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/objectMonitor.inline.hpp"
62 #include "runtime/orderAccess.hpp"
63 #include "runtime/prefetch.inline.hpp"
64 #include "runtime/sharedRuntime.hpp"
65 #include "runtime/smallRegisterMap.inline.hpp"
66 #include "runtime/stackChunkFrameStream.inline.hpp"
67 #include "runtime/stackFrameStream.inline.hpp"
68 #include "runtime/stackOverflow.hpp"
69 #include "runtime/stackWatermarkSet.inline.hpp"
70 #include "runtime/vframe.inline.hpp"
71 #include "runtime/vframe_hp.hpp"
72 #include "utilities/debug.hpp"
73 #include "utilities/exceptions.hpp"
74 #include "utilities/macros.hpp"
75 #include "utilities/vmError.hpp"
76 #if INCLUDE_ZGC
77 #include "gc/z/zStackChunkGCData.inline.hpp"
78 #endif
79 #if INCLUDE_JFR
80 #include "jfr/jfr.inline.hpp"
81 #endif
82 #ifdef COMPILER1
83 #include "c1/c1_Runtime1.hpp"
84 #endif
85 #ifdef COMPILER2
86 #include "opto/runtime.hpp"
87 #endif
88
89 /*
90 * This file contains the implementation of continuation freezing (yield) and thawing (run).
91 *
92 * This code is very latency-critical and very hot. An ordinary and well-behaved server application
93 * would likely call these operations many thousands of times per second second, on every core.
94 *
95 * Freeze might be called every time the application performs any I/O operation, every time it
96 * acquires a j.u.c. lock, every time it takes a message from a queue, and thaw can be called
97 * multiple times in each of those cases, as it is called by the return barrier, which may be
98 * invoked on method return.
99 *
100 * The amortized budget for each of those two operations is ~100-150ns. That is why, for
101 * example, every effort is made to avoid Java-VM transitions as much as possible.
102 *
103 * On the fast path, all frames are known to be compiled, and the chunk requires no barriers
104 * and so frames simply copied, and the bottom-most one is patched.
105 * On the slow path, internal pointers in interpreted frames are de/relativized to/from offsets
106 * and absolute pointers, and barriers invoked.
107 */
108
109 /************************************************
110
111 Thread-stack layout on freeze/thaw.
112 See corresponding stack-chunk layout in instanceStackChunkKlass.hpp
113
114 +----------------------------+
115 | . |
116 | . |
117 | . |
118 | carrier frames |
119 | |
120 |----------------------------|
121 | |
122 | Continuation.run |
123 | |
124 |============================|
125 | enterSpecial frame |
126 | pc |
127 | rbp |
128 | ----- |
129 ^ | int argsize | = ContinuationEntry
130 | | oopDesc* cont |
131 | | oopDesc* chunk |
132 | | ContinuationEntry* parent |
133 | | ... |
134 | |============================| <------ JavaThread::_cont_entry = entry->sp()
135 | | ? alignment word ? |
136 | |----------------------------| <--\
137 | | | |
138 | | ? caller stack args ? | | argsize (might not be 2-word aligned) words
139 Address | | | | Caller is still in the chunk.
140 | |----------------------------| |
141 | | pc (? return barrier ?) | | This pc contains the return barrier when the bottom-most frame
142 | | rbp | | isn't the last one in the continuation.
143 | | | |
144 | | frame | |
145 | | | |
146 +----------------------------| \__ Continuation frames to be frozen/thawed
147 | | /
148 | frame | |
149 | | |
150 |----------------------------| |
151 | | |
152 | frame | |
153 | | |
154 |----------------------------| <--/
155 | |
156 | doYield/safepoint stub | When preempting forcefully, we could have a safepoint stub
157 | | instead of a doYield stub
158 |============================| <- the sp passed to freeze
159 | |
160 | Native freeze/thaw frames |
161 | . |
162 | . |
163 | . |
164 +----------------------------+
165
166 ************************************************/
167
168 static const bool TEST_THAW_ONE_CHUNK_FRAME = false; // force thawing frames one-at-a-time for testing
169
170 #define CONT_JFR false // emit low-level JFR events that count slow/fast path for continuation performance debugging only
171 #if CONT_JFR
172 #define CONT_JFR_ONLY(code) code
173 #else
174 #define CONT_JFR_ONLY(code)
175 #endif
176
177 // TODO: See AbstractAssembler::generate_stack_overflow_check,
178 // Compile::bang_size_in_bytes(), m->as_SafePoint()->jvms()->interpreter_frame_size()
179 // when we stack-bang, we need to update a thread field with the lowest (farthest) bang point.
180
181 // Data invariants are defined by Continuation::debug_verify_continuation and Continuation::debug_verify_stack_chunk
182
183 // Used to just annotatate cold/hot branches
184 #define LIKELY(condition) (condition)
185 #define UNLIKELY(condition) (condition)
186
187 // debugging functions
188 #ifdef ASSERT
189 extern "C" bool dbg_is_safe(const void* p, intptr_t errvalue); // address p is readable and *(intptr_t*)p != errvalue
190
191 static void verify_continuation(oop continuation) { Continuation::debug_verify_continuation(continuation); }
192
193 static void do_deopt_after_thaw(JavaThread* thread);
194 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st);
195 static void log_frames(JavaThread* thread);
196 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp);
197 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st = tty);
198 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);
199
200 #define assert_pfl(p, ...) \
201 do { \
202 if (!(p)) { \
203 JavaThread* t = JavaThread::active(); \
204 if (t->has_last_Java_frame()) { \
205 tty->print_cr("assert(" #p ") failed:"); \
206 t->print_frame_layout(); \
207 } \
208 } \
209 vmassert(p, __VA_ARGS__); \
210 } while(0)
211
212 #else
213 static void verify_continuation(oop continuation) { }
214 #define assert_pfl(p, ...)
215 #endif
216
217 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint);
218 template<typename ConfigT, bool preempt> static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp);
219
220 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier);
221 template<typename ConfigT> static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind);
222
223
224 // Entry point to freeze. Transitions are handled manually
225 // Called from gen_continuation_yield() in sharedRuntime_<cpu>.cpp through Continuation::freeze_entry();
226 template<typename ConfigT>
227 static JRT_BLOCK_ENTRY(int, freeze(JavaThread* current, intptr_t* sp))
228 assert(sp == current->frame_anchor()->last_Java_sp(), "");
229
230 if (current->raw_cont_fastpath() > current->last_continuation()->entry_sp() || current->raw_cont_fastpath() < sp) {
231 current->set_cont_fastpath(nullptr);
232 }
233
234 return checked_cast<int>(ConfigT::freeze(current, sp));
235 JRT_END
236
237 JRT_LEAF(int, Continuation::prepare_thaw(JavaThread* thread, bool return_barrier))
238 return prepare_thaw_internal(thread, return_barrier);
239 JRT_END
240
241 template<typename ConfigT>
242 static JRT_LEAF(intptr_t*, thaw(JavaThread* thread, int kind))
243 // TODO: JRT_LEAF and NoHandleMark is problematic for JFR events.
244 // vFrameStreamCommon allocates Handles in RegisterMap for continuations.
245 // Also the preemption case with JVMTI events enabled might safepoint so
246 // undo the NoSafepointVerifier here and rely on handling by ContinuationWrapper.
247 // JRT_ENTRY instead?
248 ResetNoHandleMark rnhm;
249 DEBUG_ONLY(PauseNoSafepointVerifier pnsv(&__nsv);)
250
251 // we might modify the code cache via BarrierSetNMethod::nmethod_entry_barrier
252 MACOS_AARCH64_ONLY(ThreadWXEnable __wx(WXWrite, thread));
253 return ConfigT::thaw(thread, (Continuation::thaw_kind)kind);
254 JRT_END
255
256 JVM_ENTRY(jint, CONT_isPinned0(JNIEnv* env, jobject cont_scope)) {
257 JavaThread* thread = JavaThread::thread_from_jni_environment(env);
258 return is_pinned0(thread, JNIHandles::resolve(cont_scope), false);
259 }
260 JVM_END
261
262 ///////////
263
264 enum class oop_kind { NARROW, WIDE };
265 template <oop_kind oops, typename BarrierSetT>
266 class Config {
267 public:
268 typedef Config<oops, BarrierSetT> SelfT;
269 using OopT = std::conditional_t<oops == oop_kind::NARROW, narrowOop, oop>;
270
271 static freeze_result freeze(JavaThread* thread, intptr_t* const sp) {
272 freeze_result res = freeze_internal<SelfT, false>(thread, sp);
273 JFR_ONLY(assert((res == freeze_ok) || (res == thread->last_freeze_fail_result()), "freeze failure not set"));
274 return res;
275 }
276
277 static freeze_result freeze_preempt(JavaThread* thread, intptr_t* const sp) {
278 return freeze_internal<SelfT, true>(thread, sp);
279 }
280
281 static intptr_t* thaw(JavaThread* thread, Continuation::thaw_kind kind) {
282 return thaw_internal<SelfT>(thread, kind);
283 }
284 };
285
286 #ifdef _WINDOWS
287 static void map_stack_pages(JavaThread* thread, size_t size, address sp) {
288 address new_sp = sp - size;
289 address watermark = thread->stack_overflow_state()->shadow_zone_growth_watermark();
290
291 if (new_sp < watermark) {
292 size_t page_size = os::vm_page_size();
293 address last_touched_page = watermark - StackOverflow::stack_shadow_zone_size();
294 size_t pages_to_touch = align_up(watermark - new_sp, page_size) / page_size;
295 while (pages_to_touch-- > 0) {
296 last_touched_page -= page_size;
297 *last_touched_page = 0;
298 }
299 thread->stack_overflow_state()->set_shadow_zone_growth_watermark(new_sp);
300 }
301 }
302 #endif
303
304 static bool stack_overflow_check(JavaThread* thread, size_t size, address sp) {
305 const size_t page_size = os::vm_page_size();
306 if (size > page_size) {
307 if (sp - size < thread->stack_overflow_state()->shadow_zone_safe_limit()) {
308 return false;
309 }
310 WINDOWS_ONLY(map_stack_pages(thread, size, sp));
311 }
312 return true;
313 }
314
315 #ifdef ASSERT
316 static oop get_continuation(JavaThread* thread) {
317 assert(thread != nullptr, "");
318 assert(thread->threadObj() != nullptr, "");
319 return java_lang_Thread::continuation(thread->threadObj());
320 }
321 #endif // ASSERT
322
323 inline void clear_anchor(JavaThread* thread) {
324 thread->frame_anchor()->clear();
325 }
326
327 static void set_anchor(JavaThread* thread, intptr_t* sp, address pc) {
328 assert(pc != nullptr, "");
329
330 JavaFrameAnchor* anchor = thread->frame_anchor();
331 anchor->set_last_Java_sp(sp);
332 anchor->set_last_Java_pc(pc);
333 ContinuationHelper::set_anchor_pd(anchor, sp);
334
335 assert(thread->has_last_Java_frame(), "");
336 assert(thread->last_frame().cb() != nullptr, "");
337 }
338
339 static void set_anchor(JavaThread* thread, intptr_t* sp) {
340 address pc = ContinuationHelper::return_address_at(
341 sp - frame::sender_sp_ret_address_offset());
342 set_anchor(thread, sp, pc);
343 }
344
345 static void set_anchor_to_entry(JavaThread* thread, ContinuationEntry* entry) {
346 JavaFrameAnchor* anchor = thread->frame_anchor();
347 anchor->set_last_Java_sp(entry->entry_sp());
348 anchor->set_last_Java_pc(entry->entry_pc());
349 ContinuationHelper::set_anchor_to_entry_pd(anchor, entry);
350
351 assert(thread->has_last_Java_frame(), "");
352 assert(thread->last_frame().cb() != nullptr, "");
353 }
354
355 #if CONT_JFR
356 class FreezeThawJfrInfo : public StackObj {
357 short _e_size;
358 short _e_num_interpreted_frames;
359 public:
360
361 FreezeThawJfrInfo() : _e_size(0), _e_num_interpreted_frames(0) {}
362 inline void record_interpreted_frame() { _e_num_interpreted_frames++; }
363 inline void record_size_copied(int size) { _e_size += size << LogBytesPerWord; }
364 template<typename Event> void post_jfr_event(Event *e, oop continuation, JavaThread* jt);
365 };
366
367 template<typename Event> void FreezeThawJfrInfo::post_jfr_event(Event* e, oop continuation, JavaThread* jt) {
368 if (e->should_commit()) {
369 log_develop_trace(continuations)("JFR event: iframes: %d size: %d", _e_num_interpreted_frames, _e_size);
370 e->set_carrierThread(JFR_JVM_THREAD_ID(jt));
371 e->set_continuationClass(continuation->klass());
372 e->set_interpretedFrames(_e_num_interpreted_frames);
373 e->set_size(_e_size);
374 e->commit();
375 }
376 }
377 #endif // CONT_JFR
378
379 /////////////// FREEZE ////
380
381 class FreezeBase : public StackObj {
382 protected:
383 JavaThread* const _thread;
384 ContinuationWrapper& _cont;
385 bool _barriers; // only set when we allocate a chunk
386
387 intptr_t* _bottom_address;
388
389 // Used for preemption only
390 const bool _preempt;
391 frame _last_frame;
392
393 // Used to support freezing with held monitors
394 int _monitors_in_lockstack;
395
396 int _freeze_size; // total size of all frames plus metadata in words.
397 int _total_align_size;
398
399 intptr_t* _cont_stack_top;
400 intptr_t* _cont_stack_bottom;
401
402 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
403
404 #ifdef ASSERT
405 intptr_t* _orig_chunk_sp;
406 int _fast_freeze_size;
407 bool _empty;
408 #endif
409
410 JvmtiSampledObjectAllocEventCollector* _jvmti_event_collector;
411
412 NOT_PRODUCT(int _frames;)
413 DEBUG_ONLY(intptr_t* _last_write;)
414
415 inline FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp, bool preempt);
416
417 public:
418 NOINLINE freeze_result freeze_slow();
419 void freeze_fast_existing_chunk();
420
421 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
422 void set_jvmti_event_collector(JvmtiSampledObjectAllocEventCollector* jsoaec) { _jvmti_event_collector = jsoaec; }
423
424 inline int size_if_fast_freeze_available();
425
426 inline frame& last_frame() { return _last_frame; }
427
428 #ifdef ASSERT
429 bool check_valid_fast_path();
430 #endif
431
432 protected:
433 inline void init_rest();
434 void throw_stack_overflow_on_humongous_chunk();
435
436 // fast path
437 inline void copy_to_chunk(intptr_t* from, intptr_t* to, int size);
438 inline void unwind_frames();
439 inline void patch_stack_pd(intptr_t* frame_sp, intptr_t* heap_sp);
440
441 // slow path
442 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) = 0;
443
444 int cont_size() { return pointer_delta_as_int(_cont_stack_bottom, _cont_stack_top); }
445
446 private:
447 // slow path
448 frame freeze_start_frame();
449 frame freeze_start_frame_on_preempt();
450 NOINLINE freeze_result recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top);
451 inline frame freeze_start_frame_yield_stub();
452 template<typename FKind>
453 inline freeze_result recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize);
454 inline void before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame);
455 inline void after_freeze_java_frame(const frame& hf, bool is_bottom_frame);
456 freeze_result finalize_freeze(const frame& callee, frame& caller, int argsize);
457 void patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame);
458 NOINLINE freeze_result recurse_freeze_interpreted_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
459 freeze_result recurse_freeze_compiled_frame(frame& f, frame& caller, int callee_argsize, bool callee_interpreted);
460 NOINLINE freeze_result recurse_freeze_stub_frame(frame& f, frame& caller);
461 NOINLINE freeze_result recurse_freeze_native_frame(frame& f, frame& caller);
462 NOINLINE void finish_freeze(const frame& f, const frame& top);
463
464 void freeze_lockstack(stackChunkOop chunk);
465
466 inline bool stack_overflow();
467
468 static frame sender(const frame& f) { return f.is_interpreted_frame() ? sender<ContinuationHelper::InterpretedFrame>(f)
469 : sender<ContinuationHelper::NonInterpretedUnknownFrame>(f); }
470 template<typename FKind> static inline frame sender(const frame& f);
471 template<typename FKind> frame new_heap_frame(frame& f, frame& caller, int size_adjust = 0);
472 inline void set_top_frame_metadata_pd(const frame& hf);
473 inline void patch_pd(frame& callee, const frame& caller, bool is_bottom_frame);
474 inline void patch_pd_unused(intptr_t* sp);
475 void adjust_interpreted_frame_unextended_sp(frame& f);
476 inline void prepare_freeze_interpreted_top_frame(frame& f);
477 static inline void relativize_interpreted_frame_metadata(const frame& f, const frame& hf);
478
479 protected:
480 void freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated));
481 bool freeze_fast_new_chunk(stackChunkOop chunk);
482 };
483
484 template <typename ConfigT>
485 class Freeze : public FreezeBase {
486 private:
487 stackChunkOop allocate_chunk(size_t stack_size, int argsize_md);
488
489 public:
490 inline Freeze(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt)
491 : FreezeBase(thread, cont, frame_sp, preempt) {}
492
493 freeze_result try_freeze_fast();
494
495 protected:
496 virtual stackChunkOop allocate_chunk_slow(size_t stack_size, int argsize_md) override { return allocate_chunk(stack_size, argsize_md); }
497 };
498
499 FreezeBase::FreezeBase(JavaThread* thread, ContinuationWrapper& cont, intptr_t* frame_sp, bool preempt) :
500 _thread(thread), _cont(cont), _barriers(false), _preempt(preempt), _last_frame(false /* no initialization */) {
501 DEBUG_ONLY(_jvmti_event_collector = nullptr;)
502
503 assert(_thread != nullptr, "");
504 assert(_thread->last_continuation()->entry_sp() == _cont.entrySP(), "");
505
506 DEBUG_ONLY(_cont.entry()->verify_cookie();)
507
508 assert(!Interpreter::contains(_cont.entryPC()), "");
509
510 _bottom_address = _cont.entrySP() - _cont.entry_frame_extension();
511 #ifdef _LP64
512 if (((intptr_t)_bottom_address & 0xf) != 0) {
513 _bottom_address--;
514 }
515 assert(is_aligned(_bottom_address, frame::frame_alignment), "");
516 #endif
517
518 log_develop_trace(continuations)("bottom_address: " INTPTR_FORMAT " entrySP: " INTPTR_FORMAT " argsize: " PTR_FORMAT,
519 p2i(_bottom_address), p2i(_cont.entrySP()), (_cont.entrySP() - _bottom_address) << LogBytesPerWord);
520 assert(_bottom_address != nullptr, "");
521 assert(_bottom_address <= _cont.entrySP(), "");
522 DEBUG_ONLY(_last_write = nullptr;)
523
524 assert(_cont.chunk_invariant(), "");
525 assert(!Interpreter::contains(_cont.entryPC()), "");
526 #if !defined(PPC64) || defined(ZERO)
527 static const int doYield_stub_frame_size = frame::metadata_words;
528 #else
529 static const int doYield_stub_frame_size = frame::native_abi_reg_args_size >> LogBytesPerWord;
530 #endif
531 // With preemption doYield() might not have been resolved yet
532 assert(_preempt || SharedRuntime::cont_doYield_stub()->frame_size() == doYield_stub_frame_size, "");
533
534 if (preempt) {
535 _last_frame = _thread->last_frame();
536 }
537
538 // properties of the continuation on the stack; all sizes are in words
539 _cont_stack_top = frame_sp + (!preempt ? doYield_stub_frame_size : 0); // we don't freeze the doYield stub frame
540 _cont_stack_bottom = _cont.entrySP() + (_cont.argsize() == 0 ? frame::metadata_words_at_top : 0)
541 - ContinuationHelper::frame_align_words(_cont.argsize()); // see alignment in thaw
542
543 log_develop_trace(continuations)("freeze size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
544 cont_size(), _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
545 assert(cont_size() > 0, "");
546
547 _monitors_in_lockstack = _thread->lock_stack().monitor_count();
548 }
549
550 void FreezeBase::init_rest() { // we want to postpone some initialization after chunk handling
551 _freeze_size = 0;
552 _total_align_size = 0;
553 NOT_PRODUCT(_frames = 0;)
554 }
555
556 void FreezeBase::freeze_lockstack(stackChunkOop chunk) {
557 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "no room for lockstack");
558
559 _thread->lock_stack().move_to_address((oop*)chunk->start_address());
560 chunk->set_lockstack_size(checked_cast<uint8_t>(_monitors_in_lockstack));
561 chunk->set_has_lockstack(true);
562 }
563
564 void FreezeBase::copy_to_chunk(intptr_t* from, intptr_t* to, int size) {
565 stackChunkOop chunk = _cont.tail();
566 chunk->copy_from_stack_to_chunk(from, to, size);
567 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
568
569 #ifdef ASSERT
570 if (_last_write != nullptr) {
571 assert(_last_write == to + size, "Missed a spot: _last_write: " INTPTR_FORMAT " to+size: " INTPTR_FORMAT
572 " stack_size: %d _last_write offset: " PTR_FORMAT " to+size: " PTR_FORMAT, p2i(_last_write), p2i(to+size),
573 chunk->stack_size(), _last_write-chunk->start_address(), to+size-chunk->start_address());
574 _last_write = to;
575 }
576 #endif
577 }
578
579 static void assert_frames_in_continuation_are_safe(JavaThread* thread) {
580 #ifdef ASSERT
581 StackWatermark* watermark = StackWatermarkSet::get(thread, StackWatermarkKind::gc);
582 if (watermark == nullptr) {
583 return;
584 }
585 ContinuationEntry* ce = thread->last_continuation();
586 RegisterMap map(thread,
587 RegisterMap::UpdateMap::include,
588 RegisterMap::ProcessFrames::skip,
589 RegisterMap::WalkContinuation::skip);
590 map.set_include_argument_oops(false);
591 for (frame f = thread->last_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map)) {
592 watermark->assert_is_frame_safe(f);
593 }
594 #endif // ASSERT
595 }
596
597 // Called _after_ the last possible safepoint during the freeze operation (chunk allocation)
598 void FreezeBase::unwind_frames() {
599 ContinuationEntry* entry = _cont.entry();
600 entry->flush_stack_processing(_thread);
601 assert_frames_in_continuation_are_safe(_thread);
602 JFR_ONLY(Jfr::check_and_process_sample_request(_thread);)
603 set_anchor_to_entry(_thread, entry);
604 }
605
606 template <typename ConfigT>
607 freeze_result Freeze<ConfigT>::try_freeze_fast() {
608 assert(_thread->thread_state() == _thread_in_vm, "");
609 assert(_thread->cont_fastpath(), "");
610
611 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
612 assert(_fast_freeze_size == 0, "");
613
614 stackChunkOop chunk = allocate_chunk(cont_size() + frame::metadata_words + _monitors_in_lockstack, _cont.argsize() + frame::metadata_words_at_top);
615 if (freeze_fast_new_chunk(chunk)) {
616 return freeze_ok;
617 }
618 if (_thread->has_pending_exception()) {
619 return freeze_exception;
620 }
621
622 // TODO R REMOVE when deopt change is fixed
623 assert(!_thread->cont_fastpath() || _barriers, "");
624 log_develop_trace(continuations)("-- RETRYING SLOW --");
625 return freeze_slow();
626 }
627
628 // Returns size needed if the continuation fits, otherwise 0.
