Revert 274326 "Lobotomize the GPU memory manager"

trunk/src/content/common/gpu/gpu_memory_manager.cc

Index: trunk/src/content/common/gpu/gpu_memory_manager.cc
===================================================================
--- trunk/src/content/common/gpu/gpu_memory_manager.cc	(revision 274878)
+++ trunk/src/content/common/gpu/gpu_memory_manager.cc	(working copy)
@@ -35,22 +35,72 @@
   *total_size += (new_size - old_size);
 }
 
+template<typename T>
+T RoundUp(T n, T mul) {
+  return ((n + mul - 1) / mul) * mul;
 }
 
+template<typename T>
+T RoundDown(T n, T mul) {
+  return (n / mul) * mul;
+}
+
+}
+
 GpuMemoryManager::GpuMemoryManager(
     GpuChannelManager* channel_manager,
     uint64 max_surfaces_with_frontbuffer_soft_limit)
     : channel_manager_(channel_manager),
       manage_immediate_scheduled_(false),
-      disable_schedule_manage_(false),
       max_surfaces_with_frontbuffer_soft_limit_(
           max_surfaces_with_frontbuffer_soft_limit),
-      client_hard_limit_bytes_(0),
+      priority_cutoff_(MemoryAllocation::CUTOFF_ALLOW_EVERYTHING),
+      bytes_available_gpu_memory_(0),
+      bytes_available_gpu_memory_overridden_(false),
+      bytes_minimum_per_client_(0),
+      bytes_default_per_client_(0),
       bytes_allocated_managed_current_(0),
       bytes_allocated_unmanaged_current_(0),
-      bytes_allocated_historical_max_(0)
-{ }
+      bytes_allocated_historical_max_(0),
+      bytes_allocated_unmanaged_high_(0),
+      bytes_allocated_unmanaged_low_(0),
+      bytes_unmanaged_limit_step_(kBytesAllocatedUnmanagedStep),
+      disable_schedule_manage_(false)
+{
+  CommandLine* command_line = CommandLine::ForCurrentProcess();
 
+  // Use a more conservative memory allocation policy on Linux and Mac because
+  // the platform is unstable when under memory pressure.
+  // http://crbug.com/145600 (Linux)
+  // http://crbug.com/141377 (Mac)
+#if defined(OS_MACOSX) || (defined(OS_LINUX) && !defined(OS_CHROMEOS))
+  priority_cutoff_ = MemoryAllocation::CUTOFF_ALLOW_NICE_TO_HAVE;
+#endif
+
+#if defined(OS_ANDROID)
+  bytes_default_per_client_ = 8 * 1024 * 1024;
+  bytes_minimum_per_client_ = 8 * 1024 * 1024;
+#elif defined(OS_CHROMEOS)
+  bytes_default_per_client_ = 64 * 1024 * 1024;
+  bytes_minimum_per_client_ = 4 * 1024 * 1024;
+#elif defined(OS_MACOSX)
+  bytes_default_per_client_ = 128 * 1024 * 1024;
+  bytes_minimum_per_client_ = 128 * 1024 * 1024;
+#else
+  bytes_default_per_client_ = 64 * 1024 * 1024;
+  bytes_minimum_per_client_ = 64 * 1024 * 1024;
+#endif
+
+  if (command_line->HasSwitch(switches::kForceGpuMemAvailableMb)) {
+    base::StringToUint64(
+        command_line->GetSwitchValueASCII(switches::kForceGpuMemAvailableMb),
+        &bytes_available_gpu_memory_);
+    bytes_available_gpu_memory_ *= 1024 * 1024;
+    bytes_available_gpu_memory_overridden_ = true;
+  } else
+    bytes_available_gpu_memory_ = GetDefaultAvailableGpuMemory();
+}
+
 GpuMemoryManager::~GpuMemoryManager() {
   DCHECK(tracking_groups_.empty());
   DCHECK(clients_visible_mru_.empty());
@@ -60,19 +110,60 @@
   DCHECK(!bytes_allocated_unmanaged_current_);
 }
 
