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