629 int FreezeBase::size_if_fast_freeze_available() {
630 stackChunkOop chunk = _cont.tail();
631 if (chunk == nullptr || chunk->is_gc_mode() || chunk->requires_barriers() || chunk->has_mixed_frames()) {
632 log_develop_trace(continuations)("chunk available %s", chunk == nullptr ? "no chunk" : "chunk requires barriers");
633 return 0;
634 }
635
636 int total_size_needed = cont_size();
637 const int chunk_sp = chunk->sp();
638
639 // argsize can be nonzero if we have a caller, but the caller could be in a non-empty parent chunk,
640 // so we subtract it only if we overlap with the caller, i.e. the current chunk isn't empty.
641 // Consider leaving the chunk's argsize set when emptying it and removing the following branch,
642 // although that would require changing stackChunkOopDesc::is_empty
643 if (!chunk->is_empty()) {
644 total_size_needed -= _cont.argsize() + frame::metadata_words_at_top;
645 }
646
647 total_size_needed += _monitors_in_lockstack;
648
649 int chunk_free_room = chunk_sp - frame::metadata_words_at_bottom;
650 bool available = chunk_free_room >= total_size_needed;
651 log_develop_trace(continuations)("chunk available: %s size: %d argsize: %d top: " INTPTR_FORMAT " bottom: " INTPTR_FORMAT,
652 available ? "yes" : "no" , total_size_needed, _cont.argsize(), p2i(_cont_stack_top), p2i(_cont_stack_bottom));
653 return available ? total_size_needed : 0;
654 }
655
656 void FreezeBase::freeze_fast_existing_chunk() {
657 stackChunkOop chunk = _cont.tail();
658
659 DEBUG_ONLY(_fast_freeze_size = size_if_fast_freeze_available();)
660 assert(_fast_freeze_size > 0, "");
661
662 if (!chunk->is_empty()) { // we are copying into a non-empty chunk
663 DEBUG_ONLY(_empty = false;)
664 DEBUG_ONLY(_orig_chunk_sp = chunk->sp_address();)
665 #ifdef ASSERT
666 {
667 intptr_t* retaddr_slot = (chunk->sp_address()
668 - frame::sender_sp_ret_address_offset());
669 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
670 "unexpected saved return address");
671 }
672 #endif
673
674 // the chunk's sp before the freeze, adjusted to point beyond the stack-passed arguments in the topmost frame
675 // we overlap; we'll overwrite the chunk's top frame's callee arguments
676 const int chunk_start_sp = chunk->sp() + _cont.argsize() + frame::metadata_words_at_top;
677 assert(chunk_start_sp <= chunk->stack_size(), "sp not pointing into stack");
678
679 // increase max_size by what we're freezing minus the overlap
680 chunk->set_max_thawing_size(chunk->max_thawing_size() + cont_size() - _cont.argsize() - frame::metadata_words_at_top);
681
682 intptr_t* const bottom_sp = _cont_stack_bottom - _cont.argsize() - frame::metadata_words_at_top;
683 assert(bottom_sp == _bottom_address, "");
684 // Because the chunk isn't empty, we know there's a caller in the chunk, therefore the bottom-most frame
685 // should have a return barrier (installed back when we thawed it).
686 #ifdef ASSERT
687 {
688 intptr_t* retaddr_slot = (bottom_sp
689 - frame::sender_sp_ret_address_offset());
690 assert(ContinuationHelper::return_address_at(retaddr_slot)
691 == StubRoutines::cont_returnBarrier(),
692 "should be the continuation return barrier");
693 }
694 #endif
695 // We copy the fp from the chunk back to the stack because it contains some caller data,
696 // including, possibly, an oop that might have gone stale since we thawed.
697 patch_stack_pd(bottom_sp, chunk->sp_address());
698 // we don't patch the return pc at this time, so as not to make the stack unwalkable for async walks
699
700 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
701 } else { // the chunk is empty
702 const int chunk_start_sp = chunk->stack_size();
703
704 DEBUG_ONLY(_empty = true;)
705 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
706
707 chunk->set_max_thawing_size(cont_size());
708 chunk->set_bottom(chunk_start_sp - _cont.argsize() - frame::metadata_words_at_top);
709 chunk->set_sp(chunk->bottom());
710
711 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA false));
712 }
713 }
714
715 bool FreezeBase::freeze_fast_new_chunk(stackChunkOop chunk) {
716 DEBUG_ONLY(_empty = true;)
717
718 // Install new chunk
719 _cont.set_tail(chunk);
720
721 if (UNLIKELY(chunk == nullptr || !_thread->cont_fastpath() || _barriers)) { // OOME/probably humongous
722 log_develop_trace(continuations)("Retrying slow. Barriers: %d", _barriers);
723 return false;
724 }
725
726 chunk->set_max_thawing_size(cont_size());
727
728 // in a fresh chunk, we freeze *with* the bottom-most frame's stack arguments.
729 // They'll then be stored twice: in the chunk and in the parent chunk's top frame
730 const int chunk_start_sp = cont_size() + frame::metadata_words + _monitors_in_lockstack;
731 assert(chunk_start_sp == chunk->stack_size(), "");
732
733 DEBUG_ONLY(_orig_chunk_sp = chunk->start_address() + chunk_start_sp;)
734
735 freeze_fast_copy(chunk, chunk_start_sp CONT_JFR_ONLY(COMMA true));
736
737 return true;
738 }
739
740 void FreezeBase::freeze_fast_copy(stackChunkOop chunk, int chunk_start_sp CONT_JFR_ONLY(COMMA bool chunk_is_allocated)) {
741 assert(chunk != nullptr, "");
742 assert(!chunk->has_mixed_frames(), "");
743 assert(!chunk->is_gc_mode(), "");
744 assert(!chunk->has_bitmap(), "");
745 assert(!chunk->requires_barriers(), "");
746 assert(chunk == _cont.tail(), "");
747
748 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
749 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
750 // will either see no continuation on the stack, or a consistent chunk.
751 unwind_frames();
752
753 log_develop_trace(continuations)("freeze_fast start: chunk " INTPTR_FORMAT " size: %d orig sp: %d argsize: %d",
754 p2i((oopDesc*)chunk), chunk->stack_size(), chunk_start_sp, _cont.argsize());
755 assert(chunk_start_sp <= chunk->stack_size(), "");
756 assert(chunk_start_sp >= cont_size(), "no room in the chunk");
757
758 const int chunk_new_sp = chunk_start_sp - cont_size(); // the chunk's new sp, after freeze
759 assert(!(_fast_freeze_size > 0) || (_orig_chunk_sp - (chunk->start_address() + chunk_new_sp)) == (_fast_freeze_size - _monitors_in_lockstack), "");
760
761 intptr_t* chunk_top = chunk->start_address() + chunk_new_sp;
762 #ifdef ASSERT
763 if (!_empty) {
764 intptr_t* retaddr_slot = (_orig_chunk_sp
765 - frame::sender_sp_ret_address_offset());
766 assert(ContinuationHelper::return_address_at(retaddr_slot) == chunk->pc(),
767 "unexpected saved return address");
768 }
769 #endif
770
771 log_develop_trace(continuations)("freeze_fast start: " INTPTR_FORMAT " sp: %d chunk_top: " INTPTR_FORMAT,
772 p2i(chunk->start_address()), chunk_new_sp, p2i(chunk_top));
773
774 int adjust = frame::metadata_words_at_bottom;
775 #if INCLUDE_ASAN && defined(AARCH64)
776 // Reading at offset frame::metadata_words_at_bottom from _cont_stack_top
777 // will accesss memory at the callee frame, which on preemption cases will
778 // be the VM native method being called. The Arm 64-bit ABI doesn't specify
779 // a location where the frame record (returnpc+fp) has to be stored within
780 // a stack frame, and GCC currently chooses to save it at the top of the
781 // frame (lowest address). ASan treats this memory access in the callee as
782 // an overflow access to one of the locals stored in that frame. For these
783 // preemption cases we don't need to read these words anyways so we avoid it.
784 if (_preempt) {
785 adjust = 0;
786 }
787 #endif
788 intptr_t* from = _cont_stack_top - adjust;
789 intptr_t* to = chunk_top - adjust;
790 copy_to_chunk(from, to, cont_size() + adjust);
791 // Because we're not patched yet, the chunk is now in a bad state
792
793 // patch return pc of the bottom-most frozen frame (now in the chunk)
794 // with the actual caller's return address
795 intptr_t* chunk_bottom_retaddr_slot = (chunk_top + cont_size()
796 - _cont.argsize()
797 - frame::metadata_words_at_top
798 - frame::sender_sp_ret_address_offset());
799 #ifdef ASSERT
800 if (!_empty) {
801 assert(ContinuationHelper::return_address_at(chunk_bottom_retaddr_slot)
802 == StubRoutines::cont_returnBarrier(),
803 "should be the continuation return barrier");
804 }
805 #endif
806 ContinuationHelper::patch_return_address_at(chunk_bottom_retaddr_slot,
807 chunk->pc());
808
809 // We're always writing to a young chunk, so the GC can't see it until the next safepoint.
810 chunk->set_sp(chunk_new_sp);
811
812 // set chunk->pc to the return address of the topmost frame in the chunk
813 if (_preempt) {
814 // On aarch64/riscv64, the return pc of the top frame won't necessarily be at sp[-1].
815 // Also, on x64, if the top frame is the native wrapper frame, sp[-1] will not
816 // be the pc we used when creating the oopmap. Get the top's frame last pc from
817 // the anchor instead.
818 address last_pc = _last_frame.pc();
819 ContinuationHelper::patch_return_address_at(chunk_top - frame::sender_sp_ret_address_offset(), last_pc);
820 chunk->set_pc(last_pc);
821 // For stub/native frames the fp is not used while frozen, and will be constructed
822 // again when thawing the frame (see ThawBase::handle_preempted_continuation). We
823 // patch it with a special bad address to help with debugging, particularly when
824 // inspecting frames and identifying invalid accesses.
825 patch_pd_unused(chunk_top);
826 } else {
827 chunk->set_pc(ContinuationHelper::return_address_at(
828 _cont_stack_top - frame::sender_sp_ret_address_offset()));
829 }
830
831 if (_monitors_in_lockstack > 0) {
832 freeze_lockstack(chunk);
833 }
834
835 _cont.write();
836
837 log_develop_trace(continuations)("FREEZE CHUNK #" INTPTR_FORMAT " (young)", _cont.hash());
838 LogTarget(Trace, continuations) lt;
839 if (lt.develop_is_enabled()) {
840 LogStream ls(lt);
841 chunk->print_on(true, &ls);
842 }
843
844 // Verification
845 assert(_cont.chunk_invariant(), "");
846 chunk->verify();
847
848 #if CONT_JFR
849 EventContinuationFreezeFast e;
850 if (e.should_commit()) {
851 e.set_id(cast_from_oop<u8>(chunk));
852 DEBUG_ONLY(e.set_allocate(chunk_is_allocated);)
853 e.set_size(cont_size() << LogBytesPerWord);
854 e.commit();
855 }
856 #endif
857 }
858
859 NOINLINE freeze_result FreezeBase::freeze_slow() {
860 #ifdef ASSERT
861 ResourceMark rm;
862 #endif
863
864 log_develop_trace(continuations)("freeze_slow #" INTPTR_FORMAT, _cont.hash());
865 assert(_thread->thread_state() == _thread_in_vm || _thread->thread_state() == _thread_blocked, "");
866
867 #if CONT_JFR
868 EventContinuationFreezeSlow e;
869 if (e.should_commit()) {
870 e.set_id(cast_from_oop<u8>(_cont.continuation()));
871 e.commit();
872 }
873 #endif
874
875 init_rest();
876
877 HandleMark hm(Thread::current());
878
879 frame f = freeze_start_frame();
880
881 LogTarget(Debug, continuations) lt;
882 if (lt.develop_is_enabled()) {
883 LogStream ls(lt);
884 f.print_on(&ls);
885 }
886
887 frame caller; // the frozen caller in the chunk
888 freeze_result res = recurse_freeze(f, caller, 0, false, true);
889
890 if (res == freeze_ok) {
891 finish_freeze(f, caller);
892 _cont.write();
893 }
894
895 return res;
896 }
897
898 frame FreezeBase::freeze_start_frame() {
899 if (LIKELY(!_preempt)) {
900 return freeze_start_frame_yield_stub();
901 } else {
902 return freeze_start_frame_on_preempt();
903 }
904 }
905
906 frame FreezeBase::freeze_start_frame_yield_stub() {
907 frame f = _thread->last_frame();
908 assert(SharedRuntime::cont_doYield_stub()->contains(f.pc()), "must be");
909 f = sender<ContinuationHelper::NonInterpretedUnknownFrame>(f);
910 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), f), "");
911 return f;
912 }
913
914 frame FreezeBase::freeze_start_frame_on_preempt() {
915 assert(_last_frame.sp() == _thread->last_frame().sp(), "_last_frame should be already initialized");
916 assert(Continuation::is_frame_in_continuation(_thread->last_continuation(), _last_frame), "");
917 return _last_frame;
918 }
919
920 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
921 NOINLINE freeze_result FreezeBase::recurse_freeze(frame& f, frame& caller, int callee_argsize, bool callee_interpreted, bool top) {
922 assert(f.unextended_sp() < _bottom_address, ""); // see recurse_freeze_java_frame
923 assert(f.is_interpreted_frame() || ((top && _preempt) == ContinuationHelper::Frame::is_stub(f.cb()))
924 || ((top && _preempt) == f.is_native_frame()), "");
925
926 if (stack_overflow()) {
927 return freeze_exception;
928 }
929
930 if (f.is_compiled_frame()) {
931 if (UNLIKELY(f.oop_map() == nullptr)) {
932 // special native frame
933 return freeze_pinned_native;
934 }
935 return recurse_freeze_compiled_frame(f, caller, callee_argsize, callee_interpreted);
936 } else if (f.is_interpreted_frame()) {
937 assert(!f.interpreter_frame_method()->is_native() || (top && _preempt), "");
938 return recurse_freeze_interpreted_frame(f, caller, callee_argsize, callee_interpreted);
939 } else if (top && _preempt) {
940 assert(f.is_native_frame() || f.is_runtime_frame(), "");
941 return f.is_native_frame() ? recurse_freeze_native_frame(f, caller) : recurse_freeze_stub_frame(f, caller);
942 } else {
943 // Frame can't be frozen. Most likely the call_stub or upcall_stub
944 // which indicates there are further natives frames up the stack.
