Split image decode cache limits into "working set" vs "cache" limits

cc/tiles/gpu_image_decode_cache.cc

 // Copyright 2016 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 "cc/tiles/gpu_image_decode_cache.h"
 
 #include <inttypes.h>
 
 #include "base/auto_reset.h"
 #include "base/debug/alias.h"
@@ skipping 324 matching lines @@
   DCHECK_EQ(0u, upload.ref_count);
   DCHECK_EQ(0u, decode.ref_count);
   DCHECK_EQ(false, decode.is_locked());
   // This should always be cleaned up before deleting the image, as it needs to
   // be freed with the GL context lock held.
   DCHECK(!upload.image());
 }
 
 GpuImageDecodeCache::GpuImageDecodeCache(ContextProvider* context,
                                          ResourceFormat decode_format,
-                                         size_t max_gpu_image_bytes)
+                                         size_t max_working_set_bytes,
+                                         size_t max_cache_bytes)
     : format_(decode_format),
       context_(context),
       persistent_cache_(PersistentCache::NO_AUTO_EVICT),
-      normal_max_gpu_image_bytes_(max_gpu_image_bytes) {
+      max_working_set_bytes_(max_working_set_bytes),
+      normal_max_cache_bytes_(max_cache_bytes) {
+  DCHECK_GT(max_working_set_bytes_, normal_max_cache_bytes_);
+
   // Acquire the context_lock so that we can safely retrieve the
   // GrContextThreadSafeProxy. This proxy can then be used with no lock held.
   {
     ContextProvider::ScopedContextLock context_lock(context_);
     context_threadsafe_proxy_ = sk_sp<GrContextThreadSafeProxy>(
         context->GrContext()->threadSafeProxy());
   }
 
   // In certain cases, ThreadTaskRunnerHandle isn't set (Android Webview).
   // Don't register a dump provider in these cases.
@@ skipping 209 matching lines @@
     // We want to keep as little in our cache as possible. Set our memory limit
     // to zero and EnsureCapacity to clean up memory.
     cached_bytes_limit_ = kSuspendedOrInvisibleMaxGpuImageBytes;
     EnsureCapacity(0);
 
     // We are holding the context lock, so finish cleaning up deleted images
     // now.
     DeletePendingImages();
   } else {
     base::AutoLock lock(lock_);
-    cached_bytes_limit_ = normal_max_gpu_image_bytes_;
+    cached_bytes_limit_ = normal_max_cache_bytes_;
   }
 }
 
 bool GpuImageDecodeCache::OnMemoryDump(
     const base::trace_event::MemoryDumpArgs& args,
     base::trace_event::ProcessMemoryDump* pmd) {
   using base::trace_event::MemoryAllocatorDump;
   using base::trace_event::MemoryAllocatorDumpGuid;
   using base::trace_event::MemoryDumpLevelOfDetail;
 
@@ skipping 265 matching lines @@
   // re-locking discardable (rather than requiring a full upload like GPU
   // images).
   if (image_data->mode == DecodedDataMode::CPU && !has_any_refs) {
     images_pending_deletion_.push_back(image_data->upload.image());
     image_data->upload.SetImage(nullptr);
   }
 
   if (image_data->is_at_raster && !has_any_refs) {
     // We have an at-raster image which has reached zero refs. If it won't fit
     // in our cache, delete the image to allow it to fit.
-    if (image_data->upload.image() && !CanFitSize(image_data->size)) {
+    if (image_data->upload.image() && !CanFitInCache(image_data->size)) {
       images_pending_deletion_.push_back(image_data->upload.image());
       image_data->upload.SetImage(nullptr);
     }
 
     // We now have an at-raster image which will fit in our cache. Convert it
     // to not-at-raster.
     image_data->is_at_raster = false;
     if (image_data->upload.image()) {
       bytes_used_ += image_data->size;
       image_data->upload.budgeted = true;
     }
   }
 
   // If we have image refs on a non-at-raster image, it must be budgeted, as it
   // is either uploaded or pending upload.
   if (image_data->upload.ref_count > 0 && !image_data->upload.budgeted &&
       !image_data->is_at_raster) {
     // We should only be taking non-at-raster refs on images that fit in cache.
-    DCHECK(CanFitSize(image_data->size));
+    DCHECK(CanFitInWorkingSet(image_data->size));
 
     bytes_used_ += image_data->size;
     image_data->upload.budgeted = true;
   }
 
   // If we have no image refs on an image, it should only be budgeted if it has
   // an uploaded image. If no image exists (upload was cancelled), we should
   // un-budget the image.
   if (image_data->upload.ref_count == 0 && image_data->upload.budgeted &&
       !image_data->upload.image()) {
     DCHECK_GE(bytes_used_, image_data->size);
     bytes_used_ -= image_data->size;
     image_data->upload.budgeted = false;
   }
 
   // We should unlock the discardable memory for the image in two cases:
   // 1) The image is no longer being used (no decode or upload refs).
   // 2) This is a GPU backed image that has already been uploaded (no decode
   //    refs, and we actually already have an image).
   bool should_unlock_discardable =
       !has_any_refs ||
       (image_data->mode == DecodedDataMode::GPU &&
        !image_data->decode.ref_count && image_data->upload.image());
 
   if (should_unlock_discardable && image_data->decode.is_locked()) {
     DCHECK(image_data->decode.data());
     image_data->decode.Unlock();
   }
 
