Support "shared" histograms between processes.

base/metrics/histogram_persistence.cc

+// Copyright (c) 2015 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/metrics/histogram_persistence.h"
+
+#include "base/logging.h"
+#include "base/memory/scoped_ptr.h"
+#include "base/metrics/histogram.h"
+#include "base/metrics/histogram_base.h"
+#include "base/metrics/histogram_samples.h"
+#include "base/metrics/statistics_recorder.h"
+#include "base/synchronization/lock.h"
+
+namespace base {
+
+const Feature kPersistentHistogramsFeature = {
+  "PersistentMetrics", FEATURE_DISABLED_BY_DEFAULT
+};
+
+// Type identifiers used when storing in persistent memory so they can be
+// identified during extraction. A "version number" is added to the base
+// so that, if the structure of that object changes, stored older versions
+// will be safely ignored.
+enum : uint32_t {
+  kTypeIdHistogram   = 0xF1645910 + 1,  // SHA1(Histogram) v1
+  kTypeIdRangesArray = 0xBCEA225A + 1,  // SHA1(RangesArray) v1
+  kTypeIdCountsArray = 0x53215530 + 1,  // SHA1(CountsArray) v1
+};
+
+// This data must be held in persistent memory in order for processes to
+// locate and use histograms created elsewhere.
+struct PersistentHistogramData {
+  int histogram_type;
+  int flags;
+  int minimum;
+  int maximum;
+  size_t bucket_count;
+  PersistentMemoryAllocator::Reference ranges_ref;
+  uint32_t ranges_checksum;
+  PersistentMemoryAllocator::Reference counts_ref;
+  HistogramSamples::Metadata samples_metadata;
+
+  // Space for the histogram name will be added during the actual allocation
+  // request. This must be the last field of the structure. A zero-size array
+  // or a "flexible" array would be preferred but is not (yet) valid C++.
+  char name[1];
+};
+
+scoped_ptr<PersistentMemoryAllocator> allocator_;

[Alexei Svitkine (slow), 2016/01/08 18:52:48]

If this won't be used outside of this file, it should be in an anon namespace.
Same for the things above.

Also, conventions for naming such things is g_allocator.

Finally, I think this would create a static initializer. You may need to use a
base::LazyInstance here, since static initializers are not allowed.

But seems like you're also declaring it in the header. That doesn't look right.

How about declaring a new class - e.g. PersistentHistogramAllocator that has all
these currently free-standing functions as members and itself holding a member
PersistentMemoryAllocator. Then, the PersistentHistogramAllocator can be held by
the StatisticsRecorder.

[bcwhite, 2016/01/12 16:34:57]

There didn't seem any difference between a class with only static methods and
global functions.  I can change it if you want but don't see any benefit in it.

The "visibility" issue is annoying.  CreatePersistentHistogram needs to be
defined publicly because it has to be a friend of the Histogram classes in order
to call their non-public constructors.  The PesistentHistogramData class is only
visible because it's passed to CreatePersistentHistogram from within
histogram_persistence.cc.

Let me see what I can do to clean that up...

[Alexei Svitkine (slow), 2016/01/12 21:54:44]

I guess the main benefit is it allows you to have multiple such allocators -
which provides flexibility in terms of use and likely testing. However, if
keeping instance local to the .cc file works and we can hide everything behind
the static functions, I guess I'm OK with this approach initially. We can always
refactor if we find we need the extra flexibility. Thanks.