945 return freeze_pinned_native;
946 }
947 }
948
949 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
950 // See also StackChunkFrameStream<frame_kind>::frame_size()
951 template<typename FKind>
952 inline freeze_result FreezeBase::recurse_freeze_java_frame(const frame& f, frame& caller, int fsize, int argsize) {
953 assert(FKind::is_instance(f), "");
954
955 assert(fsize > 0, "");
956 assert(argsize >= 0, "");
957 _freeze_size += fsize;
958 NOT_PRODUCT(_frames++;)
959
960 assert(FKind::frame_bottom(f) <= _bottom_address, "");
961
962 // We don't use FKind::frame_bottom(f) == _bottom_address because on x64 there's sometimes an extra word between
963 // enterSpecial and an interpreted frame
964 if (FKind::frame_bottom(f) >= _bottom_address - 1) {
965 return finalize_freeze(f, caller, argsize); // recursion end
966 } else {
967 frame senderf = sender<FKind>(f);
968 assert(FKind::interpreted || senderf.sp() == senderf.unextended_sp(), "");
969 freeze_result result = recurse_freeze(senderf, caller, argsize, FKind::interpreted, false); // recursive call
970 return result;
971 }
972 }
973
974 inline void FreezeBase::before_freeze_java_frame(const frame& f, const frame& caller, int fsize, int argsize, bool is_bottom_frame) {
975 LogTarget(Trace, continuations) lt;
976 if (lt.develop_is_enabled()) {
977 LogStream ls(lt);
978 ls.print_cr("======== FREEZING FRAME interpreted: %d bottom: %d", f.is_interpreted_frame(), is_bottom_frame);
979 ls.print_cr("fsize: %d argsize: %d", fsize, argsize);
980 f.print_value_on(&ls);
981 }
982 assert(caller.is_interpreted_frame() == Interpreter::contains(caller.pc()), "");
983 }
984
985 inline void FreezeBase::after_freeze_java_frame(const frame& hf, bool is_bottom_frame) {
986 LogTarget(Trace, continuations) lt;
987 if (lt.develop_is_enabled()) {
988 LogStream ls(lt);
989 DEBUG_ONLY(hf.print_value_on(&ls);)
990 assert(hf.is_heap_frame(), "should be");
991 DEBUG_ONLY(print_frame_layout(hf, false, &ls);)
992 if (is_bottom_frame) {
993 ls.print_cr("bottom h-frame:");
994 hf.print_on(&ls);
995 }
996 }
997 }
998
999 // The parameter argsize_md includes metadata that has to be part of caller/callee overlap.
1000 // See also StackChunkFrameStream<frame_kind>::frame_size()
1001 freeze_result FreezeBase::finalize_freeze(const frame& callee, frame& caller, int argsize_md) {
1002 int argsize = argsize_md - frame::metadata_words_at_top;
1003 assert(callee.is_interpreted_frame()
1004 || ContinuationHelper::Frame::is_stub(callee.cb())
1005 || callee.cb()->as_nmethod()->is_osr_method()
1006 || argsize == _cont.argsize(), "argsize: %d cont.argsize: %d", argsize, _cont.argsize());
1007 log_develop_trace(continuations)("bottom: " INTPTR_FORMAT " count %d size: %d argsize: %d",
1008 p2i(_bottom_address), _frames, _freeze_size << LogBytesPerWord, argsize);
1009
1010 LogTarget(Trace, continuations) lt;
1011
1012 #ifdef ASSERT
1013 bool empty = _cont.is_empty();
1014 log_develop_trace(continuations)("empty: %d", empty);
1015 #endif
1016
1017 stackChunkOop chunk = _cont.tail();
1018
1019 assert(chunk == nullptr || (chunk->max_thawing_size() == 0) == chunk->is_empty(), "");
1020
1021 _freeze_size += frame::metadata_words; // for top frame's metadata
1022
1023 int overlap = 0; // the args overlap the caller -- if there is one in this chunk and is of the same kind
1024 int unextended_sp = -1;
1025 if (chunk != nullptr) {
1026 if (!chunk->is_empty()) {
1027 StackChunkFrameStream<ChunkFrames::Mixed> last(chunk);
1028 unextended_sp = chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp());
1029 bool top_interpreted = Interpreter::contains(chunk->pc());
1030 if (callee.is_interpreted_frame() == top_interpreted) {
1031 overlap = argsize_md;
1032 }
1033 } else {
1034 unextended_sp = chunk->stack_size() - frame::metadata_words_at_top;
1035 }
1036 }
1037
1038 log_develop_trace(continuations)("finalize _size: %d overlap: %d unextended_sp: %d", _freeze_size, overlap, unextended_sp);
1039
1040 _freeze_size -= overlap;
1041 assert(_freeze_size >= 0, "");
1042
1043 assert(chunk == nullptr || chunk->is_empty()
1044 || unextended_sp == chunk->to_offset(StackChunkFrameStream<ChunkFrames::Mixed>(chunk).unextended_sp()), "");
1045 assert(chunk != nullptr || unextended_sp < _freeze_size, "");
1046
1047 _freeze_size += _monitors_in_lockstack;
1048
1049 // _barriers can be set to true by an allocation in freeze_fast, in which case the chunk is available
1050 bool allocated_old_in_freeze_fast = _barriers;
1051 assert(!allocated_old_in_freeze_fast || (unextended_sp >= _freeze_size && chunk->is_empty()),
1052 "Chunk allocated in freeze_fast is of insufficient size "
1053 "unextended_sp: %d size: %d is_empty: %d", unextended_sp, _freeze_size, chunk->is_empty());
1054 assert(!allocated_old_in_freeze_fast || (!UseZGC && !UseG1GC), "Unexpected allocation");
1055
1056 DEBUG_ONLY(bool empty_chunk = true);
1057 if (unextended_sp < _freeze_size || chunk->is_gc_mode() || (!allocated_old_in_freeze_fast && chunk->requires_barriers())) {
1058 // ALLOCATE NEW CHUNK
1059
1060 if (lt.develop_is_enabled()) {
1061 LogStream ls(lt);
1062 if (chunk == nullptr) {
1063 ls.print_cr("no chunk");
1064 } else {
1065 ls.print_cr("chunk barriers: %d _size: %d free size: %d",
1066 chunk->requires_barriers(), _freeze_size, chunk->sp() - frame::metadata_words);
1067 chunk->print_on(&ls);
1068 }
1069 }
1070
1071 _freeze_size += overlap; // we're allocating a new chunk, so no overlap
1072 // overlap = 0;
1073
1074 chunk = allocate_chunk_slow(_freeze_size, argsize_md);
1075 if (chunk == nullptr) {
1076 return freeze_exception;
1077 }
1078
1079 // Install new chunk
1080 _cont.set_tail(chunk);
1081 assert(chunk->is_empty(), "");
1082 } else {
1083 // REUSE EXISTING CHUNK
1084 log_develop_trace(continuations)("Reusing chunk mixed: %d empty: %d", chunk->has_mixed_frames(), chunk->is_empty());
1085 if (chunk->is_empty()) {
1086 int sp = chunk->stack_size() - argsize_md;
1087 chunk->set_sp(sp);
1088 chunk->set_bottom(sp);
1089 _freeze_size += overlap;
1090 assert(chunk->max_thawing_size() == 0, "");
1091 } DEBUG_ONLY(else empty_chunk = false;)
1092 }
1093 assert(!chunk->is_gc_mode(), "");
1094 assert(!chunk->has_bitmap(), "");
1095 chunk->set_has_mixed_frames(true);
1096
1097 assert(chunk->requires_barriers() == _barriers, "");
1098 assert(!_barriers || chunk->is_empty(), "");
1099
1100 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).is_done(), "");
1101 assert(!chunk->is_empty() || StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame().is_empty(), "");
1102
1103 if (_preempt) {
1104 frame top_frame = _thread->last_frame();
1105 if (top_frame.is_interpreted_frame()) {
1106 // Some platforms do not save the last_sp in the top interpreter frame on VM calls.
1107 // We need it so that on resume we can restore the sp to the right place, since
1108 // thawing might add an alignment word to the expression stack (see finish_thaw()).
1109 // We do it now that we know freezing will be successful.
1110 prepare_freeze_interpreted_top_frame(top_frame);
1111 }
1112
1113 // Do this now so should_process_args_at_top() is set before calling finish_freeze
1114 // in case we might need to apply GC barriers to frames in this stackChunk.
1115 if (_thread->at_preemptable_init()) {
1116 assert(top_frame.is_interpreted_frame(), "only InterpreterRuntime::_new/resolve_from_cache allowed");
1117 chunk->set_at_klass_init(true);
1118 methodHandle m(_thread, top_frame.interpreter_frame_method());
1119 Bytecode_invoke call = Bytecode_invoke_check(m, top_frame.interpreter_frame_bci());
1120 assert(!call.is_valid() || call.is_invokestatic(), "only invokestatic allowed");
1121 if (call.is_invokestatic() && call.size_of_parameters() > 0) {
1122 assert(top_frame.interpreter_frame_expression_stack_size() > 0, "should have parameters in exp stack");
1123 chunk->set_has_args_at_top(true);
1124 }
1125 }
1126 }
1127
1128 // We unwind frames after the last safepoint so that the GC will have found the oops in the frames, but before
1129 // writing into the chunk. This is so that an asynchronous stack walk (not at a safepoint) that suspends us here
1130 // will either see no continuation or a consistent chunk.
1131 unwind_frames();
1132
1133 chunk->set_max_thawing_size(chunk->max_thawing_size() + _freeze_size - _monitors_in_lockstack - frame::metadata_words);
1134
1135 if (lt.develop_is_enabled()) {
1136 LogStream ls(lt);
1137 ls.print_cr("top chunk:");
1138 chunk->print_on(&ls);
1139 }
1140
1141 if (_monitors_in_lockstack > 0) {
1142 freeze_lockstack(chunk);
1143 }
1144
1145 // The topmost existing frame in the chunk; or an empty frame if the chunk is empty
1146 caller = StackChunkFrameStream<ChunkFrames::Mixed>(chunk).to_frame();
1147
1148 DEBUG_ONLY(_last_write = caller.unextended_sp() + (empty_chunk ? argsize_md : overlap);)
1149
1150 assert(chunk->is_in_chunk(_last_write - _freeze_size),
1151 "last_write-size: " INTPTR_FORMAT " start: " INTPTR_FORMAT, p2i(_last_write-_freeze_size), p2i(chunk->start_address()));
1152 #ifdef ASSERT
1153 if (lt.develop_is_enabled()) {
1154 LogStream ls(lt);
1155 ls.print_cr("top hframe before (freeze):");
1156 assert(caller.is_heap_frame(), "should be");
1157 caller.print_on(&ls);
1158 }
1159
1160 assert(!empty || Continuation::is_continuation_entry_frame(callee, nullptr), "");
1161
1162 frame entry = sender(callee);
1163
1164 assert((!empty && Continuation::is_return_barrier_entry(entry.pc())) || (empty && Continuation::is_continuation_enterSpecial(entry)), "");
1165 assert(callee.is_interpreted_frame() || entry.sp() == entry.unextended_sp(), "");
1166 #endif
1167
1168 return freeze_ok_bottom;
1169 }
1170
1171 // After freezing a frame we need to possibly adjust some values related to the caller frame.
1172 void FreezeBase::patch(const frame& f, frame& hf, const frame& caller, bool is_bottom_frame) {
1173 if (is_bottom_frame) {
1174 // If we're the bottom frame, we need to replace the return barrier with the real
1175 // caller's pc.
1176 address last_pc = caller.pc();
1177 assert((last_pc == nullptr) == _cont.tail()->is_empty(), "");
1178 ContinuationHelper::Frame::patch_pc(caller, last_pc);
1179 } else {
1180 assert(!caller.is_empty(), "");
1181 }
1182
1183 patch_pd(hf, caller, is_bottom_frame);
1184
1185 if (f.is_interpreted_frame()) {
1186 assert(hf.is_heap_frame(), "should be");
1187 ContinuationHelper::InterpretedFrame::patch_sender_sp(hf, caller);
1188 }
1189
1190 #ifdef ASSERT
1191 if (hf.is_compiled_frame()) {
1192 if (f.is_deoptimized_frame()) { // TODO DEOPT: long term solution: unroll on freeze and patch pc
1193 log_develop_trace(continuations)("Freezing deoptimized frame");
1194 assert(f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
1195 assert(f.cb()->as_nmethod()->is_deopt_pc(ContinuationHelper::Frame::real_pc(f)), "");
1196 }
1197 }
1198 #endif
1199 }
1200
1201 #ifdef ASSERT
1202 static void verify_frame_top(const frame& f, intptr_t* top) {
1203 ResourceMark rm;
1204 InterpreterOopMap mask;
1205 f.interpreted_frame_oop_map(&mask);
1206 assert(top <= ContinuationHelper::InterpretedFrame::frame_top(f, &mask),
1207 "frame_top: " INTPTR_FORMAT " Interpreted::frame_top: " INTPTR_FORMAT,
1208 p2i(top), p2i(ContinuationHelper::InterpretedFrame::frame_top(f, &mask)));
1209 }
1210 #endif // ASSERT
1211
1212 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1213 // See also StackChunkFrameStream<frame_kind>::frame_size()
1214 NOINLINE freeze_result FreezeBase::recurse_freeze_interpreted_frame(frame& f, frame& caller,
1215 int callee_argsize /* incl. metadata */,
1216 bool callee_interpreted) {
1217 adjust_interpreted_frame_unextended_sp(f);
1218
1219 // The frame's top never includes the stack arguments to the callee
1220 intptr_t* const stack_frame_top = ContinuationHelper::InterpretedFrame::frame_top(f, callee_argsize, callee_interpreted);
1221 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
1222 const int fsize = pointer_delta_as_int(stack_frame_bottom, stack_frame_top);
1223
1224 DEBUG_ONLY(verify_frame_top(f, stack_frame_top));
1225
1226 Method* frame_method = ContinuationHelper::Frame::frame_method(f);
1227 // including metadata between f and its args
1228 const int argsize = ContinuationHelper::InterpretedFrame::stack_argsize(f) + frame::metadata_words_at_top;
1229
1230 log_develop_trace(continuations)("recurse_freeze_interpreted_frame %s _size: %d fsize: %d argsize: %d callee_interpreted: %d",
1231 frame_method->name_and_sig_as_C_string(), _freeze_size, fsize, argsize, callee_interpreted);
1232 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1233 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1234
1235 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::InterpretedFrame>(f, caller, fsize, argsize);
1236 if (UNLIKELY(result > freeze_ok_bottom)) {
1237 return result;
1238 }
1239
1240 bool is_bottom_frame = result == freeze_ok_bottom;
1241 assert(!caller.is_empty() || is_bottom_frame, "");
1242
1243 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, is_bottom_frame);)
1244
1245 frame hf = new_heap_frame<ContinuationHelper::InterpretedFrame>(f, caller);
1246 _total_align_size += frame::align_wiggle; // add alignment room for internal interpreted frame alignment on AArch64/PPC64
1247
1248 intptr_t* heap_frame_top = ContinuationHelper::InterpretedFrame::frame_top(hf, callee_argsize, callee_interpreted);
1249 intptr_t* heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
1250 assert(heap_frame_bottom == heap_frame_top + fsize, "");
1251
1252 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
1253 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
1254 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1255 assert(!is_bottom_frame || !caller.is_interpreted_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1256
1257 relativize_interpreted_frame_metadata(f, hf);
1258
1259 patch(f, hf, caller, is_bottom_frame);
1260
1261 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
1262 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1263 caller = hf;
1264
1265 // Mark frame_method's GC epoch for class redefinition on_stack calculation.
1266 frame_method->record_gc_epoch();
1267
1268 return freeze_ok;
1269 }
1270
1271 // The parameter callee_argsize includes metadata that has to be part of caller/callee overlap.
1272 // See also StackChunkFrameStream<frame_kind>::frame_size()
1273 freeze_result FreezeBase::recurse_freeze_compiled_frame(frame& f, frame& caller,
1274 int callee_argsize /* incl. metadata */,
1275 bool callee_interpreted) {
1276 // The frame's top never includes the stack arguments to the callee
1277 intptr_t* const stack_frame_top = ContinuationHelper::CompiledFrame::frame_top(f, callee_argsize, callee_interpreted);
1278 intptr_t* const stack_frame_bottom = ContinuationHelper::CompiledFrame::frame_bottom(f);
1279 // including metadata between f and its stackargs
1280 int argsize = ContinuationHelper::CompiledFrame::stack_argsize(f) + frame::metadata_words_at_top;
1281 int fsize = pointer_delta_as_int(stack_frame_bottom + argsize, stack_frame_top);
1282
1283 int real_frame_size = 0;
1284 bool augmented = f.was_augmented_on_entry(real_frame_size);
1285 if (augmented) {
1286 // The args reside inside the frame so clear argsize. If the caller is compiled,
1287 // this will cause the stack arguments passed by the caller to be freezed when
1288 // freezing the caller frame itself. If the caller is interpreted this will have
1289 // the effect of discarding the arg area created in the i2c stub.
1290 argsize = 0;
1291 fsize = real_frame_size - (callee_interpreted ? 0 : callee_argsize);
1292 #ifdef ASSERT
1293 nmethod* nm = f.cb()->as_nmethod();
1294 Method* method = nm->method();
1295 address return_pc = ContinuationHelper::CompiledFrame::return_pc(f);
1296 CodeBlob* caller_cb = CodeCache::find_blob_fast(return_pc);
1297 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");
1298 assert((!caller_cb->is_nmethod() && nm->is_compiled_by_c2()) ||
1299 (nm->compiler_type() != caller_cb->as_nmethod()->compiler_type()) ||
1300 (nm->is_compiled_by_c2() && !method->is_static() && method->method_holder()->is_inline_klass()),
1301 "frame should not be extended");
1302 #endif
1303 }
1304
1305 log_develop_trace(continuations)("recurse_freeze_compiled_frame %s _size: %d fsize: %d argsize: %d augmented: %d",
1306 ContinuationHelper::Frame::frame_method(f) != nullptr ?
1307 ContinuationHelper::Frame::frame_method(f)->name_and_sig_as_C_string() : "",
1308 _freeze_size, fsize, argsize, augmented);
1309 // we'd rather not yield inside methods annotated with @JvmtiMountTransition
1310 assert(!ContinuationHelper::Frame::frame_method(f)->jvmti_mount_transition(), "");
1311
1312 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::CompiledFrame>(f, caller, fsize, argsize);
1313 if (UNLIKELY(result > freeze_ok_bottom)) {
1314 return result;
1315 }
1316
1317 bool is_bottom_frame = result == freeze_ok_bottom;
1318 assert(!caller.is_empty() || is_bottom_frame, "");
1319 assert(!is_bottom_frame || !augmented, "thaw extended frame without caller?");
1320
1321 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, is_bottom_frame);)
1322
1323 frame hf = new_heap_frame<ContinuationHelper::CompiledFrame>(f, caller, augmented ? real_frame_size - f.cb()->as_nmethod()->frame_size() : 0);
1324
1325 intptr_t* heap_frame_top = ContinuationHelper::CompiledFrame::frame_top(hf, callee_argsize, callee_interpreted);
1326
1327 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1328 assert(!is_bottom_frame || !caller.is_compiled_frame() || (heap_frame_top + fsize) == (caller.unextended_sp() + argsize), "");
1329
1330 if (caller.is_interpreted_frame()) {
1331 // When thawing the frame we might need to add alignment (see Thaw::align)
1332 _total_align_size += frame::align_wiggle;
1333 }
1334
1335 patch(f, hf, caller, is_bottom_frame);
1336
1337 assert(is_bottom_frame || Interpreter::contains(ContinuationHelper::CompiledFrame::real_pc(caller)) == caller.is_interpreted_frame(), "");
1338
1339 DEBUG_ONLY(after_freeze_java_frame(hf, is_bottom_frame);)
1340 caller = hf;
1341 return freeze_ok;
1342 }
1343
1344 NOINLINE freeze_result FreezeBase::recurse_freeze_stub_frame(frame& f, frame& caller) {
1345 DEBUG_ONLY(frame fsender = sender(f);)
1346 assert(fsender.is_compiled_frame(), "sender should be compiled frame");
1347
1348 intptr_t* const stack_frame_top = ContinuationHelper::StubFrame::frame_top(f);
1349 const int fsize = f.cb()->frame_size();
1350
1351 log_develop_trace(continuations)("recurse_freeze_stub_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1352 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1353
1354 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::StubFrame>(f, caller, fsize, 0);
1355 if (UNLIKELY(result > freeze_ok_bottom)) {
1356 return result;
1357 }
1358
1359 assert(result == freeze_ok, "should have caller");
1360 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, 0, false /*is_bottom_frame*/);)
1361
1362 frame hf = new_heap_frame<ContinuationHelper::StubFrame>(f, caller);
1363 intptr_t* heap_frame_top = ContinuationHelper::StubFrame::frame_top(hf);
1364
1365 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1366
1367 patch(f, hf, caller, false /*is_bottom_frame*/);
1368
1369 DEBUG_ONLY(after_freeze_java_frame(hf, false /*is_bottom_frame*/);)
1370
1371 caller = hf;
1372 return freeze_ok;
1373 }
1374
1375 NOINLINE freeze_result FreezeBase::recurse_freeze_native_frame(frame& f, frame& caller) {
1376 if (!f.cb()->as_nmethod()->method()->is_object_wait0()) {
1377 assert(f.cb()->as_nmethod()->method()->is_synchronized(), "");
1378 // Synchronized native method case. Unlike the interpreter native wrapper, the compiled
1379 // native wrapper tries to acquire the monitor after marshalling the arguments from the
1380 // caller into the native convention. This is so that we have a valid oopMap in case of
1381 // having to block in the slow path. But that would require freezing those registers too
1382 // and then fixing them back on thaw in case of oops. To avoid complicating things and
1383 // given that this would be a rare case anyways just pin the vthread to the carrier.