+uint64 GpuMemoryManager::GetAvailableGpuMemory() const {
+  // Allow unmanaged allocations to over-subscribe by at most (high_ - low_)
+  // before restricting managed (compositor) memory based on unmanaged usage.
+  if (bytes_allocated_unmanaged_low_ > bytes_available_gpu_memory_)
+    return 0;
+  return bytes_available_gpu_memory_ - bytes_allocated_unmanaged_low_;
+}
+
+uint64 GpuMemoryManager::GetDefaultAvailableGpuMemory() const {
+#if defined(OS_ANDROID)
+  return 16 * 1024 * 1024;
+#elif defined(OS_CHROMEOS)
+  return 1024 * 1024 * 1024;
+#else
+  return 256 * 1024 * 1024;
+#endif
+}
+
+uint64 GpuMemoryManager::GetMaximumTotalGpuMemory() const {
+#if defined(OS_ANDROID)
+  return 256 * 1024 * 1024;
+#else
+  return 1024 * 1024 * 1024;
+#endif
+}
+
+uint64 GpuMemoryManager::GetMaximumClientAllocation() const {
+#if defined(OS_ANDROID) || defined(OS_CHROMEOS)
+  return bytes_available_gpu_memory_;
+#else
+  // This is to avoid allowing a single page on to use a full 256MB of memory
+  // (the current total limit). Long-scroll pages will hit this limit,
+  // resulting in instability on some platforms (e.g, issue 141377).
+  return bytes_available_gpu_memory_ / 2;
+#endif
+}
+
+uint64 GpuMemoryManager::CalcAvailableFromGpuTotal(uint64 total_gpu_memory) {
+#if defined(OS_ANDROID)
+  // We don't need to reduce the total on Android, since
+  // the total is an estimate to begin with.
+  return total_gpu_memory;
+#else
+  // Allow Chrome to use 75% of total GPU memory, or all-but-64MB of GPU
+  // memory, whichever is less.
+  return std::min(3 * total_gpu_memory / 4, total_gpu_memory - 64*1024*1024);
+#endif
+}
+
 void GpuMemoryManager::UpdateAvailableGpuMemory() {
-  // If the value was overridden on the command line, use the specified value.
-  static bool client_hard_limit_bytes_overridden =
-      CommandLine::ForCurrentProcess()->HasSwitch(
-          switches::kForceGpuMemAvailableMb);
-  if (client_hard_limit_bytes_overridden) {
-    base::StringToUint64(
-        CommandLine::ForCurrentProcess()->GetSwitchValueASCII(
-            switches::kForceGpuMemAvailableMb),
-        &client_hard_limit_bytes_);
-    client_hard_limit_bytes_ *= 1024 * 1024;
+  // If the amount of video memory to use was specified at the command
+  // line, never change it.
+  if (bytes_available_gpu_memory_overridden_)
     return;
-  }
 
   // On non-Android, we use an operating system query when possible.
   // We do not have a reliable concept of multiple GPUs existing in
@@ -100,21 +191,29 @@
   if (!bytes_min)
     return;
 
-  client_hard_limit_bytes_ = bytes_min;
+  bytes_available_gpu_memory_ = CalcAvailableFromGpuTotal(bytes_min);
 
-#if defined(OS_ANDROID)
-  // Clamp the observed value to a specific range on Android.
-  client_hard_limit_bytes_ = std::max(client_hard_limit_bytes_,
-                                      static_cast<uint64>(16 * 1024 * 1024));
-  client_hard_limit_bytes_ = std::min(client_hard_limit_bytes_,
-                                      static_cast<uint64>(256 * 1024 * 1024));
-#else
-  // Ignore what the system said and give all clients the same maximum
-  // allocation on desktop platforms.
-  client_hard_limit_bytes_ = 256 * 1024 * 1024;
-#endif
+  // Never go below the default allocation
+  bytes_available_gpu_memory_ = std::max(bytes_available_gpu_memory_,
+                                         GetDefaultAvailableGpuMemory());
+
+  // Never go above the maximum.
+  bytes_available_gpu_memory_ = std::min(bytes_available_gpu_memory_,
+                                         GetMaximumTotalGpuMemory());
 }
 