+
+void SetPersistentHistogramMemoryAllocator(
+    PersistentMemoryAllocator* allocator) {
+  // Releasing or changing an allocator is extremely dangerous because it
+  // likely has histograms stored within it. If the backing memory is also
+  // also released, future accesses to those histograms will seg-fault.
+  // It's not a fatal CHECK() because tests do this knowing that all
+  // such persistent histograms have already been forgotten.
+  if (allocator_) {
+    LOG(WARNING) << "Active PersistentMemoryAllocator has been released."
+                 << " Some existing histogram pointers may be invalid.";
+  }
+  allocator_.reset(allocator);
+}
+
+PersistentMemoryAllocator* GetPersistentHistogramMemoryAllocator() {
+  return allocator_.get();
+}
+
+PersistentMemoryAllocator* ReleasePersistentHistogramMemoryAllocator() {
+  return allocator_.release();
+};
+
+// Extract a histogram from persistent memory. Unfortunately, the above "pickle"
+// methods cannot be used as part of the persistance because the deserialization
+// methods always create local count data (these must referenced the persistent
+// counts) and always add it to the local list of known histograms (these may
+// be simple references to histograms in other processes).

[Alexei Svitkine (slow), 2016/01/08 18:52:48]

There shouldn't be both a comment in the .h and in the .cc file above the
function. If you want to talk more about the implementation, the comment should
be within the function.

[bcwhite, 2016/01/13 13:38:29]

Done.

+HistogramBase* GetPersistentHistogram(
+    PersistentMemoryAllocator* allocator,
+    int32_t ref) {
+  PersistentHistogramData* histogram_data =
+      allocator->GetAsObject<PersistentHistogramData>(ref, kTypeIdHistogram);
+  size_t length = allocator->GetAllocSize(ref);
+  if (!histogram_data ||
+      reinterpret_cast<char*>(histogram_data)[length - 1] != '\0') {
+    LOG(WARNING) << "Persistent histogram data was invalid or had invalid name"
+                 << "; skipped.";
+    NOTREACHED();
+    return nullptr;
+  }
+  return CreatePersistentHistogram(allocator, histogram_data);
+}
+
+HistogramBase* GetNextPersistentHistogram(
+    PersistentMemoryAllocator* allocator,
+    PersistentMemoryAllocator::Iterator* iter) {
+  PersistentMemoryAllocator::Reference ref;
+  uint32_t type_id;
+  while ((ref = allocator->GetNextIterable(iter, &type_id)) != 0) {
+    if (type_id == kTypeIdHistogram)
+      return GetPersistentHistogram(allocator, ref);
+  }
+  return nullptr;
+}
+
+HistogramBase* CreatePersistentHistogram(
+    PersistentMemoryAllocator* allocator,
+    PersistentHistogramData* histogram_data_ptr) {
+  if (!histogram_data_ptr) {
+    LOG(WARNING) << "Persistent histogram data was not valid; skipped.";
+    NOTREACHED();
+    return nullptr;
+  }
+
+  // Copy the histogram_data to local storage because anything in persistent
+  // memory cannot be trusted as it could be changed at any moment by a
+  // malicious actor that shares access. The contents of histogram_data are
+  // validated below; the local copy is to ensure that the contents cannot
+  // be externally changed between validation and use.
+  PersistentHistogramData histogram_data = *histogram_data_ptr;
+  std::string name(histogram_data_ptr->name);
+
+  HistogramBase::Sample* ranges_data =
+      allocator->GetAsObject<HistogramBase::Sample>(histogram_data.ranges_ref,
+                                                    kTypeIdRangesArray);
+  if (!ranges_data || histogram_data.bucket_count < 2 ||
+      histogram_data.bucket_count + 1 >
+          std::numeric_limits<size_t>::max() / sizeof(HistogramBase::Sample) ||
+      allocator->GetAllocSize(histogram_data.ranges_ref) <
+          (histogram_data.bucket_count + 1) * sizeof(HistogramBase::Sample)) {
+    LOG(WARNING) << "Persistent histogram referenced invalid ranges array"
+                 << "; skipped.";
+    NOTREACHED();
+    return nullptr;
+  }
+  // To avoid racy destruction at shutdown, the following will be leaked.
+  BucketRanges* ranges = new BucketRanges(histogram_data.bucket_count + 1);
+  bool bad_ranges = false;
+  for (size_t i = 0; i < ranges->size(); ++i) {
+    if (i > 0 && ranges_data[i] <= ranges_data[i - 1])
+      bad_ranges = true;
+    ranges->set_range(i, ranges_data[i]);
+  }
+  ranges->ResetChecksum();
+  if (bad_ranges || ranges->checksum() != histogram_data.ranges_checksum) {
+    LOG(WARNING) << "Persistent histogram referenced invalid ranges array"
+                 << "; skipped.";
+    NOTREACHED();
+    return nullptr;
+  }
+  const BucketRanges* registered_ranges =
+      StatisticsRecorder::RegisterOrDeleteDuplicateRanges(ranges);
+
+  HistogramBase::AtomicCount* counts_data =
+      allocator->GetAsObject<HistogramBase::AtomicCount>(
+          histogram_data.counts_ref, kTypeIdCountsArray);
+  if (!counts_data ||
+      allocator->GetAllocSize(histogram_data.counts_ref) <
+          histogram_data.bucket_count * sizeof(HistogramBase::AtomicCount)) {
+    LOG(WARNING) << "Persistent histogram referenced invalid counts array"
+                 << "; skipped.";
+    NOTREACHED();
+    return nullptr;
+  }
+
+  HistogramBase* histogram = nullptr;
+  switch (histogram_data.histogram_type) {
+  case HISTOGRAM:
+    histogram = new Histogram(