1384 return freeze_pinned_native;
1385 }
1386
1387 intptr_t* const stack_frame_top = ContinuationHelper::NativeFrame::frame_top(f);
1388 // There are no stackargs but argsize must include the metadata
1389 const int argsize = frame::metadata_words_at_top;
1390 const int fsize = f.cb()->frame_size() + argsize;
1391
1392 log_develop_trace(continuations)("recurse_freeze_native_frame %s _size: %d fsize: %d :: " INTPTR_FORMAT " - " INTPTR_FORMAT,
1393 f.cb()->name(), _freeze_size, fsize, p2i(stack_frame_top), p2i(stack_frame_top+fsize));
1394
1395 freeze_result result = recurse_freeze_java_frame<ContinuationHelper::NativeFrame>(f, caller, fsize, argsize);
1396 if (UNLIKELY(result > freeze_ok_bottom)) {
1397 return result;
1398 }
1399
1400 assert(result == freeze_ok, "should have caller frame");
1401 DEBUG_ONLY(before_freeze_java_frame(f, caller, fsize, argsize, false /* is_bottom_frame */);)
1402
1403 frame hf = new_heap_frame<ContinuationHelper::NativeFrame>(f, caller);
1404 intptr_t* heap_frame_top = ContinuationHelper::NativeFrame::frame_top(hf);
1405
1406 copy_to_chunk(stack_frame_top, heap_frame_top, fsize);
1407
1408 if (caller.is_interpreted_frame()) {
1409 // When thawing the frame we might need to add alignment (see Thaw::align)
1410 _total_align_size += frame::align_wiggle;
1411 }
1412
1413 patch(f, hf, caller, false /* is_bottom_frame */);
1414
1415 DEBUG_ONLY(after_freeze_java_frame(hf, false /* is_bottom_frame */);)
1416
1417 caller = hf;
1418 return freeze_ok;
1419 }
1420
1421 NOINLINE void FreezeBase::finish_freeze(const frame& f, const frame& top) {
1422 stackChunkOop chunk = _cont.tail();
1423
1424 LogTarget(Trace, continuations) lt;
1425 if (lt.develop_is_enabled()) {
1426 LogStream ls(lt);
1427 assert(top.is_heap_frame(), "should be");
1428 top.print_on(&ls);
1429 }
1430
1431 set_top_frame_metadata_pd(top);
1432
1433 chunk->set_sp(chunk->to_offset(top.sp()));
1434 chunk->set_pc(top.pc());
1435
1436 chunk->set_max_thawing_size(chunk->max_thawing_size() + _total_align_size);
1437
1438 assert(chunk->sp_address() - chunk->start_address() >= _monitors_in_lockstack, "clash with lockstack");
1439
1440 // At this point the chunk is consistent
1441
1442 if (UNLIKELY(_barriers)) {
1443 log_develop_trace(continuations)("do barriers on old chunk");
1444 // Serial and Parallel GC can allocate objects directly into the old generation.
1445 // Then we want to relativize the derived pointers eagerly so that
1446 // old chunks are all in GC mode.
1447 assert(!UseG1GC, "G1 can not deal with allocating outside of eden");
1448 assert(!UseZGC, "ZGC can not deal with allocating chunks visible to marking");
1449 if (UseShenandoahGC) {
1450 _cont.tail()->relativize_derived_pointers_concurrently();
1451 } else {
1452 ContinuationGCSupport::transform_stack_chunk(_cont.tail());
1453 }
1454 // For objects in the old generation we must maintain the remembered set
1455 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>();
1456 }
1457
1458 log_develop_trace(continuations)("finish_freeze: has_mixed_frames: %d", chunk->has_mixed_frames());
1459 if (lt.develop_is_enabled()) {
1460 LogStream ls(lt);
1461 chunk->print_on(true, &ls);
1462 }
1463
1464 if (lt.develop_is_enabled()) {
1465 LogStream ls(lt);
1466 ls.print_cr("top hframe after (freeze):");
1467 assert(_cont.last_frame().is_heap_frame(), "should be");
1468 _cont.last_frame().print_on(&ls);
1469 DEBUG_ONLY(print_frame_layout(top, false, &ls);)
1470 }
1471
1472 assert(_cont.chunk_invariant(), "");
1473 }
1474
1475 inline bool FreezeBase::stack_overflow() { // detect stack overflow in recursive native code
1476 JavaThread* t = !_preempt ? _thread : JavaThread::current();
1477 assert(t == JavaThread::current(), "");
1478 if (os::current_stack_pointer() < t->stack_overflow_state()->shadow_zone_safe_limit()) {
1479 if (!_preempt) {
1480 ContinuationWrapper::SafepointOp so(t, _cont); // could also call _cont.done() instead
1481 Exceptions::_throw_msg(t, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Stack overflow while freezing");
1482 }
1483 return true;
1484 }
1485 return false;
1486 }
1487
1488 class StackChunkAllocator : public MemAllocator {
1489 const size_t _stack_size;
1490 int _argsize_md;
1491 ContinuationWrapper& _continuation_wrapper;
1492 JvmtiSampledObjectAllocEventCollector* const _jvmti_event_collector;
1493 mutable bool _took_slow_path;
1494
1495 // Does the minimal amount of initialization needed for a TLAB allocation.
1496 // We don't need to do a full initialization, as such an allocation need not be immediately walkable.
1497 virtual oop initialize(HeapWord* mem) const override {
1498 assert(_stack_size > 0, "");
1499 assert(_stack_size <= max_jint, "");
1500 assert(_word_size > _stack_size, "");
1501
1502 // zero out fields (but not the stack)
1503 const size_t hs = oopDesc::header_size();
1504 if (oopDesc::has_klass_gap()) {
1505 oopDesc::set_klass_gap(mem, 0);
1506 }
1507 Copy::fill_to_aligned_words(mem + hs, vmClasses::StackChunk_klass()->size_helper() - hs);
1508
1509 int bottom = (int)_stack_size - _argsize_md;
1510
1511 jdk_internal_vm_StackChunk::set_size(mem, (int)_stack_size);
1512 jdk_internal_vm_StackChunk::set_bottom(mem, bottom);
1513 jdk_internal_vm_StackChunk::set_sp(mem, bottom);
1514
1515 return finish(mem);
1516 }
1517
1518 stackChunkOop allocate_fast() const {
1519 if (!UseTLAB) {
1520 return nullptr;
1521 }
1522
1523 HeapWord* const mem = MemAllocator::mem_allocate_inside_tlab_fast();
1524 if (mem == nullptr) {
1525 return nullptr;
1526 }
1527
1528 oop obj = initialize(mem);
1529 return stackChunkOopDesc::cast(obj);
1530 }
1531
1532 public:
1533 StackChunkAllocator(Klass* klass,
1534 size_t word_size,
1535 Thread* thread,
1536 size_t stack_size,
1537 int argsize_md,
1538 ContinuationWrapper& continuation_wrapper,
1539 JvmtiSampledObjectAllocEventCollector* jvmti_event_collector)
1540 : MemAllocator(klass, word_size, thread),
1541 _stack_size(stack_size),
1542 _argsize_md(argsize_md),
1543 _continuation_wrapper(continuation_wrapper),
1544 _jvmti_event_collector(jvmti_event_collector),
1545 _took_slow_path(false) {}
1546
1547 // Provides it's own, specialized allocation which skips instrumentation
1548 // if the memory can be allocated without going to a slow-path.
1549 stackChunkOop allocate() const {
1550 // First try to allocate without any slow-paths or instrumentation.
1551 stackChunkOop obj = allocate_fast();
1552 if (obj != nullptr) {
1553 return obj;
1554 }
1555
1556 // Now try full-blown allocation with all expensive operations,
1557 // including potentially safepoint operations.
1558 _took_slow_path = true;
1559
1560 // Protect unhandled Loom oops
1561 ContinuationWrapper::SafepointOp so(_thread, _continuation_wrapper);
1562
1563 // Can safepoint
1564 _jvmti_event_collector->start();
1565
1566 // Can safepoint
1567 return stackChunkOopDesc::cast(MemAllocator::allocate());
1568 }
1569
1570 bool took_slow_path() const {
1571 return _took_slow_path;
1572 }
1573 };
1574
1575 template <typename ConfigT>
1576 stackChunkOop Freeze<ConfigT>::allocate_chunk(size_t stack_size, int argsize_md) {
1577 log_develop_trace(continuations)("allocate_chunk allocating new chunk");
1578
1579 InstanceStackChunkKlass* klass = InstanceStackChunkKlass::cast(vmClasses::StackChunk_klass());
1580 size_t size_in_words = klass->instance_size(stack_size);
1581
1582 if (CollectedHeap::stack_chunk_max_size() > 0 && size_in_words >= CollectedHeap::stack_chunk_max_size()) {
1583 if (!_preempt) {
1584 throw_stack_overflow_on_humongous_chunk();
1585 }
1586 return nullptr;
1587 }
1588
1589 JavaThread* current = _preempt ? JavaThread::current() : _thread;
1590 assert(current == JavaThread::current(), "should be current");
1591
1592 // Allocate the chunk.
1593 //
1594 // This might safepoint while allocating, but all safepointing due to
1595 // instrumentation have been deferred. This property is important for
1596 // some GCs, as this ensures that the allocated object is in the young
1597 // generation / newly allocated memory.
1598 StackChunkAllocator allocator(klass, size_in_words, current, stack_size, argsize_md, _cont, _jvmti_event_collector);
1599 stackChunkOop chunk = allocator.allocate();
1600
1601 if (chunk == nullptr) {
1602 return nullptr; // OOME
1603 }
1604
1605 // assert that chunk is properly initialized
1606 assert(chunk->stack_size() == (int)stack_size, "");
1607 assert(chunk->size() >= stack_size, "chunk->size(): %zu size: %zu", chunk->size(), stack_size);
1608 assert(chunk->sp() == chunk->bottom(), "");
1609 assert((intptr_t)chunk->start_address() % 8 == 0, "");
1610 assert(chunk->max_thawing_size() == 0, "");
1611 assert(chunk->pc() == nullptr, "");
1612 assert(chunk->is_empty(), "");
1613 assert(chunk->flags() == 0, "");
1614 assert(chunk->is_gc_mode() == false, "");
1615 assert(chunk->lockstack_size() == 0, "");
1616
1617 // fields are uninitialized
1618 chunk->set_parent_access<IS_DEST_UNINITIALIZED>(_cont.last_nonempty_chunk());
1619 chunk->set_cont_access<IS_DEST_UNINITIALIZED>(_cont.continuation());
1620
1621 #if INCLUDE_ZGC
1622 if (UseZGC) {
1623 ZStackChunkGCData::initialize(chunk);
1624 assert(!chunk->requires_barriers(), "ZGC always allocates in the young generation");
1625 _barriers = false;
1626 } else
1627 #endif
1628 #if INCLUDE_SHENANDOAHGC
1629 if (UseShenandoahGC) {
1630 _barriers = chunk->requires_barriers();
1631 } else
1632 #endif
1633 {
1634 if (!allocator.took_slow_path()) {
1635 // Guaranteed to be in young gen / newly allocated memory
1636 assert(!chunk->requires_barriers(), "Unfamiliar GC requires barriers on TLAB allocation");
1637 _barriers = false;
1638 } else {
1639 // Some GCs could put direct allocations in old gen for slow-path
1640 // allocations; need to explicitly check if that was the case.
1641 _barriers = chunk->requires_barriers();
1642 }
1643 }
1644
1645 if (_barriers) {
1646 log_develop_trace(continuations)("allocation requires barriers");
1647 }
1648
1649 assert(chunk->parent() == nullptr || chunk->parent()->is_stackChunk(), "");
1650
1651 return chunk;
1652 }
1653
1654 void FreezeBase::throw_stack_overflow_on_humongous_chunk() {
1655 ContinuationWrapper::SafepointOp so(_thread, _cont); // could also call _cont.done() instead
1656 Exceptions::_throw_msg(_thread, __FILE__, __LINE__, vmSymbols::java_lang_StackOverflowError(), "Humongous stack chunk");
1657 }
1658
1659 class AnchorMark : public StackObj {
1660 JavaThread* _current;
1661 frame& _top_frame;
1662 intptr_t* _last_sp_from_frame;
1663 bool _is_interpreted;
1664
1665 public:
1666 AnchorMark(JavaThread* current, frame& f) : _current(current), _top_frame(f), _is_interpreted(false) {
1667 intptr_t* sp = anchor_mark_set_pd();
1668 set_anchor(_current, sp);
1669 }
1670 ~AnchorMark() {
1671 clear_anchor(_current);
1672 anchor_mark_clear_pd();
1673 }
1674 inline intptr_t* anchor_mark_set_pd();
1675 inline void anchor_mark_clear_pd();
1676 };
1677
1678 #if INCLUDE_JVMTI
1679 static int num_java_frames(ContinuationWrapper& cont) {
1680 ResourceMark rm; // used for scope traversal in num_java_frames(nmethod*, address)
1681 int count = 0;
1682 for (stackChunkOop chunk = cont.tail(); chunk != nullptr; chunk = chunk->parent()) {
1683 count += chunk->num_java_frames();
1684 }
1685 return count;
1686 }
1687
1688 static void invalidate_jvmti_stack(JavaThread* thread) {
1689 JvmtiThreadState *state = thread->jvmti_thread_state();
1690 if (state != nullptr) {
1691 state->invalidate_cur_stack_depth();
1692 }
1693 }
1694
1695 static void jvmti_yield_cleanup(JavaThread* thread, ContinuationWrapper& cont) {
1696 if (!cont.entry()->is_virtual_thread()) {
1697 if (JvmtiExport::has_frame_pops(thread)) {
1698 int num_frames = num_java_frames(cont);
1699
1700 ContinuationWrapper::SafepointOp so(Thread::current(), cont);
1701 JvmtiExport::continuation_yield_cleanup(thread, num_frames);
1702 }
1703 invalidate_jvmti_stack(thread);
1704 }
1705 }
1706
1707 static void jvmti_mount_end(JavaThread* current, ContinuationWrapper& cont, frame top, Continuation::preempt_kind pk) {
1708 assert(current->vthread() != nullptr, "must be");
1709
1710 HandleMarkCleaner hm(current); // Cleanup all handles (including so._conth) before returning to Java.
1711 Handle vth(current, current->vthread());
1712 ContinuationWrapper::SafepointOp so(current, cont);
1713 AnchorMark am(current, top); // Set anchor so that the stack is walkable.
1714
1715 JRT_BLOCK
1716 JvmtiVTMSTransitionDisabler::VTMS_vthread_mount((jthread)vth.raw_value(), false);
1717
1718 if (current->pending_contended_entered_event()) {
1719 // No monitor JVMTI events for ObjectLocker case.
1720 if (pk != Continuation::object_locker) {
1721 JvmtiExport::post_monitor_contended_entered(current, current->contended_entered_monitor());
1722 }
1723 current->set_contended_entered_monitor(nullptr);
1724 }
1725 JRT_BLOCK_END
1726 }
1727 #endif // INCLUDE_JVMTI
1728
1729 #ifdef ASSERT
1730 // There are no interpreted frames if we're not called from the interpreter and we haven't ancountered an i2c
1731 // adapter or called Deoptimization::unpack_frames. As for native frames, upcalls from JNI also go through the
1732 // interpreter (see JavaCalls::call_helper), while the UpcallLinker explicitly sets cont_fastpath.
1733 bool FreezeBase::check_valid_fast_path() {
1734 ContinuationEntry* ce = _thread->last_continuation();
1735 RegisterMap map(_thread,
1736 RegisterMap::UpdateMap::skip,
1737 RegisterMap::ProcessFrames::skip,
1738 RegisterMap::WalkContinuation::skip);
1739 map.set_include_argument_oops(false);
1740 bool is_top_frame = true;
1741 for (frame f = freeze_start_frame(); Continuation::is_frame_in_continuation(ce, f); f = f.sender(&map), is_top_frame = false) {
1742 if (!((f.is_compiled_frame() && !f.is_deoptimized_frame()) || (is_top_frame && (f.is_runtime_frame() || f.is_native_frame())))) {
1743 return false;
1744 }
1745 }
1746 return true;
1747 }
1748
1749 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) {
1750 Method* m;
1751 const char* code_name;
1752 int bci;
1753 if (preempt_kind == Continuation::monitorenter) {
1754 assert(top.is_interpreted_frame() || top.is_runtime_frame(), "unexpected %sframe",
1755 top.is_compiled_frame() ? "compiled " : top.is_native_frame() ? "native " : "");
1756 bool at_sync_method;
1757 if (top.is_interpreted_frame()) {
1758 m = top.interpreter_frame_method();
1759 assert(!m->is_native() || m->is_synchronized(), "invalid method %s", m->external_name());
1760 address bcp = top.interpreter_frame_bcp();
1761 assert(bcp != 0 || m->is_native(), "");
1762 at_sync_method = m->is_synchronized() && (bcp == 0 || bcp == m->code_base());
1763 // bcp is advanced on monitorenter before making the VM call, adjust for that.
