Change ClassFileLoadHook to lazily compute dex file

Creating a dex file from the quickened or compact-dex'd internal
format for calling the JVMTI ClassFileLoadHook is quite expensive.
This meant that agents could not generally listen for this event
without causing unacceptable performance problems.

Since agents will generally not touch the buffer, needing to
instrument only a handful of classes we will fix this problem by doing
the de-quickening lazily. This is done by mmaping an empty buffer with
PROT_NONE and then filling it in when the program has a SEGV in the
appropriate address range. This should improve the performance of any
agent that listens for the ClassFileLoadHook but does not commonly do
anything to the buffer.

This does have the disadvantage that we can no longer ensure that the
buffer size we pass down in class_data_len might no longer be fully
accurate. Some agents that assert that class_data_len is exactly the
same as the dex-file will need to be updated.

Bug: 72402467
Bug: 72064989
Test: ./test.py --host -j50
Test: ./test.py --host --redefine-stress -j50

Change-Id: I39354837f1417ae10a57c5b42cbb4f38f8a563dc
diff --git a/openjdkjvmti/transform.cc b/openjdkjvmti/transform.cc
index 8445eca..d98b385 100644
--- a/openjdkjvmti/transform.cc
+++ b/openjdkjvmti/transform.cc
@@ -29,6 +29,9 @@
  * questions.
  */
 
+#include <stddef.h>
+#include <sys/types.h>
+
 #include <unordered_map>
 #include <unordered_set>
 
@@ -40,6 +43,7 @@
 #include "dex/dex_file.h"
 #include "dex/dex_file_types.h"
 #include "events-inl.h"
+#include "fault_handler.h"
 #include "gc_root-inl.h"
 #include "globals.h"
 #include "jni_env_ext-inl.h"
@@ -63,6 +67,174 @@
 
