Experimental unwinding - WIP

Ensure we use elf addresses instead of absolute addresses

include/async-profiler/codeCache.h

@@ -17,7 +17,7 @@
 #ifndef _CODECACHE_H
 #define _CODECACHE_H
 
-// #include <jvmti.h>
+#include <stdint.h>
 
 #define NO_MIN_ADDRESS ((const void *)-1)
 #define NO_MAX_ADDRESS ((const void *)0)
@@ -78,7 +78,7 @@ class CodeCache {
   short _lib_index;
   const void *_min_address;
   const void *_max_address;
-  const char *_text_base;
+  const void *_text_base;
 
   void **_got_start;
   void **_got_end;
@@ -113,7 +113,7 @@ class CodeCache {
 
   void setTextBase(const char *text_base) { _text_base = text_base; }
 
-  const char *getTextBase() { return _text_base; }
+  const void *getTextBase() { return _text_base; }
 
   void **gotStart() const { return _got_start; }
 
@@ -136,7 +136,7 @@ class CodeCache {
   void makeGotPatchable();
 
   void setDwarfTable(FrameDesc *table, int length);
-  FrameDesc *findFrameDesc(const void *pc);
+  FrameDesc *findFrameDesc(uint64_t elf_address);
 };
 
 class CodeCacheArray {

src/async-profiler/codeCache.cpp

@@ -20,10 +20,13 @@
 #include "codeCache.h"
 #include "dwarf.h"
 #include "os.h"
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <string.h>
 #include <sys/mman.h>
+#include <limits>
+#include <cassert>
 
 char *NativeFunc::create(const char *name, short lib_index) {
   NativeFunc *f = (NativeFunc *)malloc(sizeof(NativeFunc) + 1 + strlen(name));
@@ -216,8 +219,9 @@ void CodeCache::setDwarfTable(FrameDesc *table, int length) {
   _dwarf_table_length = length;
 }
 
-FrameDesc *CodeCache::findFrameDesc(const void *pc) {
-  u32 target_loc = (const char *)pc - _text_base;
+FrameDesc *CodeCache::findFrameDesc(uintptr_t elf_address) {
+  assert(elf_address < std::numeric_limits<u32>::max());
+  const u32 target_loc = (const u32)elf_address;
   int low = 0;
   int high = _dwarf_table_length - 1;
 

src/async-profiler/stackWalker.cpp

@@ -87,7 +87,7 @@ bool stepStackContext(ap::StackContext &sc, const ap::StackBuffer &buffer,
                       CodeCacheArray *cache) {
   FrameDesc *f;
   CodeCache *cc = findLibraryByAddress(cache, sc.pc);
-  if (cc == NULL || (f = cc->findFrameDesc(sc.pc)) == NULL) {
+  if (cc == NULL || (f = cc->findFrameDesc(static_cast<const char*>(sc.pc) - static_cast<const char*>(cc->getTextBase()))) == NULL) {
     f = &FrameDesc::default_frame;
   }
 //  const char *sym = cc?cc->binarySearch(sc.pc):"unknown";

src/async-profiler/symbols_linux.cpp

@@ -678,6 +678,9 @@ void Symbols::parsePidLibraries(pid_t pid, CodeCacheArray *array,
           printf("offset from get_elf_offset: %lx \n", elf_offset);
           printf("last readable: %lx \n", last_readable_base);
         }
+        else {
+          printf("Failed to read elf offsets \n");
+        }
 
         // Do not parse the same executable twice, e.g. on Alpine Linux
         if (parsed_inodes.insert(map.dev() | inode << 16).second) {

