been resolved yet
533 assert(_preempt || SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size, "");
534
535 if (preempt) {
536 _last_frame = _thread->last_frame();
537 }
538
539 // properties of the continuation on the stack; all sizes are in words
540 _cont_stack_top = frame_sp + (!preempt ? doYield_stub_frame_size : 0); // we don't freeze the doYield stub frame
541 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
542 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
543
544 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
545 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
546 assert(cont_size() > 0, "");
547
548 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
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 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
581 #ifdef ASSERT
582 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
583 if (watermark == nullptr) {
584 return;
585 }
586 ContinuationEntry* ce = thread->last_continuation();
587 RegisterMap map(thread,
588 RegisterMap::UpdateMap::include,
589 RegisterMap::ProcessFrames::skip,
590 RegisterMap::WalkContinuation::skip);
591 map.set_include_argument_oops(false);
592 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
593 watermark->assert_is_frame_safe(f);
594 }
595 #endif // ASSERT
596 }
597
598 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
599 void FreezeBase::unwind_frames() {
600 ContinuationEntry* entry = _cont.entry();
601 entry->flush_stack_processing(_thread);
602 assert_frames_in_continuation_are_safe(_thread);
603 JFR_ONLY(Jfr::check_and_process_sample_request(_thread);)
604 set_anchor_to_entry(_thread, entry);
605 }
606
607 template <typename ConfigT>
608 freeze_result Freeze<ConfigT>::try_freeze_fast() {
609 assert(_thread->thread_state() == _thread_in_vm, "");
610 assert(_thread->cont_fastpath(), "");
611
612 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
613 assert(_fast_freeze_size == 0, "");
614
615 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
616 if (freeze_fast_new_chunk(chunk)) {
617 return freeze_ok;
618 }
619 if (_thread->has_pending_exception()) {
620 return freeze_exception;
621 }
622
623 // TODO R REMOVE when deopt change is fixed
624 assert(!_thread->cont_fastpath() || _barriers, "");
625 log_develop_trace(continuations)("-- RETRYING SLOW --");
626 return freeze_slow();
627 }
628
629 // Returns size needed if the continuation fits, otherwise 0.
630 int FreezeBase::size_if_fast_freeze_available() {
631 stackChunkOop chunk = _cont.tail();
632 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
633 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
634 return 0;
635 }
636
637 int total_size_needed = cont_size();
638 const int chunk_sp = chunk->sp();
639
640 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
641 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
642 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
643 // although that would require changing stackChunkOopDesc::is_empty
644 if (!chunk->is_empty()) {
645 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
646 }
647
648 total_size_needed += _monitors_in_lockstack;
649
650 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
651 bool available = chunk_free_room >= total_size_needed;
652 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
653 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
654 return available ? total_size_needed : 0;
655 }
656
657 void FreezeBase::freeze_fast_existing_chunk() {
658 stackChunkOop chunk = _cont.tail();
659
660 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
661 assert(_fast_freeze_size > 0, "");
662
663 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
664 DEBUG_ONLY(_empty = false;)
665 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
666 #ifdef ASSERT
667 {
668 intptr_t* retaddr_slot = (chunk->sp_address()
669 - frame::sender_sp_ret_address_offset());
670 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
671 "unexpected saved return address");
672 }
673 #endif
674
675 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
676 // we overlap; we'll overwrite the chunk's top frame's callee arguments
677 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
678 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
679
680 // increase max_size by what we're freezing minus the overlap
681 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
682
683 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
684 assert(bottom_sp == _bottom_address, "");
685 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
686 // should have a return barrier (installed back when we thawed it).
687 #ifdef ASSERT
688 {
689 intptr_t* retaddr_slot = (bottom_sp
690 - frame::sender_sp_ret_address_offset());
691 assert(ContinuationHelper::return_address_at(retaddr_slot)
692 == StubRoutines::cont_returnBarrier(),
693 "should be the continuation return barrier");
694 }
695 #endif
696 // We copy the fp from the chunk back to the stack because it contains some caller data,
697 // including, possibly, an oop that might have gone stale since we thawed.
698 patch_stack_pd(bottom_sp, chunk->sp_address());
699 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
700
701 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
702 } else { // the chunk is empty
703 const int chunk_start_sp = chunk->stack_size();
704
705 DEBUG_ONLY(_empty = true;)
706 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
707
708 chunk->set_max_thawing_size(cont_size());
709 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
710 chunk->set_sp(chunk->bottom());
711
712 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
713 }
714 }
715
716 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
717 DEBUG_ONLY(_empty = true;)
718
719 // Install new chunk
720 _cont.set_tail(chunk);
721
722 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
723 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
724 return false;
725 }
726
727 chunk->set_max_thawing_size(cont_size());
728
729 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
730 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
731 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
732 assert(chunk_start_sp == chunk->stack_size(), "");
733
734 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
735
736 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
737
738 return true;
739 }
740
741 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
742 assert(chunk != nullptr, "");
743 assert(!chunk->has_mixed_frames(), "");
744 assert(!chunk->is_gc_mode(), "");
745 assert(!chunk->has_bitmap(), "");
746 assert(!chunk->requires_barriers(), "");
747 assert(chunk == _cont.tail(), "");
748
749 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
750 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
751 // will either see no continuation on the stack, or a consistent chunk.
752 unwind_frames();
753
754 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
755 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
756 assert(chunk_start_sp <= chunk->stack_size(), "");
757 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
758
759 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
760 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
761
762 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
763 #ifdef ASSERT
764 if (!_empty) {
765 intptr_t* retaddr_slot = (_orig_chunk_sp
766 - frame::sender_sp_ret_address_offset());
767 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
768 "unexpected saved return address");
769 }
770 #endif
771
772 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
773 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
774
775 int adjust = frame::metadata_words_at_bottom;
776 #if INCLUDE_ASAN && defined(AARCH64)
777 // Reading at offset frame::metadata_words_at_bottom from _cont_stack_top
778 // will accesss memory at the callee frame, which on preemption cases will
779 // be the VM native method being called. The Arm 64-bit ABI doesn't specify
780 // a location where the frame record (returnpc+fp) has to be stored within
781 // a stack frame, and GCC currently chooses to save it at the top of the
782 // frame (lowest address). ASan treats this memory access in the callee as
783 // an overflow access to one of the locals stored in that frame. For these
784 // preemption cases we don't need to read these words anyways so we avoid it.
785 if (_preempt) {
786 adjust = 0;
787 }
788 #endif
789 intptr_t* from = _cont_stack_top - adjust;
790 intptr_t* to = chunk_top - adjust;
791 copy_to_chunk(from, to, cont_size() + adjust);
792 // Because we're not patched yet, the chunk is now in a bad state
793
794 // patch return pc of the bottom-most frozen frame (now in the chunk)
795 // with the actual caller's return address
796 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
797 - _cont.argsize()
798 - frame::metadata_words_at_top
799 - frame::sender_sp_ret_address_offset());
800 #ifdef ASSERT
801 if (!_empty) {
802 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
803 == StubRoutines::cont_returnBarrier(),
804 "should be the continuation return barrier");
805 }
806 #endif
807 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
808 chunk->pc());
809
810 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
811 chunk->set_sp(chunk_new_sp);
812
813 // set chunk->pc to the return address of the topmost frame in the chunk
814 if (_preempt) {
815 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
816 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
817 // be the pc we used when creating the oopmap. Get the top's frame last pc from
818 // the anchor instead.
819 address last_pc = _last_frame.pc();
820 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
821 chunk->set_pc(last_pc);
822 // For stub/native frames the fp is not used while frozen, and will be constructed
823 // again when thawing the frame (see ThawBase::handle_preempted_continuation). We
824 // patch it with a special bad address to help with debugging, particularly when
825 // inspecting frames and identifying invalid accesses.
826 patch_pd_unused(chunk_top);
827 } else {
828 chunk->set_pc(ContinuationHelper::return_address_at(
829 _cont_stack_top - frame::sender_sp_ret_address_offset()));
830 }
831
832 if (_monitors_in_lockstack > 0) {
833 freeze_lockstack(chunk);
834 }
835
836 _cont.write();
837
838 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
839 LogTarget(Trace, continuations) lt;
840 if (lt.develop_is_enabled()) {
841 LogStream ls(lt);
842 chunk->print_on(true, &ls);
843 }
844
845 // Verification
846 assert(_cont.chunk_invariant(), "");
847 chunk->verify();
848
849 #if CONT_JFR
850 EventContinuationFreezeFast e;
851 if (e.should_commit()) {
852 e.set_id(cast_from_oop<u8>(chunk));
853 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
854 e.set_size(cont_size() << LogBytesPerWord);
855 e.commit();
856 }
857 #endif
858 }
859
860 NOINLINE freeze_result FreezeBase::freeze_slow() {
861 #ifdef ASSERT
862 ResourceMark rm;
863 #endif
864
865 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
866 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
867
868 #if CONT_JFR
869 EventContinuationFreezeSlow e;
870 if (e.should_commit()) {
871 e.set_id(cast_from_oop<u8>(_cont.continuation()));
872 e.commit();
873 }
874 #endif
875
876 init_rest();
877
878 HandleMark hm(Thread::current());
879
880 frame f = freeze_start_frame();
881
882 LogTarget(Debug, continuations) lt;
883 if (lt.develop_is_enabled()) {
884 LogStream ls(lt);
885 f.print_on(&ls);
886 }
887
888 frame caller; // the frozen caller in the chunk
889 freeze_result res = recurse_freeze(f, caller, 0, false, true);
890
891 if (res == freeze_ok) {
892 finish_freeze(f, caller);
893 _cont.write();
894 }
895
896 return res;
897 }
898
899 frame FreezeBase::freeze_start_frame() {
900 if (LIKELY(!_preempt)) {
901 return freeze_start_frame_yield_stub();
902 } else {
903 return freeze_start_frame_on_preempt();
904 }
905 }
906
907 frame FreezeBase::freeze_start_frame_yield_stub() {
908 frame f = _thread->last_frame();
909 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be");
910 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
911 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
912 return f;
913 }
914
915 frame FreezeBase::freeze_start_frame_on_preempt() {
916 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
917 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
918 return _last_frame;
919 }
920
921 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
922 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
923 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
924 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
925 || ((top && _preempt) == f.is_native_frame()), "");
926
927 if (stack_overflow()) {
928 return freeze_exception;
929 }
930
931 if (f.is_compiled_frame()) {
932 if (UNLIKELY(f.oop_map() == nullptr)) {
933 // special native frame
934 return freeze_pinned_native;
935 }
936 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
937 } else if (f.is_interpreted_frame()) {
938 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
939 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
940 } else if (top && _preempt) {
941 assert(f.is_native_frame() || f.is_runtime_frame(), "");
942 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
943 } else {
944 // Frame can't be frozen. Most likely the call_stub or upcall_stub
945 // which indicates there are further natives frames up the stack.