[Alexei Svitkine (slow), 2016/01/08 18:52:48]

I don't like that all the classes need to be friends with this function. Can
this be avoided? e.g. by this class taking a pointer to a Factory object? The
Factory object can then be responsible for calling the specific ctors.

[bcwhite, 2016/01/12 16:34:57]

The Factory object, like the FactoryGet methods, are what call this method (and
do other preparation).  I could create new FactoryGet methods that take all
these parameters and make them public; that would solve the visibility issue and
thus clean up other things.

Okay with that?

+        name,
+        histogram_data.minimum,
+        histogram_data.maximum,
+        registered_ranges,
+        counts_data,
+        histogram_data.bucket_count,
+        &histogram_data_ptr->samples_metadata);
+    break;
+  case LINEAR_HISTOGRAM:
+    histogram = new LinearHistogram(
+        name,
+        histogram_data.minimum,
+        histogram_data.maximum,
+        registered_ranges,
+        counts_data,
+        histogram_data.bucket_count,
+        &histogram_data_ptr->samples_metadata);
+    break;
+  case BOOLEAN_HISTOGRAM:
+    histogram = new BooleanHistogram(
+        name,
+        registered_ranges,
+        counts_data,
+        &histogram_data_ptr->samples_metadata);
+    break;
+  case CUSTOM_HISTOGRAM:
+    histogram = new CustomHistogram(
+        name,
+        registered_ranges,
+        counts_data,
+        histogram_data.bucket_count,
+        &histogram_data_ptr->samples_metadata);
+    break;
+  }
+
+  if (histogram) {
+    DCHECK_EQ(histogram_data.histogram_type, histogram->GetHistogramType());
+    histogram->SetFlags(histogram_data.flags);
+  }
+
+  return histogram;
+}
+
+void FinalizePersistentHistogram(PersistentMemoryAllocator::Reference ref,
+                                 bool registered) {
+  // If the created persistent histogram was registered then it needs to
+  // be marked as "iterable" in order to be found by other processes.
+  if (registered)
+    GetPersistentHistogramMemoryAllocator()->MakeIterable(ref);
+  // If it wasn't registered then a race condition must have caused
+  // two to be created. The allocator does not support releasing the
+  // acquired memory so just change the type to be empty.
+  else
+    GetPersistentHistogramMemoryAllocator()->SetType(ref, 0);
+}
+
+HistogramBase* AllocatePersistentHistogram(
+    PersistentMemoryAllocator* allocator,
+    HistogramType histogram_type,
+    const std::string& name,
+    int minimum,
+    int maximum,
+    const BucketRanges* bucket_ranges,
+    int32_t flags,
+    PersistentMemoryAllocator::Reference* ref_ptr) {
+  if (allocator) {
+    size_t bucket_count = bucket_ranges->bucket_count();
+    CHECK(bucket_count <= std::numeric_limits<int32_t>::max() /
+                              sizeof(HistogramBase::AtomicCount));
+    size_t counts_memory = bucket_count * sizeof(HistogramBase::AtomicCount);
+    size_t ranges_memory = (bucket_count + 1) * sizeof(HistogramBase::Sample);
+    PersistentMemoryAllocator::Reference ranges_ref =