test/CMakeLists.txt

@@ -391,22 +391,22 @@ add_unit_test(
 
 add_unit_test(ddprof_module_lib-ut ddprof_module_lib-ut.cc ../src/ddprof_module_lib.cc
               ../src/build_id.cc ../src/dso.cc LIBRARIES ${ELFUTILS_LIBRARIES})
-#add_unit_test(
-#  dwarf_unwind-ut
-#  dwarf_unwind-ut.cc
-#  ../src/lib/pthread_fixes.cc
-#  ../src/lib/savecontext.cc
-#  ../src/lib/saveregisters.cc
-#  # ../src/lib/allocation_tracker.cc
-#  ../src/ringbuffer_utils.cc
-#  ../src/perf_ringbuffer.cc
-#  ../src/perf.cc
-#  ../src/pevent_lib.cc
-#  ../src/sys_utils.cc
-#  ../src/user_override.cc
-#  ../src/perf_watcher.cc
-#  LIBRARIES ${ELFUTILS_LIBRARIES} DDProf::AsyncProf)
-#target_include_directories(dwarf_unwind-ut PRIVATE ${ASYNC_PROFILER_INCLUDE})
+add_unit_test(
+  dwarf_unwind-ut
+  dwarf_unwind-ut.cc
+  ../src/lib/pthread_fixes.cc
+  ../src/lib/savecontext.cc
+  ../src/lib/saveregisters.cc
+  # ../src/lib/allocation_tracker.cc
+  ../src/ringbuffer_utils.cc
+  ../src/perf_ringbuffer.cc
+  ../src/perf.cc
+  ../src/pevent_lib.cc
+  ../src/sys_utils.cc
+  ../src/user_override.cc
+  ../src/perf_watcher.cc
+  LIBRARIES ${ELFUTILS_LIBRARIES} DDProf::AsyncProf)
+target_include_directories(dwarf_unwind-ut PRIVATE ${ASYNC_PROFILER_INCLUDE})
 
 add_benchmark(savecontext-bench savecontext-bench.cc ../src/lib/pthread_fixes.cc
               ../src/lib/savecontext.cc ../src/lib/saveregisters.cc LIBRARIES llvm-demangle)