+  // EnsureCapacity to make sure we are under our cache limits.
+  EnsureCapacity(0);
+
 #if DCHECK_IS_ON()
   // Sanity check the above logic.
   if (image_data->upload.image()) {
     DCHECK(image_data->is_at_raster || image_data->upload.budgeted);
     if (image_data->mode == DecodedDataMode::CPU)
       DCHECK(image_data->decode.is_locked());
   } else {
     DCHECK(!image_data->upload.budgeted || image_data->upload.ref_count > 0);
   }
 #endif
 }
 
 // Ensures that we can fit a new image of size |required_size| in our cache. In

[vmpstr, 2017/03/28 18:41:31]

in our working set?

[ericrk, 2017/03/28 20:10:10]

Done.

 // doing so, this function will free unreferenced image data as necessary to
 // create rooom.
 bool GpuImageDecodeCache::EnsureCapacity(size_t required_size) {
   TRACE_EVENT0(TRACE_DISABLED_BY_DEFAULT("cc.debug"),
                "GpuImageDecodeCache::EnsureCapacity");
   lock_.AssertAcquired();
 
-  if (CanFitSize(required_size) && !ExceedsPreferredCount())
+  if (CanFitInCache(required_size) && !ExceedsPreferredCount())

[vmpstr, 2017/03/28 18:41:30]

This could use a comment about why you're checking cache vs working set.

[ericrk, 2017/03/28 20:10:10]

Done.

     return true;
 
   // While we are over memory or preferred item capacity, we iterate through
   // our set of cached image data in LRU order. For each image, we can do two
   // things: 1) We can free the uploaded image, reducing the memory usage of
   // the cache and 2) we can remove the entry entirely, reducing the count of
   // elements in the cache.
   for (auto it = persistent_cache_.rbegin(); it != persistent_cache_.rend();) {
     if (it->second->decode.ref_count != 0 ||
         it->second->upload.ref_count != 0) {
@@ skipping 18 matching lines @@
       it->second->upload.budgeted = false;
     }
 
     // Free the entire entry if necessary.
     if (ExceedsPreferredCount()) {
       it = persistent_cache_.Erase(it);
     } else {
       ++it;
     }
 
-    if (CanFitSize(required_size) && !ExceedsPreferredCount())
+    if (CanFitInCache(required_size) && !ExceedsPreferredCount())
       return true;
   }
 
-  // Preferred count is only used as a guideline when triming the cache. Allow
-  // new elements to be added as long as we are below our size limit.
-  return CanFitSize(required_size);
+  // While we always try to bring the cache size within bounds above, we only

[vmpstr, 2017/03/28 18:41:31]

This comment can probably migrate to the top of the function to just explain the
whole thing in one go

[ericrk, 2017/03/28 20:10:10]

Done.

+  // really care about whether the new element can fit in our working set.
+  return CanFitInWorkingSet(required_size);
 }
 
-bool GpuImageDecodeCache::CanFitSize(size_t size) const {
+bool GpuImageDecodeCache::CanFitInCache(size_t size) const {
   lock_.AssertAcquired();
 
   size_t bytes_limit;
   if (memory_state_ == base::MemoryState::NORMAL) {
     bytes_limit = cached_bytes_limit_;
   } else if (memory_state_ == base::MemoryState::THROTTLED) {
     bytes_limit = cached_bytes_limit_ / kThrottledCacheSizeReductionFactor;
   } else {
     DCHECK_EQ(base::MemoryState::SUSPENDED, memory_state_);
     bytes_limit = kSuspendedOrInvisibleMaxGpuImageBytes;
   }
 
   base::CheckedNumeric<uint32_t> new_size(bytes_used_);
   new_size += size;
   return new_size.IsValid() && new_size.ValueOrDie() <= bytes_limit;
 }
 
+bool GpuImageDecodeCache::CanFitInWorkingSet(size_t size) const {
+  lock_.AssertAcquired();
+
+  base::CheckedNumeric<uint32_t> new_size(bytes_used_);
+  new_size += size;
+  return new_size.IsValid() && new_size.ValueOrDie() <= max_working_set_bytes_;
+}
+
 bool GpuImageDecodeCache::ExceedsPreferredCount() const {
   lock_.AssertAcquired();
 
   size_t items_limit;
   if (memory_state_ == base::MemoryState::NORMAL) {
     items_limit = kNormalMaxItemsInCache;
   } else if (memory_state_ == base::MemoryState::THROTTLED) {
     items_limit = kThrottledMaxItemsInCache;
   } else {
     DCHECK_EQ(base::MemoryState::SUSPENDED, memory_state_);
@@ skipping 262 matching lines @@
 
 void GpuImageDecodeCache::OnPurgeMemory() {
   base::AutoLock lock(lock_);
   // Temporary changes |memory_state_| to free up cache as much as possible.
   base::AutoReset<base::MemoryState> reset(&memory_state_,
                                            base::MemoryState::SUSPENDED);
   EnsureCapacity(0);
 }
 
 }  // namespace cc