perftools.rb 0.4.0 → 0.4.2

data/README

@@ -29,6 +29,10 @@ google-perftools for ruby code
 
     $ CPUPROFILE_REALTIME=1 ruby my_app.rb
 
+  Profile object allocations instead of cpu/wall time:
+
+    $ CPUPROFILE_OBJECTS=1 ruby my_app.rb
+
 
 === Reporting
 

data/ext/extconf.rb

@@ -46,7 +46,8 @@ Dir.chdir('src') do
       ['perftools-osx', RUBY_PLATFORM =~ /darwin/],
       ['perftools-osx-106', RUBY_PLATFORM =~ /darwin10/],
       ['perftools-debug', true],
-      ['perftools-realtime', true]
+      ['perftools-realtime', true],
+      ['perftools-objects', true]
     ].each do |patch, apply|
       if apply
         sys("patch -p1 < ../../../patches/#{patch}.patch")
@@ -80,7 +81,7 @@ def add_define(name)
 end
 
 case RUBY_PLATFORM
-when /darwin/, /linux/
+when /darwin/, /linux/, /freebsd/
   CONFIG['LDSHARED'] = "$(CXX) " + CONFIG['LDSHARED'].split[1..-1].join(' ')
 end
 
@@ -88,6 +89,7 @@ if RUBY_VERSION >= "1.9"
   add_define 'RUBY19'
 
   hdrs = proc {
+    have_header("method.h") # exists on 1.9.2
     have_header("vm_core.h") and
     have_header("iseq.h") and
     have_header("insns.inc") and

data/ext/perftools.c

@@ -5,6 +5,7 @@ void ProfilerGcMark(void (*cb)(VALUE));
 int ProfilerStart(const char*);
 void ProfilerStop();
 void ProfilerFlush();
+void ProfilerRecord(int, void*, void*);
 
 static VALUE Iallocate;
 static VALUE I__send__;
@@ -147,8 +148,13 @@ static VALUE Isend;
         case VM_FRAME_MAGIC_METHOD:
         case VM_FRAME_MAGIC_CFUNC:
           self = cfp->self;
+#ifdef HAVE_METHOD_H
+          klass = cfp->me->klass;
+          method = cfp->me->called_id;
+#else
           klass = cfp->method_class;
           method = cfp->method_id;
+#endif
           SAVE_FRAME();
           break;
       }
@@ -193,18 +199,23 @@ static VALUE Isend;
   #endif
 #endif
 
+static VALUE objprofiler_setup();
+static VALUE objprofiler_teardown();
+
+/* CpuProfiler */
+
 static VALUE cPerfTools;
 static VALUE cCpuProfiler;
 static VALUE bProfilerRunning;
 static VALUE gc_hook;
 
-VALUE
+static VALUE
 cpuprofiler_running_p(VALUE self)
 {
   return bProfilerRunning;
 }
 
-VALUE
+static VALUE
 cpuprofiler_stop(VALUE self)
 {
   if (!bProfilerRunning)
@@ -213,10 +224,14 @@ cpuprofiler_stop(VALUE self)
   bProfilerRunning = Qfalse;
   ProfilerStop();
   ProfilerFlush();
+
+  if (getenv("CPUPROFILE_OBJECTS"))
+    objprofiler_teardown();
+
   return Qtrue;
 }
 
-VALUE
+static VALUE
 cpuprofiler_start(VALUE self, VALUE filename)
 {
   StringValue(filename);
@@ -224,6 +239,9 @@ cpuprofiler_start(VALUE self, VALUE filename)
   if (bProfilerRunning)
     return Qfalse;
 
+  if (getenv("CPUPROFILE_OBJECTS"))
+    objprofiler_setup();
+
   ProfilerStart(RSTRING_PTR(filename));
   bProfilerRunning = Qtrue;
 