1764 bool at_sync_bytecode = bcp > m->code_base() && Bytecode(m, bcp - 1).code() == Bytecodes::Code::_monitorenter;
1765 assert(at_sync_method || at_sync_bytecode, "");
1766 bci = at_sync_method ? -1 : top.interpreter_frame_bci();
1767 } else {
1768 JavaThread* current = JavaThread::current();
1769 ResourceMark rm(current);
1770 CodeBlob* cb = top.cb();
1771 RegisterMap reg_map(current,
1772 RegisterMap::UpdateMap::skip,
1773 RegisterMap::ProcessFrames::skip,
1774 RegisterMap::WalkContinuation::include);
1775 if (top.is_heap_frame()) {
1776 assert(chunk != nullptr, "");
1777 reg_map.set_stack_chunk(chunk);
1778 top = chunk->relativize(top);
1779 top.set_frame_index(0);
1780 }
1781 frame fr = top.sender(®_map);
1782 vframe* vf = vframe::new_vframe(&fr, ®_map, current);
1783 compiledVFrame* cvf = compiledVFrame::cast(vf);
1784 m = cvf->method();
1785 bci = cvf->scope()->bci();
1786 at_sync_method = bci == SynchronizationEntryBCI;
1787 assert(!at_sync_method || m->is_synchronized(), "bci is %d but method %s is not synchronized", bci, m->external_name());
1788 bool is_c1_monitorenter = false, is_c2_monitorenter = false;
1789 COMPILER1_PRESENT(is_c1_monitorenter = cb == Runtime1::blob_for(StubId::c1_monitorenter_id) ||
1790 cb == Runtime1::blob_for(StubId::c1_monitorenter_nofpu_id);)
1791 COMPILER2_PRESENT(is_c2_monitorenter = cb == CodeCache::find_blob(OptoRuntime::complete_monitor_locking_Java());)
1792 assert(is_c1_monitorenter || is_c2_monitorenter, "wrong runtime stub frame");
1793 }
1794 code_name = at_sync_method ? "synchronized method" : "monitorenter";
1795 } else if (preempt_kind == Continuation::object_wait) {
1796 assert(top.is_interpreted_frame() || top.is_native_frame(), "");
1797 m = top.is_interpreted_frame() ? top.interpreter_frame_method() : top.cb()->as_nmethod()->method();
1798 assert(m->is_object_wait0(), "");
1799 bci = 0;
1800 code_name = "";
1801 } else {
1802 assert(preempt_kind == Continuation::object_locker, "invalid preempt kind");
1803 assert(top.is_interpreted_frame(), "");
1804 m = top.interpreter_frame_method();
1805 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
1806 Bytecodes::Code code = current_bytecode.code();
1807 assert(code == Bytecodes::Code::_new || code == Bytecodes::Code::_invokestatic ||
1808 (code == Bytecodes::Code::_getstatic || code == Bytecodes::Code::_putstatic), "invalid bytecode");
1809 bci = top.interpreter_frame_bci();
1810 code_name = Bytecodes::name(current_bytecode.code());
1811 }
1812 assert(bci >= 0 || m->is_synchronized(), "invalid bci:%d at method %s", bci, m->external_name());
1813
1814 if (m_ptr != nullptr) {
1815 *m_ptr = m;
1816 *code_name_ptr = code_name;
1817 *bci_ptr = bci;
1818 }
1819 }
1820
1821 static void log_preempt_after_freeze(const ContinuationWrapper& cont) {
1822 JavaThread* current = cont.thread();
1823 int64_t tid = current->monitor_owner_id();
1824
1825 StackChunkFrameStream<ChunkFrames::Mixed> sfs(cont.tail());
1826 frame top_frame = sfs.to_frame();
1827 bool at_init = current->at_preemptable_init();
1828 bool at_enter = current->current_pending_monitor() != nullptr;
1829 bool at_wait = current->current_waiting_monitor() != nullptr;
1830 assert((at_enter && !at_wait) || (!at_enter && at_wait), "");
1831 Continuation::preempt_kind pk = at_init ? Continuation::object_locker : at_enter ? Continuation::monitorenter : Continuation::object_wait;
1832
1833 Method* m = nullptr;
1834 const char* code_name = nullptr;
1835 int bci = InvalidFrameStateBci;
1836 verify_frame_kind(top_frame, pk, &m, &code_name, &bci, cont.tail());
1837 assert(m != nullptr && code_name != nullptr && bci != InvalidFrameStateBci, "should be set");
1838
1839 ResourceMark rm(current);
1840 if (bci < 0) {
1841 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " while synchronizing on %smethod %s", tid, m->is_native() ? "native " : "", m->external_name());
1842 } else if (m->is_object_wait0()) {
1843 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at native method %s", tid, m->external_name());
1844 } else {
1845 Klass* k = current->preempt_init_klass();
1846 assert(k != nullptr || !at_init, "");
1847 log_trace(continuations, preempt)("Preempted " INT64_FORMAT " at %s(bci:%d) in method %s %s%s", tid, code_name, bci,
1848 m->external_name(), at_init ? "trying to initialize klass " : "", at_init ? k->external_name() : "");
1849 }
1850 }
1851 #endif // ASSERT
1852
1853 static inline freeze_result freeze_epilog(ContinuationWrapper& cont) {
1854 verify_continuation(cont.continuation());
1855 assert(!cont.is_empty(), "");
1856
1857 log_develop_debug(continuations)("=== End of freeze cont ### #" INTPTR_FORMAT, cont.hash());
1858 return freeze_ok;
1859 }
1860
1861 static freeze_result freeze_epilog(JavaThread* thread, ContinuationWrapper& cont, freeze_result res) {
1862 if (UNLIKELY(res != freeze_ok)) {
1863 JFR_ONLY(thread->set_last_freeze_fail_result(res);)
1864 verify_continuation(cont.continuation());
1865 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1866 return res;
1867 }
1868
1869 JVMTI_ONLY(jvmti_yield_cleanup(thread, cont)); // can safepoint
1870 return freeze_epilog(cont);
1871 }
1872
1873 static freeze_result preempt_epilog(ContinuationWrapper& cont, freeze_result res, frame& old_last_frame) {
1874 if (UNLIKELY(res != freeze_ok)) {
1875 verify_continuation(cont.continuation());
1876 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1877 return res;
1878 }
1879
1880 // Set up things so that on return to Java we jump to preempt stub.
1881 patch_return_pc_with_preempt_stub(old_last_frame);
1882 cont.tail()->set_preempted(true);
1883 DEBUG_ONLY(log_preempt_after_freeze(cont);)
1884 return freeze_epilog(cont);
1885 }
1886
1887 template<typename ConfigT, bool preempt>
1888 static inline freeze_result freeze_internal(JavaThread* current, intptr_t* const sp) {
1889 assert(!current->has_pending_exception(), "");
1890
1891 #ifdef ASSERT
1892 log_trace(continuations)("~~~~ freeze sp: " INTPTR_FORMAT "JavaThread: " INTPTR_FORMAT, p2i(current->last_continuation()->entry_sp()), p2i(current));
1893 log_frames(current);
1894 #endif
1895
1896 CONT_JFR_ONLY(EventContinuationFreeze event;)
1897
1898 ContinuationEntry* entry = current->last_continuation();
1899
1900 oop oopCont = entry->cont_oop(current);
1901 assert(oopCont == current->last_continuation()->cont_oop(current), "");
1902 assert(ContinuationEntry::assert_entry_frame_laid_out(current), "");
1903
1904 verify_continuation(oopCont);
1905 ContinuationWrapper cont(current, oopCont);
1906 log_develop_debug(continuations)("FREEZE #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
1907
1908 assert(entry->is_virtual_thread() == (entry->scope(current) == java_lang_VirtualThread::vthread_scope()), "");
1909
1910 if (entry->is_pinned()) {
1911 log_develop_debug(continuations)("PINNED due to critical section");
1912 verify_continuation(cont.continuation());
1913 const freeze_result res = freeze_pinned_cs;
1914 if (!preempt) {
1915 JFR_ONLY(current->set_last_freeze_fail_result(res);)
1916 }
1917 log_develop_trace(continuations)("=== end of freeze (fail %d)", res);
1918 // Avoid Thread.yield() loops without safepoint polls.
1919 if (SafepointMechanism::should_process(current) && !preempt) {
1920 cont.done(); // allow safepoint
1921 ThreadInVMfromJava tivmfj(current);
1922 }
1923 return res;
1924 }
1925
1926 Freeze<ConfigT> freeze(current, cont, sp, preempt);
1927
1928 assert(!current->cont_fastpath() || freeze.check_valid_fast_path(), "");
1929 bool fast = UseContinuationFastPath && current->cont_fastpath();
1930 if (fast && freeze.size_if_fast_freeze_available() > 0) {
1931 freeze.freeze_fast_existing_chunk();
1932 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1933 return !preempt ? freeze_epilog(cont) : preempt_epilog(cont, freeze_ok, freeze.last_frame());
1934 }
1935
1936 if (preempt) {
1937 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1938 freeze.set_jvmti_event_collector(&jsoaec);
1939
1940 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1941
1942 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1943 preempt_epilog(cont, res, freeze.last_frame());
1944 return res;
1945 }
1946
1947 log_develop_trace(continuations)("chunk unavailable; transitioning to VM");
1948 assert(current == JavaThread::current(), "must be current thread");
1949 JRT_BLOCK
1950 // delays a possible JvmtiSampledObjectAllocEventCollector in alloc_chunk
1951 JvmtiSampledObjectAllocEventCollector jsoaec(false);
1952 freeze.set_jvmti_event_collector(&jsoaec);
1953
1954 freeze_result res = fast ? freeze.try_freeze_fast() : freeze.freeze_slow();
1955
1956 CONT_JFR_ONLY(freeze.jfr_info().post_jfr_event(&event, oopCont, current);)
1957 freeze_epilog(current, cont, res);
1958 cont.done(); // allow safepoint in the transition back to Java
1959 return res;
1960 JRT_BLOCK_END
1961 }
1962
1963 static freeze_result is_pinned0(JavaThread* thread, oop cont_scope, bool safepoint) {
1964 ContinuationEntry* entry = thread->last_continuation();
1965 if (entry == nullptr) {
1966 return freeze_ok;
1967 }
1968 if (entry->is_pinned()) {
1969 return freeze_pinned_cs;
1970 }
1971
1972 RegisterMap map(thread,
1973 RegisterMap::UpdateMap::include,
1974 RegisterMap::ProcessFrames::skip,
1975 RegisterMap::WalkContinuation::skip);
1976 map.set_include_argument_oops(false);
1977 frame f = thread->last_frame();
1978
1979 if (!safepoint) {
1980 f = f.sender(&map); // this is the yield frame
1981 } else { // safepoint yield
1982 #if (defined(X86) || defined(AARCH64) || defined(RISCV64)) && !defined(ZERO)
1983 f.set_fp(f.real_fp()); // Instead of this, maybe in ContinuationWrapper::set_last_frame always use the real_fp?
1984 #else
1985 Unimplemented();
1986 #endif
1987 if (!Interpreter::contains(f.pc())) {
1988 assert(ContinuationHelper::Frame::is_stub(f.cb()), "must be");
1989 assert(f.oop_map() != nullptr, "must be");
1990 f.oop_map()->update_register_map(&f, &map); // we have callee-save registers in this case
1991 }
1992 }
1993
1994 while (true) {
1995 if ((f.is_interpreted_frame() && f.interpreter_frame_method()->is_native()) || f.is_native_frame()) {
1996 return freeze_pinned_native;
1997 }
1998
1999 f = f.sender(&map);
2000 if (!Continuation::is_frame_in_continuation(entry, f)) {
2001 oop scope = jdk_internal_vm_Continuation::scope(entry->cont_oop(thread));
2002 if (scope == cont_scope) {
2003 break;
2004 }
2005 entry = entry->parent();
2006 if (entry == nullptr) {
2007 break;
2008 }
2009 if (entry->is_pinned()) {
2010 return freeze_pinned_cs;
2011 }
2012 }
2013 }
2014 return freeze_ok;
2015 }
2016
2017 /////////////// THAW ////
2018
2019 static int thaw_size(stackChunkOop chunk) {
2020 int size = chunk->max_thawing_size();
2021 size += frame::metadata_words; // For the top pc+fp in push_return_frame or top = stack_sp - frame::metadata_words in thaw_fast
2022 size += 2*frame::align_wiggle; // in case of alignments at the top and bottom
2023 return size;
2024 }
2025
2026 // make room on the stack for thaw
2027 // returns the size in bytes, or 0 on failure
2028 static inline int prepare_thaw_internal(JavaThread* thread, bool return_barrier) {
2029 log_develop_trace(continuations)("~~~~ prepare_thaw return_barrier: %d", return_barrier);
2030
2031 assert(thread == JavaThread::current(), "");
2032
2033 ContinuationEntry* ce = thread->last_continuation();
2034 assert(ce != nullptr, "");
2035 oop continuation = ce->cont_oop(thread);
2036 assert(continuation == get_continuation(thread), "");
2037 verify_continuation(continuation);
2038
2039 stackChunkOop chunk = jdk_internal_vm_Continuation::tail(continuation);
2040 assert(chunk != nullptr, "");
2041
2042 // The tail can be empty because it might still be available for another freeze.
2043 // However, here we want to thaw, so we get rid of it (it will be GCed).
2044 if (UNLIKELY(chunk->is_empty())) {
2045 chunk = chunk->parent();
2046 assert(chunk != nullptr, "");
2047 assert(!chunk->is_empty(), "");
2048 jdk_internal_vm_Continuation::set_tail(continuation, chunk);
2049 }
2050
2051 // Verification
2052 chunk->verify();
2053 assert(chunk->max_thawing_size() > 0, "chunk invariant violated; expected to not be empty");
2054
2055 // Only make space for the last chunk because we only thaw from the last chunk
2056 int size = thaw_size(chunk) << LogBytesPerWord;
2057
2058 const address bottom = (address)thread->last_continuation()->entry_sp();
2059 // 300 is an estimate for stack size taken for this native code, in addition to StackShadowPages
2060 // for the Java frames in the check below.
2061 if (!stack_overflow_check(thread, size + 300, bottom)) {
2062 return 0;
2063 }
2064
2065 log_develop_trace(continuations)("prepare_thaw bottom: " INTPTR_FORMAT " top: " INTPTR_FORMAT " size: %d",
2066 p2i(bottom), p2i(bottom - size), size);
2067 return size;
2068 }
2069
2070 class ThawBase : public StackObj {
2071 protected:
2072 JavaThread* _thread;
2073 ContinuationWrapper& _cont;
2074 CONT_JFR_ONLY(FreezeThawJfrInfo _jfr_info;)
2075
2076 intptr_t* _fastpath;
2077 bool _barriers;
2078 bool _preempted_case;
2079 bool _process_args_at_top;
2080 intptr_t* _top_unextended_sp_before_thaw;
2081 int _align_size;
2082 DEBUG_ONLY(intptr_t* _top_stack_address);
2083
2084 // Only used for preemption on ObjectLocker
2085 ObjectMonitor* _init_lock;
2086
2087 StackChunkFrameStream<ChunkFrames::Mixed> _stream;
2088
2089 NOT_PRODUCT(int _frames;)
2090
2091 protected:
2092 ThawBase(JavaThread* thread, ContinuationWrapper& cont) :
2093 _thread(thread), _cont(cont),
2094 _fastpath(nullptr) {
2095 DEBUG_ONLY(_top_unextended_sp_before_thaw = nullptr;)
2096 assert (cont.tail() != nullptr, "no last chunk");
2097 DEBUG_ONLY(_top_stack_address = _cont.entrySP() - thaw_size(cont.tail());)
2098 }
2099
2100 void clear_chunk(stackChunkOop chunk);
2101 template<bool check_stub>
2102 int remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize);
2103 int remove_scalarized_frames(StackChunkFrameStream<ChunkFrames::CompiledOnly>& scfs, int &argsize);
2104 void copy_from_chunk(intptr_t* from, intptr_t* to, int size);
2105
2106 void thaw_lockstack(stackChunkOop chunk);
2107
2108 // fast path
2109 inline void prefetch_chunk_pd(void* start, int size_words);
2110 void patch_return(intptr_t* sp, bool is_last);
2111
2112 intptr_t* handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case);
2113 inline intptr_t* push_cleanup_continuation();
2114 inline intptr_t* push_preempt_adapter();
2115 intptr_t* redo_vmcall(JavaThread* current, frame& top);
2116 void throw_interrupted_exception(JavaThread* current, frame& top);
2117
2118 void recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case);
2119 void finish_thaw(frame& f);
2120
2121 private:
2122 template<typename FKind> bool recurse_thaw_java_frame(frame& caller, int num_frames);
2123 void finalize_thaw(frame& entry, int argsize);
2124
2125 inline bool seen_by_gc();
2126
2127 inline void before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame);
2128 inline void after_thaw_java_frame(const frame& f, bool bottom);
2129 inline void patch(frame& f, const frame& caller, bool bottom, bool augmented = false);
2130 void clear_bitmap_bits(address start, address end);
2131
2132 NOINLINE void recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top);
2133 void recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller);
2134 void recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames);
2135 void recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames);
2136
2137 void push_return_frame(const frame& f);
2138 inline frame new_entry_frame();
2139 template<typename FKind> frame new_stack_frame(const frame& hf, frame& caller, bool bottom, int size_adjust = 0);
2140 inline void patch_pd(frame& f, const frame& sender);
2141 inline void patch_pd(frame& f, intptr_t* caller_sp);
2142 inline intptr_t* align(const frame& hf, intptr_t* frame_sp, frame& caller, bool bottom);
2143
2144 void maybe_set_fastpath(intptr_t* sp) { if (sp > _fastpath) _fastpath = sp; }
2145
2146 static inline void derelativize_interpreted_frame_metadata(const frame& hf, const frame& f);
2147
2148 public:
2149 CONT_JFR_ONLY(FreezeThawJfrInfo& jfr_info() { return _jfr_info; })
2150 };
2151
2152 template <typename ConfigT>
2153 class Thaw : public ThawBase {
2154 public:
2155 Thaw(JavaThread* thread, ContinuationWrapper& cont) : ThawBase(thread, cont) {}
2156
2157 inline bool can_thaw_fast(stackChunkOop chunk) {
2158 return !_barriers
2159 && _thread->cont_fastpath_thread_state()
2160 && !chunk->has_thaw_slowpath_condition()
2161 && !PreserveFramePointer;
2162 }
2163
2164 inline intptr_t* thaw(Continuation::thaw_kind kind);
2165 template<bool check_stub = false>
2166 NOINLINE intptr_t* thaw_fast(stackChunkOop chunk);
2167 NOINLINE intptr_t* thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind);
2168 inline void patch_caller_links(intptr_t* sp, intptr_t* bottom);
2169 };
2170
2171 template <typename ConfigT>
2172 inline intptr_t* Thaw<ConfigT>::thaw(Continuation::thaw_kind kind) {
2173 verify_continuation(_cont.continuation());
2174 assert(!jdk_internal_vm_Continuation::done(_cont.continuation()), "");
2175 assert(!_cont.is_empty(), "");
2176
2177 stackChunkOop chunk = _cont.tail();
2178 assert(chunk != nullptr, "guaranteed by prepare_thaw");
2179 assert(!chunk->is_empty(), "guaranteed by prepare_thaw");
2180
2181 _barriers = chunk->requires_barriers();
2182 return (LIKELY(can_thaw_fast(chunk))) ? thaw_fast(chunk)
2183 : thaw_slow(chunk, kind);
2184 }
2185
2186 class ReconstructedStack : public StackObj {
2187 intptr_t* _base; // _cont.entrySP(); // top of the entry frame
2188 int _thaw_size;
2189 int _argsize;
2190 public:
2191 ReconstructedStack(intptr_t* base, int thaw_size, int argsize)
2192 : _base(base), _thaw_size(thaw_size - (argsize == 0 ? frame::metadata_words_at_top : 0)), _argsize(argsize) {
2193 // The only possible source of misalignment is stack-passed arguments b/c compiled frames are 16-byte aligned.