946 return freeze_pinned_native;
947 }
948 }
949
950 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
951 // See also StackChunkFrameStream<frame_kind>::frame_size()
952 template<typename FKind>
953 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
954 assert(FKind::is_instance(f), "");
955
956 assert(fsize > 0, "");
957 assert(argsize >= 0, "");
958 _freeze_size += fsize;
959 NOT_PRODUCT(_frames++;)
960
961 assert(FKind::frame_bottom(f) <= _bottom_address, "");
962
963 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
964 // enterSpecial and an interpreted frame
965 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
966 return finalize_freeze(f, caller, argsize); // recursion end
967 } else {
968 frame senderf = sender<FKind>(f);
969 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
970 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
971 return result;
972 }
973 }
974
975 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
976 LogTarget(Trace, continuations) lt;
977 if (lt.develop_is_enabled()) {
978 LogStream ls(lt);
979 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
980 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
981 f.print_value_on(&ls);
982 }
983 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
984 }
985
986 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
987 LogTarget(Trace, continuations) lt;
988 if (lt.develop_is_enabled()) {
989 LogStream ls(lt);
990 DEBUG_ONLY(hf.print_value_on(&ls);)
991 assert(hf.is_heap_frame(), "should be");
992 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
993 if (is_bottom_frame) {
994 ls.print_cr("bottom h-frame:");
995 hf.print_on(&ls);
996 }
997 }
998 }
999
1000 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
1001 // See also StackChunkFrameStream<frame_kind>::frame_size()
1002 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
1003 int argsize = argsize_md - frame::metadata_words_at_top;
1004 assert(callee.is_interpreted_frame()
1005 || ContinuationHelper::Frame::is_stub(callee.cb())
1006 || callee.cb()->as_nmethod()->is_osr_method()
1007 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
1008 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
1009 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
1010
1011 LogTarget(Trace, continuations) lt;
1012
1013 #ifdef ASSERT
1014 bool empty = _cont.is_empty();
1015 log_develop_trace(continuations)("empty: %d", empty);
1016 #endif
1017
1018 stackChunkOop chunk = _cont.tail();
1019
1020 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1021
1022 _freeze_size += frame::metadata_words; // for top frame's metadata
1023
1024 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1025 int unextended_sp = -1;
1026 if (chunk != nullptr) {
1027 if (!chunk->is_empty()) {
1028 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1029 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1030 bool top_interpreted = Interpreter::contains(chunk->pc());
1031 if (callee.is_interpreted_frame() == top_interpreted) {
1032 overlap = argsize_md;
1033 }
1034 } else {
1035 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1036 }
1037 }
1038
1039 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1040
1041 _freeze_size -= overlap;
1042 assert(_freeze_size >= 0, "");
1043
1044 assert(chunk == nullptr || chunk->is_empty()
1045 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1046 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1047
1048 _freeze_size += _monitors_in_lockstack;
1049
1050 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1051 bool allocated_old_in_freeze_fast = _barriers;
1052 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1053 "Chunk allocated in freeze_fast is of insufficient size "
1054 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1055 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1056
1057 DEBUG_ONLY(bool empty_chunk = true);
1058 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1059 // ALLOCATE NEW CHUNK
1060
1061 if (lt.develop_is_enabled()) {
1062 LogStream ls(lt);
1063 if (chunk == nullptr) {
1064 ls.print_cr("no chunk");
1065 } else {
1066 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1067 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1068 chunk->print_on(&ls);
1069 }
1070 }
1071
1072 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1073 // overlap = 0;
1074
1075 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1076 if (chunk == nullptr) {
1077 return freeze_exception;
1078 }
1079
1080 // Install new chunk
1081 _cont.set_tail(chunk);
1082 assert(chunk->is_empty(), "");
1083 } else {
1084 // REUSE EXISTING CHUNK
1085 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1086 if (chunk->is_empty()) {
1087 int sp = chunk->stack_size() - argsize_md;
1088 chunk->set_sp(sp);
1089 chunk->set_bottom(sp);
1090 _freeze_size += overlap;
1091 assert(chunk->max_thawing_size() == 0, "");
1092 } DEBUG_ONLY(else empty_chunk = false;)
1093 }
1094 assert(!chunk->is_gc_mode(), "");
1095 assert(!chunk->has_bitmap(), "");
1096 chunk->set_has_mixed_frames(true);
1097
1098 assert(chunk->requires_barriers() == _barriers, "");
1099 assert(!_barriers || chunk->is_empty(), "");
1100
1101 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1102 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1103
1104 if (_preempt) {
1105 frame top_frame = _thread->last_frame();
1106 if (top_frame.is_interpreted_frame()) {
1107 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1108 // We need it so that on resume we can restore the sp to the right place, since
1109 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1110 // We do it now that we know freezing will be successful.
1111 prepare_freeze_interpreted_top_frame(top_frame);
1112 }
1113
1114 // Do this now so should_process_args_at_top() is set before calling finish_freeze
1115 // in case we might need to apply GC barriers to frames in this stackChunk.
1116 if (_thread->at_preemptable_init()) {
1117 assert(top_frame.is_interpreted_frame(), "only InterpreterRuntime::_new/resolve_from_cache allowed");
1118 chunk->set_at_klass_init(true);
1119 methodHandle m(_thread, top_frame.interpreter_frame_method());
1120 Bytecode_invoke call = Bytecode_invoke_check(m, top_frame.interpreter_frame_bci());
1121 assert(!call.is_valid() || call.is_invokestatic(), "only invokestatic allowed");
1122 if (call.is_invokestatic() && call.size_of_parameters() > 0) {
1123 assert(top_frame.interpreter_frame_expression_stack_size() > 0, "should have parameters in exp stack");
1124 chunk->set_has_args_at_top(true);
1125 }
1126 }
1127 }
1128
1129 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1130 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1131 // will either see no continuation or a consistent chunk.
1132 unwind_frames();
1133
1134 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1135
1136 if (lt.develop_is_enabled()) {
1137 LogStream ls(lt);
1138 ls.print_cr("top chunk:");
1139 chunk->print_on(&ls);
1140 }
1141
1142 if (_monitors_in_lockstack > 0) {
1143 freeze_lockstack(chunk);
1144 }
1145
1146 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1147 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1148
1149 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1150
1151 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1152 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1153 #ifdef ASSERT
1154 if (lt.develop_is_enabled()) {
1155 LogStream ls(lt);
1156 ls.print_cr("top hframe before (freeze):");
1157 assert(caller.is_heap_frame(), "should be");
1158 caller.print_on(&ls);
1159 }
1160
1161 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1162
1163 frame entry = sender(callee);
1164
1165 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1166 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1167 #endif
1168
1169 return freeze_ok_bottom;
1170 }
1171
1172 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1173 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1174 if (is_bottom_frame) {
1175 // If we're the bottom frame, we need to replace the return barrier with the real
1176 // caller's pc.
1177 address last_pc = caller.pc();
1178 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1179 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1180 } else {
1181 assert(!caller.is_empty(), "");
1182 }
1183
1184 patch_pd(hf, caller, is_bottom_frame);
1185
1186 if (f.is_interpreted_frame()) {
1187 assert(hf.is_heap_frame(), "should be");
1188 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1189 }
1190
1191 #ifdef ASSERT
1192 if (hf.is_compiled_frame()) {
1193 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1194 log_develop_trace(continuations)("Freezing deoptimized frame");
1195 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1196 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1197 }
1198 }
1199 #endif
1200 }
1201
1202 #ifdef ASSERT
1203 static void verify_frame_top(const frame& f, intptr_t* top) {
1204 ResourceMark rm;
1205 InterpreterOopMap mask;
1206 f.interpreted_frame_oop_map(&mask);
1207 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1208 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1209 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1210 }
1211 #endif // ASSERT
1212
1213 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1214 // See also StackChunkFrameStream<frame_kind>::frame_size()
1215 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1216 int callee_argsize /* incl. metadata */,
1217 bool callee_interpreted) {
1218 adjust_interpreted_frame_unextended_sp(f);
1219
1220 // The frame's top never includes the stack arguments to the callee
1221 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1222 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1223 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1224
1225 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1226
1227 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1228 // including metadata between f and its args
1229 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1230
1231 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1232 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1233 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1234 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1235
1236 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1237 if (UNLIKELY(result > freeze_ok_bottom)) {
1238 return result;
1239 }
1240
1241 bool is_bottom_frame = result == freeze_ok_bottom;
1242 assert(!caller.is_empty() || is_bottom_frame, "");
1243
1244 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1245
1246 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1247 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1248
1249 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1250 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1251 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1252
1253 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1254 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1255 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1256 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1257
1258 relativize_interpreted_frame_metadata(f, hf);
1259
1260 patch(f, hf, caller, is_bottom_frame);
1261
1262 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1263 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1264 caller = hf;
1265
1266 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1267 frame_method->record_gc_epoch();
1268
1269 return freeze_ok;
1270 }
1271
1272 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1273 // See also StackChunkFrameStream<frame_kind>::frame_size()
1274 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1275 int callee_argsize /* incl. metadata */,
1276 bool callee_interpreted) {
1277 // The frame's top never includes the stack arguments to the callee
1278 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1279 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1280 // including metadata between f and its stackargs
1281 int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1282 int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1283
1284 int real_frame_size = 0;
1285 bool augmented = f.was_augmented_on_entry(real_frame_size);
1286 if (augmented) {
1287 // The args reside inside the frame so clear argsize. If the caller is compiled,
1288 // this will cause the stack arguments passed by the caller to be freezed when
1289 // freezing the caller frame itself. If the caller is interpreted this will have
1290 // the effect of discarding the arg area created in the i2c stub.
