Update ReadProcMaps() to reflect lack of atomicity when reading /proc/self/maps.

base/debug/proc_maps_linux.cc

 // Copyright (c) 2013 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.
 
 #include "base/debug/proc_maps_linux.h"
 
+#include <fcntl.h>
+
 #if defined(OS_LINUX)
 #include <inttypes.h>
 #endif
 
 #include "base/file_util.h"
 #include "base/strings/string_split.h"
 
 #if defined(OS_ANDROID)
 // Bionic's inttypes.h defines PRI/SCNxPTR as an unsigned long int, which
 // is incompatible with Bionic's stdint.h defining uintptr_t as a unsigned int:
 // https://code.google.com/p/android/issues/detail?id=57218
 #undef SCNxPTR
 #define SCNxPTR "x"
 #endif
 
 namespace base {
 namespace debug {
 
+// Scans |proc_maps| starting from |pos| returning true if the gate VMA was
+// found, otherwise returns false.
+static bool ContainsGateVMA(std::string* proc_maps, size_t pos) {
+#if defined(ARCH_CPU_ARM_FAMILY)
+  // The gate VMA on ARM kernels is the interrupt vectors page.
+  return proc_maps->find(" [vectors]\n", pos) != std::string::npos;
+#elif defined(ARCH_CPU_X86_64)
+  // The gate VMA on x86 64-bit kernels is the virtual system call page.
+  return proc_maps->find(" [vsyscall]\n", pos) != std::string::npos;
+#else
+  // Otherwise assume there is no gate VMA in which case we shouldn't
+  // get duplicate entires.
+  return false;
+#endif
+}
+
 bool ReadProcMaps(std::string* proc_maps) {
-  FilePath proc_maps_path("/proc/self/maps");
-  return ReadFileToString(proc_maps_path, proc_maps);
+  // seq_file only writes out a page-sized amount on each call. Refer to header
+  // file for details.
+  const long kBufferSize = sysconf(_SC_PAGESIZE);
+  scoped_ptr<char[]> buffer(new char[kBufferSize]);
+
+  int fd = HANDLE_EINTR(open("/proc/self/maps", O_RDONLY));
+  if (fd == -1) {
+    DPLOG(ERROR) << "Couldn't open /proc/self/maps";
+    return false;
+  }
+  file_util::ScopedFD fd_closer(&fd);
+  proc_maps->clear();
+
+  while (true) {
+    ssize_t bytes_read = HANDLE_EINTR(read(fd, buffer.get(), kBufferSize));

[Mark Mentovai, 2013/09/05 17:52:43]

I suppose it doesn’t really matter for this non-performance-critical code, but
you could save the copy (in vector<>::append) and get rid of the distinct buffer
variable by using proc_maps here. You’d just have to extend it by kBufferSize
prior to the read and then resize it to the actual size after the read.

[scherkus (not reviewing), 2013/09/05 19:01:40]

I like it. Done.

+    if (bytes_read < 0) {
+      DPLOG(ERROR) << "Couldn't read /proc/self/maps";
+      return false;
+    }
+
+    if (bytes_read == 0)
+      break;
+
+    // Use 2-arg version of append() as |buffer| *is not* NUL-terminated. This
+    // also avoids an unnecessary length computation.
+    size_t pos = proc_maps->size();
+    proc_maps->append(buffer.get(), bytes_read);
+
+    // The gate VMA is handled as a special case after seq_file has finished
+    // iterating through all entries in the virtual memory table.
+    //
+    // Unfortunately, if additional entries are added at this point in time
+    // seq_file gets confused and the next call to read() will return duplicate
+    // entries including the gate VMA again.
+    //
+    // Avoid this by searching for the gate VMA and breaking early.
+    if (ContainsGateVMA(proc_maps, pos))
+      break;
+  }
+
+  return true;
 }
 
 bool ParseProcMaps(const std::string& input,
                    std::vector<MappedMemoryRegion>* regions_out) {
   std::vector<MappedMemoryRegion> regions;
 
   // This isn't async safe nor terribly efficient, but it doesn't need to be at
   // this point in time.
   std::vector<std::string> lines;
   SplitString(input, '\n', &lines);
@@ skipping 54 matching lines @@
     regions.push_back(region);
     regions.back().path.assign(line + path_index);
   }
 
   regions_out->swap(regions);
   return true;
 }
 
 }  // namespace debug
 }  // namespace base