Add prepareClone and atClone APIs for Verilated models (verilator#3503)…

… (verilator#4444)

This API is used if the user copies the process using `fork`
and similar OS-level mechanisms. The `at_clone` member function
ensures that all model-allocated resources are re-allocated, such
that the copied child process/model can simulate correctly.

A typical allocated resource is the thread pool, which every model
has its own pool.

docs/guide/connecting.rst

@@ -128,6 +128,43 @@ in the distribution.  These headers use Doxygen comments, `///` and `//<`,
 to indicate and document those functions that are part of the Verilated
 public API.
 
+Process-Level Clone APIs
+--------------------------
+
+Modern operating systems support process-level clone (a.k.a copying, forking)
+with system call interfaces in C/C++, e.g., :code:`fork()` in Linux.
+
+However, after cloning a parent process, some resources cannot be inherited
+in the child process. For example, in POSIX systems, when you fork a process,
+the child process inherits all the memory of the parent process. However,
+only the thread that called fork is replicated in the child process. Other
+threads are not.
+
+Therefore, to support the process-level clone mechanisms, Verilator supports
+:code:`prepareClone()` and :code:`atClone()` APIs to allow the user to manually
+re-construct the model in the child process. The two APIs handle all necessary
+resources required for releasing and re-initializing before and after cloning.
+
+The two APIs are supported in the verilated models. Here is an example of usage
+with Linux :code:`fork()` and :code:`pthread_atfork` APIs:
+
+.. code-block:: C++
+
+    // static function pointers to fit pthread_atfork
+    static auto prepareClone = [](){ topp->prepareClone(); };
+    static auto atClone = [](){ topp->atClone(); };
+
+    // in main function, register the handlers:
+    pthread_atfork(prepareClone, atClone, atClone);
+
+For better flexibility, you can also manually call the handlers before and
+after :code:`fork()`.
+
+With the process-level clone APIs, users can create process-level snapshots
+for the verilated models. While the Verilator save/restore option provides
+persistent and circuit-independent snapshots, the process-level clone APIs
+enable in-memory, circuit-transparent, and highly efficient snapshots.
+
 
 Direct Programming Interface (DPI)
 ==================================

include/verilated.cpp

@@ -2608,6 +2608,14 @@ VerilatedVirtualBase* VerilatedContext::threadPoolp() {
     return m_threadPool.get();
 }
 
+void VerilatedContext::prepareClone() { delete m_threadPool.release(); }
+
+VerilatedVirtualBase* VerilatedContext::threadPoolpOnClone() {
+    if (VL_UNLIKELY(m_threadPool)) m_threadPool.release();
+    m_threadPool = std::make_unique<VlThreadPool>(this, m_threads - 1);
+    return m_threadPool.get();
+}
+
 VerilatedVirtualBase*
 VerilatedContext::enableExecutionProfiler(VerilatedVirtualBase* (*construct)(VerilatedContext&)) {
     if (!m_executionProfiler) m_executionProfiler.reset(construct(*this));

include/verilated.h

@@ -568,6 +568,8 @@ class VerilatedContext VL_NOT_FINAL {
     void addModel(VerilatedModel*);
 
     VerilatedVirtualBase* threadPoolp();
+    void prepareClone();
+    VerilatedVirtualBase* threadPoolpOnClone();
     VerilatedVirtualBase*
     enableExecutionProfiler(VerilatedVirtualBase* (*construct)(VerilatedContext&));
 

