Create "persistent memory allocator" for persisting and sharing objects.

base/memory/shared_memory_allocator.cc

Index: base/memory/shared_memory_allocator.cc
diff --git a/base/memory/shared_memory_allocator.cc b/base/memory/shared_memory_allocator.cc
new file mode 100644
index 0000000000000000000000000000000000000000..f1f55bfebf544b4e8f69dd41e2712bb92bb5b1b7
--- /dev/null
+++ b/base/memory/shared_memory_allocator.cc
@@ -0,0 +1,417 @@
+// 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/memory/shared_memory_allocator.h"
+
+#include <assert.h>
+
+#include "base/logging.h"
+
+// All integer constants in this file are signed because Atomic32 is signed
+// and keeping all others consistent with this avoids a lot of unnecessary
+// casting to avoid signed/unsigned operations just to avoid compiler errors.
+// This means an occasonal cast of a constant from sizeof() to "int" but
+// is far simpler than the alternative.
+
+namespace {
+
+// All allocations and data-structures must be aligned to this byte boundary.
+// It shouldn't be less than 8 so that 64-bit values can be read in a single

[Dmitry Vyukov, 2015/11/03 14:06:46]

What architecture do you have in mind? And what about cache?
In C++ proper formulation of this would be "... to ensure proper alignment for
all types".

[bcwhite, 2015/11/03 16:28:20]

Comment expanded and updated.  I don't have a particular architecture in mind
other than 64-bit but cache is a consideration.  I originally had an alignment
of 16 so at least the block header is always within a cache line but it was felt
by others that it wasn't worth the extra ~4 bytes of wasted padding.

+// RAM bus access.  16 can be used so that the block header would always fall
+// within a single cache line.
+const int32_t kAllocAlignment = 8;
+
+// A constant (random) value placed in the shared metadata to identify
+// an already initialized memory segment.
+const int32_t kGlobalCookie = 0x408305DC;
+
+// The current version of the metadata.  If updates are made that change
+// the metadata, the version number can be queried to operate in a backward-
+// compatible manner until the memory segment is completely re-initalized.
+const int32_t kGlobalVersion = 1;
+
+// Constant values placed in the block headers to indicate its state.
+const int32_t kBlockCookieFree = 0;
+const int32_t kBlockCookieQueue = 1;
+const int32_t kBlockCookieWasted = -1;
+const int32_t kBlockCookieAllocated = 0xC8799269;
+
+// TODO(bcwhite): When acceptable, consider moving flags to std::atomic<char>
+// types rather than combined bitfield.
+
+enum {
+  kFlagCorrupted,
+  kFlagFull
+};
+
+bool CheckFlag(base::subtle::Atomic32* flags, int flag) {
+  base::subtle::Atomic32 loaded_flags = base::subtle::NoBarrier_Load(flags);

[Alexander Potapenko, 2015/11/03 08:12:54]

It's not immediately evident whether the currently existing flags require any
barriers.
But to be sure nothing breaks in the future, SetFlag() and CheckFlag() must
probably constitute a happens-before arc.

These two flags (or other flags introduced later) may be used to signal that
some bits of metadata have been changed, i.e.:
1. The user modifies the metadata and sets the corresponding flag
2. He then checks the flag value in another thread.
3. In the case the flag is true, the user assumes the metadata bits have already
been changed.

This assumption is quite natural, but is only possible if the flag is written
using a Release_Store (it already is) and read using an Acquire_Load(). (See
base::CancellationFlag, which has similar semantics)

[bcwhite, 2015/11/03 16:28:19]

Done.

+  return (loaded_flags & 1 << flag) != 0;
+}
+
+void SetFlag(base::subtle::Atomic32* flags, int flag, bool set) {

[Dmitry Vyukov, 2015/11/03 14:06:45]

You never pass set=false. Please delete it. This code is complex enough without
dead code.

[bcwhite, 2015/11/03 16:28:20]

Done.

+  for (;;) {
+    base::subtle::Atomic32 loaded_flags = base::subtle::NoBarrier_Load(flags);
+    base::subtle::Atomic32 new_flags =
+        (loaded_flags & ~(1 << flag)) | (set ? 1 : 0) << flag;
+    if (base::subtle::Release_CompareAndSwap(

[Dmitry Vyukov, 2015/11/03 14:06:45]

You use NoBarrier_Load to load flags, so Release_CompareAndSwap cannot possibly
be the right thing. Release/Acquire barrier must always be paired. If you don't
need synchronization use NoBarrier to both load and set; if you need
synchronization then change NoBarrier_Load to Acquire_Load.

[bcwhite, 2015/11/03 16:28:20]

Done.