1291 argsize = 0;
1292 fsize = real_frame_size - (callee_interpreted ? 0 : callee_argsize);
1293 #ifdef ASSERT
1294 nmethod* nm = f.cb()->as_nmethod();
1295 Method* method = nm->method();
1296 address return_pc = ContinuationHelper::CompiledFrame::return_pc(f);
1297 CodeBlob* caller_cb = CodeCache::find_blob_fast(return_pc);
1298 assert(nm->is_compiled_by_c2() || (caller_cb->is_nmethod() && caller_cb->as_nmethod()->is_compiled_by_c2()), "caller or callee should be c2 compiled");
1299 assert((!caller_cb->is_nmethod() && nm->is_compiled_by_c2()) ||
1300 (nm->compiler_type() != caller_cb->as_nmethod()->compiler_type()) ||
1301 (nm->is_compiled_by_c2() && !method->is_static() && method->method_holder()->is_inline_klass()),
1302 "frame should not be extended");
1303 #endif
1304 }
1305
1306 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d augmented: %d",
1307 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1308 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1309 _freeze_size, fsize, argsize, augmented);
1310 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1311 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1312
1313 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1314 if (UNLIKELY(result > freeze_ok_bottom)) {
1315 return result;
1316 }
1317
1318 bool is_bottom_frame = result == freeze_ok_bottom;
1319 assert(!caller.is_empty() || is_bottom_frame, "");
1320 assert(!is_bottom_frame || !augmented, "thaw extended frame without caller?");
1321
1322 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1323
1324 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller, augmented ? real_frame_size - f.cb()->as_nmethod()->frame_size() : 0);
1325
1326 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1327
1328 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1329 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1330
1331 if (caller.is_interpreted_frame()) {
1332 // When thawing the frame we might need to add alignment (see Thaw::align)
1333 _total_align_size += frame::align_wiggle;
1334 }
1335
1336 patch(f, hf, caller, is_bottom_frame);
1337
1338 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1339
1340 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1341 caller = hf;
1342 return freeze_ok;
1343 }
1344
1345 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1346 DEBUG_ONLY(frame fsender = sender(f);)
1347 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1348
1349 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1350 const int fsize = f.cb()->frame_size();
1351
1352 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1353 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1354
1355 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1356 if (UNLIKELY(result > freeze_ok_bottom)) {
1357 return result;
1358 }
1359
1360 assert(result == freeze_ok, "should have caller");
1361 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1362
1363 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1364 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1365
1366 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1367
1368 patch(f, hf, caller, false /*is_bottom_frame*/);
1369
1370 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1371
1372 caller = hf;
1373 return freeze_ok;
1374 }
1375
1376 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1377 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1378 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1379 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1380 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1381 // caller into the native convention. This is so that we have a valid oopMap in case of
1382 // having to block in the slow path. But that would require freezing those registers too
1383 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1384 // given that this would be a rare case anyways just pin the vthread to the carrier.
1385 return freeze_pinned_native;
1386 }
1387
1388 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1389 // There are no stackargs but argsize must include the metadata
1390 const int argsize = frame::metadata_words_at_top;
1391 const int fsize = f.cb()->frame_size() + argsize;
1392
1393 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1394 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1395
1396 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1397 if (UNLIKELY(result > freeze_ok_bottom)) {
1398 return result;
1399 }
1400
1401 assert(result == freeze_ok, "should have caller frame");
1402 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1403
1404 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1405 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1406
1407 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1408
1409 if (caller.is_interpreted_frame()) {
1410 // When thawing the frame we might need to add alignment (see Thaw::align)
1411 _total_align_size += frame::align_wiggle;
1412 }
1413
1414 patch(f, hf, caller, false /* is_bottom_frame */);
1415
1416 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1417
1418 caller = hf;
1419 return freeze_ok;
1420 }
1421
1422 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1423 stackChunkOop chunk = _cont.tail();
1424
1425 LogTarget(Trace, continuations) lt;
1426 if (lt.develop_is_enabled()) {
1427 LogStream ls(lt);
1428 assert(top.is_heap_frame(), "should be");
1429 top.print_on(&ls);
1430 }
1431
1432 set_top_frame_metadata_pd(top);
1433
1434 chunk->set_sp(chunk->to_offset(top.sp()));
1435 chunk->set_pc(top.pc());
1436
1437 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1438
1439 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1440
1441 // At this point the chunk is consistent
1442
1443 if (UNLIKELY(_barriers)) {
1444 log_develop_trace(continuations)("do barriers on old chunk");
1445 // Serial and Parallel GC can allocate objects directly into the old generation.
1446 // Then we want to relativize the derived pointers eagerly so that
1447 // old chunks are all in GC mode.
1448 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1449 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1450 if (UseShenandoahGC) {
1451 _cont.tail()->relativize_derived_pointers_concurrently();
1452 } else {
1453 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1454 }
1455 // For objects in the old generation we must maintain the remembered set
1456 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1457 }
1458
1459 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1460 if (lt.develop_is_enabled()) {
1461 LogStream ls(lt);
1462 chunk->print_on(true, &ls);
1463 }
1464
1465 if (lt.develop_is_enabled()) {
1466 LogStream ls(lt);
1467 ls.print_cr("top hframe after (freeze):");
1468 assert(_cont.last_frame().is_heap_frame(), "should be");
1469 _cont.last_frame().print_on(&ls);
1470 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1471 }
1472
1473 assert(_cont.chunk_invariant(), "");
1474 }
1475
1476 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1477 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1478 assert(t == JavaThread::current(), "");
1479 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1480 if (!_preempt) {
1481 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1482 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1483 }
1484 return true;
1485 }
1486 return false;
1487 }
1488
1489 class StackChunkAllocator : public MemAllocator {
1490 const size_t _stack_size;
1491 int _argsize_md;
1492 ContinuationWrapper& _continuation_wrapper;
1493 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1494 mutable bool _took_slow_path;
1495
1496 // Does the minimal amount of initialization needed for a TLAB allocation.
1497 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1498 virtual oop initialize(HeapWord* mem) const override {
1499 assert(_stack_size > 0, "");
1500 assert(_stack_size <= max_jint, "");
1501 assert(_word_size > _stack_size, "");
1502
1503 // zero out fields (but not the stack)
1504 const size_t hs = oopDesc::header_size();
1505 if (oopDesc::has_klass_gap()) {
1506 oopDesc::set_klass_gap(mem, 0);
1507 }
1508 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1509
1510 int bottom = (int)_stack_size - _argsize_md;
1511
1512 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1513 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1514 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1515
1516 return finish(mem);
1517 }
1518
1519 stackChunkOop allocate_fast() const {
1520 if (!UseTLAB) {
1521 return nullptr;
1522 }
1523
1524 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1525 if (mem == nullptr) {
1526 return nullptr;
1527 }
1528
1529 oop obj = initialize(mem);
1530 return stackChunkOopDesc::cast(obj);
1531 }
1532
1533 public:
1534 StackChunkAllocator(Klass* klass,
1535 size_t word_size,
1536 Thread* thread,
1537 size_t stack_size,
1538 int argsize_md,
1539 ContinuationWrapper& continuation_wrapper,
1540 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1541 : MemAllocator(klass, word_size, thread),
1542 _stack_size(stack_size),
1543 _argsize_md(argsize_md),
1544 _continuation_wrapper(continuation_wrapper),
1545 _jvmti_event_collector(jvmti_event_collector),
1546 _took_slow_path(false) {}
1547
1548 // Provides it's own, specialized allocation which skips instrumentation
1549 // if the memory can be allocated without going to a slow-path.
1550 stackChunkOop allocate() const {
1551 // First try to allocate without any slow-paths or instrumentation.
1552 stackChunkOop obj = allocate_fast();
1553 if (obj != nullptr) {
1554 return obj;
1555 }
1556
1557 // Now try full-blown allocation with all expensive operations,
1558 // including potentially safepoint operations.
1559 _took_slow_path = true;
1560
1561 // Protect unhandled Loom oops
1562 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1563
1564 // Can safepoint
1565 _jvmti_event_collector->start();
1566
1567 // Can safepoint
1568 return stackChunkOopDesc::cast(MemAllocator::allocate());
1569 }
1570
1571 bool took_slow_path() const {
1572 return _took_slow_path;
1573 }
1574 };
1575
1576 template <typename ConfigT>
1577 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1578 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1579
1580 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1581 size_t size_in_words = klass->instance_size(stack_size);
1582
1583 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1584 if (!_preempt) {
1585 throw_stack_overflow_on_humongous_chunk();
1586 }
1587 return nullptr;
1588 }
1589
1590 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1591 assert(current == JavaThread::current(), "should be current");
1592
1593 // Allocate the chunk.
1594 //
1595 // This might safepoint while allocating, but all safepointing due to
1596 // instrumentation have been deferred. This property is important for
1597 // some GCs, as this ensures that the allocated object is in the young
1598 // generation / newly allocated memory.
1599 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1600 stackChunkOop chunk = allocator.allocate();
1601
1602 if (chunk == nullptr) {
1603 return nullptr; // OOME
1604 }
1605
1606 // assert that chunk is properly initialized
1607 assert(chunk->stack_size() == (int)stack_size, "");
1608 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1609 assert(chunk->sp() == chunk->bottom(), "");
1610 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1611 assert(chunk->max_thawing_size() == 0, "");
1612 assert(chunk->pc() == nullptr, "");
1613 assert(chunk->is_empty(), "");
1614 assert(chunk->flags() == 0, "");
1615 assert(chunk->is_gc_mode() == false, "");
1616 assert(chunk->lockstack_size() == 0, "");
1617
1618 // fields are uninitialized
1619 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1620 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1621
1622 #if INCLUDE_ZGC
1623 if (UseZGC) {
1624 ZStackChunkGCData::initialize(chunk);
1625 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1626 _barriers = false;
1627 } else
1628 #endif
1629 #if INCLUDE_SHENANDOAHGC
1630 if (UseShenandoahGC) {
1631 _barriers = chunk->requires_barriers();
1632 } else
1633 #endif
1634 {
1635 if (!allocator.took_slow_path()) {
1636 // Guaranteed to be in young gen / newly allocated memory
1637 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1638 _barriers = false;
1639 } else {
1640 // Some GCs could put direct allocations in old gen for slow-path
1641 // allocations; need to explicitly check if that was the case.
1642 _barriers = chunk->requires_barriers();
1643 }
1644 }
1645
1646 if (_barriers) {
1647 log_develop_trace(continuations)("allocation requires barriers");
1648 }
1649
1650 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1651
1652 return chunk;
1653 }
1654
1655 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1656 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1657 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1658 }
1659
1660 class AnchorMark : public StackObj {
1661 JavaThread* _current;
1662 frame& _top_frame;
1663 intptr_t* _last_sp_from_frame;
1664 bool _is_interpreted;
1665
1666 public:
1667 AnchorMark(JavaThread* current, frame& f) : _current(current), _top_frame(f), _is_interpreted(false) {
1668 intptr_t* sp = anchor_mark_set_pd();
1669 set_anchor(_current, sp);
1670 }
1671 ~AnchorMark() {
1672 clear_anchor(_current);
1673 anchor_mark_clear_pd();
1674 }
1675 inline intptr_t* anchor_mark_set_pd();
1676 inline void anchor_mark_clear_pd();
1677 };
1678
1679 #if INCLUDE_JVMTI
1680 static int num_java_frames(ContinuationWrapper& cont) {
1681 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1682 int count = 0;
1683 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1684 count += chunk->num_java_frames();
1685 }
1686 return count;
1687 }
1688
1689 static void invalidate_jvmti_stack(JavaThread* thread) {
1690 JvmtiThreadState *state = thread->jvmti_thread_state();
1691 if (state != nullptr) {
1692 state->invalidate_cur_stack_depth();
1693 }
1694 }
1695
1696 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1697 if (!cont.entry()->is_virtual_thread()) {
1698 if (JvmtiExport::has_frame_pops(thread)) {
1699 int num_frames = num_java_frames(cont);
1700
1701 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1702 JvmtiExport::continuation_yield_cleanup(thread, num_frames);
1703 }
1704 invalidate_jvmti_stack(thread);
1705 }
1706 }
1707
1708 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, Continuation::preempt_kind pk) {
1709 assert(current->vthread() != nullptr, "must be");
1710
1711 HandleMarkCleaner hm(current); // Cleanup all handles (including so._conth) before returning to Java.
1712 Handle vth(current, current->vthread());
1713 ContinuationWrapper::SafepointOp so(current, cont);
1714 AnchorMark am(current, top); // Set anchor so that the stack is walkable.
1715
1716 JRT_BLOCK
1717 MountUnmountDisabler::end_transition(current, vth(), true /*is_mount*/, false /*is_thread_start*/);
1718
1719 if (current->pending_contended_entered_event()) {
1720 // No monitor JVMTI events for ObjectLocker case.