src/V3EmitCModel.cpp

@@ -234,6 +234,12 @@ class EmitCModel final : public EmitCFunc {
         puts("const char* hierName() const override final;\n");
         puts("const char* modelName() const override final;\n");
         puts("unsigned threads() const override final;\n");
+        puts("/// Prepare for cloning the model at the process level (e.g. fork in Linux)\n");
+        puts("/// Release necessary resources. Called before cloning.\n");
+        puts("void prepareClone() const;\n");
+        puts("/// Re-init after cloning the model at the process level (e.g. fork in Linux)\n");
+        puts("/// Re-allocate necessary resources. Called after cloning.\n");
+        puts("void atClone() const;\n");
         if (v3Global.opt.trace()) {
             puts("std::unique_ptr<VerilatedTraceConfig> traceConfig() const override final;\n");
         }
@@ -479,6 +485,15 @@ class EmitCModel final : public EmitCFunc {
              + "\"; }\n");
         puts("unsigned " + topClassName() + "::threads() const { return "
              + cvtToStr(std::max(1, v3Global.opt.threads())) + "; }\n");
+        puts("void " + topClassName()
+             + "::prepareClone() const { contextp()->prepareClone(); }\n");
+        puts("void " + topClassName() + "::atClone() const {\n");
+        if (v3Global.opt.threads() > 1) {
+            puts("vlSymsp->__Vm_threadPoolp = static_cast<VlThreadPool*>(");
+        }
+        puts("contextp()->threadPoolpOnClone()");
+        if (v3Global.opt.threads() > 1) puts(")");
+        puts(";\n}\n");
 
         if (v3Global.opt.trace()) {
             puts("std::unique_ptr<VerilatedTraceConfig> " + topClassName()

src/V3EmitCSyms.cpp

@@ -468,7 +468,7 @@ void EmitCSyms::emitSymHdr() {
 
     if (v3Global.opt.mtasks()) {
         puts("\n// MULTI-THREADING\n");
-        puts("VlThreadPool* const __Vm_threadPoolp;\n");
+        puts("VlThreadPool* __Vm_threadPoolp;\n");
         puts("bool __Vm_even_cycle__ico = false;\n");
         puts("bool __Vm_even_cycle__act = false;\n");
         puts("bool __Vm_even_cycle__nba = false;\n");

test_regress/t/t_flag_csplit.pl

@@ -97,6 +97,8 @@ sub check_cpp {
                 && $func !~ /::traceInit$/
                 && $func !~ /::traceFull$/
                 && $func !~ /::final$/
+                && $func !~ /::prepareClone$/
+                && $func !~ /::atClone$/
                 ) {
                 push @funcs, $func;
             }

test_regress/t/t_wrapper_clone.cpp

@@ -0,0 +1,89 @@
+//
+// DESCRIPTION: Verilator: Verilog Test module for prepareClone/atClone APIs
+//
+// This file ONLY is placed into the Public Domain, for any use,
+// without warranty, 2023 by Yinan Xu.
+// SPDX-License-Identifier: CC0-1.0
+
+#include <verilated.h>
+
+#include <unistd.h>
+
+#include <sys/wait.h>
+
+// These require the above. Comment prevents clang-format moving them
+#include "TestCheck.h"
+
+#include VM_PREFIX_INCLUDE
+
+double sc_time_stamp() { return 0; }
+
+// Note: Since the pthread_atfork API accepts only function pointers,
+// we are using a static variable for the TOP just for a simple example.
+// Without using the pthread_atfork API, the user can instead manually call
+// prepareClone and atClone before and after calling fork, and topp can be
+// allocated dynamically.
+static VM_PREFIX* topp = nullptr;
+static auto prepareClone = []() { topp->prepareClone(); };
+static auto atClone = []() { topp->atClone(); };
+
+void single_cycle(VM_PREFIX* topp) {
+    topp->clock = 1;
+    topp->eval();
+
+    topp->clock = 0;
+    topp->eval();
+}
+
+int main(int argc, char** argv) {
+    // We disable the buffering for stdout in this test.
+    // Redirecting the stdout to files with buffering causes duplicated stdout
+    // outputs in both parent and child processes, even if they are actually
+    // called before the fork.
+    setvbuf(stdout, nullptr, _IONBF, 0);
+
+    VerilatedContext* contextp = new VerilatedContext;
+    topp = new VM_PREFIX{contextp};
+
+    // To avoid resource leaks, prepareClone must be called before fork to
+    // free all the allocated resources. Though this would bring performance
+    // overhead to the parent process, we believe that fork should not be
+    // called frequently, and the overhead is minor compared to simulation.
+    pthread_atfork(prepareClone, atClone, atClone);
+
+    // If you care about critical performance, prepareClone can be avoided,
+    // with atClone being called only at the child process, as follows.
+    // It has the same functionality as the previous one, but has memory leaks.
+    // According to the sanitizer, 288 bytes are leaked for one fork call.
+    // pthread_atfork(nullptr, nullptr, atClone);
+
+    topp->reset = 1;
+    topp->is_parent = 0;
+    for (int i = 0; i < 5; i++) { single_cycle(topp); }
+
+    topp->reset = 0;
+    while (!contextp->gotFinish()) {
+        single_cycle(topp);
+
+        if (topp->do_clone) {
+            const int pid = fork();
+            if (pid < 0) {
+                printf("fork failed\n");
+            } else if (pid == 0) {
+                printf("child: here we go\n");
+            } else {
+                while (wait(nullptr) > 0)
+                    ;
+                printf("parent: here we go\n");
+                topp->is_parent = 1;
+            }
+        }
+    }
+
+    topp->final();
+
+    VL_DO_DANGLING(delete topp, topp);
+    VL_DO_DANGLING(delete contextp, contextp);
+
+    return 0;
+}