+        allocator->Allocate(ranges_memory, kTypeIdRangesArray);
+    PersistentMemoryAllocator::Reference counts_ref =
+        allocator->Allocate(counts_memory, kTypeIdCountsArray);
+    PersistentMemoryAllocator::Reference histogram_ref =
+        allocator->Allocate(offsetof(PersistentHistogramData, name) +
+                            name.length() + 1, kTypeIdHistogram);
+    HistogramBase::Sample* ranges_data =
+        allocator->GetAsObject<HistogramBase::Sample>(ranges_ref,
+                                                      kTypeIdRangesArray);
+    PersistentHistogramData* histogram_data =
+        allocator->GetAsObject<PersistentHistogramData>(histogram_ref,
+                                                        kTypeIdHistogram);
+
+    // Only continue here if all allocations were successful. If they weren't
+    // there is no way to free the space but that's not really a problem since
+    // the allocations only fail because the space is full and so any future
+    // attempts will also fail.
+    if (counts_ref && ranges_data && histogram_data) {
+      strcpy(histogram_data->name, name.c_str());
+      for (size_t i = 0; i < bucket_ranges->size(); ++i)
+        ranges_data[i] = bucket_ranges->range(i);
+
+      histogram_data->histogram_type = histogram_type;
+      histogram_data->flags = flags;
+      histogram_data->minimum = minimum;
+      histogram_data->maximum = maximum;
+      histogram_data->bucket_count = bucket_count;
+      histogram_data->ranges_ref = ranges_ref;
+      histogram_data->ranges_checksum = bucket_ranges->checksum();
+      histogram_data->counts_ref = counts_ref;
+
+      // Create the histogram using resources in persistent memory. This ends up
+      // resolving the "ref" values stored in histogram_data instad of just
+      // using what is already known above but avoids duplicating the switch
+      // statement here and serves as a double-check that everything is
+      // correct before commiting the new histogram to persistent space.
+      HistogramBase* histogram =
+          CreatePersistentHistogram(allocator, histogram_data);
+      DCHECK(histogram);
+      if (ref_ptr != nullptr)
+        *ref_ptr = histogram_ref;
+      return histogram;
+    }
+
+    LOG(WARNING) << "Could not create histogram \"" << name
+                 << "\" in persistent memory (full=" << allocator->IsFull()
+                 << ", corrupt=" << allocator->IsCorrupt() << ")";
+  }
+
+  return nullptr;
+}
+
+void ImportPersistentHistograms() {
+  // Each call resumes from where it last left off so need persistant iterator.
+  // The lock protects against concurrent access to the iterator.
+  static PersistentMemoryAllocator::Iterator iter;
+  static base::Lock lock;
+
+  if (allocator_) {
+    base::AutoLock auto_lock(lock);
+    if (iter.is_clear())
+      allocator_->CreateIterator(&iter);
+
+    for (;;) {
+      HistogramBase* h = GetNextPersistentHistogram(allocator_.get(), &iter);
+      if (!h)
+        break;
+      StatisticsRecorder::RegisterOrDeleteDuplicate(h);
+    }
+  }
+}
+
+}  // namespace base