1721 if (pk != Continuation::object_locker) {
1722 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1723 }
1724 current->set_contended_entered_monitor(nullptr);
1725 }
1726 JRT_BLOCK_END
1727 }
1728 #endif // INCLUDE_JVMTI
1729
1730 #ifdef ASSERT
1731 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1732 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1733 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1734 bool FreezeBase::check_valid_fast_path() {
1735 ContinuationEntry* ce = _thread->last_continuation();
1736 RegisterMap map(_thread,
1737 RegisterMap::UpdateMap::skip,
1738 RegisterMap::ProcessFrames::skip,
1739 RegisterMap::WalkContinuation::skip);
1740 map.set_include_argument_oops(false);
1741 bool is_top_frame = true;
1742 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1743 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1744 return false;
1745 }
1746 }
1747 return true;
1748 }
1749
1750 static void verify_frame_kind(frame& top, Continuation::preempt_kind preempt_kind, Method** m_ptr, const char** code_name_ptr, int* bci_ptr, stackChunkOop chunk) {
1751 Method* m;
1752 const char* code_name;
1753 int bci;
1754 if (preempt_kind == Continuation::monitorenter) {
1755 assert(top.is_interpreted_frame() || top.is_runtime_frame(), "unexpected %sframe",
1756 top.is_compiled_frame() ? "compiled " : top.is_native_frame() ? "native " : "");
1757 bool at_sync_method;
1758 if (top.is_interpreted_frame()) {
1759 m = top.interpreter_frame_method();
1760 assert(!m->is_native() || m->is_synchronized(), "invalid method %s", m->external_name());
1761 address bcp = top.interpreter_frame_bcp();
1762 assert(bcp != 0 || m->is_native(), "");
1763 at_sync_method = m->is_synchronized() && (bcp == 0 || bcp == m->code_base());
1764 // bcp is advanced on monitorenter before making the VM call, adjust for that.
1765 bool at_sync_bytecode = bcp > m->code_base() && Bytecode(m, bcp - 1).code() == Bytecodes::Code::_monitorenter;
1766 assert(at_sync_method || at_sync_bytecode, "");
1767 bci = at_sync_method ? -1 : top.interpreter_frame_bci();
1768 } else {
1769 JavaThread* current = JavaThread::current();
1770 ResourceMark rm(current);
1771 CodeBlob* cb = top.cb();
1772 RegisterMap reg_map(current,
1773 RegisterMap::UpdateMap::skip,
1774 RegisterMap::ProcessFrames::skip,
1775 RegisterMap::WalkContinuation::include);
1776 if (top.is_heap_frame()) {
1777 assert(chunk != nullptr, "");
1778 reg_map.set_stack_chunk(chunk);
1779 top = chunk->relativize(top);
1780 top.set_frame_index(0);
1781 }
1782 frame fr = top.sender(®_map);
1783 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1784 compiledVFrame* cvf = compiledVFrame::cast(vf);
1785 m = cvf->method();
1786 bci = cvf->scope()->bci();
1787 at_sync_method = bci == SynchronizationEntryBCI;
1788 assert(!at_sync_method || m->is_synchronized(), "bci is %d but method %s is not synchronized", bci, m->external_name());
1789 bool is_c1_monitorenter = false, is_c2_monitorenter = false;
1790 COMPILER1_PRESENT(is_c1_monitorenter = cb == Runtime1::blob_for(StubId::c1_monitorenter_id) ||
1791 cb == Runtime1::blob_for(StubId::c1_monitorenter_nofpu_id);)
1792 COMPILER2_PRESENT(is_c2_monitorenter = cb == CodeCache::find_blob(OptoRuntime::complete_monitor_locking_Java());)
1793 assert(is_c1_monitorenter || is_c2_monitorenter, "wrong runtime stub frame");
1794 }
1795 code_name = at_sync_method ? "synchronized method" : "monitorenter";
1796 } else if (preempt_kind == Continuation::object_wait) {
1797 assert(top.is_interpreted_frame() || top.is_native_frame(), "");
1798 m = top.is_interpreted_frame() ? top.interpreter_frame_method() : top.cb()->as_nmethod()->method();
1799 assert(m->is_object_wait0(), "");
1800 bci = 0;
1801 code_name = "";
1802 } else {
1803 assert(preempt_kind == Continuation::object_locker, "invalid preempt kind");
1804 assert(top.is_interpreted_frame(), "");
1805 m = top.interpreter_frame_method();
1806 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
1807 Bytecodes::Code code = current_bytecode.code();
1808 assert(code == Bytecodes::Code::_new || code == Bytecodes::Code::_invokestatic ||
1809 (code == Bytecodes::Code::_getstatic || code == Bytecodes::Code::_putstatic), "invalid bytecode");
1810 bci = top.interpreter_frame_bci();
1811 code_name = Bytecodes::name(current_bytecode.code());
1812 }
1813 assert(bci >= 0 || m->is_synchronized(), "invalid bci:%d at method %s", bci, m->external_name());
1814
1815 if (m_ptr != nullptr) {
1816 *m_ptr = m;
1817 *code_name_ptr = code_name;
1818 *bci_ptr = bci;
1819 }
1820 }
1821
1822 static void log_preempt_after_freeze(const ContinuationWrapper& cont) {
1823 JavaThread* current = cont.thread();
1824 int64_t tid = current->monitor_owner_id();
1825
1826 StackChunkFrameStream<ChunkFrames::Mixed> sfs(cont.tail());
1827 frame top_frame = sfs.to_frame();
1828 bool at_init = current->at_preemptable_init();
1829 bool at_enter = current->current_pending_monitor() != nullptr;
1830 bool at_wait = current->current_waiting_monitor() != nullptr;
1831 assert((at_enter && !at_wait) || (!at_enter && at_wait), "");
1832 Continuation::preempt_kind pk = at_init ? Continuation::object_locker : at_enter ? Continuation::monitorenter : Continuation::object_wait;
1833
1834 Method* m = nullptr;
1835 const char* code_name = nullptr;
1836 int bci = InvalidFrameStateBci;
1837 verify_frame_kind(top_frame, pk, &m, &code_name, &bci, cont.tail());
1838 assert(m != nullptr && code_name != nullptr && bci != InvalidFrameStateBci, "should be set");
1839
1840 ResourceMark rm(current);
1841 if (bci < 0) {
1842 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " while synchronizing on %smethod %s", tid, m->is_native() ? "native " : "", m->external_name());
1843 } else if (m->is_object_wait0()) {
1844 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at native method %s", tid, m->external_name());
1845 } else {
1846 Klass* k = current->preempt_init_klass();
1847 assert(k != nullptr || !at_init, "");
1848 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at %s(bci:%d) in method %s %s%s", tid, code_name, bci,
1849 m->external_name(), at_init ? "trying to initialize klass " : "", at_init ? k->external_name() : "");
1850 }
1851 }
1852 #endif // ASSERT
1853
1854 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1855 verify_continuation(cont.continuation());
1856 assert(!cont.is_empty(), "");
1857
1858 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1859 return freeze_ok;
1860 }
1861
1862 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1863 if (UNLIKELY(res != freeze_ok)) {
1864 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1865 verify_continuation(cont.continuation());
1866 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1867 return res;
1868 }
1869
1870 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1871 return freeze_epilog(cont);
1872 }
1873
1874 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1875 if (UNLIKELY(res != freeze_ok)) {
1876 verify_continuation(cont.continuation());
1877 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1878 return res;
1879 }
1880
1881 // Set up things so that on return to Java we jump to preempt stub.
1882 patch_return_pc_with_preempt_stub(old_last_frame);
1883 cont.tail()->set_preempted(true);
1884 DEBUG_ONLY(log_preempt_after_freeze(cont);)
1885 return freeze_epilog(cont);
1886 }
1887
1888 template<typename ConfigT, bool preempt>
1889 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1890 assert(!current->has_pending_exception(), "");
1891
1892 #ifdef ASSERT
1893 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1894 log_frames(current);
1895 #endif
1896
1897 CONT_JFR_ONLY(EventContinuationFreeze event;)
1898
1899 ContinuationEntry* entry = current->last_continuation();
1900
1901 oop oopCont = entry->cont_oop(current);
1902 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1903 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1904
1905 verify_continuation(oopCont);
1906 ContinuationWrapper cont(current, oopCont);
1907 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1908
1909 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1910
1911 if (entry->is_pinned()) {
1912 log_develop_debug(continuations)("PINNED due to critical section");
1913 verify_continuation(cont.continuation());
1914 const freeze_result res = freeze_pinned_cs;
1915 if (!preempt) {
1916 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1917 }
1918 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1919 // Avoid Thread.yield() loops without safepoint polls.