test_regress/t/t_wrapper_clone.out

@@ -0,0 +1,15 @@
+counter =  0
+counter =  1
+counter =  2
+counter =  3
+counter =  4
+counter =  5
+child: here we go
+counter =  6
+counter =  7
+counter =  8
+parent: here we go
+counter =  6
+counter =  7
+counter =  8
+*-* All Finished *-*

test_regress/t/t_wrapper_clone.pl

@@ -0,0 +1,25 @@
+#!/usr/bin/env perl
+if (!$::Driver) { use strict; use FindBin; exec("$FindBin::Bin/bootstrap.pl", @ARGV, $0); die; }
+# DESCRIPTION: Verilator: Verilog Test module for prepareClone/atClone APIs
+#
+# This file ONLY is placed into the Public Domain, for any use,
+# without warranty, 2023 by Yinan Xu.
+# SPDX-License-Identifier: CC0-1.0
+
+scenarios(vlt_all => 1);
+
+compile(
+    make_top_shell => 0,
+    make_main => 0,
+    verilator_flags2 => ["--exe $Self->{t_dir}/$Self->{name}.cpp",
+                         "-cc"],
+    threads => $Self->{vltmt} ? 2 : 1,
+    );
+
+execute(
+    check_finished => 1,
+    expect_filename => $Self->{golden_filename},
+    );
+
+ok(1);
+1;

test_regress/t/t_wrapper_clone.v

@@ -0,0 +1,38 @@
+// DESCRIPTION: Verilator: Verilog Test module for prepareClone/atClone APIs
+//
+// This model counts from 0 to 8. It forks a child process (in C++) at 6
+// and waits for the child to simulate and exit for resumption (of the parent).
+//
+// This file ONLY is placed into the Public Domain, for any use,
+// without warranty, 2023 by Yinan Xu.
+// SPDX-License-Identifier: CC0-1.0
+
+module top(
+  input clock,
+  input reset,
+  input is_parent,
+  output do_clone
+);
+
+reg [3:0] counter;
+
+assign do_clone = counter == 4'h6;
+
+always @(posedge clock) begin
+  if (reset) begin
+    counter <= 4'h0;
+  end
+  else begin
+    counter <= counter + 4'h1;
+    $write("counter = %d\n", counter);
+  end
+
+  if (counter[3]) begin
+    if (is_parent) begin
+      $write("*-* All Finished *-*\n");
+    end
+    $finish(0);
+  end
+end
+
+endmodule