Support delayed allocations from persistent memory.

base/metrics/persistent_memory_allocator.h

Index: base/metrics/persistent_memory_allocator.h
diff --git a/base/metrics/persistent_memory_allocator.h b/base/metrics/persistent_memory_allocator.h
index 94a7744bfb2714d74ae956b41e1ceff7483176e2..d52b6e9b32fbf6e4407852b7d0e4880dbe070c52 100644
--- a/base/metrics/persistent_memory_allocator.h
+++ b/base/metrics/persistent_memory_allocator.h
@@ -767,6 +767,96 @@ class BASE_EXPORT FilePersistentMemoryAllocator
 };
 #endif  // !defined(OS_NACL)
 
+// An allocation that is defined but not executed until required at a later
+// time. This allows for potential users of an allocation to be decoupled
+// from the logic that defines it. In addition, there can be multiple users
+// of the same allocation or any region thereof that are guaranteed to always
+// use the same space. It's okay to copy/move these objects.
+//
+// This is a top-level class instead of an inner class of the PMA so that it
+// can be forward-declared in other header files without the need to include
+// the full contents of this file.
+class BASE_EXPORT DelayedPersistentAllocation {
+ public:
+  using Reference = PersistentMemoryAllocator::Reference;
+
+  // Creates a delayed allocation using the specified |allocator|. When
+  // needed, the memory will be allocated using the specified |type| and
+  // |size|. If |offset| is given, the returned pointer will be at that

[Alexei Svitkine (slow), 2017/04/12 17:04:25]

Please expand comment to make explicit whether |size| includes offset or
doesn't.

[bcwhite, 2017/04/12 21:23:43]

Done.

+  // offset into the segment; this allows combining allocations into a
+  // single persistent segment to reduce overhead and means as "all or

[Alexei Svitkine (slow), 2017/04/12 17:04:25]

Nit: "means as" -> "means an"

[bcwhite, 2017/04/12 21:23:42]

Done.

+  // nothing" request.
+  //
+  // Once allocated, a reference to the segment will be stored at |ref|.
+  // This shared location must be initialized to zero (0); it is checked
+  // with every Get() request to see if the allocation has already been
+  // done.
+  //
+  // For conveience, methods taking both Atomic32 and std::atomic<Reference>

[Alexei Svitkine (slow), 2017/04/12 17:04:25]

Nit: convenience

[bcwhite, 2017/04/12 21:23:42]

Done.

+  // are defined.
+  DelayedPersistentAllocation(PersistentMemoryAllocator* allocator,
+                              subtle::Atomic32* ref,
+                              uint32_t type,
+                              size_t size);
+  DelayedPersistentAllocation(PersistentMemoryAllocator* allocator,
+                              subtle::Atomic32* ref,
+                              uint32_t type,
+                              size_t size,
+                              size_t offset);
+  DelayedPersistentAllocation(PersistentMemoryAllocator* allocator,
+                              std::atomic<Reference>* ref,
+                              uint32_t type,
+                              size_t size);
+  DelayedPersistentAllocation(PersistentMemoryAllocator* allocator,
+                              std::atomic<Reference>* ref,
+                              uint32_t type,
+                              size_t size,
+                              size_t offset);
+  ~DelayedPersistentAllocation();
+
+  // Make this allocation "iterable" when realized.
+  void MakeIterable();
+
+  // Gets a pointer to the defined allocation. This will realize the request
+  // and update the reference provided during construction. The memory will
+  // be zeroed the first time it is returned, after that it is shared with
+  // all other Get() requests and so shows any changes made to it elsewhere.
+  //
+  // If the allocation fails for any reason, null will be returned. This works
+  // even on "const" objects because the allocation is already defined, just
+  // delayed.
+  void* Get() const;
+
+  // Gets the internal reference value. If this returns a non-zero value then
+  // a subsequent call to Get() will do nothing but convert that reference into
+  // a memory location -- useful for accessing an existing allocation without
+  // creating one unnecessarily.
+  Reference reference() const {
+    return reference_->load(std::memory_order_relaxed);
+  }
+
+ private:
+  // The underlying object that does the actual allocation of memory. Its
+  // lifetime must exceed that of all DelayedPersistentAllocation objects
+  // that use it.
+  PersistentMemoryAllocator* const allocator_;
+
+  // The desired type and size of the allocated segment plus the offset
+  // within it for the defined request.
+  uint32_t const type_;

[Alexei Svitkine (slow), 2017/04/12 17:04:25]

Nit: I think for primitive types, it's more common to put const at the front.
Let's use that style for consistency with more code.

[bcwhite, 2017/04/12 21:23:42]

Done.

+  size_t const size_;
+  size_t const offset_;
+
+  // The location at which a reference to the allocated segment is to be
+  // stored once the allocation is complete. If multiple delayed allocations
+  // share the same pointer then an allocation on one will amount to an
+  // allocation for all.
+  volatile std::atomic<Reference>* const reference_;
+
+  // Flag indicating if allocation should be made iterable when done.
+  bool make_iterable_ = false;
+};

[Alexei Svitkine (slow), 2017/04/12 17:04:25]

DISALLOW_COPY_AND_ASSIGN()?

[bcwhite, 2017/04/12 21:23:42]

No, it's explicitly allowed and safe.  This makes it easy to pass it as a
reference and copy it to the final destination.

[Alexei Svitkine (slow), 2017/04/21 15:14:16]

Okay, add a comment about it then in the same place where that macro would be.
:)

[bcwhite, 2017/04/21 15:59:04]

Done.

+
 }  // namespace base
 
 #endif  // BASE_METRICS_PERSISTENT_MEMORY_ALLOCATOR_H_