Add statistics about free space to microdump format.

src/client/linux/microdump_writer/microdump_writer.cc

Index: src/client/linux/microdump_writer/microdump_writer.cc
diff --git a/src/client/linux/microdump_writer/microdump_writer.cc b/src/client/linux/microdump_writer/microdump_writer.cc
index d459d9ec441370c2e06e2e2f9a365bedf10e6353..7b1f1a9f2f2cc932192c7028c9ebb4f86d84b0f1 100644
--- a/src/client/linux/microdump_writer/microdump_writer.cc
+++ b/src/client/linux/microdump_writer/microdump_writer.cc
@@ -34,8 +34,6 @@
 
 #include <sys/utsname.h>
 
-#include <algorithm>
-
 #include "client/linux/dump_writer_common/thread_info.h"
 #include "client/linux/dump_writer_common/ucontext_reader.h"
 #include "client/linux/handler/exception_handler.h"
@@ -44,6 +42,7 @@
 #include "client/linux/minidump_writer/linux_ptrace_dumper.h"
 #include "common/linux/file_id.h"
 #include "common/linux/linux_libc_support.h"
+#include "common/memory.h"
 
 namespace {
 
@@ -61,6 +60,44 @@ using google_breakpad::UContextReader;
 
 const size_t kLineBufferSize = 2048;
 
+int Log2Floor(uint64_t n) {

[Primiano Tucci (use gerrit), 2016/05/06 13:58:38]

Maybe just add a comment saying:
// From chromium src/base/bits.h

actually this seems a bit different from the base one. Any reason for not
keeping that?
The reason why I asked to take the one from base is because I trusted that to be
covered and tested. I try to introduce as less new code as possible in breakpad,
because breakpad is one of those things extremely hard to detect and debug if
something goes wrong.

I just tested this out of curiosity (sorry I am a super paranoid person as
usual):
http://pastebin.com/YiRmTJ5C

This function takes 9 seconds to compute the Log2Floor(128) (as opposite to base
one which is a bunch of loops).
And if you build that with -O2 this seems to never terminate.
Which is a good way to lose all the crashes :)

+  if (!n) return -1;
+  int log = 0;
+  for (int i = 5; i >= 0; ++i) {
+    int shift = 1 << i;
+    if (n >> shift) {
+      log += shift;
+      if (!(n >>= shift)) break;
+    }
+  }
+  return log;
+}
+
+bool MappingsAreAdjacent(const MappingInfo* a, const MappingInfo* b) {
+  // Because of load biasing, we can end up with a situation where two
+  // mappings actually overlap. So we will define adjacency to also include a
+  // b start address that lies within a's address range (including starting
+  // immediately after a).
+  // Because load biasing only ever moves the start address backwards, the end
+  // address should still increase.
+  return a->start_addr <= b->start_addr &&
+         a->start_addr + a->size >= b->start_addr;
+}
+
+size_t NextOrderedMapping(
+    const google_breakpad::wasteful_vector<MappingInfo*>& mappings,
+    size_t curr) {
+  size_t best = mappings.size();
+  for (size_t next = 0; next < mappings.size(); ++next) {
+    if (mappings[curr]->start_addr < mappings[next]->start_addr &&
+        (best == mappings.size() ||
+         mappings[next]->start_addr < mappings[best]->start_addr)) {
+      best = next;
+    }
+  }
+  return best;
+}
+
 class MicrodumpWriter {
  public:
   MicrodumpWriter(const ExceptionHandler::CrashContext* context,
@@ -98,6 +135,7 @@ class MicrodumpWriter {
     DumpProductInformation();
     DumpOSInformation();
     DumpGPUInformation();
+    DumpFreeSpace();
     success = DumpCrashingThread();
     if (success)
       success = DumpMappings();
@@ -128,16 +166,21 @@ class MicrodumpWriter {
   }
 