test/dwarf_unwind-ut.cc

@@ -0,0 +1,174 @@
+#include <gtest/gtest.h>
+
+#include "savecontext.hpp"
+#include "stackWalker.h"
+#include "unwind_state.hpp"
+
+#include <array>
+
+#include "async-profiler/codeCache.h"
+#include "async-profiler/stack_context.h"
+#include "async-profiler/symbols.h"
+
+// Retrieves instruction pointer
+#define _THIS_IP_                                                              \
+  ({                                                                           \
+    __label__ __here;                                                          \
+  __here:                                                                      \
+    (unsigned long)&&__here;                                                   \
+  })
+
+// #include "ddprof_defs.hpp"
+
+// temp copy pasta
+#define PERF_SAMPLE_STACK_SIZE (4096UL * 8)
+
+std::byte stack[PERF_SAMPLE_STACK_SIZE];
+
+DDPROF_NOINLINE size_t funcA(std::array<uint64_t, ddprof::k_perf_register_count> &regs);
+DDPROF_NOINLINE size_t funcB(std::array<uint64_t, ddprof::k_perf_register_count> &regs);
+
+size_t funcB(std::array<uint64_t, ddprof::k_perf_register_count> &regs) {
+  printf("dwarf_unwind-ut:%s %lx \n", __FUNCTION__, _THIS_IP_);
+  std::span<const std::byte> bounds = ddprof::retrieve_stack_bounds();
+  size_t size = ddprof::save_context(bounds, regs, stack);
+
+  return size;
+}
+
+size_t funcA(std::array<uint64_t, ddprof::k_perf_register_count> &regs) {
+  printf("dwarf_unwind-ut:%s %lx \n", __FUNCTION__, _THIS_IP_);
+  return funcB(regs);
+}
+
+TEST(dwarf_unwind, simple) {
+  CodeCacheArray cache_arary;
+  // Load libraries
+  Symbols::parsePidLibraries(getpid(), &cache_arary, false);
+  std::array<uint64_t, ddprof::k_perf_register_count> regs;
+  size_t size_stack = funcA(regs);
+  EXPECT_TRUE(size_stack);
+
+  ap::StackContext sc = ap::from_regs(std::span(regs));
+  ap::StackBuffer buffer(stack, sc.sp, sc.sp + size_stack);
+
+  void *callchain[128];
+  int n = stackWalk(&cache_arary, sc, buffer,
+                    const_cast<const void **>(callchain), 128, 0);
+  const char *syms[128];
+  for (int i = 0; i < n; ++i) {
+    { // retrieve symbol
+      CodeCache *code_cache = findLibraryByAddress(
+          &cache_arary, reinterpret_cast<void *>(callchain[i]));
+      if (code_cache) {
+        syms[i] = code_cache->binarySearch(callchain[i]);
+        printf("IP = %p - %s\n", callchain[i], syms[i]);
+      }
+    }
+  }
+
+  // Check that we found the expected functions during unwinding
+  ASSERT_TRUE(std::string(syms[0]).find("save_context") != std::string::npos);
+  ASSERT_TRUE(std::string(syms[1]).find("funcB") != std::string::npos);
+  ASSERT_TRUE(std::string(syms[2]).find("funcA") != std::string::npos);
+}
+
+#ifdef ALLOC_TRACKER
+#include "allocation_tracker.hpp"
+#include "perf_ringbuffer.hpp"
+#include "ringbuffer_holder.hpp"
+#include "ringbuffer_utils.hpp"
+#include <span>
+#include "defer.hpp"
+
+namespace ddprof {
+static const uint64_t kSamplingRate = 1;
+
+DDPROF_NOINLINE void func_save_sleep(size_t size);
+DDPROF_NOINLINE void func_intermediate_0(size_t size);
+DDPROF_NOINLINE void func_intermediate_1(size_t size);
+
+DDPROF_NOINLINE void func_save_sleep(size_t size) {
+  ddprof::TrackerThreadLocalState *tl_state = AllocationTracker::get_tl_state();
+  assert(tl_state);
+  int i = 0;
+  while (++i < 100000) {
+
+    ddprof::AllocationTracker::track_allocation_s(0xdeadbeef, size, *tl_state);
+    // prevent tail call optimization
+    getpid();
+    usleep(100);
+    //    printf("Save context nb -- %d \n", i);
+  }
+}
+
+void func_intermediate_0(size_t size) { func_intermediate_1(size); }
+
+void func_intermediate_1(size_t size) { func_save_sleep(size); }
+
+TEST(dwarf_unwind, remote) {
+  const uint64_t rate = 1;
+  const size_t buf_size_order = 5;
+  ddprof::RingBufferHolder ring_buffer{buf_size_order,
+                                       RingBufferType::kMPSCRingBuffer};
+  AllocationTracker::allocation_tracking_init(
+      kSamplingRate,
+      AllocationTracker::kDeterministicSampling |
+          AllocationTracker::kTrackDeallocations,
+      k_default_perf_stack_sample_size, ring_buffer.get_buffer_info(), {});
+  defer { AllocationTracker::allocation_tracking_free(); };
+
+  // Fork
+  pid_t temp_pid = fork();
+  if (!temp_pid) {
+    func_intermediate_0(10);
+    //    char *const argList[] = {"sleep", "10", nullptr};
+    //    execvp("sleep", argList);
+    return;
+  }
+
+  // Load libraries from the fork - Cache array is relent to a single pid
+  CodeCacheArray cache_arary;
+  sleep(1);
+  Symbols::parsePidLibraries(temp_pid, &cache_arary, false);
+  // Establish a ring buffer ?
+
+  ddprof::MPSCRingBufferReader reader{&ring_buffer.get_ring_buffer()};
+  ASSERT_GT(reader.available_size(), 0);
+
+  auto buf = reader.read_sample();
+  ASSERT_FALSE(buf.empty());
+  const perf_event_header *hdr =
+      reinterpret_cast<const perf_event_header *>(buf.data());
+  ASSERT_EQ(hdr->type, PERF_RECORD_SAMPLE);
+
+  // convert based on mask for this watcher (default in this case)
+  perf_event_sample *sample = hdr2samp(hdr, ddprof::perf_event_default_sample_type());
+
+  std::span<const uint64_t, ddprof::k_perf_register_count> regs_span{sample->regs, ddprof::k_perf_register_count};
+  ap::StackContext sc = ap::from_regs(regs_span);
+  std::span<const std::byte> stack{
+      reinterpret_cast<const std::byte *>(sample->data_stack), sample->size_stack};
+  ap::StackBuffer buffer(stack, sc.sp, sc.sp + sample->size_stack);
+
+  void *callchain[ddprof::kMaxStackDepth];
+  int n =
+      stackWalk(&cache_arary, sc, buffer, const_cast<const void **>(callchain),
+                ddprof::kMaxStackDepth, 0);
+
+  std::array<const char *, ddprof::kMaxStackDepth> syms;
+  for (int i = 0; i < n; ++i) {
+    { // retrieve symbol
+      CodeCache *code_cache = findLibraryByAddress(
+          &cache_arary, reinterpret_cast<void *>(callchain[i]));
+      if (code_cache) {
+        syms[i] = code_cache->binarySearch(callchain[i]);
+        printf("IP = %p - %s\n", callchain[i], syms[i]);
+      }
+    }
+    // cleanup the producer fork
+    kill(temp_pid, SIGTERM);
+  }
+}
+}
+#endif