2194 assert(argsize != 0 || (_base - _thaw_size) == ContinuationHelper::frame_align_pointer(_base - _thaw_size), "");
2195 // We're at most one alignment word away from entrySP
2196 assert(_base - 1 <= top() + total_size() + frame::metadata_words_at_bottom, "missed entry frame");
2197 }
2198
2199 int entry_frame_extension() const { return _argsize + (_argsize > 0 ? frame::metadata_words_at_top : 0); }
2200
2201 // top and bottom stack pointers
2202 intptr_t* sp() const { return ContinuationHelper::frame_align_pointer(_base - _thaw_size); }
2203 intptr_t* bottom_sp() const { return ContinuationHelper::frame_align_pointer(_base - entry_frame_extension()); }
2204
2205 // several operations operate on the totality of the stack being reconstructed,
2206 // including the metadata words
2207 intptr_t* top() const { return sp() - frame::metadata_words_at_bottom; }
2208 int total_size() const { return _thaw_size + frame::metadata_words_at_bottom; }
2209 };
2210
2211 inline void ThawBase::clear_chunk(stackChunkOop chunk) {
2212 chunk->set_sp(chunk->bottom());
2213 chunk->set_max_thawing_size(0);
2214 }
2215
2216 int ThawBase::remove_scalarized_frames(StackChunkFrameStream<ChunkFrames::CompiledOnly>& f, int &argsize) {
2217 intptr_t* top = f.sp();
2218
2219 while (f.cb()->as_nmethod()->needs_stack_repair()) {
2220 f.next(SmallRegisterMap::instance_no_args(), false /* stop */);
2221 }
2222 assert(!f.is_done(), "");
2223 assert(f.is_compiled(), "");
2224
2225 intptr_t* bottom = f.sp() + f.cb()->frame_size();
2226 argsize = f.stack_argsize();
2227 return bottom - top;
2228 }
2229
2230 template<bool check_stub>
2231 int ThawBase::remove_top_compiled_frame_from_chunk(stackChunkOop chunk, int &argsize) {
2232 bool empty = false;
2233 StackChunkFrameStream<ChunkFrames::CompiledOnly> f(chunk);
2234 DEBUG_ONLY(intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();)
2235 assert(chunk_sp == f.sp(), "");
2236 assert(chunk_sp == f.unextended_sp(), "");
2237
2238 int frame_size = f.cb()->frame_size();
2239 argsize = f.stack_argsize();
2240
2241 assert(!f.is_stub() || check_stub, "");
2242 if (check_stub && f.is_stub()) {
2243 // If we don't thaw the top compiled frame too, after restoring the saved
2244 // registers back in Java, we would hit the return barrier to thaw one more
2245 // frame effectively overwriting the restored registers during that call.
2246 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2247 assert(!f.is_done(), "");
2248
2249 f.get_cb();
2250 assert(f.is_compiled(), "");
2251 if (f.cb()->as_nmethod()->is_marked_for_deoptimization()) {
2252 // The caller of the runtime stub when the continuation is preempted is not at a
2253 // Java call instruction, and so cannot rely on nmethod patching for deopt.
2254 log_develop_trace(continuations)("Deoptimizing runtime stub caller");
2255 f.to_frame().deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2256 }
2257
2258 if (f.cb()->as_nmethod()->needs_stack_repair()) {
2259 frame_size += remove_scalarized_frames(f, argsize);
2260 } else {
2261 frame_size += f.cb()->frame_size();
2262 argsize = f.stack_argsize();
2263 }
2264 } else if (f.cb()->as_nmethod()->needs_stack_repair()) {
2265 frame_size = remove_scalarized_frames(f, argsize);
2266 }
2267
2268 f.next(SmallRegisterMap::instance_no_args(), true /* stop */);
2269 empty = f.is_done();
2270 assert(!empty || argsize == chunk->argsize(), "");
2271
2272 if (empty) {
2273 clear_chunk(chunk);
2274 } else {
2275 chunk->set_sp(chunk->sp() + frame_size);
2276 chunk->set_max_thawing_size(chunk->max_thawing_size() - frame_size);
2277 // We set chunk->pc to the return pc into the next frame
2278 chunk->set_pc(f.pc());
2279 #ifdef ASSERT
2280 {
2281 intptr_t* retaddr_slot = (chunk_sp
2282 + frame_size
2283 - frame::sender_sp_ret_address_offset());
2284 assert(f.pc() == ContinuationHelper::return_address_at(retaddr_slot),
2285 "unexpected pc");
2286 }
2287 #endif
2288 }
2289 assert(empty == chunk->is_empty(), "");
2290 // returns the size required to store the frame on stack, and because it is a
2291 // compiled frame, it must include a copy of the arguments passed by the caller
2292 return frame_size + argsize + frame::metadata_words_at_top;
2293 }
2294
2295 void ThawBase::thaw_lockstack(stackChunkOop chunk) {
2296 int lockStackSize = chunk->lockstack_size();
2297 assert(lockStackSize > 0 && lockStackSize <= LockStack::CAPACITY, "");
2298
2299 oop tmp_lockstack[LockStack::CAPACITY];
2300 chunk->transfer_lockstack(tmp_lockstack, _barriers);
2301 _thread->lock_stack().move_from_address(tmp_lockstack, lockStackSize);
2302
2303 chunk->set_lockstack_size(0);
2304 chunk->set_has_lockstack(false);
2305 }
2306
2307 void ThawBase::copy_from_chunk(intptr_t* from, intptr_t* to, int size) {
2308 assert(to >= _top_stack_address, "overwrote past thawing space"
2309 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(to), p2i(_top_stack_address));
2310 assert(to + size <= _cont.entrySP(), "overwrote past thawing space");
2311 _cont.tail()->copy_from_chunk_to_stack(from, to, size);
2312 CONT_JFR_ONLY(_jfr_info.record_size_copied(size);)
2313 }
2314
2315 void ThawBase::patch_return(intptr_t* sp, bool is_last) {
2316 log_develop_trace(continuations)("thaw_fast patching -- sp: " INTPTR_FORMAT, p2i(sp));
2317
2318 address pc = !is_last ? StubRoutines::cont_returnBarrier() : _cont.entryPC();
2319 ContinuationHelper::patch_return_address_at(
2320 sp - frame::sender_sp_ret_address_offset(),
2321 pc);
2322 }
2323
2324 template <typename ConfigT>
2325 template<bool check_stub>
2326 NOINLINE intptr_t* Thaw<ConfigT>::thaw_fast(stackChunkOop chunk) {
2327 assert(chunk == _cont.tail(), "");
2328 assert(!chunk->has_mixed_frames(), "");
2329 assert(!chunk->requires_barriers(), "");
2330 assert(!chunk->has_bitmap(), "");
2331 assert(!_thread->is_interp_only_mode(), "");
2332
2333 LogTarget(Trace, continuations) lt;
2334 if (lt.develop_is_enabled()) {
2335 LogStream ls(lt);
2336 ls.print_cr("thaw_fast");
2337 chunk->print_on(true, &ls);
2338 }
2339
2340 // Below this heuristic, we thaw the whole chunk, above it we thaw just one frame.
2341 static const int threshold = 500; // words
2342
2343 const int full_chunk_size = chunk->stack_size() - chunk->sp(); // this initial size could be reduced if it's a partial thaw
2344 int argsize, thaw_size;
2345
2346 intptr_t* const chunk_sp = chunk->start_address() + chunk->sp();
2347
2348 bool partial, empty;
2349 if (LIKELY(!TEST_THAW_ONE_CHUNK_FRAME && (full_chunk_size < threshold))) {
2350 prefetch_chunk_pd(chunk->start_address(), full_chunk_size); // prefetch anticipating memcpy starting at highest address
2351
2352 partial = false;
2353 argsize = chunk->argsize(); // must be called *before* clearing the chunk
2354 clear_chunk(chunk);
2355 thaw_size = full_chunk_size;
2356 empty = true;
2357 } else { // thaw a single frame
2358 partial = true;
2359 thaw_size = remove_top_compiled_frame_from_chunk<check_stub>(chunk, argsize);
2360 empty = chunk->is_empty();
2361 }
2362
2363 // Are we thawing the last frame(s) in the continuation
2364 const bool is_last = empty && chunk->parent() == nullptr;
2365 assert(!is_last || argsize == 0, "");
2366
2367 log_develop_trace(continuations)("thaw_fast partial: %d is_last: %d empty: %d size: %d argsize: %d entrySP: " PTR_FORMAT,
2368 partial, is_last, empty, thaw_size, argsize, p2i(_cont.entrySP()));
2369
2370 ReconstructedStack rs(_cont.entrySP(), thaw_size, argsize);
2371
2372 // also copy metadata words at frame bottom
2373 copy_from_chunk(chunk_sp - frame::metadata_words_at_bottom, rs.top(), rs.total_size());
2374
2375 // update the ContinuationEntry
2376 _cont.set_argsize(argsize);
2377 log_develop_trace(continuations)("setting entry argsize: %d", _cont.argsize());
2378 assert(rs.bottom_sp() == _cont.entry()->bottom_sender_sp(), "");
2379
2380 // install the return barrier if not last frame, or the entry's pc if last
2381 patch_return(rs.bottom_sp(), is_last);
2382
2383 // insert the back links from callee to caller frames
2384 patch_caller_links(rs.top(), rs.top() + rs.total_size());
2385
2386 assert(is_last == _cont.is_empty(), "");
2387 assert(_cont.chunk_invariant(), "");
2388
2389 #if CONT_JFR
2390 EventContinuationThawFast e;
2391 if (e.should_commit()) {
2392 e.set_id(cast_from_oop<u8>(chunk));
2393 e.set_size(thaw_size << LogBytesPerWord);
2394 e.set_full(!partial);
2395 e.commit();
2396 }
2397 #endif
2398
2399 #ifdef ASSERT
2400 if (LoomDeoptAfterThaw) {
2401 frame top(rs.sp());
2402 AnchorMark am(_thread, top);
2403 log_frames(_thread);
2404 do_deopt_after_thaw(_thread);
2405 }
2406 #endif
2407
2408 return rs.sp();
2409 }
2410
2411 inline bool ThawBase::seen_by_gc() {
2412 return _barriers || _cont.tail()->is_gc_mode();
2413 }
2414
2415 static inline void relativize_chunk_concurrently(stackChunkOop chunk) {
2416 #if INCLUDE_ZGC || INCLUDE_SHENANDOAHGC
2417 if (UseZGC || UseShenandoahGC) {
2418 chunk->relativize_derived_pointers_concurrently();
2419 }
2420 #endif
2421 }
2422
2423 template <typename ConfigT>
2424 NOINLINE intptr_t* Thaw<ConfigT>::thaw_slow(stackChunkOop chunk, Continuation::thaw_kind kind) {
2425 Continuation::preempt_kind preempt_kind;
2426 bool retry_fast_path = false;
2427
2428 _process_args_at_top = false;
2429 _preempted_case = chunk->preempted();
2430 if (_preempted_case) {
2431 ObjectMonitor* mon = nullptr;
2432 ObjectWaiter* waiter = java_lang_VirtualThread::objectWaiter(_thread->vthread());
2433 if (waiter != nullptr) {
2434 // Mounted again after preemption. Resume the pending monitor operation,
2435 // which will be either a monitorenter or Object.wait() call.
2436 mon = waiter->monitor();
2437 preempt_kind = waiter->is_wait() ? Continuation::object_wait : Continuation::monitorenter;
2438
2439 bool mon_acquired = mon->resume_operation(_thread, waiter, _cont);
2440 assert(!mon_acquired || mon->has_owner(_thread), "invariant");
2441 if (!mon_acquired) {
2442 // Failed to acquire monitor. Return to enterSpecial to unmount again.
2443 log_develop_trace(continuations, preempt)("Failed to acquire monitor, unmounting again");
2444 return push_cleanup_continuation();
2445 }
2446 chunk = _cont.tail(); // reload oop in case of safepoint in resume_operation (if posting JVMTI events).
2447 JVMTI_ONLY(assert(_thread->contended_entered_monitor() == nullptr || _thread->contended_entered_monitor() == mon, ""));
2448 } else {
2449 // Preemption cancelled on moniterenter or ObjectLocker case. We
2450 // actually acquired the monitor after freezing all frames so no
2451 // need to call resume_operation. If this is the ObjectLocker case
2452 // we released the monitor already at ~ObjectLocker, so _init_lock
2453 // will be set to nullptr below since there is no monitor to release.
2454 preempt_kind = Continuation::monitorenter;
2455 }
2456
2457 // Call this first to avoid racing with GC threads later when modifying the chunk flags.
2458 relativize_chunk_concurrently(chunk);
2459
2460 if (chunk->at_klass_init()) {
2461 preempt_kind = Continuation::object_locker;
2462 chunk->set_at_klass_init(false);
2463 _process_args_at_top = chunk->has_args_at_top();
2464 if (_process_args_at_top) {
2465 // Only needed for the top frame which will be thawed.
2466 chunk->set_has_args_at_top(false);
2467 }
2468 assert(waiter == nullptr || mon != nullptr, "should have a monitor");
2469 _init_lock = mon; // remember monitor since we will need it on handle_preempted_continuation()
2470 }
2471 chunk->set_preempted(false);
2472 retry_fast_path = true;
2473 } else {
2474 relativize_chunk_concurrently(chunk);
2475 }
2476
2477 // On first thaw after freeze restore oops to the lockstack if any.
2478 assert(chunk->lockstack_size() == 0 || kind == Continuation::thaw_top, "");
2479 if (kind == Continuation::thaw_top && chunk->lockstack_size() > 0) {
2480 thaw_lockstack(chunk);
2481 retry_fast_path = true;
2482 }
2483
2484 // Retry the fast path now that we possibly cleared the FLAG_HAS_LOCKSTACK
2485 // and FLAG_PREEMPTED flags from the stackChunk.
2486 if (retry_fast_path && can_thaw_fast(chunk)) {
2487 intptr_t* sp = thaw_fast<true>(chunk);
2488 if (_preempted_case) {
2489 return handle_preempted_continuation(sp, preempt_kind, true /* fast_case */);
2490 }
2491 return sp;
2492 }
2493
2494 LogTarget(Trace, continuations) lt;
2495 if (lt.develop_is_enabled()) {
2496 LogStream ls(lt);
2497 ls.print_cr("thaw slow return_barrier: %d " INTPTR_FORMAT, kind, p2i(chunk));
2498 chunk->print_on(true, &ls);
2499 }
2500
2501 #if CONT_JFR
2502 EventContinuationThawSlow e;
2503 if (e.should_commit()) {
2504 e.set_id(cast_from_oop<u8>(_cont.continuation()));
2505 e.commit();
2506 }
2507 #endif
2508
2509 DEBUG_ONLY(_frames = 0;)
2510 _align_size = 0;
2511 int num_frames = kind == Continuation::thaw_top ? 2 : 1;
2512
2513 _stream = StackChunkFrameStream<ChunkFrames::Mixed>(chunk);
2514 _top_unextended_sp_before_thaw = _stream.unextended_sp();
2515
2516 frame heap_frame = _stream.to_frame();
2517 if (lt.develop_is_enabled()) {
2518 LogStream ls(lt);
2519 ls.print_cr("top hframe before (thaw):");
2520 assert(heap_frame.is_heap_frame(), "should have created a relative frame");
2521 heap_frame.print_value_on(&ls);
2522 }
2523
2524 frame caller; // the thawed caller on the stack
2525 recurse_thaw(heap_frame, caller, num_frames, _preempted_case);
2526 finish_thaw(caller); // caller is now the topmost thawed frame
2527 _cont.write();
2528
2529 assert(_cont.chunk_invariant(), "");
2530
2531 JVMTI_ONLY(if (!_cont.entry()->is_virtual_thread()) invalidate_jvmti_stack(_thread));
2532
2533 _thread->set_cont_fastpath(_fastpath);
2534
2535 intptr_t* sp = caller.sp();
2536
2537 if (_preempted_case) {
2538 return handle_preempted_continuation(sp, preempt_kind, false /* fast_case */);
2539 }
2540 return sp;
2541 }
2542
2543 void ThawBase::recurse_thaw(const frame& heap_frame, frame& caller, int num_frames, bool top_on_preempt_case) {
2544 log_develop_debug(continuations)("thaw num_frames: %d", num_frames);
2545 assert(!_cont.is_empty(), "no more frames");
2546 assert(num_frames > 0, "");
2547 assert(!heap_frame.is_empty(), "");
2548
2549 if (top_on_preempt_case && (heap_frame.is_native_frame() || heap_frame.is_runtime_frame())) {
2550 heap_frame.is_native_frame() ? recurse_thaw_native_frame(heap_frame, caller, 2) : recurse_thaw_stub_frame(heap_frame, caller, 2);
2551 } else if (!heap_frame.is_interpreted_frame()) {
2552 recurse_thaw_compiled_frame(heap_frame, caller, num_frames, false);
2553 } else {
2554 recurse_thaw_interpreted_frame(heap_frame, caller, num_frames, top_on_preempt_case);
2555 }
2556 }
2557
2558 template<typename FKind>
2559 bool ThawBase::recurse_thaw_java_frame(frame& caller, int num_frames) {
2560 assert(num_frames > 0, "");
2561
2562 DEBUG_ONLY(_frames++;)
2563
2564 int argsize = _stream.stack_argsize();
2565 CodeBlob* cb = _stream.cb();
2566
2567 _stream.next(SmallRegisterMap::instance_no_args());
2568 assert(_stream.to_frame().is_empty() == _stream.is_done(), "");
2569
2570 // We never leave a compiled caller of an interpreted frame as the top frame in the chunk
2571 // as it makes detecting that situation and adjusting unextended_sp tricky. We also always
2572 // thaw the caller of a frame that needs_stack_repair, as it would otherwise complicate things:
2573 // - Regardless of whether the frame was extended or not, we would need to copy the right arg
2574 // size if its greater than the one given by the normal method signature (non-scalarized).
2575 // - If the frame was indeed extended, leaving its caller as the top frame would complicate walking
2576 // the chunk (we need unextended_sp, but we only have sp).