   // Stages the hex repr. of the given int type in the current line buffer.
-  template<typename T>
-  void LogAppend(T value) {
+  enum LeadingZeros { kKeepLeadingZeros, kDiscardLeadingZeros };
+  template <typename T>
+  void LogAppend(T value, const LeadingZeros leading_zeros =
+                              LeadingZeros::kKeepLeadingZeros) {
     // Make enough room to hex encode the largest int type + NUL.
     static const char HEX[] = {'0', '1', '2', '3', '4', '5', '6', '7', '8', '9',
                                'A', 'B', 'C', 'D', 'E', 'F'};
     char hexstr[sizeof(T) * 2 + 1];
-    for (int i = sizeof(T) * 2 - 1; i >= 0; --i, value >>= 4)
-      hexstr[i] = HEX[static_cast<uint8_t>(value) & 0x0F];
-    hexstr[sizeof(T) * 2] = '\0';
-    LogAppend(hexstr);
+    int i = sizeof(T) * 2;
+    hexstr[i] = '\0';
+    do {
+      hexstr[--i] = HEX[static_cast<uint8_t>(value) & 0x0F];
+      value >>= 4;
+    } while (value || leading_zeros == kKeepLeadingZeros && i > 0);
+    LogAppend(hexstr + i);
   }
 
   // Stages the buffer content hex-encoded in the current line buffer.
@@ -391,6 +434,78 @@ class MicrodumpWriter {
     LogCommitLine();
   }
 
+  bool DumpFreeSpace() {
+    const google_breakpad::wasteful_vector<MappingInfo*>& mappings =
+        dumper_->mappings();
+    if (mappings.size() == 0) return false;
+
+    // This is complicated by the fact that mappings is not in order. It should
+    // be mostly in order, however the mapping that contains the entry point for
+    // the process is always at the front of the vector.
+
+    static const int HBITS = sizeof(size_t) * 8;
+    unsigned int hole_histogram[HBITS];
+    my_memset(hole_histogram, 0, sizeof(hole_histogram));
+
+    // Find the lowest address mapping.
+    size_t curr = 0;
+    for (size_t i = 1; i < mappings.size(); ++i) {
+      if (mappings[i]->start_addr < mappings[curr]->start_addr) curr = i;
+    }
+
+    uintptr_t lo_addr = mappings[curr]->start_addr;
+
+    unsigned int hole_cnt = 0;
+    size_t hole_max = 0;
+    size_t hole_sum = 0;
+
+    while (true) {
+      // Skip to the end of an adjacent run of mappings. This is an optimization
+      // for the fact that mappings is mostly sorted.
+      while (curr != mappings.size() - 1 &&
+             MappingsAreAdjacent(mappings[curr], mappings[curr + 1]))
+        ++curr;
+
+      size_t next = NextOrderedMapping(mappings, curr);
+      if (next == mappings.size())
+        break;
+
+      uintptr_t hole_lo = mappings[curr]->start_addr + mappings[curr]->size;
+      uintptr_t hole_hi = mappings[next]->start_addr;
+
+      if (hole_hi > hole_lo) {
+        size_t hole_sz = hole_hi - hole_lo;
+        hole_sum += hole_sz;
+        hole_max = std::max(hole_sz, hole_max);
+        ++hole_cnt;
+        ++hole_histogram[Log2Floor(hole_sz)];
+      }
+      curr = next;
+    }
+
+    uintptr_t hi_addr = mappings[curr]->start_addr + mappings[curr]->size;
+
+    LogAppend("H ");
+    LogAppend(lo_addr);
+    LogAppend(" ");
+    LogAppend(hi_addr);
+    LogAppend(" ");
+    LogAppend(hole_cnt, kDiscardLeadingZeros);
+    LogAppend(" ");
+    LogAppend(hole_max, kDiscardLeadingZeros);
+    LogAppend(" ");
+    LogAppend(hole_sum, kDiscardLeadingZeros);
+    for (unsigned int i = 0; i < HBITS; ++i) {
+      if (!hole_histogram[i]) continue;
+      LogAppend(" ");
+      LogAppend(i, kDiscardLeadingZeros);
+      LogAppend(":");
+      LogAppend(hole_histogram[i], kDiscardLeadingZeros);
+    }
+    LogCommitLine();
+    return true;
+  }
+
   // Write information about the mappings in effect.
   bool DumpMappings() {
     // First write all the mappings from the dumper