+            flags, loaded_flags, new_flags) == loaded_flags) {
+      break;
+    }
+  }
+}
+
+}  // namespace
+
+namespace base {
+
+// The block-header is placed at the top of every allocation within the
+// segment to describe the data that follows it.
+struct SharedMemoryAllocator::BlockHeader {
+  int32_t size;           // Number of bytes in this block, including header.
+  int32_t cookie;         // Constant value indicating completed allocation.
+  int32_t type;           // A number provided by caller indicating data type.
+  subtle::Atomic32 next;  // Pointer to the next block when iterating

[Dmitry Vyukov, 2015/11/03 14:06:45]

add . at the end of comment for consistency

[bcwhite, 2015/11/03 16:28:19]

Done.

+};
+
+// The shared metadata exists once at the top of the memory segment to
+// describe the state of the allocator to all processes.
+struct SharedMemoryAllocator::SharedMetadata {
+  int32_t cookie;     // Some value that indicates complete initialization.
+  int32_t size;       // Total size of memory segment.
+  int32_t page_size;  // Paging size within memory segment.
+  int32_t version;    // Version code so upgrades don't break.
+  subtle::Atomic32 freeptr;  // Offset to first free space in the segment.
+  subtle::Atomic32 flags;    // Bitfield of information flags.
+  int32_t reserved;   // Padding to ensure size is multiple of alignment.
+
+  // The "iterable" queue is an M&S Queue as described here, append-only:
+  // https://www.research.ibm.com/people/m/michael/podc-1996.pdf
+  subtle::Atomic32 tailptr;  // Last block available for iteration.
+  BlockHeader queue;  // Empty block for linked-list head/tail. (must be last)
+};
+
+// The "queue" block header is used to detect "last node" so that zero/null
+// can be used to indicate that it hasn't been added at all.  It is part of
+// the SharedMetadata structure which itself is always located at offset zero.
+// This can't be a constant because SharedMetadata is a private definition.
+#define OFFSET_QUEUE offsetof(SharedMetadata, queue)
+#define OFFSET_NULL 0  // the equivalest NULL value for an offset
+
+SharedMemoryAllocator::SharedMemoryAllocator(void* base, int32_t size,
+                                             int32_t page_size)
+  : shared_meta_(static_cast<SharedMetadata*>(base)),
+    mem_base_(static_cast<char*>(base)),
+    mem_size_(size),
+    mem_page_(page_size ? page_size : size),
+    last_seen_(0),
+    corrupted_(0) {
+  static_assert(sizeof(BlockHeader) % kAllocAlignment == 0,
+                "BlockHeader is not a multiple of kAllocAlignment");
+  static_assert(sizeof(SharedMetadata) % kAllocAlignment == 0,
+                "SharedMetadata is not a multiple of kAllocAlignment");
+
+  DCHECK(base && reinterpret_cast<uintptr_t>(base) % kAllocAlignment == 0);

[Alexander Potapenko, 2015/11/03 08:12:54]

These invariants shouldn't be checked often, should be fine to make them CHECKs.

[bcwhite, 2015/11/03 16:28:19]

Done.