 namespace openjdkjvmti {
 
+// A FaultHandler that will deal with initializing ClassDefinitions when they are actually needed.
+class TransformationFaultHandler FINAL : public art::FaultHandler {
+ public:
+  explicit TransformationFaultHandler(art::FaultManager* manager)
+      : art::FaultHandler(manager),
+        uninitialized_class_definitions_lock_("JVMTI Initialized class definitions lock",
+                                              art::LockLevel::kSignalHandlingLock),
+        class_definition_initialized_cond_("JVMTI Initialized class definitions condition",
+                                           uninitialized_class_definitions_lock_) {
+    manager->AddHandler(this, /* generated_code */ false);
+  }
+
+  ~TransformationFaultHandler() {
+    art::MutexLock mu(art::Thread::Current(), uninitialized_class_definitions_lock_);
+    uninitialized_class_definitions_.clear();
+  }
+
+  bool Action(int sig, siginfo_t* siginfo, void* context ATTRIBUTE_UNUSED) OVERRIDE {
+    DCHECK_EQ(sig, SIGSEGV);
+    art::Thread* self = art::Thread::Current();
+    if (UNLIKELY(uninitialized_class_definitions_lock_.IsExclusiveHeld(self))) {
+      if (self != nullptr) {
+        LOG(FATAL) << "Recursive call into Transformation fault handler!";
+        UNREACHABLE();
+      } else {
+        LOG(ERROR) << "Possible deadlock due to recursive signal delivery of segv.";
+      }
+    }
+    uintptr_t ptr = reinterpret_cast<uintptr_t>(siginfo->si_addr);
+    ArtClassDefinition* res = nullptr;
+
+    {
+      // NB Technically using a mutex and condition variables here is non-posix compliant but
+      // everything should be fine since both glibc and bionic implementations of mutexs and
+      // condition variables work fine so long as the thread was not interrupted during a
+      // lock/unlock (which it wasn't) on all architectures we care about.
+      art::MutexLock mu(self, uninitialized_class_definitions_lock_);
+      auto it = std::find_if(uninitialized_class_definitions_.begin(),
+                             uninitialized_class_definitions_.end(),
+                             [&](const auto op) { return op->ContainsAddress(ptr); });
+      if (it != uninitialized_class_definitions_.end()) {
+        res = *it;
+        // Remove the class definition.
+        uninitialized_class_definitions_.erase(it);
+        // Put it in the initializing list
+        initializing_class_definitions_.push_back(res);
+      } else {
+        // Wait for the ptr to be initialized (if it is currently initializing).
+        while (DefinitionIsInitializing(ptr)) {
+          WaitForClassInitializationToFinish();
+        }
+        // Return true (continue with user code) if we find that the definition has been
+        // initialized. Return false (continue on to next signal handler) if the definition is not
+        // initialized or found.
+        return std::find_if(initialized_class_definitions_.begin(),
+                            initialized_class_definitions_.end(),
+                            [&](const auto op) { return op->ContainsAddress(ptr); }) !=
+            uninitialized_class_definitions_.end();
+      }
+    }
+
+    VLOG(signals) << "Lazy initialization of dex file for transformation of " << res->GetName()
+                  << " during SEGV";
+    res->InitializeMemory();
+
+    {
+      art::MutexLock mu(self, uninitialized_class_definitions_lock_);
+      // Move to initialized state and notify waiters.
+      initializing_class_definitions_.erase(std::find(initializing_class_definitions_.begin(),
+                                                      initializing_class_definitions_.end(),
+                                                      res));
+      initialized_class_definitions_.push_back(res);
+      class_definition_initialized_cond_.Broadcast(self);
+    }
+
+    return true;
+  }
+
+  void RemoveDefinition(ArtClassDefinition* def) REQUIRES(!uninitialized_class_definitions_lock_) {
+    art::MutexLock mu(art::Thread::Current(), uninitialized_class_definitions_lock_);
+    auto it = std::find(uninitialized_class_definitions_.begin(),
+                        uninitialized_class_definitions_.end(),
+                        def);
+    if (it != uninitialized_class_definitions_.end()) {
+      uninitialized_class_definitions_.erase(it);
+      return;
+    }
+    while (std::find(initializing_class_definitions_.begin(),
+                     initializing_class_definitions_.end(),
+                     def) != initializing_class_definitions_.end()) {
+      WaitForClassInitializationToFinish();
+    }
+    it = std::find(initialized_class_definitions_.begin(),
+                   initialized_class_definitions_.end(),
+                   def);
+    CHECK(it != initialized_class_definitions_.end()) << "Could not find class definition for "
+                                                      << def->GetName();
+    initialized_class_definitions_.erase(it);
+  }
+
+  void AddArtDefinition(ArtClassDefinition* def) REQUIRES(!uninitialized_class_definitions_lock_) {
+    DCHECK(def->IsLazyDefinition());
+    art::MutexLock mu(art::Thread::Current(), uninitialized_class_definitions_lock_);
+    uninitialized_class_definitions_.push_back(def);
+  }
+
+ private:
+  bool DefinitionIsInitializing(uintptr_t ptr) REQUIRES(uninitialized_class_definitions_lock_) {
+    return std::find_if(initializing_class_definitions_.begin(),
+                        initializing_class_definitions_.end(),
+                        [&](const auto op) { return op->ContainsAddress(ptr); }) !=
+        initializing_class_definitions_.end();
+  }
+
+  void WaitForClassInitializationToFinish() REQUIRES(uninitialized_class_definitions_lock_) {
+    class_definition_initialized_cond_.Wait(art::Thread::Current());
+  }
+
+  art::Mutex uninitialized_class_definitions_lock_ ACQUIRED_BEFORE(art::Locks::abort_lock_);
+  art::ConditionVariable class_definition_initialized_cond_
+      GUARDED_BY(uninitialized_class_definitions_lock_);
+
+  // A list of the class definitions that have a non-readable map.
+  std::vector<ArtClassDefinition*> uninitialized_class_definitions_
+      GUARDED_BY(uninitialized_class_definitions_lock_);
+
+  // A list of class definitions that are currently undergoing unquickening. Threads should wait
+  // until the definition is no longer in this before returning.
+  std::vector<ArtClassDefinition*> initializing_class_definitions_
+      GUARDED_BY(uninitialized_class_definitions_lock_);
+
+  // A list of class definitions that are already unquickened. Threads should immediately return if
+  // it is here.
+  std::vector<ArtClassDefinition*> initialized_class_definitions_
+      GUARDED_BY(uninitialized_class_definitions_lock_);
+};
+
+static TransformationFaultHandler* gTransformFaultHandler = nullptr;
+
+void Transformer::Setup() {
+  // Although we create this the fault handler is actually owned by the 'art::fault_manager' which
+  // will take care of destroying it.
+  if (art::MemMap::kCanReplaceMapping && ArtClassDefinition::kEnableOnDemandDexDequicken) {
+    gTransformFaultHandler = new TransformationFaultHandler(&art::fault_manager);
+  }
+}
+
+// Simple helper to add and remove the class definition from the fault handler.
+class ScopedDefinitionHandler {
+ public:
+  explicit ScopedDefinitionHandler(ArtClassDefinition* def)
+      : def_(def), is_lazy_(def_->IsLazyDefinition()) {
+    if (is_lazy_) {
+      gTransformFaultHandler->AddArtDefinition(def_);
+    }
+  }
+
+  ~ScopedDefinitionHandler() {
+    if (is_lazy_) {
+      gTransformFaultHandler->RemoveDefinition(def_);
+    }
+  }
+
+ private:
+  ArtClassDefinition* def_;
+  bool is_lazy_;
+};
+
 // Initialize templates.
 template
 void Transformer::TransformSingleClassDirect<ArtJvmtiEvent::kClassFileLoadHookNonRetransformable>(
@@ -78,6 +250,7 @@
   static_assert(kEvent == ArtJvmtiEvent::kClassFileLoadHookNonRetransformable ||
                 kEvent == ArtJvmtiEvent::kClassFileLoadHookRetransformable,
                 "bad event type");
+  ScopedDefinitionHandler handler(def);
   jint new_len = -1;
   unsigned char* new_data = nullptr;
   art::ArrayRef<const unsigned char> dex_data = def->GetDexData();