2577 if (num_frames == 1 && !_stream.is_done() && ((FKind::interpreted && _stream.is_compiled()) || (FKind::compiled && cb->as_nmethod_or_null()->needs_stack_repair()))) {
2578 log_develop_trace(continuations)("thawing extra compiled frame to not leave a compiled interpreted-caller at top");
2579 num_frames++;
2580 }
2581
2582 if (num_frames == 1 || _stream.is_done()) { // end recursion
2583 finalize_thaw(caller, FKind::interpreted ? 0 : argsize);
2584 return true; // bottom
2585 } else { // recurse
2586 recurse_thaw(_stream.to_frame(), caller, num_frames - 1, false /* top_on_preempt_case */);
2587 return false;
2588 }
2589 }
2590
2591 void ThawBase::finalize_thaw(frame& entry, int argsize) {
2592 stackChunkOop chunk = _cont.tail();
2593
2594 if (!_stream.is_done()) {
2595 assert(_stream.sp() >= chunk->sp_address(), "");
2596 chunk->set_sp(chunk->to_offset(_stream.sp()));
2597 chunk->set_pc(_stream.pc());
2598 } else {
2599 chunk->set_sp(chunk->bottom());
2600 chunk->set_pc(nullptr);
2601 }
2602 assert(_stream.is_done() == chunk->is_empty(), "");
2603
2604 int total_thawed = pointer_delta_as_int(_stream.unextended_sp(), _top_unextended_sp_before_thaw);
2605 chunk->set_max_thawing_size(chunk->max_thawing_size() - total_thawed);
2606
2607 _cont.set_argsize(argsize);
2608 entry = new_entry_frame();
2609
2610 assert(entry.sp() == _cont.entrySP(), "");
2611 assert(Continuation::is_continuation_enterSpecial(entry), "");
2612 assert(_cont.is_entry_frame(entry), "");
2613 }
2614
2615 inline void ThawBase::before_thaw_java_frame(const frame& hf, const frame& caller, bool bottom, int num_frame) {
2616 LogTarget(Trace, continuations) lt;
2617 if (lt.develop_is_enabled()) {
2618 LogStream ls(lt);
2619 ls.print_cr("======== THAWING FRAME: %d", num_frame);
2620 assert(hf.is_heap_frame(), "should be");
2621 hf.print_value_on(&ls);
2622 }
2623 assert(bottom == _cont.is_entry_frame(caller), "bottom: %d is_entry_frame: %d", bottom, _cont.is_entry_frame(hf));
2624 }
2625
2626 inline void ThawBase::after_thaw_java_frame(const frame& f, bool bottom) {
2627 #ifdef ASSERT
2628 LogTarget(Trace, continuations) lt;
2629 if (lt.develop_is_enabled()) {
2630 LogStream ls(lt);
2631 ls.print_cr("thawed frame:");
2632 print_frame_layout(f, false, &ls); // f.print_on(&ls);
2633 }
2634 #endif
2635 }
2636
2637 inline void ThawBase::patch(frame& f, const frame& caller, bool bottom, bool augmented) {
2638 assert(!bottom || caller.fp() == _cont.entryFP(), "");
2639 if (bottom) {
2640 ContinuationHelper::Frame::patch_pc(caller, _cont.is_empty() ? caller.pc()
2641 : StubRoutines::cont_returnBarrier());
2642 } else if (caller.is_compiled_frame()){
2643 // caller might have been deoptimized during thaw but we've overwritten the return address when copying f from the heap.
2644 // If the caller is not deoptimized, pc is unchanged.
2645 ContinuationHelper::Frame::patch_pc(caller, caller.raw_pc(), augmented /*callee_augmented*/);
2646 }
2647
2648 patch_pd(f, caller);
2649
2650 if (f.is_interpreted_frame()) {
2651 ContinuationHelper::InterpretedFrame::patch_sender_sp(f, caller);
2652 }
2653
2654 assert(!bottom || !_cont.is_empty() || Continuation::is_continuation_entry_frame(f, nullptr), "");
2655 assert(!bottom || (_cont.is_empty() != Continuation::is_cont_barrier_frame(f)), "");
2656 }
2657
2658 void ThawBase::clear_bitmap_bits(address start, address end) {
2659 assert(is_aligned(start, wordSize), "should be aligned: " PTR_FORMAT, p2i(start));
2660 assert(is_aligned(end, VMRegImpl::stack_slot_size), "should be aligned: " PTR_FORMAT, p2i(end));
2661
2662 // we need to clear the bits that correspond to arguments as they reside in the caller frame
2663 // or they will keep objects that are otherwise unreachable alive.
2664
2665 // Align `end` if UseCompressedOops is not set to avoid UB when calculating the bit index, since
2666 // `end` could be at an odd number of stack slots from `start`, i.e might not be oop aligned.
2667 // If that's the case the bit range corresponding to the last stack slot should not have bits set
2668 // anyways and we assert that before returning.
2669 address effective_end = UseCompressedOops ? end : align_down(end, wordSize);
2670 log_develop_trace(continuations)("clearing bitmap for " INTPTR_FORMAT " - " INTPTR_FORMAT, p2i(start), p2i(effective_end));
2671 stackChunkOop chunk = _cont.tail();
2672 chunk->bitmap().clear_range(chunk->bit_index_for(start), chunk->bit_index_for(effective_end));
2673 assert(effective_end == end || !chunk->bitmap().at(chunk->bit_index_for(effective_end)), "bit should not be set");
2674 }
2675
2676 intptr_t* ThawBase::handle_preempted_continuation(intptr_t* sp, Continuation::preempt_kind preempt_kind, bool fast_case) {
2677 frame top(sp);
2678 assert(top.pc() == *(address*)(sp - frame::sender_sp_ret_address_offset()), "");
2679 DEBUG_ONLY(verify_frame_kind(top, preempt_kind);)
2680 NOT_PRODUCT(int64_t tid = _thread->monitor_owner_id();)
2681
2682 #if INCLUDE_JVMTI
2683 // Finish the VTMS transition.
2684 assert(_thread->is_in_VTMS_transition(), "must be");
2685 bool is_vthread = Continuation::continuation_scope(_cont.continuation()) == java_lang_VirtualThread::vthread_scope();
2686 if (is_vthread) {
2687 if (JvmtiVTMSTransitionDisabler::VTMS_notify_jvmti_events()) {
2688 jvmti_mount_end(_thread, _cont, top, preempt_kind);
2689 } else {
2690 _thread->set_is_in_VTMS_transition(false);
2691 java_lang_Thread::set_is_in_VTMS_transition(_thread->vthread(), false);
2692 }
2693 }
2694 #endif
2695
2696 if (fast_case) {
2697 // If we thawed in the slow path the runtime stub/native wrapper frame already
2698 // has the correct fp (see ThawBase::new_stack_frame). On the fast path though,
2699 // we copied the fp patched during freeze, which will now have to be fixed.
2700 assert(top.is_runtime_frame() || top.is_native_frame(), "");
2701 int fsize = top.cb()->frame_size();
2702 patch_pd(top, sp + fsize);
2703 }
2704
2705 if (preempt_kind == Continuation::object_wait) {
2706 // Check now if we need to throw IE exception.
2707 bool throw_ie = _thread->pending_interrupted_exception();
2708 if (throw_ie) {
2709 throw_interrupted_exception(_thread, top);
2710 _thread->set_pending_interrupted_exception(false);
2711 }
2712 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT" after preemption on Object.wait%s", tid, throw_ie ? "(throwing IE)" : "");
2713 } else if (preempt_kind == Continuation::monitorenter) {
2714 if (top.is_runtime_frame()) {
2715 // The continuation might now run on a different platform thread than the previous time so
2716 // we need to adjust the current thread saved in the stub frame before restoring registers.
2717 JavaThread** thread_addr = frame::saved_thread_address(top);
2718 if (thread_addr != nullptr) *thread_addr = _thread;
2719 }
2720 log_develop_trace(continuations, preempt)("Resuming " INT64_FORMAT " after preemption on monitorenter", tid);
2721 } else {
2722 // We need to redo the original call into the VM. First though, we need
2723 // to exit the monitor we just acquired (except on preemption cancelled
2724 // case where it was already released).
2725 assert(preempt_kind == Continuation::object_locker, "");
2726 if (_init_lock != nullptr) _init_lock->exit(_thread);
2727 sp = redo_vmcall(_thread, top);
2728 }
2729 return sp;
2730 }
2731
2732 intptr_t* ThawBase::redo_vmcall(JavaThread* current, frame& top) {
2733 assert(!current->preempting(), "");
2734 NOT_PRODUCT(int64_t tid = current->monitor_owner_id();)
2735 intptr_t* sp = top.sp();
2736
2737 {
2738 HandleMarkCleaner hmc(current); // Cleanup all handles (including so._conth) before returning to Java.
2739 ContinuationWrapper::SafepointOp so(current, _cont);
2740 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2741
2742 Method* m = top.interpreter_frame_method();
2743 Bytecode current_bytecode = Bytecode(m, top.interpreter_frame_bcp());
2744 Bytecodes::Code code = current_bytecode.code();
2745 log_develop_trace(continuations, preempt)("Redoing InterpreterRuntime::%s for " INT64_FORMAT, code == Bytecodes::Code::_new ? "_new" : "resolve_from_cache", tid);
2746
2747 // These InterpreterRuntime entry points use JRT_ENTRY which uses a HandleMarkCleaner.
2748 // Create a HandeMark to avoid destroying so._conth.
2749 HandleMark hm(current);
2750 DEBUG_ONLY(JavaThread::AtRedoVMCall apvmc(current);)
2751 if (code == Bytecodes::Code::_new) {
2752 InterpreterRuntime::_new(current, m->constants(), current_bytecode.get_index_u2(code));
2753 } else {
2754 InterpreterRuntime::resolve_from_cache(current, code);
2755 }
2756 }
2757
2758 if (current->preempting()) {
2759 // Preempted again so we just arrange to return to preempt stub to unmount.
2760 sp = push_preempt_adapter();
2761 current->set_preempt_alternate_return(nullptr);
2762 bool cancelled = current->preemption_cancelled();
2763 if (cancelled) {
2764 // Since preemption was cancelled, the thread will call thaw again from the preempt
2765 // stub. These retries could happen several times due to contention on the init_lock,
2766 // so just let the vthread umount to give a chance for other vthreads to run.
2767 current->set_preemption_cancelled(false);
2768 oop vthread = current->vthread();
2769 assert(java_lang_VirtualThread::state(vthread) == java_lang_VirtualThread::RUNNING, "wrong state for vthread");
2770 java_lang_VirtualThread::set_state(vthread, java_lang_VirtualThread::YIELDING);
2771 #if INCLUDE_JVMTI
2772 if (current->contended_entered_monitor() != nullptr) {
2773 current->set_contended_entered_monitor(nullptr);
2774 }
2775 #endif
2776 }
2777 log_develop_trace(continuations, preempt)("Preempted " INT64_FORMAT " again%s", tid, cancelled ? "(preemption cancelled, setting state to YIELDING)" : "");
2778 } else {
2779 log_develop_trace(continuations, preempt)("Call succesful, resuming " INT64_FORMAT, tid);
2780 }
2781 return sp;
2782 }
2783
2784 void ThawBase::throw_interrupted_exception(JavaThread* current, frame& top) {
2785 HandleMarkCleaner hm(current); // Cleanup all handles (including so._conth) before returning to Java.
2786 ContinuationWrapper::SafepointOp so(current, _cont);
2787 AnchorMark am(current, top); // Set the anchor so that the stack is walkable.
2788 JRT_BLOCK
2789 THROW(vmSymbols::java_lang_InterruptedException());
2790 JRT_BLOCK_END
2791 }
2792
2793 NOINLINE void ThawBase::recurse_thaw_interpreted_frame(const frame& hf, frame& caller, int num_frames, bool is_top) {
2794 assert(hf.is_interpreted_frame(), "");
2795
2796 if (UNLIKELY(seen_by_gc())) {
2797 if (is_top && _process_args_at_top) {
2798 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_with_args());
2799 } else {
2800 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2801 }
2802 }
2803
2804 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::InterpretedFrame>(caller, num_frames);
2805
2806 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2807
2808 _align_size += frame::align_wiggle; // possible added alignment for internal interpreted frame alignment om AArch64
2809
2810 frame f = new_stack_frame<ContinuationHelper::InterpretedFrame>(hf, caller, is_bottom_frame);
2811
2812 intptr_t* const stack_frame_top = f.sp() + frame::metadata_words_at_top;
2813 intptr_t* const stack_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(f);
2814 intptr_t* const heap_frame_top = hf.unextended_sp() + frame::metadata_words_at_top;
2815 intptr_t* const heap_frame_bottom = ContinuationHelper::InterpretedFrame::frame_bottom(hf);
2816
2817 assert(hf.is_heap_frame(), "should be");
2818 assert(!f.is_heap_frame(), "should not be");
2819
2820 const int fsize = pointer_delta_as_int(heap_frame_bottom, heap_frame_top);
2821 assert((stack_frame_bottom == stack_frame_top + fsize), "");
2822
2823 // Some architectures (like AArch64/PPC64/RISC-V) add padding between the locals and the fixed_frame to keep the fp 16-byte-aligned.
2824 // On those architectures we freeze the padding in order to keep the same fp-relative offsets in the fixed_frame.
2825 copy_from_chunk(heap_frame_top, stack_frame_top, fsize);
2826
2827 // Make sure the relativized locals is already set.
2828 assert(f.interpreter_frame_local_at(0) == stack_frame_bottom - 1, "invalid frame bottom");
2829
2830 derelativize_interpreted_frame_metadata(hf, f);
2831 patch(f, caller, is_bottom_frame);
2832
2833 assert(f.is_interpreted_frame_valid(_cont.thread()), "invalid thawed frame");
2834 assert(stack_frame_bottom <= ContinuationHelper::Frame::frame_top(caller), "");
2835
2836 CONT_JFR_ONLY(_jfr_info.record_interpreted_frame();)
2837
2838 maybe_set_fastpath(f.sp());
2839
2840 Method* m = hf.interpreter_frame_method();
2841 assert(!m->is_native() || !is_bottom_frame, "should be top frame of thaw_top case; missing caller frame");
2842 const int locals = m->max_locals();
2843
2844 if (!is_bottom_frame) {
2845 // can only fix caller once this frame is thawed (due to callee saved regs)
2846 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2847 } else if (_cont.tail()->has_bitmap() && locals > 0) {
2848 assert(hf.is_heap_frame(), "should be");
2849 address start = (address)(heap_frame_bottom - locals);
2850 address end = (address)heap_frame_bottom;
2851 clear_bitmap_bits(start, end);
2852 }
2853
2854 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2855 caller = f;
2856 }
2857
2858 void ThawBase::recurse_thaw_compiled_frame(const frame& hf, frame& caller, int num_frames, bool stub_caller) {
2859 assert(hf.is_compiled_frame(), "");
2860 assert(_preempted_case || !stub_caller, "stub caller not at preemption");
2861
2862 if (!stub_caller && UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2863 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2864 }
2865
2866 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::CompiledFrame>(caller, num_frames);
2867
2868 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
2869
2870 assert(caller.sp() == caller.unextended_sp(), "");
2871
2872 if ((!is_bottom_frame && caller.is_interpreted_frame()) || (is_bottom_frame && Interpreter::contains(_cont.tail()->pc()))) {
2873 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_compiled_frame
2874 }
2875
2876 int fsize = 0;
2877 int added_argsize = 0;
2878 bool augmented = hf.was_augmented_on_entry(fsize);
2879 if (!augmented) {
2880 added_argsize = (is_bottom_frame || caller.is_interpreted_frame()) ? hf.compiled_frame_stack_argsize() : 0;
2881 fsize += added_argsize;
2882 }
2883 assert(!is_bottom_frame || !augmented, "");
2884
2885
2886 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
2887 // yet laid out in the stack, and so the original_pc is not stored in it.
2888 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
2889 frame f = new_stack_frame<ContinuationHelper::CompiledFrame>(hf, caller, is_bottom_frame, augmented ? fsize - hf.cb()->frame_size() : 0);
2890 assert(f.cb()->frame_size() == (int)(caller.sp() - f.sp()), "");
2891
2892 intptr_t* const stack_frame_top = f.sp();
2893 intptr_t* const heap_frame_top = hf.unextended_sp();
2894 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
2895 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
2896 // copy metadata, except the metadata at the top of the (unextended) entry frame
2897 int sz = fsize + frame::metadata_words_at_bottom + (is_bottom_frame && added_argsize == 0 ? 0 : frame::metadata_words_at_top);
2898
2899 // If we're the bottom-most thawed frame, we're writing to within one word from entrySP
2900 // (we might have one padding word for alignment)
2901 assert(!is_bottom_frame || (_cont.entrySP() - 1 <= to + sz && to + sz <= _cont.entrySP()), "");
2902 assert(!is_bottom_frame || hf.compiled_frame_stack_argsize() != 0 || (to + sz && to + sz == _cont.entrySP()), "");
2903
2904 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
2905
2906 patch(f, caller, is_bottom_frame, augmented);
2907
2908 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
2909 assert(!f.is_deoptimized_frame(), "");
2910 if (hf.is_deoptimized_frame()) {
2911 maybe_set_fastpath(f.sp());
2912 } else if (_thread->is_interp_only_mode()
2913 || (stub_caller && f.cb()->as_nmethod()->is_marked_for_deoptimization())) {
2914 // The caller of the safepoint stub when the continuation is preempted is not at a call instruction, and so
2915 // cannot rely on nmethod patching for deopt.
2916 assert(_thread->is_interp_only_mode() || stub_caller, "expected a stub-caller");
2917
2918 log_develop_trace(continuations)("Deoptimizing thawed frame");
2919 DEBUG_ONLY(ContinuationHelper::Frame::patch_pc(f, nullptr));
2920
2921 f.deoptimize(nullptr); // the null thread simply avoids the assertion in deoptimize which we're not set up for
2922 assert(f.is_deoptimized_frame(), "");
2923 assert(ContinuationHelper::Frame::is_deopt_return(f.raw_pc(), f), "");
2924 maybe_set_fastpath(f.sp());
2925 }
2926
2927 if (!is_bottom_frame) {
2928 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
2929 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
2930 } else if (_cont.tail()->has_bitmap() && added_argsize > 0) {
2931 address start = (address)(heap_frame_top + ContinuationHelper::CompiledFrame::size(hf) + frame::metadata_words_at_top);
2932 int stack_args_slots = f.cb()->as_nmethod()->num_stack_arg_slots(false /* rounded */);
2933 int argsize_in_bytes = stack_args_slots * VMRegImpl::stack_slot_size;
2934 clear_bitmap_bits(start, start + argsize_in_bytes);
2935 }
2936
2937 DEBUG_ONLY(after_thaw_java_frame(f, is_bottom_frame);)
2938 caller = f;
2939 }
2940
2941 void ThawBase::recurse_thaw_stub_frame(const frame& hf, frame& caller, int num_frames) {
2942 DEBUG_ONLY(_frames++;)
2943
2944 if (UNLIKELY(seen_by_gc())) {
2945 // Process the stub's caller here since we might need the full map.