+  DCHECK(size >= 1 << 10 && size <= 1 << 20 &&  // 1 KiB <= size <= 1 MiB
+         size % kAllocAlignment == 0);
+  DCHECK(page_size >= 0 && (page_size == 0 || size % page_size == 0));
+
+  if (shared_meta_->cookie != kGlobalCookie) {
+    // This block is only executed when a completely new memory segment is
+    // being initialized.  It's unshared and single-threaded...
+    const BlockHeader* first_block = reinterpret_cast<BlockHeader*>(
+        mem_base_ + sizeof(SharedMetadata));
+    if (shared_meta_->cookie != 0 ||
+        shared_meta_->size != 0 ||
+        shared_meta_->version != 0 ||
+        subtle::NoBarrier_Load(&shared_meta_->freeptr) != 0 ||
+        subtle::NoBarrier_Load(&shared_meta_->flags) != 0 ||
+        shared_meta_->tailptr != 0 ||
+        shared_meta_->queue.cookie != 0 ||
+        subtle::NoBarrier_Load(&shared_meta_->queue.next) != 0 ||
+        first_block->size != 0 ||
+        first_block->cookie != 0 ||
+        first_block->type != 0 ||
+        first_block->next != 0) {
+      // ...or something malicious has been playing with the metadata.
+      NOTREACHED();
+      SetCorrupted();
+    }
+
+    // This is still safe to do even if corruption has been detected.
+    shared_meta_->cookie = kGlobalCookie;
+    shared_meta_->size = size;
+    shared_meta_->page_size = page_size;
+    shared_meta_->version = kGlobalVersion;
+    subtle::NoBarrier_Store(&shared_meta_->freeptr, sizeof(SharedMetadata));
+
+    // Set up the queue of iterable allocations.
+    shared_meta_->queue.size = sizeof(BlockHeader);
+    shared_meta_->queue.cookie = kBlockCookieQueue;
+    subtle::NoBarrier_Store(&shared_meta_->queue.next, OFFSET_QUEUE);
+    subtle::NoBarrier_Store(&shared_meta_->tailptr, OFFSET_QUEUE);
+  } else {
+    // The allocator is attaching to a previously initialized segment of
+    // memory.  Make sure the embedded data matches what has been passed.
+    if (shared_meta_->size != size ||
+        shared_meta_->page_size != page_size) {
+      NOTREACHED();
+      SetCorrupted();
+    }
+  }
+}
+
+SharedMemoryAllocator::~SharedMemoryAllocator() {
+}
+
+int32_t SharedMemoryAllocator::Allocate(int32_t size, int32_t type) {
+  if (size < 0) {
+    NOTREACHED();
+    return OFFSET_NULL;
+  }
+
+  // Round up the requested size, plus header, to the next allocation alignment.
+  size += sizeof(BlockHeader);
+  size = (size + (kAllocAlignment - 1)) & ~(kAllocAlignment - 1);
+  if (size > mem_page_)
+    return OFFSET_NULL;
+
+  // Allocation is lockless so we do all our caculation and then, if saving
+  // indicates a change has occurred since we started, scrap everything and
+  // start over.
+  for (;;) {
+    if (IsCorrupted())
+      return OFFSET_NULL;
+
+    int32_t freeptr = subtle::Acquire_Load(&shared_meta_->freeptr);
+    if (freeptr + size > mem_size_) {
+      SetFlag(&shared_meta_->flags, kFlagFull, true);
+      return OFFSET_NULL;
+    }
+
+    // Get pointer to the "free" block.  It doesn't even have a header; pass
+    // -sizeof(header) so accouting for that will yield an expected size of
+    // zero which is what will be stored at that location.  If something
+    // has been allocated since the load of freeptr above, it is still safe
+    // as nothing will be written to that location until after the CAS below.
+    BlockHeader* block = GetBlock(freeptr, 0, -(int)sizeof(BlockHeader), true);
+    if (!block) {
+      SetCorrupted();
+      return OFFSET_NULL;
+    }
+
+    // An allocation cannot cross page boundaries.  If it would, create a
+    // "wasted" block and begin again at the top of the next page.
+    int32_t page_free = mem_page_ - freeptr % mem_page_;
+    if (size > page_free) {
+      int32_t new_freeptr = freeptr + page_free;
+      if (subtle::Release_CompareAndSwap(
+              &shared_meta_->freeptr, freeptr, new_freeptr) == freeptr) {
+        block->size = page_free;
+        block->cookie = kBlockCookieWasted;
+      }
+      continue;
+    }
+
+    // Don't leave a slice at the end of a page too small for anything.  This
+    // can result in an allocation up to two alignment-sizes greater than the
+    // minimum required by requested-size + header + alignment.
+    if (page_free - size < (int)(sizeof(BlockHeader) + kAllocAlignment))
+      size = page_free;
+
+    int32_t new_freeptr = freeptr + size;
+    if (new_freeptr > mem_size_) {
+      SetCorrupted();
+      return OFFSET_NULL;
+    }
+
+    if (subtle::Release_CompareAndSwap(
+            &shared_meta_->freeptr, freeptr, new_freeptr) != freeptr) {
+      // Another thread must have completed an allocation while we were working.
+      // Try again.
+      continue;
+    }
+
+    // Given that all memory was zeroed before ever being given to an instance
+    // of this class and given that we only allocate in a monotomic fashion
+    // going forward, it must be that the newly allocated block is completely
+    // full of zeros.  If we find anything in the block header that is NOT a
+    // zero then something must have previously run amuck through memory,
+    // writing beyond the allocated space and into unallocated space.
+    if (block->size != 0 ||
+        block->cookie != kBlockCookieFree ||
+        block->type != 0 ||
+        subtle::NoBarrier_Load(&block->next) != 0) {
+      SetCorrupted();
+      return OFFSET_NULL;
+    }
+
+    block->size = size;
+    block->cookie = kBlockCookieAllocated;
+    block->type = type;
+    return freeptr;
+  }
+}
+
+void SharedMemoryAllocator::GetMemoryInfo(MemoryInfo* meminfo) {
+  int32_t remaining =
+      mem_size_ - subtle::NoBarrier_Load(&shared_meta_->freeptr);
+  meminfo->total = mem_size_;
+  meminfo->free = IsCorrupted() ? 0 : remaining - sizeof(BlockHeader);
+}
+
+void SharedMemoryAllocator::MakeIterable(int32_t offset) {
+  if (IsCorrupted())
+    return;
+  BlockHeader* block = GetBlock(offset, 0, 0, false);
+  if (!block)  // invalid offset
+    return;
+  if (subtle::NoBarrier_Load(&block->next) != 0)  // previously set iterable
+    return;
+  subtle::NoBarrier_Store(&block->next, OFFSET_QUEUE);  // will be tail block