+void GpuMemoryManager::UpdateUnmanagedMemoryLimits() {
+  // Set the limit to be [current_, current_ + step_ / 4), with the endpoints
+  // of the intervals rounded down and up to the nearest step_, to avoid
+  // thrashing the interval.
+  bytes_allocated_unmanaged_high_ = RoundUp(
+      bytes_allocated_unmanaged_current_ + bytes_unmanaged_limit_step_ / 4,
+      bytes_unmanaged_limit_step_);
+  bytes_allocated_unmanaged_low_ = RoundDown(
+      bytes_allocated_unmanaged_current_,
+      bytes_unmanaged_limit_step_);
+}
+
 void GpuMemoryManager::ScheduleManage(
     ScheduleManageTime schedule_manage_time) {
   if (disable_schedule_manage_)
@@ -164,8 +263,14 @@
                    GetCurrentUsage());
   }
 
-  if (GetCurrentUsage() > bytes_allocated_historical_max_ +
-                          kBytesAllocatedUnmanagedStep) {
+  // If we've gone past our current limit on unmanaged memory, schedule a
+  // re-manage to take int account the unmanaged memory.
+  if (bytes_allocated_unmanaged_current_ >= bytes_allocated_unmanaged_high_)
+    ScheduleManage(kScheduleManageNow);
+  if (bytes_allocated_unmanaged_current_ < bytes_allocated_unmanaged_low_)
+    ScheduleManage(kScheduleManageLater);
+
+  if (GetCurrentUsage() > bytes_allocated_historical_max_) {
       bytes_allocated_historical_max_ = GetCurrentUsage();
       // If we're blowing into new memory usage territory, spam the browser
       // process with the most up-to-date information about our memory usage.
@@ -214,8 +319,27 @@
 
 void GpuMemoryManager::SetClientStateManagedMemoryStats(
     GpuMemoryManagerClientState* client_state,
-    const ManagedMemoryStats& stats) {
-  // TODO(ccameron): delete this from the full stack.
+    const ManagedMemoryStats& stats)
+{
+  client_state->managed_memory_stats_ = stats;
+
+  // If this is the first time that stats have been received for this
+  // client, use them immediately.
+  if (!client_state->managed_memory_stats_received_) {
+    client_state->managed_memory_stats_received_ = true;
+    ScheduleManage(kScheduleManageNow);
+    return;
+  }
+
+  // If these statistics sit outside of the range that we used in our
+  // computation of memory allocations then recompute the allocations.
+  if (client_state->managed_memory_stats_.bytes_nice_to_have >
+      client_state->bytes_nicetohave_limit_high_) {
+    ScheduleManage(kScheduleManageNow);
+  } else if (client_state->managed_memory_stats_.bytes_nice_to_have <
+             client_state->bytes_nicetohave_limit_low_) {
+    ScheduleManage(kScheduleManageLater);
+  }
 }
 
 uint64 GpuMemoryManager::GetClientMemoryUsage(
@@ -271,6 +395,9 @@
   // Update the amount of GPU memory available on the system.
   UpdateAvailableGpuMemory();
 
+  // Update the limit on unmanaged memory.
+  UpdateUnmanagedMemoryLimits();
+
   // Determine which clients are "hibernated" (which determines the
   // distribution of frontbuffers and memory among clients that don't have
   // surfaces).
@@ -285,7 +412,243 @@
   SendUmaStatsToBrowser();
 }
 