@@ -241,20 +259,153 @@ cpuprofiler_gc_mark()
   ProfilerGcMark(rb_gc_mark);
 }
 
+/* ObjProfiler */
+
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+#ifndef _XOPEN_SOURCE
+#define _XOPEN_SOURCE 600
+#endif
+
+#include <assert.h>
+#include <ucontext.h>
+#include <unistd.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+
+static VALUE bObjProfilerRunning;
+#define NUM_ORIG_BYTES 2
+
+struct {
+  void *location;
+  unsigned char value;
+} orig_bytes[NUM_ORIG_BYTES];
+
+static inline void *
+page_align(void *addr) {
+  assert(addr != NULL);
+  return (void *)((size_t)addr & ~(0xFFFF));
+}
+
+static void
+copy_instructions(void *dest, void *src, size_t count) {
+  assert(dest != NULL);
+  assert(src != NULL);
+
+  void *aligned_addr = page_align(dest);
+  if (mprotect(aligned_addr, (dest - aligned_addr) + count, PROT_READ|PROT_WRITE|PROT_EXEC) != 0)
+    perror("mprotect");
+  memcpy(dest, src, count);
+}
+
+static inline void**
+uc_get_ip(ucontext_t *uc) {
+  #if defined(__FreeBSD__)
+    return (void**)&uc->uc_mcontext.mc_rip;
+  #elif defined(__dietlibc__)
+    return (void**)&uc->uc_mcontext.rip;
+  #elif defined(__APPLE__)
+    return (void**)&uc->uc_mcontext->__ss.__rip;
+  #else
+    return (void**)&uc->uc_mcontext.gregs[REG_RIP];
+  #endif
+}
+
+static void
+trap_handler(int sig, siginfo_t *info, void *data) {
+  int i;
+  ucontext_t *uc = (ucontext_t *)data;
+  void **ip = uc_get_ip(uc);
+
+  // printf("signal: %d, addr: %p, ip: %p\n", signal, info->si_addr, *ip);
+
+  for (i=0; i<NUM_ORIG_BYTES; i++) {
+    if (orig_bytes[i].location == *ip-1) {
+      // restore original byte
+      copy_instructions(orig_bytes[i].location, &orig_bytes[i].value, 1);
+
+      // setup next breakpoint
+      copy_instructions(orig_bytes[(i+1)%NUM_ORIG_BYTES].location, "\xCC", 1);
+
+      // first breakpoint is the notification
+      if (i == 0)
+        ProfilerRecord(sig, info, data);
+
+      // reset instruction pointer
+      *ip -= 1;
+
+      break;
+    }
+  }
+}
+
+static VALUE
+objprofiler_setup()
+{
+  if (bObjProfilerRunning)
+    return Qtrue;
+
+  int i;
+  struct sigaction sig = { .sa_sigaction = trap_handler, .sa_flags = SA_SIGINFO };
+  sigemptyset(&sig.sa_mask);
+  sigaction(SIGTRAP, &sig, NULL);
+
+  for (i=0; i<NUM_ORIG_BYTES; i++) {
+    orig_bytes[i].location = rb_newobj + i;
+    orig_bytes[i].value    = ((unsigned char*)rb_newobj)[i];
+    copy_instructions(rb_newobj + i, "\xCC", 1);
+  }
+
+  // setenv("CPUPROFILE_OBJECTS", "1", 1);
+  bObjProfilerRunning = Qtrue;
+  return Qtrue;
+}
+
+static VALUE
+objprofiler_teardown()
+{
+  if (!bObjProfilerRunning)
+    return Qfalse;
+
+  int i;
+  struct sigaction sig = { .sa_handler = SIG_IGN };
+  sigemptyset(&sig.sa_mask);
+  sigaction(SIGTRAP, &sig, NULL);
+
+  for (i=0; i<NUM_ORIG_BYTES; i++) {
+    copy_instructions(orig_bytes[i].location, &orig_bytes[i].value, 1);
+  }
+
+  // unsetenv("CPUPROFILE_OBJECTS");
+  bObjProfilerRunning = Qfalse;
+  return Qtrue;
+}
+
+/* Init */
+
 void
 Init_perftools()
 {
   cPerfTools = rb_define_class("PerfTools", rb_cObject);
   cCpuProfiler = rb_define_class_under(cPerfTools, "CpuProfiler", rb_cObject);
-  bProfilerRunning = Qfalse;
+
   Iallocate = rb_intern("allocate");
   I__send__ = rb_intern("__send__");
   Isend = rb_intern("send");
 
