Implement a tool called link_limiter to impose a global restriction on how many

tools/win/link_limiter/limiter.cc

Index: tools/win/link_limiter/limiter.cc
diff --git a/tools/win/link_limiter/limiter.cc b/tools/win/link_limiter/limiter.cc
new file mode 100644
index 0000000000000000000000000000000000000000..d1218ef4ed583daa6acfbb715fd4298000508320
--- /dev/null
+++ b/tools/win/link_limiter/limiter.cc
@@ -0,0 +1,337 @@
+// Copyright (c) 2012 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#define NOMINMAX
+#include <windows.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include <algorithm>
+#include <iterator>
+#include <string>
+#include <sstream>
+#include <vector>

[cmp, 2012/08/01 23:36:42]

alphanum sort lines 11-15

+
+typedef std::basic_string<TCHAR> tstring;
+
+const bool g_is_debug = (_wgetenv(L"LIMITER_DEBUG") != NULL);
+const int  g_wait_time = 30 * 1000; // 30 seconds

[cmp, 2012/08/01 23:36:42]

probably should remove one of the extra spaces after 'int'

+
+// Don't use stderr for errors because VS has large buffers on them, leading
+// to confusing error output.
+static void Error(const tstring& msg, ...) {
+  tstring new_msg = L"limiter fatal error: "+msg+L"\n";
+  va_list args;
+  va_start(args, msg);
+  vwprintf(new_msg.c_str(), args);
+  va_end(args);
+}
+static void Warn(const tstring& msg, ...) {

[cmp, 2012/08/01 23:36:42]

insert an empty line before line 31

+  if(!g_is_debug)

[cmp, 2012/08/01 23:36:42]

one space after if

+    return;
+  tstring new_msg = L"limiter warning: "+msg+L"\n";
+  va_list args;
+  va_start(args, msg);
+  vwprintf(new_msg.c_str(), args);
+  va_end(args);
+}
+
+static tstring ErrorMessageToString(DWORD err) {
+  TCHAR* msg_buf = NULL;
+  DWORD rc = FormatMessage(
+    FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM,
+    NULL,
+    err,
+    MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT),
+    reinterpret_cast<LPTSTR>(&msg_buf),
+    0,
+    NULL);
+  if (!rc)
+    return L"unknown error";
+  tstring ret(msg_buf);
+  LocalFree(msg_buf);
+  return ret;
+}
+
+static DWORD RunExe(const tstring& exe_name) {
+  STARTUPINFO startup_info = { sizeof(STARTUPINFO) };
+  PROCESS_INFORMATION process_info;
+  DWORD exit_code;
+
+  GetStartupInfo(&startup_info);
+  tstring cmdline = tstring(GetCommandLine());
+
+  size_t first_space = cmdline.find(' ');
+  if (first_space == -1) {
+    // I'm not sure why this would ever happen, but just in case...
+    cmdline = exe_name;
+  } else {
+    cmdline = exe_name + cmdline.substr(first_space+1);
+  }
+
+  if (!CreateProcess(NULL,
+                     const_cast<TCHAR*>(cmdline.c_str()),
+                     NULL,
+                     NULL,
+                     TRUE,
+                     0,
+                     NULL,
+                     NULL,
+                     &startup_info, &process_info)) {
+    Error(L"Error in CreateProcess[%s]: %s",
+      cmdline, ErrorMessageToString(GetLastError()).c_str());
+    return MAXDWORD;
+  }
+  CloseHandle(process_info.hThread);
+  WaitForSingleObject(process_info.hProcess, INFINITE);
+  GetExitCodeProcess(process_info.hProcess, &exit_code);
+  CloseHandle(process_info.hProcess);
+  return exit_code;
+}
+
+// Returns 0 if there was an error
+static int CpuConcurrencyMetric(const tstring& envvar_name) {
+  int max_concurrent = 0;
+  std::vector<char> buffer(1);
+  BOOL ok = false;
+  DWORD last_error = 0;
+  do {
+    DWORD bufsize = buffer.size();
+    ok = GetLogicalProcessorInformation(
+      reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION>(&buffer[0]),
+      &bufsize);
+    last_error = GetLastError();
+    if (!ok && last_error == ERROR_INSUFFICIENT_BUFFER &&
+      bufsize > buffer.size())