1920 if (SafepointMechanism::should_process(current) && !preempt) {
1921 cont.done(); // allow safepoint
1922 ThreadInVMfromJava tivmfj(current);
1923 }
1924 return res;
1925 }
1926
1927 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1928
1929 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1930 bool fast = UseContinuationFastPath && current->cont_fastpath();
1931 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1932 freeze.freeze_fast_existing_chunk();
1933 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1934 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1935 }
1936
1937 if (preempt) {
1938 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1939 freeze.set_jvmti_event_collector(&jsoaec);
1940
1941 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1942
1943 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1944 preempt_epilog(cont, res, freeze.last_frame());
1945 return res;
1946 }
1947
1948 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1949 assert(current == JavaThread::current(), "must be current thread");
1950 JRT_BLOCK
1951 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1952 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1953 freeze.set_jvmti_event_collector(&jsoaec);
1954
1955 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1956
1957 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1958 freeze_epilog(current, cont, res);
1959 cont.done(); // allow safepoint in the transition back to Java
1960 return res;
1961 JRT_BLOCK_END
1962 }
1963
1964 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1965 ContinuationEntry* entry = thread->last_continuation();
1966 if (entry == nullptr) {
1967 return freeze_ok;
1968 }
1969 if (entry->is_pinned()) {
1970 return freeze_pinned_cs;
1971 }
1972
1973 RegisterMap map(thread,
1974 RegisterMap::UpdateMap::include,
1975 RegisterMap::ProcessFrames::skip,
1976 RegisterMap::WalkContinuation::skip);
1977 map.set_include_argument_oops(false);
1978 frame f = thread->last_frame();
1979
1980 if (!safepoint) {
1981 f = f.sender(&map); // this is the yield frame
1982 } else { // safepoint yield
1983 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1984 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1985 #else
1986 Unimplemented();
1987 #endif
1988 if (!Interpreter::contains(f.pc())) {
1989 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1990 assert(f.oop_map() != nullptr, "must be");
1991 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1992 }
1993 }
1994
1995 while (true) {
1996 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1997 return freeze_pinned_native;
1998 }
1999
2000 f = f.sender(&map);
2001 if (!Continuation::is_frame_in_continuation(entry, f)) {
2002 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
2003 if (scope == cont_scope) {
2004 break;
2005 }
2006 entry = entry->parent();
2007 if (entry == nullptr) {
2008 break;
2009 }
2010 if (entry->is_pinned()) {
2011 return freeze_pinned_cs;
2012 }
2013 }
2014 }
2015 return freeze_ok;
2016 }
2017
2018 /////////////// THAW ////
2019
2020 static int thaw_size(stackChunkOop chunk) {
2021 int size = chunk->max_thawing_size();
2022 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
2023 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
2024 return size;
2025 }
2026
2027 // make room on the stack for thaw
2028 // returns the size in bytes, or 0 on failure
2029 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
2030 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
2031
2032 assert(thread == JavaThread::current(), "");
2033
2034 ContinuationEntry* ce = thread->last_continuation();
2035 assert(ce != nullptr, "");
2036 oop continuation = ce->cont_oop(thread);
2037 assert(continuation == get_continuation(thread), "");
2038 verify_continuation(continuation);
2039
2040 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
2041 assert(chunk != nullptr, "");
2042
2043 // The tail can be empty because it might still be available for another freeze.
2044 // However, here we want to thaw, so we get rid of it (it will be GCed).
2045 if (UNLIKELY(chunk->is_empty())) {
2046 chunk = chunk->parent();
2047 assert(chunk != nullptr, "");
2048 assert(!chunk->is_empty(), "");
2049 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
2050 }
2051
2052 // Verification
2053 chunk->verify();
2054 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
2055
2056 // Only make space for the last chunk because we only thaw from the last chunk
2057 int size = thaw_size(chunk) << LogBytesPerWord;
2058
2059 const address bottom = (address)thread->last_continuation()->entry_sp();
2060 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
2061 // for the Java frames in the check below.
2062 if (!stack_overflow_check(thread, size + 300, bottom)) {
2063 return 0;
2064 }
2065
2066 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
2067 p2i(bottom), p2i(bottom - size), size);
2068 return size;
2069 }
2070
2071 class ThawBase : public StackObj {
2072 protected:
2073 JavaThread* _thread;
2074 ContinuationWrapper& _cont;
2075 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
2076
2077 intptr_t* _fastpath;
2078 bool _barriers;
2079 bool _preempted_case;
2080 bool _process_args_at_top;
2081 intptr_t* _top_unextended_sp_before_thaw;
2082 int _align_size;
2083 DEBUG_ONLY(intptr_t* _top_stack_address);
2084
2085 // Only used for preemption on ObjectLocker
2086 ObjectMonitor* _init_lock;
2087
2088 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
2089
2090 NOT_PRODUCT(int _frames;)
2091
2092 protected:
2093 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2094 _thread(thread), _cont(cont),
2095 _fastpath(nullptr) {
2096 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2097 assert (cont.tail() != nullptr, "no last chunk");
2098 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2099 }
2100
2101 void clear_chunk(stackChunkOop chunk);
2102 template<bool check_stub>
2103 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2104 int remove_scalarized_frames(StackChunkFrameStream<ChunkFrames::CompiledOnly>& scfs, int &argsize);
2105 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2106
2107 void thaw_lockstack(stackChunkOop chunk);
2108
2109 // fast path
2110 inline void prefetch_chunk_pd(void* start, int size_words);
2111 void patch_return(intptr_t* sp, bool is_last);
2112
2113 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
2114 inline intptr_t* push_cleanup_continuation();
2115 inline intptr_t* push_preempt_adapter();
2116 intptr_t* redo_vmcall(JavaThread* current, frame& top);
2117 void throw_interrupted_exception(JavaThread* current, frame& top);
2118
2119 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
2120 void finish_thaw(frame& f);
2121
2122 private:
2123 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2124 void finalize_thaw(frame& entry, int argsize);
2125
2126 inline bool seen_by_gc();
2127
2128 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2129 inline void after_thaw_java_frame(const frame& f, bool bottom);
2130 inline void patch(frame& f, const frame& caller, bool bottom, bool augmented = false);
2131 void clear_bitmap_bits(address start, address end);
2132
2133 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top);
2134 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2135 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2136 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
2137
2138 void push_return_frame(const frame& f);
2139 inline frame new_entry_frame();
2140 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom, int size_adjust = 0);
2141 inline void patch_pd(frame& f, const frame& sender);
2142 inline void patch_pd(frame& f, intptr_t* caller_sp);
2143 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2144
2145 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2146
2147 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2148
2149 public:
2150 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2151 };
2152
2153 template <typename ConfigT>
2154 class Thaw : public ThawBase {
2155 public:
2156 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2157
2158 inline bool can_thaw_fast(stackChunkOop chunk) {
2159 return !_barriers
2160 && _thread->cont_fastpath_thread_state()
2161 && !chunk->has_thaw_slowpath_condition()
2162 && !PreserveFramePointer;
2163 }
2164
2165 inline intptr_t* thaw(Continuation::thaw_kind kind);
2166 template<bool check_stub = false>
2167 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2168 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2169 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2170 };
2171
2172 template <typename ConfigT>
2173 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2174 verify_continuation(_cont.continuation());
2175 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2176 assert(!_cont.is_empty(), "");
2177
2178 stackChunkOop chunk = _cont.tail();
2179 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2180 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2181
2182 _barriers = chunk->requires_barriers();
2183 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2184 : thaw_slow(chunk, kind);
2185 }
2186
2187 class ReconstructedStack : public StackObj {
2188 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2189 int _thaw_size;
2190 int _argsize;
2191 public:
2192 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2193 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2194 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2195 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2196 // We're at most one alignment word away from entrySP
2197 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2198 }
2199
2200 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2201
2202 // top and bottom stack pointers
2203 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2204 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2205
2206 // several operations operate on the totality of the stack being reconstructed,
2207 // including the metadata words
2208 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2209 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2210 };
2211
2212 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2213 chunk->set_sp(chunk->bottom());
2214 chunk->set_max_thawing_size(0);
2215 }
2216
2217 int ThawBase::remove_scalarized_frames(StackChunkFrameStream<ChunkFrames::CompiledOnly>& f, int &argsize) {
2218 intptr_t* top = f.sp();
2219
2220 while (f.cb()->as_nmethod()->needs_stack_repair()) {
2221 f.next(SmallRegisterMap::instance_no_args(), false /* stop */);
2222 }
2223 assert(!f.is_done(), "");
2224 assert(f.is_compiled(), "");
2225
2226 intptr_t* bottom = f.sp() + f.cb()->frame_size();
2227 argsize = f.stack_argsize();
2228 return bottom - top;
2229 }
2230
2231 template<bool check_stub>
2232 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2233 bool empty = false;
2234 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2235 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2236 assert(chunk_sp == f.sp(), "");
2237 assert(chunk_sp == f.unextended_sp(), "");
2238
2239 int frame_size = f.cb()->frame_size();
2240 argsize = f.stack_argsize();
2241
2242 assert(!f.is_stub() || check_stub, "");
2243 if (check_stub && f.is_stub()) {
2244 // If we don't thaw the top compiled frame too, after restoring the saved
2245 // registers back in Java, we would hit the return barrier to thaw one more
2246 // frame effectively overwriting the restored registers during that call.