+// static
+uint64 GpuMemoryManager::ComputeCap(
+    std::vector<uint64> bytes, uint64 bytes_sum_limit)
+{
+  size_t bytes_size = bytes.size();
+  uint64 bytes_sum = 0;
+
+  if (bytes_size == 0)
+    return std::numeric_limits<uint64>::max();
+
+  // Sort and add up all entries
+  std::sort(bytes.begin(), bytes.end());
+  for (size_t i = 0; i < bytes_size; ++i)
+    bytes_sum += bytes[i];
+
+  // As we go through the below loop, let bytes_partial_sum be the
+  // sum of bytes[0] + ... + bytes[bytes_size - i - 1]
+  uint64 bytes_partial_sum = bytes_sum;
+
+  // Try using each entry as a cap, and see where we get cut off.
+  for (size_t i = 0; i < bytes_size; ++i) {
+    // Try limiting cap to bytes[bytes_size - i - 1]
+    uint64 test_cap = bytes[bytes_size - i - 1];
+    uint64 bytes_sum_with_test_cap = i * test_cap + bytes_partial_sum;
+
+    // If that fits, raise test_cap to give an even distribution to the
+    // last i entries.
+    if (bytes_sum_with_test_cap <= bytes_sum_limit) {
+      if (i == 0)
+        return std::numeric_limits<uint64>::max();
+      else
+        return test_cap + (bytes_sum_limit - bytes_sum_with_test_cap) / i;
+    } else {
+      bytes_partial_sum -= test_cap;
+    }
+  }
+
+  // If we got here, then we can't fully accommodate any of the clients,
+  // so distribute bytes_sum_limit evenly.
+  return bytes_sum_limit / bytes_size;
+}
+
+uint64 GpuMemoryManager::ComputeClientAllocationWhenVisible(
+    GpuMemoryManagerClientState* client_state,
+    uint64 bytes_above_required_cap,
+    uint64 bytes_above_minimum_cap,
+    uint64 bytes_overall_cap) {
+  ManagedMemoryStats* stats = &client_state->managed_memory_stats_;
+
+  if (!client_state->managed_memory_stats_received_)
+    return GetDefaultClientAllocation();
+
+  uint64 bytes_required = 9 * stats->bytes_required / 8;
+  bytes_required = std::min(bytes_required, GetMaximumClientAllocation());
+  bytes_required = std::max(bytes_required, GetMinimumClientAllocation());
+
+  uint64 bytes_nicetohave = 4 * stats->bytes_nice_to_have / 3;
+  bytes_nicetohave = std::min(bytes_nicetohave, GetMaximumClientAllocation());
+  bytes_nicetohave = std::max(bytes_nicetohave, GetMinimumClientAllocation());
+  bytes_nicetohave = std::max(bytes_nicetohave, bytes_required);
+
+  uint64 allocation = GetMinimumClientAllocation();
+  allocation += std::min(bytes_required - GetMinimumClientAllocation(),
+                         bytes_above_minimum_cap);
+  allocation += std::min(bytes_nicetohave - bytes_required,
+                         bytes_above_required_cap);
+  allocation = std::min(allocation,
+                        bytes_overall_cap);
+  return allocation;
+}
+
+void GpuMemoryManager::ComputeVisibleSurfacesAllocations() {
+  uint64 bytes_available_total = GetAvailableGpuMemory();
+  uint64 bytes_above_required_cap = std::numeric_limits<uint64>::max();
+  uint64 bytes_above_minimum_cap = std::numeric_limits<uint64>::max();
+  uint64 bytes_overall_cap_visible = GetMaximumClientAllocation();
+
+  // Compute memory usage at three levels
+  // - painting everything that is nicetohave for visible clients
+  // - painting only what that is visible
+  // - giving every client the minimum allocation
+  uint64 bytes_nicetohave_visible = 0;
+  uint64 bytes_required_visible = 0;
+  uint64 bytes_minimum_visible = 0;
+  for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+       it != clients_visible_mru_.end();
+       ++it) {
+    GpuMemoryManagerClientState* client_state = *it;
+    client_state->bytes_allocation_ideal_nicetohave_ =
+        ComputeClientAllocationWhenVisible(
+            client_state,
+            bytes_above_required_cap,
+            bytes_above_minimum_cap,
+            bytes_overall_cap_visible);
+    client_state->bytes_allocation_ideal_required_ =
+        ComputeClientAllocationWhenVisible(
+            client_state,
+            0,
+            bytes_above_minimum_cap,
+            bytes_overall_cap_visible);
+    client_state->bytes_allocation_ideal_minimum_ =
+        ComputeClientAllocationWhenVisible(
+            client_state,
+            0,
+            0,
+            bytes_overall_cap_visible);
+
+    bytes_nicetohave_visible +=
+        client_state->bytes_allocation_ideal_nicetohave_;
+    bytes_required_visible +=
+        client_state->bytes_allocation_ideal_required_;
+    bytes_minimum_visible +=
+        client_state->bytes_allocation_ideal_minimum_;
+  }
+
+  // Determine which of those three points we can satisfy, and limit
+  // bytes_above_required_cap and bytes_above_minimum_cap to not go
+  // over the limit.
+  if (bytes_minimum_visible > bytes_available_total) {
+    bytes_above_required_cap = 0;
+    bytes_above_minimum_cap = 0;
+  } else if (bytes_required_visible > bytes_available_total) {
+    std::vector<uint64> bytes_to_fit;
+    for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+         it != clients_visible_mru_.end();
+         ++it) {
+      GpuMemoryManagerClientState* client_state = *it;