[cmp, 2012/08/01 23:36:42]

this indent should line the first b in bufsize under the ! in the line above
(aka "4 space indent when spanning lines in a conditional")

+    {
+      buffer.resize(bufsize);
+    }
+  } while(!ok && last_error == ERROR_INSUFFICIENT_BUFFER);

[cmp, 2012/08/01 23:36:42]

insert a space after while

+
+  if (!ok) {
+    Warn(L"Error while getting number of cores. Try setting the "
+         L" environment variable '%s'  to (num_cores-1): %s",
+         envvar_name.c_str(), ErrorMessageToString(last_error).c_str());
+    return 0;
+  }
+
+  PSYSTEM_LOGICAL_PROCESSOR_INFORMATION pproc_info =
+    reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION>(&buffer[0]);

[cmp, 2012/08/01 23:36:42]

increase indent on line 121 by 2 spaces

+  int num_entries = buffer.size() /
+    sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION);

[cmp, 2012/08/01 23:36:42]

increase indent on line 123 by 2 spaces

+
+  for(int i = 0; i < num_entries; ++i) {

[cmp, 2012/08/01 23:36:42]

one space after for

+    SYSTEM_LOGICAL_PROCESSOR_INFORMATION &info = pproc_info[i];
+    if(info.Relationship == RelationProcessorCore) {

[cmp, 2012/08/01 23:36:42]

one space after if

+      max_concurrent += 1;
+    }
+  }
+
+  // Leave one core for other tasks
+  return max_concurrent-1;
+}
+
+
+// TODO(defaults): Create a better heuristic than # of CPUs. It seems likely
+// that the right value will, in fact, be based on the memory capacity of the
+// machine, not on the number of CPUs.
+enum ConcurrencyMetricEnum {
+  CONCURRENCY_METRIC_ONE,
+  CONCURRENCY_METRIC_CPU,
+  CONCURRENCY_METRIC_DEFAULT = CONCURRENCY_METRIC_CPU
+};
+
+static int GetMaxConcurrency(
+  const tstring& base_pipename,
+  ConcurrencyMetricEnum metric = CONCURRENCY_METRIC_DEFAULT)
+{
+  static int max_concurrent = -1;
+
+  if (max_concurrent == -1) {

[cmp, 2012/08/01 23:36:42]

could max_concurrent ever not equal -1?  i'm thinking this if condition could go
away and the code in it could be de-indented 2 spaces

[iannucci1, 2012/08/01 23:52:39]

It's static. Multiple calls will get the cached value. (and since this
application is single threaded, static is safe here).

This used to be more important in a previous flavors of the code (this function
was called multiple times). I could take it or leave it.

[cmp, 2012/08/02 00:04:23]

Never mind, then.

+    tstring envvar_name = base_pipename+L"_MAXCONCURRENCY";
+
+    const LPTSTR max_concurrent_str = _wgetenv(envvar_name.c_str());
+    max_concurrent = _wtoi(max_concurrent_str ? max_concurrent_str : L"0");
+
+    if (max_concurrent == 0) {
+      switch(metric) {

[cmp, 2012/08/01 23:36:42]

one space after switch

+      case CONCURRENCY_METRIC_CPU:

[cmp, 2012/08/01 23:36:42]

each case in a switch should be indented by 2 spaces

+        max_concurrent = CpuConcurrencyMetric(envvar_name);
+        if(max_concurrent)

[cmp, 2012/08/01 23:36:42]

one space after if

+          break;
+        // else fallthrough
+      case CONCURRENCY_METRIC_ONE:
+        max_concurrent = 1;
+        break;
+      }
+    }
+
+    max_concurrent = std::min(std::max(max_concurrent, 1),
+                              PIPE_UNLIMITED_INSTANCES);
+  }
+
+  return max_concurrent;
+}
+
+

[cmp, 2012/08/01 23:36:42]

let's be consistent about 2 empty lines between functions or 1 empty line.  i
couldn't find anything in the style guide about this and example files i found
in base/ only use 1.  so i guess stick with 1?