[Dmitry Vyukov, 2015/11/03 14:06:46]

Why don't you use a lock-free stack?
The stack algorithm is much simpler. It will also eliminate the
SharedMetadata.queue, OFFSET_QUEUE, special flag in GetBlock and just greatly
simplify lots of things.

[bcwhite, 2015/11/03 16:28:19]

Honestly, because an M&S Queue was what Alexander recommended when I was doing
the design.  But, I also need to be able to iterate through the queue and
continue iteration after stopping if new items get added to the collection. 
That means "forward" pointers from the head whereas a Stack has only "backward"
pointers from the tail.

[Alexander Potapenko, 2015/11/03 17:25:15]

Was that another Alexander? I never mentioned M&S queues (of which I first heard
from the design doc)

[Dmitry Vyukov, 2015/11/03 18:15:07]

Makes sense.

I guess we could do something along of the following lines for stack iteration.
Namely, re-read stack head, and it some new elements get added we iterate from
new head to the previous head. But at this point I am not sure it worth it.

struct Iter {
	Node*	pos;
	Node*	head;
	Node*	next_head;
	int32	n;
};

void Alloc::IterInit(Iter *iter)
{
	iter->pos = 0;
	iter->head = 0;
	iter->last_head = 0;
	iter->n = 0;
}

void* Alloc::IterNext(Iter *iter)
{
	if (iter->pos == iter->head) {
		iter->head = iter->next_head;
		iter->pos = iter->next_head = head_;
	}
	if (iter->pos == iter->head)
		return NULL;
	void *cur = iter->pos;
	iter->pos = iter->pos->next;
	return cur;
}


I will look at the queue algorithm tomorrow.

+
+  // Try to add this block to the tail of the queue.  May take multiple tries.
+  int32_t tail;
+  for (;;) {
+    tail = subtle::Acquire_Load(&shared_meta_->tailptr);
+    block = GetBlock(tail, 0, 0, true);
+    if (!block) {
+      SetCorrupted();
+      return;
+    }
+    int32_t next = subtle::NoBarrier_Load(&block->next);
+
+    // Ensure that the tail pointer didn't change while reading next.  Only
+    // the read of the tail pointer is atomic but we need to read both the
+    // tail pointer and the next pointer from it in an atomic fashion.  The
+    // way to do this is to read both non-atomically and then verify after
+    // the second read that the first read is still valid/unchanged.
+    if (tail == subtle::Release_Load(&shared_meta_->tailptr)) {
+      // Check if the found block is truely the last in the queue (i.e. it
+      // points back to the "queue" node).
+      if (next == OFFSET_QUEUE) {
+        // Yes.  Try to append the passed block after the current tail block.
+        if (subtle::Release_CompareAndSwap(
+                &block->next, OFFSET_QUEUE, offset) == OFFSET_QUEUE) {
+          // Success!  The block is enqueued; need to update the tail pointer.
+          break;
+        }
+      } else {
+        // No.  Another thread has stopped between the block-next update
+        // and the tail-pointer update.  Try to update tailptr past the
+        // found block.  That other thread may complete it first or it
+        // may have crashed.  Be fail-safe.
+        subtle::Release_CompareAndSwap(&shared_meta_->tailptr, tail, next);
+      }
+    }
+  }
+
+  // Block has been enqueued.  Now update the tail-pointer past it.  This
+  // could fail if another thread has already completed the operation as
+  // part of being fail-safe.
+  subtle::Release_CompareAndSwap(&shared_meta_->tailptr, tail, offset);
+}
+
+void SharedMemoryAllocator::CreateIterator(Iterator* state) {
+  state->last = OFFSET_QUEUE;
+  state->loop_detector = OFFSET_QUEUE;
+}
+
+int32_t SharedMemoryAllocator::GetNextIterable(Iterator* state, int32_t* type) {
+  const BlockHeader* block = GetBlock(state->last, 0, 0, true);
+  if (!block)  // invalid iterator state
+    return OFFSET_NULL;
+  int32_t next = subtle::NoBarrier_Load(&block->next);
+  block = GetBlock(next, 0, 0, false);
+  if (!block)  // no next allocation in queue
+    return OFFSET_NULL;
+  if (next == state->loop_detector) {
+    SetCorrupted();
+    return OFFSET_NULL;
+  }
+
+  state->last = next;
+  *type = block->type;
+
+  // Memory corruption could cause a loop in the list.  We need to detect
+  // that so as to not cause an infinite loop in the caller.  This is done
+  // by having a second pointer that double-increments through the list.
+  // If it ever comes around to match "last" then we have a loop and need
+  // to stop iterating.  It's possible to not iterate through all items and
+  // it's possible to loop multiple times before the loop is detected but at
+  // least it stops.
+  if (state->loop_detector == OFFSET_QUEUE)