2247 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2248 assert(!f.is_done(), "");
2249
2250 f.get_cb();
2251 assert(f.is_compiled(), "");
2252 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2253 // The caller of the runtime stub when the continuation is preempted is not at a
2254 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2255 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2256 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2257 }
2258
2259 if (f.cb()->as_nmethod()->needs_stack_repair()) {
2260 frame_size += remove_scalarized_frames(f, argsize);
2261 } else {
2262 frame_size += f.cb()->frame_size();
2263 argsize = f.stack_argsize();
2264 }
2265 } else if (f.cb()->as_nmethod()->needs_stack_repair()) {
2266 frame_size = remove_scalarized_frames(f, argsize);
2267 }
2268
2269 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2270 empty = f.is_done();
2271 assert(!empty || argsize == chunk->argsize(), "");
2272
2273 if (empty) {
2274 clear_chunk(chunk);
2275 } else {
2276 chunk->set_sp(chunk->sp() + frame_size);
2277 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2278 // We set chunk->pc to the return pc into the next frame
2279 chunk->set_pc(f.pc());
2280 #ifdef ASSERT
2281 {
2282 intptr_t* retaddr_slot = (chunk_sp
2283 + frame_size
2284 - frame::sender_sp_ret_address_offset());
2285 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2286 "unexpected pc");
2287 }
2288 #endif
2289 }
2290 assert(empty == chunk->is_empty(), "");
2291 // returns the size required to store the frame on stack, and because it is a
2292 // compiled frame, it must include a copy of the arguments passed by the caller
2293 return frame_size + argsize + frame::metadata_words_at_top;
2294 }
2295
2296 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2297 int lockStackSize = chunk->lockstack_size();
2298 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2299
2300 oop tmp_lockstack[LockStack::CAPACITY];
2301 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2302 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2303
2304 chunk->set_lockstack_size(0);
2305 chunk->set_has_lockstack(false);
2306 }
2307
2308 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2309 assert(to >= _top_stack_address, "overwrote past thawing space"
2310 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2311 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2312 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2313 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2314 }
2315
2316 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2317 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2318
2319 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2320 ContinuationHelper::patch_return_address_at(
2321 sp - frame::sender_sp_ret_address_offset(),
2322 pc);
2323 }
2324
2325 template <typename ConfigT>
2326 template<bool check_stub>
2327 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2328 assert(chunk == _cont.tail(), "");
2329 assert(!chunk->has_mixed_frames(), "");
2330 assert(!chunk->requires_barriers(), "");
2331 assert(!chunk->has_bitmap(), "");
2332 assert(!_thread->is_interp_only_mode(), "");
2333
2334 LogTarget(Trace, continuations) lt;
2335 if (lt.develop_is_enabled()) {
2336 LogStream ls(lt);
2337 ls.print_cr("thaw_fast");
2338 chunk->print_on(true, &ls);
2339 }
2340
2341 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2342 static const int threshold = 500; // words
2343
2344 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2345 int argsize, thaw_size;
2346
2347 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2348
2349 bool partial, empty;
2350 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2351 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2352
2353 partial = false;
2354 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2355 clear_chunk(chunk);
2356 thaw_size = full_chunk_size;
2357 empty = true;
2358 } else { // thaw a single frame
2359 partial = true;
2360 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2361 empty = chunk->is_empty();
2362 }
2363
2364 // Are we thawing the last frame(s) in the continuation
2365 const bool is_last = empty && chunk->parent() == nullptr;
2366 assert(!is_last || argsize == 0, "");
2367
2368 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2369 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2370
2371 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2372
2373 // also copy metadata words at frame bottom
2374 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2375
2376 // update the ContinuationEntry
2377 _cont.set_argsize(argsize);
2378 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2379 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2380
2381 // install the return barrier if not last frame, or the entry's pc if last
2382 patch_return(rs.bottom_sp(), is_last);
2383
2384 // insert the back links from callee to caller frames
2385 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2386
2387 assert(is_last == _cont.is_empty(), "");
2388 assert(_cont.chunk_invariant(), "");
2389
2390 #if CONT_JFR
2391 EventContinuationThawFast e;
2392 if (e.should_commit()) {
2393 e.set_id(cast_from_oop<u8>(chunk));
2394 e.set_size(thaw_size << LogBytesPerWord);
2395 e.set_full(!partial);
2396 e.commit();
2397 }
2398 #endif
2399
2400 #ifdef ASSERT
2401 if (LoomDeoptAfterThaw) {
2402 frame top(rs.sp());
2403 AnchorMark am(_thread, top);
2404 log_frames(_thread);
2405 do_deopt_after_thaw(_thread);
2406 }
2407 #endif
2408
2409 return rs.sp();
2410 }
2411
2412 inline bool ThawBase::seen_by_gc() {
2413 return _barriers || _cont.tail()->is_gc_mode();
2414 }
2415
2416 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2417 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2418 if (UseZGC || UseShenandoahGC) {
2419 chunk->relativize_derived_pointers_concurrently();
2420 }
2421 #endif
2422 }
2423
2424 template <typename ConfigT>
2425 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2426 Continuation::preempt_kind preempt_kind;
2427 bool retry_fast_path = false;
2428
2429 _process_args_at_top = false;
2430 _preempted_case = chunk->preempted();
2431 if (_preempted_case) {
2432 ObjectMonitor* mon = nullptr;
2433 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2434 if (waiter != nullptr) {
2435 // Mounted again after preemption. Resume the pending monitor operation,
2436 // which will be either a monitorenter or Object.wait() call.
2437 mon = waiter->monitor();
2438 preempt_kind = waiter->is_wait() ? Continuation::object_wait : Continuation::monitorenter;
2439
2440 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2441 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2442 if (!mon_acquired) {
2443 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2444 log_develop_trace(continuations, preempt)("Failed to acquire monitor, unmounting again");
2445 return push_cleanup_continuation();
2446 }
2447 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2448 JVMTI_ONLY(assert(_thread->contended_entered_monitor() == nullptr || _thread->contended_entered_monitor() == mon, ""));
2449 } else {
2450 // Preemption cancelled on moniterenter or ObjectLocker case. We
2451 // actually acquired the monitor after freezing all frames so no
2452 // need to call resume_operation. If this is the ObjectLocker case
2453 // we released the monitor already at ~ObjectLocker, so _init_lock
2454 // will be set to nullptr below since there is no monitor to release.
2455 preempt_kind = Continuation::monitorenter;
2456 }
2457
2458 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2459 relativize_chunk_concurrently(chunk);
2460
2461 if (chunk->at_klass_init()) {
2462 preempt_kind = Continuation::object_locker;
2463 chunk->set_at_klass_init(false);
2464 _process_args_at_top = chunk->has_args_at_top();
2465 if (_process_args_at_top) {
2466 // Only needed for the top frame which will be thawed.
2467 chunk->set_has_args_at_top(false);
2468 }
2469 assert(waiter == nullptr || mon != nullptr, "should have a monitor");
2470 _init_lock = mon; // remember monitor since we will need it on handle_preempted_continuation()
2471 }
2472 chunk->set_preempted(false);
2473 retry_fast_path = true;
2474 } else {
2475 relativize_chunk_concurrently(chunk);
2476 }
2477
2478 // On first thaw after freeze restore oops to the lockstack if any.
2479 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2480 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2481 thaw_lockstack(chunk);
2482 retry_fast_path = true;
2483 }
2484
2485 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2486 // and FLAG_PREEMPTED flags from the stackChunk.
2487 if (retry_fast_path && can_thaw_fast(chunk)) {
2488 intptr_t* sp = thaw_fast<true>(chunk);
2489 if (_preempted_case) {
2490 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2491 }
2492 return sp;
2493 }
2494
2495 LogTarget(Trace, continuations) lt;
2496 if (lt.develop_is_enabled()) {
2497 LogStream ls(lt);
2498 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2499 chunk->print_on(true, &ls);
2500 }
2501
2502 #if CONT_JFR
2503 EventContinuationThawSlow e;
2504 if (e.should_commit()) {
2505 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2506 e.commit();
2507 }
2508 #endif
2509
2510 DEBUG_ONLY(_frames = 0;)
2511 _align_size = 0;
2512 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2513
2514 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2515 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2516
2517 frame heap_frame = _stream.to_frame();
2518 if (lt.develop_is_enabled()) {
2519 LogStream ls(lt);
2520 ls.print_cr("top hframe before (thaw):");
2521 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2522 heap_frame.print_value_on(&ls);
2523 }
2524
2525 frame caller; // the thawed caller on the stack
2526 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2527 finish_thaw(caller); // caller is now the topmost thawed frame
2528 _cont.write();
2529
2530 assert(_cont.chunk_invariant(), "");
2531
2532 JVMTI_ONLY(if (!_cont.entry()->is_virtual_thread()) invalidate_jvmti_stack(_thread));
2533
2534 _thread->set_cont_fastpath(_fastpath);
2535
2536 intptr_t* sp = caller.sp();
2537
2538 if (_preempted_case) {
2539 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2540 }
2541 return sp;
2542 }
2543
2544 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2545 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2546 assert(!_cont.is_empty(), "no more frames");
2547 assert(num_frames > 0, "");
2548 assert(!heap_frame.is_empty(), "");
2549
2550 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2551 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2552 } else if (!heap_frame.is_interpreted_frame()) {
2553 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2554 } else {
2555 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames, top_on_preempt_case);
2556 }
2557 }
2558
2559 template<typename FKind>
2560 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2561 assert(num_frames > 0, "");
2562
2563 DEBUG_ONLY(_frames++;)
2564
2565 int argsize = _stream.stack_argsize();
2566 CodeBlob* cb = _stream.cb();
2567
2568 _stream.next(SmallRegisterMap::instance_no_args());
2569 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2570
2571 // We never leave a compiled caller of an interpreted frame as the top frame in the chunk
2572 // as it makes detecting that situation and adjusting unextended_sp tricky. We also always
2573 // thaw the caller of a frame that needs_stack_repair, as it would otherwise complicate things:
2574 // - Regardless of whether the frame was extended or not, we would need to copy the right arg
2575 // size if its greater than the one given by the normal method signature (non-scalarized).
2576 // - If the frame was indeed extended, leaving its caller as the top frame would complicate walking
2577 // the chunk (we need unextended_sp, but we only have sp).
2578 if (num_frames == 1 && !_stream.is_done() && ((FKind::interpreted && _stream.is_compiled()) || (FKind::compiled && cb->as_nmethod_or_null()->needs_stack_repair()))) {
2579 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2580 num_frames++;
2581 }
2582
2583 if (num_frames == 1 || _stream.is_done()) { // end recursion
2584 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2585 return true; // bottom
2586 } else { // recurse
2587 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2588 return false;
2589 }
2590 }
2591
2592 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2593 stackChunkOop chunk = _cont.tail();
2594
2595 if (!_stream.is_done()) {
2596 assert(_stream.sp() >= chunk->sp_address(), "");
2597 chunk->set_sp(chunk->to_offset(_stream.sp()));
2598 chunk->set_pc(_stream.pc());
2599 } else {
2600 chunk->set_sp(chunk->bottom());
2601 chunk->set_pc(nullptr);
2602 }
2603 assert(_stream.is_done() == chunk->is_empty(), "");
2604
2605 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2606 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2607
2608 _cont.set_argsize(argsize);
2609 entry = new_entry_frame();
2610
2611 assert(entry.sp() == _cont.entrySP(), "");
2612 assert(Continuation::is_continuation_enterSpecial(entry), "");
2613 assert(_cont.is_entry_frame(entry), "");
2614 }
2615
2616 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2617 LogTarget(Trace, continuations) lt;
2618 if (lt.develop_is_enabled()) {
2619 LogStream ls(lt);
2620 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2621 assert(hf.is_heap_frame(), "should be");
2622 hf.print_value_on(&ls);
2623 }
2624 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2625 }
2626
2627 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2628 #ifdef ASSERT
2629 LogTarget(Trace, continuations) lt;
2630 if (lt.develop_is_enabled()) {
2631 LogStream ls(lt);
2632 ls.print_cr("thawed frame:");
2633 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2634 }
2635 #endif
2636 }
2637
2638 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom, bool augmented) {
2639 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2640 if (bottom) {
2641 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2642 : StubRoutines::cont_returnBarrier());
2643 } else if (caller.is_compiled_frame()){
2644 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2645 // If the caller is not deoptimized, pc is unchanged.
2646 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc(), augmented /*callee_augmented*/);
2647 }
2648
2649 patch_pd(f, caller);
2650
2651 if (f.is_interpreted_frame()) {
2652 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2653 }
2654
2655 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2656 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2657 }
2658
2659 void ThawBase::clear_bitmap_bits(address start, address end) {
2660 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2661 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2662
2663 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2664 // or they will keep objects that are otherwise unreachable alive.
2665
2666 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2667 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2668 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2669 // anyways and we assert that before returning.
2670 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2671 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2672 stackChunkOop chunk = _cont.tail();
2673 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2674 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2675 }
2676
2677 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2678 frame top(sp);
2679 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2680 DEBUG_ONLY(verify_frame_kind(top, preempt_kind);)
2681 NOT_PRODUCT(int64_t tid = _thread->monitor_owner_id();)
2682
2683 // Finish the VTMS transition.