+      bytes_to_fit.push_back(client_state->bytes_allocation_ideal_required_ -
+                             client_state->bytes_allocation_ideal_minimum_);
+    }
+    bytes_above_required_cap = 0;
+    bytes_above_minimum_cap = ComputeCap(
+        bytes_to_fit, bytes_available_total - bytes_minimum_visible);
+  } else if (bytes_nicetohave_visible > bytes_available_total) {
+    std::vector<uint64> bytes_to_fit;
+    for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+         it != clients_visible_mru_.end();
+         ++it) {
+      GpuMemoryManagerClientState* client_state = *it;
+      bytes_to_fit.push_back(client_state->bytes_allocation_ideal_nicetohave_ -
+                             client_state->bytes_allocation_ideal_required_);
+    }
+    bytes_above_required_cap = ComputeCap(
+        bytes_to_fit, bytes_available_total - bytes_required_visible);
+    bytes_above_minimum_cap = std::numeric_limits<uint64>::max();
+  }
+
+  // Given those computed limits, set the actual memory allocations for the
+  // visible clients, tracking the largest allocation and the total allocation
+  // for future use.
+  uint64 bytes_allocated_visible = 0;
+  uint64 bytes_allocated_max_client_allocation = 0;
+  for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+       it != clients_visible_mru_.end();
+       ++it) {
+    GpuMemoryManagerClientState* client_state = *it;
+    client_state->bytes_allocation_when_visible_ =
+        ComputeClientAllocationWhenVisible(
+            client_state,
+            bytes_above_required_cap,
+            bytes_above_minimum_cap,
+            bytes_overall_cap_visible);
+    bytes_allocated_visible += client_state->bytes_allocation_when_visible_;
+    bytes_allocated_max_client_allocation = std::max(
+        bytes_allocated_max_client_allocation,
+        client_state->bytes_allocation_when_visible_);
+  }
+
+  // Set the limit for nonvisible clients for when they become visible.
+  // Use the same formula, with a lowered overall cap in case any of the
+  // currently-nonvisible clients are much more resource-intensive than any
+  // of the existing clients.
+  uint64 bytes_overall_cap_nonvisible = bytes_allocated_max_client_allocation;
+  if (bytes_available_total > bytes_allocated_visible) {
+    bytes_overall_cap_nonvisible +=
+        bytes_available_total - bytes_allocated_visible;
+  }
+  bytes_overall_cap_nonvisible = std::min(bytes_overall_cap_nonvisible,
+                                          GetMaximumClientAllocation());
+  for (ClientStateList::const_iterator it = clients_nonvisible_mru_.begin();
+       it != clients_nonvisible_mru_.end();
+       ++it) {
+    GpuMemoryManagerClientState* client_state = *it;
+    client_state->bytes_allocation_when_visible_ =
+        ComputeClientAllocationWhenVisible(
+            client_state,
+            bytes_above_required_cap,
+            bytes_above_minimum_cap,
+            bytes_overall_cap_nonvisible);
+  }
+}
+
+void GpuMemoryManager::DistributeRemainingMemoryToVisibleSurfaces() {
+  uint64 bytes_available_total = GetAvailableGpuMemory();
+  uint64 bytes_allocated_total = 0;
+
+  for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+       it != clients_visible_mru_.end();
+       ++it) {
+    GpuMemoryManagerClientState* client_state = *it;
+    bytes_allocated_total += client_state->bytes_allocation_when_visible_;
+  }
+
+  if (bytes_allocated_total >= bytes_available_total)
+    return;
+
+  std::vector<uint64> bytes_extra_requests;
+  for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+       it != clients_visible_mru_.end();
+       ++it) {
+    GpuMemoryManagerClientState* client_state = *it;
+    CHECK(GetMaximumClientAllocation() >=
+          client_state->bytes_allocation_when_visible_);
+    uint64 bytes_extra = GetMaximumClientAllocation() -
+                         client_state->bytes_allocation_when_visible_;
+    bytes_extra_requests.push_back(bytes_extra);
+  }
+  uint64 bytes_extra_cap = ComputeCap(
+      bytes_extra_requests, bytes_available_total - bytes_allocated_total);
+  for (ClientStateList::const_iterator it = clients_visible_mru_.begin();
+       it != clients_visible_mru_.end();
+       ++it) {
+    GpuMemoryManagerClientState* client_state = *it;
+    uint64 bytes_extra = GetMaximumClientAllocation() -
+                         client_state->bytes_allocation_when_visible_;
+    client_state->bytes_allocation_when_visible_ += std::min(
+        bytes_extra, bytes_extra_cap);
+  }
+}
+
 void GpuMemoryManager::AssignSurfacesAllocations() {
+  // Compute allocation when for all clients.
+  ComputeVisibleSurfacesAllocations();
+
+  // Distribute the remaining memory to visible clients.
+  DistributeRemainingMemoryToVisibleSurfaces();
+
   // Send that allocation to the clients.
   ClientStateList clients = clients_visible_mru_;
   clients.insert(clients.end(),
@@ -296,24 +659,20 @@
        ++it) {
     GpuMemoryManagerClientState* client_state = *it;
 