[Dmitry Vyukov, 2015/11/03 14:06:46]

There is a simpler way to do it:

count number of iterations in state->niter
if state->niter > freeptr/sizeof(BlockHeader), we have a loop

[bcwhite, 2015/11/03 16:28:19]

Great!  Though mine worked, it had to change anyway because it occurred to me
that two loops could go undetected.

+    state->loop_detector = next;
+  block = GetBlock(state->loop_detector, 0, 0, false);
+  if (block) {
+    state->loop_detector = subtle::NoBarrier_Load(&block->next);
+    block = GetBlock(state->loop_detector, 0, 0, false);
+    if (block)
+      state->loop_detector  = subtle::NoBarrier_Load(&block->next);
+  }
+
+  return next;
+}
+
+void SharedMemoryAllocator::SetCorrupted() {
+  LOG(ERROR) << "Corruption detected in shared-memory segment.";
+  subtle::NoBarrier_Store(&corrupted_, 1);

[Alexander Potapenko, 2015/11/03 08:12:54]

Why do you need both corrupted_ and kFlagCorrupted?
Isn't any of the two enough?

[bcwhite, 2015/11/03 16:28:19]

The shared flag can't be trusted since a malicious actor could overwrite it. 
The local version is immune to foreign actors.  Thus, if we see it, we copy it
locally and once known, always restore it globally.  I'll add a comment.

+  SetFlag(&shared_meta_->flags, kFlagCorrupted, true);
+}
+
+bool SharedMemoryAllocator::IsCorrupted() {
+  if (subtle::NoBarrier_Load(&corrupted_) ||
+      CheckFlag(&shared_meta_->flags, kFlagCorrupted)) {
+    SetCorrupted();  // Make sure all indicators are set.
+    return true;
+  }
+  return false;
+}
+
+bool SharedMemoryAllocator::IsFull() {
+  return CheckFlag(&shared_meta_->flags, kFlagFull);
+}
+
+// Dereference a block |offset| and ensure that it's valid for the desired
+// |type| and |size|.  |special| indicates that we may try to access block
+// headers not available to callers but still accessed by this module.  By
+// having internal dereferences go through this same function, the allocator
+// is hardened against corruption.
+SharedMemoryAllocator::BlockHeader* SharedMemoryAllocator::GetBlock(
+    int32_t offset, int32_t type, int32_t size, bool special) {
+  // Validation of parameters.
+  if (offset % kAllocAlignment != 0)
+    return nullptr;
+  if (offset < (int)(special ? OFFSET_QUEUE : sizeof(SharedMetadata)))
+    return nullptr;
+  size += sizeof(BlockHeader);
+  if (offset + size > mem_size_)
+    return nullptr;
+  int32_t freeptr = subtle::NoBarrier_Load(&shared_meta_->freeptr);
+  if (offset + size > freeptr)
+    return nullptr;
+
+  // Validation of referenced block-header.
+  const BlockHeader* block = reinterpret_cast<BlockHeader*>(mem_base_ + offset);
+  if (block->size < size)
+    return nullptr;
+  if (!special && block->cookie != kBlockCookieAllocated)
+    return nullptr;
+  if (type != 0 && block->type != type)
+    return nullptr;
+
+  // Return pointer to block data.
+  return reinterpret_cast<BlockHeader*>(mem_base_ + offset);
+}
+
+void* SharedMemoryAllocator::GetBlockData(int32_t offset, int32_t type,
+                                          int32_t size, bool special) {
+  DCHECK(size > 0);
+  BlockHeader* block = GetBlock(offset, type, size, special);
+  if (!block)
+    return nullptr;
+  return reinterpret_cast<char*>(block) + sizeof(BlockHeader);
+}
+
+}  // namespace base