+static HANDLE wait_for_pipe(const tstring& pipename,
+                            HANDLE hEvent,
+                            int max_concurrency)
+{
+  // We're using a named pipe instead of a semaphore so the Kernel can clean up
+  // after us if we crash while holding onto the pipe (A real semaphore will
+  // not release on process termination).
+  HANDLE hPipe = INVALID_HANDLE_VALUE;

[cmp, 2012/08/01 23:36:42]

i believe the 'no hungarian notation' rule applies here

+  for(;;) {

[cmp, 2012/08/01 23:36:42]

one space after for

+    hPipe = CreateNamedPipe(
+      pipename.c_str(),

[cmp, 2012/08/01 23:36:42]

these should be indented by an extra 2 spaces per line

+      PIPE_ACCESS_DUPLEX,
+      PIPE_TYPE_BYTE,
+      max_concurrency,
+      1, // nOutBufferSize
+      1, // nInBufferSize
+      0, // nDefaultTimeOut
+      NULL // Default security attributes (noinherit)
+      );

[cmp, 2012/08/01 23:36:42]

i think this line should be de-indented by 2 spaces

+    if (hPipe != INVALID_HANDLE_VALUE)
+      break;
+
+    DWORD error = GetLastError();
+    if (error == ERROR_PIPE_BUSY) {
+      if (hEvent) {
+        WaitForSingleObject(hEvent, g_wait_time);
+      } else {
+        Sleep(g_wait_time);
+      }
+    } else {
+      Warn(L"Got error %d while waiting for pipe: %s", error,
+           ErrorMessageToString(error).c_str());
+      return INVALID_HANDLE_VALUE;
+    }
+  }
+
+  return hPipe;
+}
+
+static int wait_and_run(const tstring& shimmed_exe,
+                        const tstring& base_pipename)
+{
+  ULONGLONG start_time = 0, end_time = 0;
+  tstring pipename = L"\\\\.\\pipe\\"+base_pipename;
+  tstring event_name = L"Local\\EVENT_"+base_pipename;
+
+  // This event lets us do better than strict polling, but we don't rely on it
+  // (in case a process crashes before signalling the event).
+  HANDLE hEvent = CreateEvent(
+    NULL, // Default security attributes
+    FALSE, // Manual reset
+    FALSE, // Initial state
+    event_name.c_str());
+
+  if (g_is_debug)
+    start_time = GetTickCount64();
+
+  HANDLE hPipe =

[cmp, 2012/08/01 23:36:42]

hungarian notation should not be used (see C++ style guide section on 'windows
code')