+  bObjProfilerRunning = bProfilerRunning = Qfalse;
+
   rb_define_singleton_method(cCpuProfiler, "running?", cpuprofiler_running_p, 0);
   rb_define_singleton_method(cCpuProfiler, "start", cpuprofiler_start, 1);
   rb_define_singleton_method(cCpuProfiler, "stop", cpuprofiler_stop, 0);
 
   gc_hook = Data_Wrap_Struct(cCpuProfiler, cpuprofiler_gc_mark, NULL, NULL);
   rb_global_variable(&gc_hook);
+
+  if (getenv("CPUPROFILE") && getenv("CPUPROFILE_OBJECTS"))
+    objprofiler_setup();
 }

data/objalloc_tests/Makefile

@@ -0,0 +1,7 @@
+default: trap
+
+trap: trap.c
+	gcc -o trap trap.c -ggdb -O0 -Wall
+
+clean:
+	rm -rf *.o trap *.dSYM

data/objalloc_tests/trap.c

@@ -0,0 +1,103 @@
+#define _GNU_SOURCE
+#define _XOPEN_SOURCE 600
+
+#include <assert.h>
+#include <ucontext.h>
+#include <unistd.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+
+
+void
+func() {
+  printf("hi\n");
+}
+
+
+static inline void *
+page_align(void *addr) {
+  assert(addr != NULL);
+  return (void *)((size_t)addr & ~(0xFFFF));
+}
+
+static void
+copy_instructions(void *dest, void *src, size_t count) {
+  assert(dest != NULL);
+  assert(src != NULL);
+
+  void *aligned_addr = page_align(dest);
+  if (mprotect(aligned_addr, (dest - aligned_addr) + count, PROT_READ|PROT_WRITE|PROT_EXEC) != 0)
+    perror("mprotect");
+  memcpy(dest, src, count);
+}
+
+#define NUM_ORIG_BYTES 2
+struct {
+  void *location;
+  unsigned char value;
+} orig_bytes[NUM_ORIG_BYTES];
+
+static inline void**
+uc_get_ip(ucontext_t *uc) {
+  #if defined(__FreeBSD__)
+    return (void**)&uc->uc_mcontext.mc_rip;
+  #elif defined(__dietlibc__)
+    return (void**)&uc->uc_mcontext.rip;
+  #elif defined(__APPLE__)
+    return (void**)&uc->uc_mcontext->__ss.__rip;
+  #else
+    return (void**)&uc->uc_mcontext.gregs[REG_RIP];
+  #endif
+}
+
+static void
+trap_handler(int signal, siginfo_t *info, void *data) {
+  int i;
+  ucontext_t *uc = (ucontext_t *)data;
+  void **ip = uc_get_ip(uc);
+
+  // printf("signal: %d, addr: %p, ip: %p\n", signal, info->si_addr, *ip);
+
+  for (i=0; i<NUM_ORIG_BYTES; i++) {
+    if (orig_bytes[i].location == *ip-1) {
+      // restore original byte
+      copy_instructions(orig_bytes[i].location, &orig_bytes[i].value, 1);
+
+      // setup next breakpoint
+      copy_instructions(orig_bytes[(i+1)%NUM_ORIG_BYTES].location, "\xCC", 1);
+
+      // first breakpoint is the notification
+      if (i == 0)
+        printf(" ---> YOU'RE CALLING FUNC()\n");
+
+      // reset instruction pointer
+      *ip -= 1;
+
+      break;
+    }
+  }
+}
+
+int
+main() {
+  int i;
+  struct sigaction sig = { .sa_sigaction = trap_handler, .sa_flags = SA_SIGINFO };
+  sigemptyset(&sig.sa_mask);
+  sigaction(SIGTRAP, &sig, NULL);
+
+  for (i=0; i<NUM_ORIG_BYTES; i++) {
+    orig_bytes[i].location = func + i;
+    orig_bytes[i].value    = ((unsigned char*)func)[i];
+    copy_instructions(func + i, "\xCC", 1);
+  }
+
+  printf("func: %p\n", func);
+
+  for (i=0; i<10; i++)
+    func();
+
+  return 0;
+}