2946 RegisterMap map(nullptr,
2947 RegisterMap::UpdateMap::include,
2948 RegisterMap::ProcessFrames::skip,
2949 RegisterMap::WalkContinuation::skip);
2950 map.set_include_argument_oops(false);
2951 _stream.next(&map);
2952 assert(!_stream.is_done(), "");
2953 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, &map);
2954 } else {
2955 _stream.next(SmallRegisterMap::instance_no_args());
2956 assert(!_stream.is_done(), "");
2957 }
2958
2959 recurse_thaw_compiled_frame(_stream.to_frame(), caller, num_frames, true);
2960
2961 assert(caller.is_compiled_frame(), "");
2962 assert(caller.sp() == caller.unextended_sp(), "");
2963
2964 DEBUG_ONLY(before_thaw_java_frame(hf, caller, false /*is_bottom_frame*/, num_frames);)
2965
2966 frame f = new_stack_frame<ContinuationHelper::StubFrame>(hf, caller, false);
2967 intptr_t* stack_frame_top = f.sp();
2968 intptr_t* heap_frame_top = hf.sp();
2969 int fsize = ContinuationHelper::StubFrame::size(hf);
2970
2971 copy_from_chunk(heap_frame_top - frame::metadata_words, stack_frame_top - frame::metadata_words,
2972 fsize + frame::metadata_words);
2973
2974 patch(f, caller, false /*is_bottom_frame*/);
2975
2976 // can only fix caller once this frame is thawed (due to callee saved regs)
2977 RegisterMap map(nullptr,
2978 RegisterMap::UpdateMap::include,
2979 RegisterMap::ProcessFrames::skip,
2980 RegisterMap::WalkContinuation::skip);
2981 map.set_include_argument_oops(false);
2982 f.oop_map()->update_register_map(&f, &map);
2983 ContinuationHelper::update_register_map_with_callee(caller, &map);
2984 _cont.tail()->fix_thawed_frame(caller, &map);
2985
2986 DEBUG_ONLY(after_thaw_java_frame(f, false /*is_bottom_frame*/);)
2987 caller = f;
2988 }
2989
2990 void ThawBase::recurse_thaw_native_frame(const frame& hf, frame& caller, int num_frames) {
2991 assert(hf.is_native_frame(), "");
2992 assert(_preempted_case && hf.cb()->as_nmethod()->method()->is_object_wait0(), "");
2993
2994 if (UNLIKELY(seen_by_gc())) { // recurse_thaw_stub_frame already invoked our barriers with a full regmap
2995 _cont.tail()->do_barriers<stackChunkOopDesc::BarrierType::Store>(_stream, SmallRegisterMap::instance_no_args());
2996 }
2997
2998 const bool is_bottom_frame = recurse_thaw_java_frame<ContinuationHelper::NativeFrame>(caller, num_frames);
2999 assert(!is_bottom_frame, "");
3000
3001 DEBUG_ONLY(before_thaw_java_frame(hf, caller, is_bottom_frame, num_frames);)
3002
3003 assert(caller.sp() == caller.unextended_sp(), "");
3004
3005 if (caller.is_interpreted_frame()) {
3006 _align_size += frame::align_wiggle; // we add one whether or not we've aligned because we add it in recurse_freeze_native_frame
3007 }
3008
3009 // new_stack_frame must construct the resulting frame using hf.pc() rather than hf.raw_pc() because the frame is not
3010 // yet laid out in the stack, and so the original_pc is not stored in it.
3011 // As a result, f.is_deoptimized_frame() is always false and we must test hf to know if the frame is deoptimized.
3012 frame f = new_stack_frame<ContinuationHelper::NativeFrame>(hf, caller, false /* bottom */);
3013 intptr_t* const stack_frame_top = f.sp();
3014 intptr_t* const heap_frame_top = hf.unextended_sp();
3015
3016 int fsize = ContinuationHelper::NativeFrame::size(hf);
3017 assert(fsize <= (int)(caller.unextended_sp() - f.unextended_sp()), "");
3018
3019 intptr_t* from = heap_frame_top - frame::metadata_words_at_bottom;
3020 intptr_t* to = stack_frame_top - frame::metadata_words_at_bottom;
3021 int sz = fsize + frame::metadata_words_at_bottom;
3022
3023 copy_from_chunk(from, to, sz); // copying good oops because we invoked barriers above
3024
3025 patch(f, caller, false /* bottom */);
3026
3027 // f.is_deoptimized_frame() is always false and we must test hf.is_deoptimized_frame() (see comment above)
3028 assert(!f.is_deoptimized_frame(), "");
3029 assert(!hf.is_deoptimized_frame(), "");
3030 assert(!f.cb()->as_nmethod()->is_marked_for_deoptimization(), "");
3031
3032 // can only fix caller once this frame is thawed (due to callee saved regs); this happens on the stack
3033 _cont.tail()->fix_thawed_frame(caller, SmallRegisterMap::instance_no_args());
3034
3035 DEBUG_ONLY(after_thaw_java_frame(f, false /* bottom */);)
3036 caller = f;
3037 }
3038
3039 void ThawBase::finish_thaw(frame& f) {
3040 stackChunkOop chunk = _cont.tail();
3041
3042 if (chunk->is_empty()) {
3043 // Only remove chunk from list if it can't be reused for another freeze
3044 if (seen_by_gc()) {
3045 _cont.set_tail(chunk->parent());
3046 } else {
3047 chunk->set_has_mixed_frames(false);
3048 }
3049 chunk->set_max_thawing_size(0);
3050 } else {
3051 chunk->set_max_thawing_size(chunk->max_thawing_size() - _align_size);
3052 }
3053 assert(chunk->is_empty() == (chunk->max_thawing_size() == 0), "");
3054
3055 if (!is_aligned(f.sp(), frame::frame_alignment)) {
3056 assert(f.is_interpreted_frame(), "");
3057 f.set_sp(align_down(f.sp(), frame::frame_alignment));
3058 }
3059 push_return_frame(f);
3060 // can only fix caller after push_return_frame (due to callee saved regs)
3061 if (_process_args_at_top) {
3062 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_with_args());
3063 } else {
3064 chunk->fix_thawed_frame(f, SmallRegisterMap::instance_no_args());
3065 }
3066
3067 assert(_cont.is_empty() == _cont.last_frame().is_empty(), "");
3068
3069 log_develop_trace(continuations)("thawed %d frames", _frames);
3070
3071 LogTarget(Trace, continuations) lt;
3072 if (lt.develop_is_enabled()) {
3073 LogStream ls(lt);
3074 ls.print_cr("top hframe after (thaw):");
3075 _cont.last_frame().print_value_on(&ls);
3076 }
3077 }
3078
3079 void ThawBase::push_return_frame(const frame& f) { // see generate_cont_thaw
3080 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == f.cb()->as_nmethod()->is_deopt_pc(f.raw_pc()), "");
3081 assert(!f.is_compiled_frame() || f.is_deoptimized_frame() == (f.pc() != f.raw_pc()), "");
3082
3083 LogTarget(Trace, continuations) lt;
3084 if (lt.develop_is_enabled()) {
3085 LogStream ls(lt);
3086 ls.print_cr("push_return_frame");
3087 f.print_value_on(&ls);
3088 }
3089
3090 assert(f.sp() - frame::metadata_words_at_bottom >= _top_stack_address, "overwrote past thawing space"
3091 " to: " INTPTR_FORMAT " top_address: " INTPTR_FORMAT, p2i(f.sp() - frame::metadata_words), p2i(_top_stack_address));
3092 ContinuationHelper::Frame::patch_pc(f, f.raw_pc()); // in case we want to deopt the frame in a full transition, this is checked.
3093 ContinuationHelper::push_pd(f);
3094
3095 assert(ContinuationHelper::Frame::assert_frame_laid_out(f), "");
3096 }
3097
3098 // returns new top sp
3099 // called after preparations (stack overflow check and making room)
3100 template<typename ConfigT>
3101 static inline intptr_t* thaw_internal(JavaThread* thread, const Continuation::thaw_kind kind) {
3102 assert(thread == JavaThread::current(), "Must be current thread");
3103
3104 CONT_JFR_ONLY(EventContinuationThaw event;)
3105
3106 log_develop_trace(continuations)("~~~~ thaw kind: %d sp: " INTPTR_FORMAT, kind, p2i(thread->last_continuation()->entry_sp()));
3107
3108 ContinuationEntry* entry = thread->last_continuation();
3109 assert(entry != nullptr, "");
3110 oop oopCont = entry->cont_oop(thread);
3111
3112 assert(!jdk_internal_vm_Continuation::done(oopCont), "");
3113 assert(oopCont == get_continuation(thread), "");
3114 verify_continuation(oopCont);
3115
3116 assert(entry->is_virtual_thread() == (entry->scope(thread) == java_lang_VirtualThread::vthread_scope()), "");
3117
3118 ContinuationWrapper cont(thread, oopCont);
3119 log_develop_debug(continuations)("THAW #" INTPTR_FORMAT " " INTPTR_FORMAT, cont.hash(), p2i((oopDesc*)oopCont));
3120
3121 #ifdef ASSERT
3122 set_anchor_to_entry(thread, cont.entry());
3123 log_frames(thread);
3124 clear_anchor(thread);
3125 #endif
3126
3127 Thaw<ConfigT> thw(thread, cont);
3128 intptr_t* const sp = thw.thaw(kind);
3129 assert(is_aligned(sp, frame::frame_alignment), "");
3130 DEBUG_ONLY(log_frames_after_thaw(thread, cont, sp);)
3131
3132 CONT_JFR_ONLY(thw.jfr_info().post_jfr_event(&event, cont.continuation(), thread);)
3133
3134 verify_continuation(cont.continuation());
3135 log_develop_debug(continuations)("=== End of thaw #" INTPTR_FORMAT, cont.hash());
3136
3137 return sp;
3138 }
3139
3140 #ifdef ASSERT
3141 static void do_deopt_after_thaw(JavaThread* thread) {
3142 int i = 0;
3143 StackFrameStream fst(thread, true, false);
3144 fst.register_map()->set_include_argument_oops(false);
3145 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3146 for (; !fst.is_done(); fst.next()) {
3147 if (fst.current()->cb()->is_nmethod()) {
3148 nmethod* nm = fst.current()->cb()->as_nmethod();
3149 if (!nm->method()->is_continuation_native_intrinsic()) {
3150 nm->make_deoptimized();
3151 }
3152 }
3153 }
3154 }
3155
3156 class ThawVerifyOopsClosure: public OopClosure {
3157 intptr_t* _p;
3158 outputStream* _st;
3159 bool is_good_oop(oop o) {
3160 return dbg_is_safe(o, -1) && dbg_is_safe(o->klass(), -1) && oopDesc::is_oop(o) && o->klass()->is_klass();
3161 }
3162 public:
3163 ThawVerifyOopsClosure(outputStream* st) : _p(nullptr), _st(st) {}
3164 intptr_t* p() { return _p; }
3165 void reset() { _p = nullptr; }
3166
3167 virtual void do_oop(oop* p) {
3168 oop o = *p;
3169 if (o == nullptr || is_good_oop(o)) {
3170 return;
3171 }
3172 _p = (intptr_t*)p;
3173 _st->print_cr("*** non-oop " PTR_FORMAT " found at " PTR_FORMAT, p2i(*p), p2i(p));
3174 }
3175 virtual void do_oop(narrowOop* p) {
3176 oop o = RawAccess<>::oop_load(p);
3177 if (o == nullptr || is_good_oop(o)) {
3178 return;
3179 }
3180 _p = (intptr_t*)p;
3181 _st->print_cr("*** (narrow) non-oop %x found at " PTR_FORMAT, (int)(*p), p2i(p));
3182 }
3183 };
3184
3185 static bool do_verify_after_thaw(JavaThread* thread, stackChunkOop chunk, outputStream* st) {
3186 assert(thread->has_last_Java_frame(), "");
3187
3188 ResourceMark rm;
3189 ThawVerifyOopsClosure cl(st);
3190 NMethodToOopClosure cf(&cl, false);
3191
3192 StackFrameStream fst(thread, true, false);
3193 fst.register_map()->set_include_argument_oops(false);
3194 ContinuationHelper::update_register_map_with_callee(*fst.current(), fst.register_map());
3195 for (; !fst.is_done() && !Continuation::is_continuation_enterSpecial(*fst.current()); fst.next()) {
3196 if (fst.current()->cb()->is_nmethod() && fst.current()->cb()->as_nmethod()->is_marked_for_deoptimization()) {
3197 st->print_cr(">>> do_verify_after_thaw deopt");
3198 fst.current()->deoptimize(nullptr);
3199 fst.current()->print_on(st);
3200 }
3201
3202 fst.current()->oops_do(&cl, &cf, fst.register_map());
3203 if (cl.p() != nullptr) {
3204 frame fr = *fst.current();
3205 st->print_cr("Failed for frame barriers: %d",chunk->requires_barriers());
3206 fr.print_on(st);
3207 if (!fr.is_interpreted_frame()) {
3208 st->print_cr("size: %d argsize: %d",
3209 ContinuationHelper::NonInterpretedUnknownFrame::size(fr),
3210 ContinuationHelper::NonInterpretedUnknownFrame::stack_argsize(fr));
3211 }
3212 VMReg reg = fst.register_map()->find_register_spilled_here(cl.p(), fst.current()->sp());
3213 if (reg != nullptr) {
3214 st->print_cr("Reg %s %d", reg->name(), reg->is_stack() ? (int)reg->reg2stack() : -99);
3215 }
3216 cl.reset();
3217 DEBUG_ONLY(thread->print_frame_layout();)
3218 if (chunk != nullptr) {
3219 chunk->print_on(true, st);
3220 }
3221 return false;
3222 }
3223 }
3224 return true;
3225 }
3226
3227 static void log_frames(JavaThread* thread) {
3228 const static int show_entry_callers = 3;
3229 LogTarget(Trace, continuations) lt;
3230 if (!lt.develop_is_enabled()) {
3231 return;
3232 }
3233 LogStream ls(lt);
3234
3235 ls.print_cr("------- frames --------- for thread " INTPTR_FORMAT, p2i(thread));
3236 if (!thread->has_last_Java_frame()) {
3237 ls.print_cr("NO ANCHOR!");
3238 }
3239
3240 RegisterMap map(thread,
3241 RegisterMap::UpdateMap::include,
3242 RegisterMap::ProcessFrames::include,
3243 RegisterMap::WalkContinuation::skip);
3244 map.set_include_argument_oops(false);
3245
3246 if (false) {
3247 for (frame f = thread->last_frame(); !f.is_entry_frame(); f = f.sender(&map)) {
3248 f.print_on(&ls);
3249 }
3250 } else {
3251 map.set_skip_missing(true);
3252 ResetNoHandleMark rnhm;
3253 ResourceMark rm;
3254 HandleMark hm(Thread::current());
3255 FrameValues values;
3256
3257 int i = 0;
3258 int post_entry = -1;
3259 for (frame f = thread->last_frame(); !f.is_first_frame(); f = f.sender(&map), i++) {
3260 f.describe(values, i, &map, i == 0);
3261 if (post_entry >= 0 || Continuation::is_continuation_enterSpecial(f))
3262 post_entry++;
3263 if (post_entry >= show_entry_callers)
3264 break;
3265 }
3266 values.print_on(thread, &ls);
3267 }
3268
3269 ls.print_cr("======= end frames =========");
3270 }
3271
3272 static void log_frames_after_thaw(JavaThread* thread, ContinuationWrapper& cont, intptr_t* sp) {
3273 intptr_t* sp0 = sp;
3274 address pc0 = *(address*)(sp - frame::sender_sp_ret_address_offset());
3275
3276 bool preempted = false;
3277 stackChunkOop tail = cont.tail();
3278 if (tail != nullptr && tail->preempted()) {
3279 // Still preempted (monitor not acquired) so no frames were thawed.
3280 set_anchor(thread, cont.entrySP(), cont.entryPC());
3281 preempted = true;
3282 } else {
3283 set_anchor(thread, sp0);
3284 }
3285
3286 log_frames(thread);
3287 if (LoomVerifyAfterThaw) {
3288 assert(do_verify_after_thaw(thread, cont.tail(), tty), "");
3289 }
3290 assert(preempted || ContinuationEntry::assert_entry_frame_laid_out(thread), "");
3291 clear_anchor(thread);
3292
3293 LogTarget(Trace, continuations) lt;
3294 if (lt.develop_is_enabled()) {
3295 LogStream ls(lt);
3296 ls.print_cr("Jumping to frame (thaw):");
3297 frame(sp).print_value_on(&ls);
3298 }
3299 }
3300 #endif // ASSERT
3301
3302 #include CPU_HEADER_INLINE(continuationFreezeThaw)
3303
3304 #ifdef ASSERT
3305 static void print_frame_layout(const frame& f, bool callee_complete, outputStream* st) {
3306 ResourceMark rm;
3307 FrameValues values;
3308 assert(f.get_cb() != nullptr, "");
3309 RegisterMap map(f.is_heap_frame() ?
3310 nullptr :
3311 JavaThread::current(),
3312 RegisterMap::UpdateMap::include,
3313 RegisterMap::ProcessFrames::skip,
3314 RegisterMap::WalkContinuation::skip);
3315 map.set_include_argument_oops(false);
3316 map.set_skip_missing(true);
3317 if (callee_complete) {
3318 frame::update_map_with_saved_link(&map, ContinuationHelper::Frame::callee_link_address(f));
3319 }
3320 const_cast<frame&>(f).describe(values, 0, &map, true);
3321 values.print_on(static_cast<JavaThread*>(nullptr), st);
3322 }
3323 #endif
3324
3325 static address thaw_entry = nullptr;
3326 static address freeze_entry = nullptr;
3327 static address freeze_preempt_entry = nullptr;
3328
3329 address Continuation::thaw_entry() {
3330 return ::thaw_entry;
3331 }
3332
3333 address Continuation::freeze_entry() {
3334 return ::freeze_entry;
3335 }
3336
3337 address Continuation::freeze_preempt_entry() {
3338 return ::freeze_preempt_entry;
3339 }
3340
3341 class ConfigResolve {
3342 public:
3343 static void resolve() { resolve_compressed(); }
3344
3345 static void resolve_compressed() {
3346 UseCompressedOops ? resolve_gc<true>()
3347 : resolve_gc<false>();
3348 }
3349
3350 private:
3351 template <bool use_compressed>
3352 static void resolve_gc() {
3353 BarrierSet* bs = BarrierSet::barrier_set();
3354 assert(bs != nullptr, "freeze/thaw invoked before BarrierSet is set");
3355 switch (bs->kind()) {
3356 #define BARRIER_SET_RESOLVE_BARRIER_CLOSURE(bs_name) \
3357 case BarrierSet::bs_name: { \
3358 resolve<use_compressed, typename BarrierSet::GetType<BarrierSet::bs_name>::type>(); \
3359 } \
3360 break;
3361 FOR_EACH_BARRIER_SET_DO(BARRIER_SET_RESOLVE_BARRIER_CLOSURE)
3362 #undef BARRIER_SET_RESOLVE_BARRIER_CLOSURE
3363
3364 default:
3365 fatal("BarrierSet resolving not implemented");
3366 };
3367 }
3368
3369 template <bool use_compressed, typename BarrierSetT>
3370 static void resolve() {
3371 typedef Config<use_compressed ? oop_kind::NARROW : oop_kind::WIDE, BarrierSetT> SelectedConfigT;
3372
3373 freeze_entry = (address)freeze<SelectedConfigT>;
3374 freeze_preempt_entry = (address)SelectedConfigT::freeze_preempt;
3375
3376 // If we wanted, we could templatize by kind and have three different thaw entries
3377 thaw_entry = (address)thaw<SelectedConfigT>;
3378 }
3379 };
3380
3381 void Continuation::init() {
3382 ConfigResolve::resolve();
3383 }