+    wait_for_pipe(pipename, hEvent, GetMaxConcurrency(base_pipename));
+
+  if (g_is_debug) {
+    end_time = GetTickCount64();
+    wprintf(L"  took %.2fs to acquire semaphore.\n",
+      (end_time - start_time) / 1000.0);
+  }
+
+  DWORD ret = RunExe(shimmed_exe);
+
+  if (hPipe != INVALID_HANDLE_VALUE) { CloseHandle(hPipe); }
+  if (hEvent != NULL) { SetEvent(hEvent); }
+
+  return ret;
+}
+
+void usage(const tstring& msg) {
+  std::wostringstream strm;
+  strm << msg << L"\n";
+  strm << L"Usage: SHIMED_NAME__SEMAPHORE_NAME\n"
+       << L"\n"
+       << L"  SHIMMED_NAME   - ex. 'link.exe' or 'lib.exe'\n"
+       << L"                 - can be exe, bat, or com\n"
+       << L"                 - must exist in PATH\n"
+       << L"\n"
+       << L"  SEMAPHORE_NAME - ex. 'SOME_NAME' or 'GROOVY_SEMAPHORE'\n"
+       << L"\n"
+       << L"  Example:\n"
+       << L"    link.exe__LINK_LIMITER.exe\n"
+       << L"    lib.exe__LINK_LIMITER.exe\n"
+       << L"      * Both will limit on the same semaphore\n"
+       << L"\n"
+       << L"    link.exe__LINK_LIMITER.exe\n"
+       << L"    lib.exe__LIB_LIMITER.exe\n"
+       << L"      * Both will limit on independent semaphores\n"
+       << L"\n"
+       << L"  This program is meant to be run after renaming it into the\n"
+       << L"  above format. Once you have done so, executing it will block\n"
+       << L"  on the availability of the semaphore SEMAPHORE_NAME. Once\n"
+       << L"  the semaphore is obtained, it will execute SHIMMED_NAME, \n"
+       << L"  passing through all arguments as-is.\n"
+       << L"\n"
+       << L"  The maximum concurrency can be manually set by setting the\n"
+       << L"  environment variable <SEMAPHORE_NAME>_MAXCONCURRENCY to an\n"
+       << L"  integer value (1, 254).\n"
+       << L"    * This value must be set the same for ALL invocations.\n"
+       << L"    * If the value is not set, it defaults to (num_cores-1).\n"
+       << L"\n"
+       << L"  The semaphore is automatically released when the program\n"
+       << L"  completes normally, OR if the program crashes (or even if\n"
+       << L"  limiter itself crashes).\n";
+  Error(strm.str().c_str());
+  exit(-1);
+}
+
+// Input command line is assumed to be of the form:
+//
+// thing.exe__PIPE_NAME.exe ...
+//
+// Specifically, wait for a semaphore (whose concurrency is specified by
+// LIMITER_MAXCONCURRENT), and then pass through everything once we have
+// acquired the semaphore.
+//
+// argv[0] is parsed for:
+//   * exe_to_shim_including_extension.exe
+//     * This could also be a bat or com. Anything that CreateProcess will
+//       accept.
+//   * "__"
+//     * We search for this separator from the end of argv[0], so the exe name
+//       could contain a double underscore if necessary.
+//   * PIPE_NAME
+//     * Can only contain single underscores, not a double underscore.
+//     * i.e. HELLO_WORLD_PIPE will work, but HELLO__WORLD_PIPE will not.
+//     * This would allow the shimmed exe to contain arbitrary numbers of
+//       underscores. We control the pipe name, but not necessarily the thing
+//       we're shimming.
+//
+int wmain(int, wchar_t** argv) {
+  tstring shimmed_plus_pipename = argv[0];
+  size_t last_slash = shimmed_plus_pipename.find_last_of(L"/\\");
+  if (last_slash != tstring::npos) {
+    shimmed_plus_pipename = shimmed_plus_pipename.substr(last_slash+1);
+  }
+
+  size_t separator = shimmed_plus_pipename.rfind(L"__");
+  if (separator == tstring::npos) {
+    usage(L"Cannot parse argv[0]. No '__' found. "
+          L"Should be like '[...(\\|/)]link.exe__PIPE_NAME.exe'");
+  }
+  tstring shimmed_exe   = shimmed_plus_pipename.substr(0, separator);
+  tstring base_pipename = shimmed_plus_pipename.substr(separator+2);
+
+  size_t dot = base_pipename.find(L'.');

[cmp, 2012/08/01 23:36:42]

for consistency, use " here

[iannucci1, 2012/08/01 23:52:39]

The implementation of find is (marginally) more efficient for chars than for
c-strings. Make it a string for real?

[cmp, 2012/08/02 00:04:23]

Nah, character here, then.  Can you document that before line 329?

+  if (dot == tstring::npos) {
+    usage(L"Hunh? No '.' in argv[0]? Are we on windows?");

[cmp, 2012/08/01 23:36:42]

can you make this more clear?  (eg. 'ERROR: Required . in argv[0], but none
found.')

[iannucci1, 2012/08/01 23:52:39]

Well... Since argv[0] must be the executable name, and this program can only run
on windows, this is (completely?) impossible, since windows can't execute
binaries which don't have an extension (AFAIK).

Maybe, "Expected an executable extension in argv[0]. No '.' found." ?

[cmp, 2012/08/02 00:04:23]

SGTM

+  }
+  base_pipename = base_pipename.substr(0, dot);
+
+  return wait_and_run(shimmed_exe, base_pipename);
+}
+

[cmp, 2012/08/01 23:36:42]

nit: remove empty line