+    // Re-assign memory limits to this client when its "nice to have" bucket
+    // grows or shrinks by 1/4.
+    client_state->bytes_nicetohave_limit_high_ =
+        5 * client_state->managed_memory_stats_.bytes_nice_to_have / 4;
+    client_state->bytes_nicetohave_limit_low_ =
+        3 * client_state->managed_memory_stats_.bytes_nice_to_have / 4;
+
     // Populate and send the allocation to the client
     MemoryAllocation allocation;
-    allocation.bytes_limit_when_visible = client_hard_limit_bytes_;
-#if defined(OS_ANDROID)
-    // On Android, because there is only one visible tab at any time, allow
-    // that renderer to cache as much as it can.
-    allocation.priority_cutoff_when_visible =
-        MemoryAllocation::CUTOFF_ALLOW_EVERYTHING;
-#else
-    // On desktop platforms, instruct the renderers to cache only a smaller
-    // set, to play nice with other renderers and other applications. If this
-    // if not done, then the system can become unstable.
-    // http://crbug.com/145600 (Linux)
-    // http://crbug.com/141377 (Mac)
-    allocation.priority_cutoff_when_visible =
-        MemoryAllocation::CUTOFF_ALLOW_NICE_TO_HAVE;
-#endif
 
+    allocation.bytes_limit_when_visible =
+        client_state->bytes_allocation_when_visible_;
+    allocation.priority_cutoff_when_visible = priority_cutoff_;
+
     client_state->client_->SetMemoryAllocation(allocation);
     client_state->client_->SuggestHaveFrontBuffer(!client_state->hibernated_);
   }
@@ -327,7 +686,8 @@
     MemoryAllocation allocation;
 
     if (!client_state->hibernated_) {
-      allocation.bytes_limit_when_visible = 1;
+      allocation.bytes_limit_when_visible =
+          GetMinimumClientAllocation();
       allocation.priority_cutoff_when_visible =
           MemoryAllocation::CUTOFF_ALLOW_EVERYTHING;
     }
@@ -384,7 +744,7 @@
   GPUMemoryUmaStats params;
   params.bytes_allocated_current = GetCurrentUsage();
   params.bytes_allocated_max = bytes_allocated_historical_max_;
-  params.bytes_limit = client_hard_limit_bytes_;
+  params.bytes_limit = bytes_available_gpu_memory_;
   params.client_count = clients_visible_mru_.size() +
                         clients_nonvisible_mru_.size() +
                         clients_nonsurface_.size();