2684 assert(_thread->is_in_vthread_transition(), "must be");
2685 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2686 if (is_vthread) {
2687 #if INCLUDE_JVMTI
2688 if (MountUnmountDisabler::notify_jvmti_events()) {
2689 jvmti_mount_end(_thread, _cont, top, preempt_kind);
2690 } else
2691 #endif
2692 { // Faster version of MountUnmountDisabler::end_transition() to avoid
2693 // unnecessary extra instructions from jvmti_mount_end().
2694 java_lang_Thread::set_is_in_vthread_transition(_thread->vthread(), false);
2695 _thread->set_is_in_vthread_transition(false);
2696 }
2697 }
2698
2699 if (fast_case) {
2700 // If we thawed in the slow path the runtime stub/native wrapper frame already
2701 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2702 // we copied the fp patched during freeze, which will now have to be fixed.
2703 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2704 int fsize = top.cb()->frame_size();
2705 patch_pd(top, sp + fsize);
2706 }
2707
2708 if (preempt_kind == Continuation::object_wait) {
2709 // Check now if we need to throw IE exception.
2710 bool throw_ie = _thread->pending_interrupted_exception();
2711 if (throw_ie) {
2712 throw_interrupted_exception(_thread, top);
2713 _thread->set_pending_interrupted_exception(false);
2714 }
2715 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT" after preemption on Object.wait%s", tid, throw_ie ? "(throwing IE)" : "");
2716 } else if (preempt_kind == Continuation::monitorenter) {
2717 if (top.is_runtime_frame()) {
2718 // The continuation might now run on a different platform thread than the previous time so
2719 // we need to adjust the current thread saved in the stub frame before restoring registers.
2720 JavaThread** thread_addr = frame::saved_thread_address(top);
2721 if (thread_addr != nullptr) *thread_addr = _thread;
2722 }
2723 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT " after preemption on monitorenter", tid);
2724 } else {
2725 // We need to redo the original call into the VM. First though, we need
2726 // to exit the monitor we just acquired (except on preemption cancelled
2727 // case where it was already released).
2728 assert(preempt_kind == Continuation::object_locker, "");
2729 if (_init_lock != nullptr) _init_lock->exit(_thread);
2730 sp = redo_vmcall(_thread, top);
2731 }
2732 return sp;
2733 }
2734
2735 intptr_t* ThawBase::redo_vmcall(JavaThread* current, frame& top) {
2736 assert(!current->preempting(), "");
2737 NOT_PRODUCT(int64_t tid = current->monitor_owner_id();)
2738 intptr_t* sp = top.sp();
2739
2740 {
2741 HandleMarkCleaner hmc(current); // Cleanup all handles (including so._conth) before returning to Java.
2742 ContinuationWrapper::SafepointOp so(current, _cont);
2743 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2744
2745 Method* m = top.interpreter_frame_method();
2746 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
2747 Bytecodes::Code code = current_bytecode.code();
2748 log_develop_trace(continuations, preempt)("Redoing InterpreterRuntime::%s for " INT64_FORMAT, code == Bytecodes::Code::_new ? "_new" : "resolve_from_cache", tid);
2749
2750 // These InterpreterRuntime entry points use JRT_ENTRY which uses a HandleMarkCleaner.
2751 // Create a HandeMark to avoid destroying so._conth.
2752 HandleMark hm(current);
2753 DEBUG_ONLY(JavaThread::AtRedoVMCall apvmc(current);)
2754 if (code == Bytecodes::Code::_new) {
2755 InterpreterRuntime::_new(current, m->constants(), current_bytecode.get_index_u2(code));
2756 } else {
2757 InterpreterRuntime::resolve_from_cache(current, code);
2758 }
2759 }
2760
2761 if (current->preempting()) {
2762 // Preempted again so we just arrange to return to preempt stub to unmount.
2763 sp = push_preempt_adapter();
2764 current->set_preempt_alternate_return(nullptr);
2765 bool cancelled = current->preemption_cancelled();
2766 if (cancelled) {
2767 // Since preemption was cancelled, the thread will call thaw again from the preempt
2768 // stub. These retries could happen several times due to contention on the init_lock,
2769 // so just let the vthread umount to give a chance for other vthreads to run.
2770 current->set_preemption_cancelled(false);
2771 oop vthread = current->vthread();
2772 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
2773 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::YIELDING);
2774 #if INCLUDE_JVMTI
2775 if (current->contended_entered_monitor() != nullptr) {
2776 current->set_contended_entered_monitor(nullptr);
2777 }
2778 #endif
2779 }
2780 log_develop_trace(continuations, preempt)("Preempted " INT64_FORMAT " again%s", tid, cancelled ? "(preemption cancelled, setting state to YIELDING)" : "");
2781 } else {
2782 log_develop_trace(continuations, preempt)("Call succesful, resuming " INT64_FORMAT, tid);
2783 }
2784 return sp;
2785 }
2786
2787 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2788 HandleMarkCleaner hm(current); // Cleanup all handles (including so._conth) before returning to Java.
2789 ContinuationWrapper::SafepointOp so(current, _cont);
2790 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2791 JRT_BLOCK
2792 THROW(vmSymbols::java_lang_InterruptedException());
2793 JRT_BLOCK_END
2794 }
2795
2796 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top) {
2797 assert(hf.is_interpreted_frame(), "");
2798
2799 if (UNLIKELY(seen_by_gc())) {
2800 if (is_top && _process_args_at_top) {
2801 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_with_args());
2802 } else {
2803 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2804 }
2805 }
2806
2807 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2808
2809 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2810
2811 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2812
2813 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2814
2815 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2816 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2817 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2818 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2819
2820 assert(hf.is_heap_frame(), "should be");
2821 assert(!f.is_heap_frame(), "should not be");
2822
2823 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2824 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2825
2826 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2827 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2828 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2829
2830 // Make sure the relativized locals is already set.
2831 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2832
2833 derelativize_interpreted_frame_metadata(hf, f);
2834 patch(f, caller, is_bottom_frame);
2835
2836 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2837 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2838
2839 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2840
2841 maybe_set_fastpath(f.sp());
2842
2843 Method* m = hf.interpreter_frame_method();
2844 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2845 const int locals = m->max_locals();
2846
2847 if (!is_bottom_frame) {
2848 // can only fix caller once this frame is thawed (due to callee saved regs)
2849 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2850 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2851 assert(hf.is_heap_frame(), "should be");
2852 address start = (address)(heap_frame_bottom - locals);
2853 address end = (address)heap_frame_bottom;
2854 clear_bitmap_bits(start, end);
2855 }
2856
2857 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2858 caller = f;
2859 }
2860
2861 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2862 assert(hf.is_compiled_frame(), "");
2863 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2864
2865 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2866 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2867 }
2868
2869 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2870
2871 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2872
2873 assert(caller.sp() == caller.unextended_sp(), "");
2874
2875 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2876 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2877 }
2878
2879 int fsize = 0;
2880 int added_argsize = 0;
2881 bool augmented = hf.was_augmented_on_entry(fsize);
2882 if (!augmented) {
2883 added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2884 fsize += added_argsize;
2885 }
2886 assert(!is_bottom_frame || !augmented, "");
2887
2888
2889 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2890 // yet laid out in the stack, and so the original_pc is not stored in it.
2891 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2892 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame, augmented ? fsize - hf.cb()->frame_size() : 0);
2893 assert(f.cb()->frame_size() == (int)(caller.sp() - f.sp()), "");
2894
2895 intptr_t* const stack_frame_top = f.sp();
2896 intptr_t* const heap_frame_top = hf.unextended_sp();
2897 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2898 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2899 // copy metadata, except the metadata at the top of the (unextended) entry frame
2900 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2901
2902 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2903 // (we might have one padding word for alignment)
2904 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2905 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2906
2907 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2908
2909 patch(f, caller, is_bottom_frame, augmented);
2910
2911 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2912 assert(!f.is_deoptimized_frame(), "");
2913 if (hf.is_deoptimized_frame()) {
2914 maybe_set_fastpath(f.sp());
2915 } else if (_thread->is_interp_only_mode()
2916 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2917 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2918 // cannot rely on nmethod patching for deopt.
2919 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2920
2921 log_develop_trace(continuations)("Deoptimizing thawed frame");
2922 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2923
2924 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2925 assert(f.is_deoptimized_frame(), "");
2926 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2927 maybe_set_fastpath(f.sp());
2928 }
2929
2930 if (!is_bottom_frame) {
2931 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2932 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2933 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2934 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2935 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2936 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2937 clear_bitmap_bits(start, start + argsize_in_bytes);
2938 }
2939
2940 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2941 caller = f;
2942 }
2943
2944 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2945 DEBUG_ONLY(_frames++;)
2946
2947 if (UNLIKELY(seen_by_gc())) {
2948 // Process the stub's caller here since we might need the full map.
2949 RegisterMap map(nullptr,
2950 RegisterMap::UpdateMap::include,
2951 RegisterMap::ProcessFrames::skip,
2952 RegisterMap::WalkContinuation::skip);
2953 map.set_include_argument_oops(false);
2954 _stream.next(&map);
2955 assert(!_stream.is_done(), "");
2956 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2957 } else {
2958 _stream.next(SmallRegisterMap::instance_no_args());
2959 assert(!_stream.is_done(), "");
2960 }
2961
2962 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2963
2964 assert(caller.is_compiled_frame(), "");
2965 assert(caller.sp() == caller.unextended_sp(), "");
2966
2967 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2968
2969 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2970 intptr_t* stack_frame_top = f.sp();
2971 intptr_t* heap_frame_top = hf.sp();
2972 int fsize = ContinuationHelper::StubFrame::size(hf);
2973
2974 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2975 fsize + frame::metadata_words);
2976
2977 patch(f, caller, false /*is_bottom_frame*/);
2978
2979 // can only fix caller once this frame is thawed (due to callee saved regs)
2980 RegisterMap map(nullptr,
2981 RegisterMap::UpdateMap::include,
2982 RegisterMap::ProcessFrames::skip,
2983 RegisterMap::WalkContinuation::skip);
2984 map.set_include_argument_oops(false);
2985 f.oop_map()->update_register_map(&f, &map);
2986 ContinuationHelper::update_register_map_with_callee(caller, &map);
2987 _cont.tail()->fix_thawed_frame(caller, &map);
2988
2989 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2990 caller = f;
2991 }
2992
2993 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2994 assert(hf.is_native_frame(), "");
2995 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2996
2997 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2998 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2999 }
3000
3001 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
3002 assert(!is_bottom_frame, "");
3003
3004 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
3005
3006 assert(caller.sp() == caller.unextended_sp(), "");
3007
3008 if (caller.is_interpreted_frame()) {
3009 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
3010 }
3011
3012 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
3013 // yet laid out in the stack, and so the original_pc is not stored in it.