data/patches/perftools-objects.patch

@@ -0,0 +1,91 @@
+commit 5ffd87871619a750ad19e247670a0887a6d38fbd
+Author: Aman Gupta <aman@tmm1.net>
+Date:   Sun Aug 1 18:50:38 2010 -0700
+
+    perftools-objects
+
+diff --git a/src/profile-handler.cc b/src/profile-handler.cc
+index 619b980..2ecf2c0 100644
+--- a/src/profile-handler.cc
++++ b/src/profile-handler.cc
+@@ -402,6 +402,8 @@ void ProfileHandler::GetState(ProfileHandlerState* state) {
+ }
+ 
+ void ProfileHandler::StartTimer() {
++  if (getenv("CPUPROFILE_OBJECTS")) return;
++
+   struct itimerval timer;
+   timer.it_interval.tv_sec = 0;
+   timer.it_interval.tv_usec = 1000000 / frequency_;
+@@ -410,12 +412,16 @@ void ProfileHandler::StartTimer() {
+ }
+ 
+ void ProfileHandler::StopTimer() {
++  if (getenv("CPUPROFILE_OBJECTS")) return;
++
+   struct itimerval timer;
+   memset(&timer, 0, sizeof timer);
+   setitimer(realtime_ ? ITIMER_REAL : ITIMER_PROF, &timer, 0);
+ }
+ 
+ bool ProfileHandler::IsTimerRunning() {
++  if (getenv("CPUPROFILE_OBJECTS")) return false;
++
+   struct itimerval current_timer;
+   RAW_CHECK(0 == getitimer(realtime_ ? ITIMER_REAL : ITIMER_PROF, &current_timer), "getitimer");
+   return (current_timer.it_value.tv_sec != 0 ||
+@@ -423,6 +429,8 @@ bool ProfileHandler::IsTimerRunning() {
+ }
+ 
+ void ProfileHandler::EnableHandler() {
++  if (getenv("CPUPROFILE_OBJECTS")) return;
++
+   struct sigaction sa;
+   sa.sa_sigaction = SignalHandler;
+   sa.sa_flags = SA_RESTART | SA_SIGINFO;
+@@ -431,6 +439,8 @@ void ProfileHandler::EnableHandler() {
+ }
+ 
+ void ProfileHandler::DisableHandler() {
++  if (getenv("CPUPROFILE_OBJECTS")) return;
++
+   struct sigaction sa;
+   sa.sa_handler = SIG_IGN;
+   sa.sa_flags = SA_RESTART;
+diff --git a/src/profiler.cc b/src/profiler.cc
+index 37234b2..905288f 100644
+--- a/src/profiler.cc
++++ b/src/profiler.cc
+@@ -97,6 +97,10 @@ class CpuProfiler {
+ 
+   static CpuProfiler instance_;
+ 
++  // Signal handler that records the interrupted pc in the profile data.
++  static void prof_handler(int sig, siginfo_t*, void* signal_ucontext,
++                           void* cpu_profiler);
++
+  private:
+   // This lock implements the locking requirements described in the ProfileData
+   // documentation, specifically:
+@@ -125,10 +129,6 @@ class CpuProfiler {
+ 
+   // Disables receiving SIGPROF interrupt.
+   void DisableHandler();
+-
+-  // Signal handler that records the interrupted pc in the profile data.
+-  static void prof_handler(int sig, siginfo_t*, void* signal_ucontext,
+-                           void* cpu_profiler);
+ };
+ 
+ // Profile data structure singleton: Constructor will check to see if
+@@ -318,6 +318,10 @@ extern "C" PERFTOOLS_DLL_DECL void ProfilerFlush() {
+ extern "C" PERFTOOLS_DLL_DECL void ProfilerGcMark(void (*cb)(VALUE)) {
+   CpuProfiler::instance_.GcMark(cb);
+ }
++
++extern "C" PERFTOOLS_DLL_DECL void ProfilerRecord(int sig, siginfo_t* info, void* signal_ucontext) {
++  CpuProfiler::prof_handler(sig, info, signal_ucontext, (void*)&CpuProfiler::instance_);
++}
+ #endif
+ 
+ extern "C" PERFTOOLS_DLL_DECL int ProfilingIsEnabledForAllThreads() {