3014 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
3015 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
3016 intptr_t* const stack_frame_top = f.sp();
3017 intptr_t* const heap_frame_top = hf.unextended_sp();
3018
3019 int fsize = ContinuationHelper::NativeFrame::size(hf);
3020 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
3021
3022 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
3023 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
3024 int sz = fsize + frame::metadata_words_at_bottom;
3025
3026 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
3027
3028 patch(f, caller, false /* bottom */);
3029
3030 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
3031 assert(!f.is_deoptimized_frame(), "");
3032 assert(!hf.is_deoptimized_frame(), "");
3033 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
3034
3035 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
3036 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
3037
3038 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
3039 caller = f;
3040 }
3041
3042 void ThawBase::finish_thaw(frame& f) {
3043 stackChunkOop chunk = _cont.tail();
3044
3045 if (chunk->is_empty()) {
3046 // Only remove chunk from list if it can't be reused for another freeze
3047 if (seen_by_gc()) {
3048 _cont.set_tail(chunk->parent());
3049 } else {
3050 chunk->set_has_mixed_frames(false);
3051 }
3052 chunk->set_max_thawing_size(0);
3053 } else {
3054 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
3055 }
3056 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
3057
3058 if (!is_aligned(f.sp(), frame::frame_alignment)) {
3059 assert(f.is_interpreted_frame(), "");
3060 f.set_sp(align_down(f.sp(), frame::frame_alignment));
3061 }
3062 push_return_frame(f);
3063 // can only fix caller after push_return_frame (due to callee saved regs)
3064 if (_process_args_at_top) {
3065 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_with_args());
3066 } else {
3067 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_no_args());
3068 }
3069
3070 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
3071
3072 log_develop_trace(continuations)("thawed %d frames", _frames);
3073
3074 LogTarget(Trace, continuations) lt;
3075 if (lt.develop_is_enabled()) {
3076 LogStream ls(lt);
3077 ls.print_cr("top hframe after (thaw):");
3078 _cont.last_frame().print_value_on(&ls);
3079 }
3080 }
3081
3082 void ThawBase::push_return_frame(const frame& f) { // see generate_cont_thaw
3083 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
3084 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
3085
3086 LogTarget(Trace, continuations) lt;
3087 if (lt.develop_is_enabled()) {
3088 LogStream ls(lt);
3089 ls.print_cr("push_return_frame");
3090 f.print_value_on(&ls);
3091 }
3092
3093 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
3094 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
3095 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
3096 ContinuationHelper::push_pd(f);
3097
3098 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
3099 }
3100
3101 // returns new top sp
3102 // called after preparations (stack overflow check and making room)
3103 template<typename ConfigT>
3104 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
3105 assert(thread == JavaThread::current(), "Must be current thread");
3106
3107 CONT_JFR_ONLY(EventContinuationThaw event;)
3108
3109 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
3110
3111 ContinuationEntry* entry = thread->last_continuation();
3112 assert(entry != nullptr, "");
3113 oop oopCont = entry->cont_oop(thread);
3114
3115 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
3116 assert(oopCont == get_continuation(thread), "");
3117 verify_continuation(oopCont);
3118
3119 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
3120
3121 ContinuationWrapper cont(thread, oopCont);
3122 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
3123
3124 #ifdef ASSERT
3125 set_anchor_to_entry(thread, cont.entry());
3126 log_frames(thread);
3127 clear_anchor(thread);
3128 #endif
3129
3130 Thaw<ConfigT> thw(thread, cont);
3131 intptr_t* const sp = thw.thaw(kind);
3132 assert(is_aligned(sp, frame::frame_alignment), "");
3133 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp);)
3134
3135 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
3136
3137 verify_continuation(cont.continuation());
3138 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
3139
3140 return sp;
3141 }
3142
3143 #ifdef ASSERT
3144 static void do_deopt_after_thaw(JavaThread* thread) {
3145 int i = 0;
3146 StackFrameStream fst(thread, true, false);
3147 fst.register_map()->set_include_argument_oops(false);
3148 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3149 for (; !fst.is_done(); fst.next()) {
3150 if (fst.current()->cb()->is_nmethod()) {
3151 nmethod* nm = fst.current()->cb()->as_nmethod();
3152 if (!nm->method()->is_continuation_native_intrinsic()) {
3153 nm->make_deoptimized();
3154 }
3155 }
3156 }
3157 }
3158
3159 class ThawVerifyOopsClosure: public OopClosure {
3160 intptr_t* _p;
3161 outputStream* _st;
3162 bool is_good_oop(oop o) {
3163 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
3164 }
3165 public:
3166 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
3167 intptr_t* p() { return _p; }
3168 void reset() { _p = nullptr; }
3169
3170 virtual void do_oop(oop* p) {
3171 oop o = *p;
3172 if (o == nullptr || is_good_oop(o)) {
3173 return;
3174 }
3175 _p = (intptr_t*)p;
3176 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
3177 }
3178 virtual void do_oop(narrowOop* p) {
3179 oop o = RawAccess<>::oop_load(p);
3180 if (o == nullptr || is_good_oop(o)) {
3181 return;
3182 }
3183 _p = (intptr_t*)p;
3184 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
3185 }
3186 };
3187
3188 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
3189 assert(thread->has_last_Java_frame(), "");
3190
3191 ResourceMark rm;
3192 ThawVerifyOopsClosure cl(st);
3193 NMethodToOopClosure cf(&cl, false);
3194
3195 StackFrameStream fst(thread, true, false);
3196 fst.register_map()->set_include_argument_oops(false);
3197 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3198 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
3199 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
3200 st->print_cr(">>> do_verify_after_thaw deopt");
3201 fst.current()->deoptimize(nullptr);
3202 fst.current()->print_on(st);
3203 }
3204
3205 fst.current()->oops_do(&cl, &cf, fst.register_map());
3206 if (cl.p() != nullptr) {
3207 frame fr = *fst.current();
3208 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
3209 fr.print_on(st);
3210 if (!fr.is_interpreted_frame()) {
3211 st->print_cr("size: %d argsize: %d",
3212 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
3213 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
3214 }
3215 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
3216 if (reg != nullptr) {
3217 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
3218 }
3219 cl.reset();
3220 DEBUG_ONLY(thread->print_frame_layout();)
3221 if (chunk != nullptr) {
3222 chunk->print_on(true, st);
3223 }
3224 return false;
3225 }
3226 }
3227 return true;
3228 }
3229
3230 static void log_frames(JavaThread* thread) {
3231 const static int show_entry_callers = 3;
3232 LogTarget(Trace, continuations) lt;
3233 if (!lt.develop_is_enabled()) {
3234 return;
3235 }
3236 LogStream ls(lt);
3237
3238 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3239 if (!thread->has_last_Java_frame()) {
3240 ls.print_cr("NO ANCHOR!");
3241 }
3242
3243 RegisterMap map(thread,
3244 RegisterMap::UpdateMap::include,
3245 RegisterMap::ProcessFrames::include,
3246 RegisterMap::WalkContinuation::skip);
3247 map.set_include_argument_oops(false);
3248
3249 if (false) {
3250 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3251 f.print_on(&ls);
3252 }
3253 } else {
3254 map.set_skip_missing(true);
3255 ResetNoHandleMark rnhm;
3256 ResourceMark rm;
3257 HandleMark hm(Thread::current());
3258 FrameValues values;
3259
3260 int i = 0;
3261 int post_entry = -1;
3262 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3263 f.describe(values, i, &map, i == 0);
3264 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3265 post_entry++;
3266 if (post_entry >= show_entry_callers)
3267 break;
3268 }
3269 values.print_on(thread, &ls);
3270 }
3271
3272 ls.print_cr("======= end frames =========");
3273 }
3274
3275 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp) {
3276 intptr_t* sp0 = sp;
3277 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
3278
3279 bool preempted = false;
3280 stackChunkOop tail = cont.tail();
3281 if (tail != nullptr && tail->preempted()) {
3282 // Still preempted (monitor not acquired) so no frames were thawed.
3283 set_anchor(thread, cont.entrySP(), cont.entryPC());
3284 preempted = true;
3285 } else {
3286 set_anchor(thread, sp0);
3287 }
3288
3289 log_frames(thread);
3290 if (LoomVerifyAfterThaw) {
3291 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
3292 }
3293 assert(preempted || ContinuationEntry::assert_entry_frame_laid_out(thread), "");
3294 clear_anchor(thread);
3295
3296 LogTarget(Trace, continuations) lt;
3297 if (lt.develop_is_enabled()) {
3298 LogStream ls(lt);
3299 ls.print_cr("Jumping to frame (thaw):");
3300 frame(sp).print_value_on(&ls);
3301 }
3302 }
3303 #endif // ASSERT
3304
3305 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3306
3307 #ifdef ASSERT
3308 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3309 ResourceMark rm;
3310 FrameValues values;
3311 assert(f.get_cb() != nullptr, "");
3312 RegisterMap map(f.is_heap_frame() ?
3313 nullptr :
3314 JavaThread::current(),
3315 RegisterMap::UpdateMap::include,
3316 RegisterMap::ProcessFrames::skip,
3317 RegisterMap::WalkContinuation::skip);
3318 map.set_include_argument_oops(false);
3319 map.set_skip_missing(true);
3320 if (callee_complete) {
3321 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3322 }
3323 const_cast<frame&>(f).describe(values, 0, &map, true);
3324 values.print_on(static_cast<JavaThread*>(nullptr), st);
3325 }
3326 #endif
3327
3328 static address thaw_entry = nullptr;
3329 static address freeze_entry = nullptr;
3330 static address freeze_preempt_entry = nullptr;
3331
3332 address Continuation::thaw_entry() {
3333 return ::thaw_entry;
3334 }
3335
3336 address Continuation::freeze_entry() {
3337 return ::freeze_entry;
3338 }
3339
3340 address Continuation::freeze_preempt_entry() {
3341 return ::freeze_preempt_entry;
3342 }
3343
3344 class ConfigResolve {
3345 public:
3346 static void resolve() { resolve_compressed(); }
3347
3348 static void resolve_compressed() {
3349 UseCompressedOops ? resolve_gc<true>()
3350 : resolve_gc<false>();
3351 }
3352
3353 private:
3354 template <bool use_compressed>
3355 static void resolve_gc() {
3356 BarrierSet* bs = BarrierSet::barrier_set();
3357 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3358 switch (bs->kind()) {
3359 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3360 case BarrierSet::bs_name: { \
3361 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3362 } \
3363 break;
3364 FOR_EACH_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3365 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3366
3367 default:
3368 fatal("BarrierSet resolving not implemented");
3369 };
3370 }
3371
3372 template <bool use_compressed, typename BarrierSetT>
3373 static void resolve() {
3374 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3375
3376 freeze_entry = (address)freeze<SelectedConfigT>;
3377 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3378
3379 // If we wanted, we could templatize by kind and have three different thaw entries
3380 thaw_entry = (address)thaw<SelectedConfigT>;
3381 }
3382 };
3383
3384 void Continuation::init() {
3385 ConfigResolve::resolve();
3386 }