data/perftools.rb.gemspec

@@ -1,7 +1,7 @@
 spec = Gem::Specification.new do |s|
   s.name = 'perftools.rb'
-  s.version = '0.4.0'
-  s.date = '2009-11-16'
+  s.version = '0.4.2'
+  s.date = '2010-08-01'
   s.rubyforge_project = 'perftools-rb'
   s.summary = 'google-perftools for ruby code'
   s.description = 'A sampling profiler for ruby code based on patches to google-perftools'
@@ -17,20 +17,5 @@ spec = Gem::Specification.new do |s|
   s.executables << 'pprof.rb'
 
   # ruby -rpp -e' pp `git ls-files | grep -v examples`.split("\n").sort '
-  s.files = [
-    "README",
-    "bin/pprof.rb",
-    "ext/extconf.rb",
-    "ext/perftools.c",
-    "ext/src/google-perftools-1.4.tar.gz",
-    "patches/perftools-debug.patch",
-    "patches/perftools-gc.patch",
-    "patches/perftools-osx-106.patch",
-    "patches/perftools-osx.patch",
-    "patches/perftools-pprof.patch",
-    "patches/perftools-realtime.patch",
-    "patches/perftools-notests.patch",
-    "patches/perftools.patch",
-    "perftools.rb.gemspec"
-  ]
+  s.files = `git ls-files`.split("\n").reject{ |f| f =~ /^examples/ }
 end

metadata

@@ -1,7 +1,13 @@
 --- !ruby/object:Gem::Specification 
 name: perftools.rb
 version: !ruby/object:Gem::Version 
-  version: 0.4.0
+  hash: 11
+  prerelease: false
+  segments: 
+  - 0
+  - 4
+  - 2
+  version: 0.4.2
 platform: ruby
 authors: 
 - Aman Gupta
@@ -9,7 +15,7 @@ autorequire:
 bindir: bin
 cert_chain: []
 
-date: 2009-11-16 00:00:00 -08:00
+date: 2010-08-01 00:00:00 -07:00
 default_executable: 
 dependencies: []
 
@@ -27,13 +33,16 @@ files:
 - ext/extconf.rb
 - ext/perftools.c
 - ext/src/google-perftools-1.4.tar.gz
+- objalloc_tests/Makefile
+- objalloc_tests/trap.c
 - patches/perftools-debug.patch
 - patches/perftools-gc.patch
+- patches/perftools-notests.patch
+- patches/perftools-objects.patch
 - patches/perftools-osx-106.patch
 - patches/perftools-osx.patch
 - patches/perftools-pprof.patch
 - patches/perftools-realtime.patch
-- patches/perftools-notests.patch
 - patches/perftools.patch
 - perftools.rb.gemspec
 has_rdoc: true
@@ -46,21 +55,27 @@ rdoc_options: []
 require_paths: 
 - lib
 required_ruby_version: !ruby/object:Gem::Requirement 
+  none: false
   requirements: 
   - - ">="
     - !ruby/object:Gem::Version 
+      hash: 3
+      segments: 
+      - 0
       version: "0"
-  version: 
 required_rubygems_version: !ruby/object:Gem::Requirement 
+  none: false
   requirements: 
   - - ">="
     - !ruby/object:Gem::Version 
+      hash: 3
+      segments: 
+      - 0
       version: "0"
-  version: 
 requirements: []
 
 rubyforge_project: perftools-rb
-rubygems_version: 1.3.5
+rubygems_version: 1.3.7
 signing_key: 
 specification_version: 3
 summary: google-perftools for ruby code