Disk cache: Don't evict anything for the first five

net/disk_cache/backend_impl.cc

Index: net/disk_cache/backend_impl.cc
===================================================================
--- net/disk_cache/backend_impl.cc	(revision 97896)
+++ net/disk_cache/backend_impl.cc	(working copy)
@@ -47,6 +47,9 @@
 const int kBaseTableLen = 64 * 1024;
 const int kDefaultCacheSize = 80 * 1024 * 1024;
 
+// Avoid trimming the cache for the first 5 minutes (10 timer ticks).
+const int kTrimDelay = 10;

[gavinp, 2011/08/31 03:01:20]

I'm also made a bit grumpy that these are in multiples of 30 seconds, without
being calculated.  I'd rather this was 5 * kTicksPerMinute or 300 /
kSecondsPerTick.  I haven't looked to see if either constant is available
though.

[rvargas (doing something else), 2011/08/31 18:06:35]

I see your point, but right now 30 seconds is kind of embedded in the code
(mostly relevant for UMA) and I don't want to give the impression that changing
the timer interval is as easy as changing a constant at the top of the file :(

+
 int DesiredIndexTableLen(int32 storage_size) {
   if (storage_size <= k64kEntriesStore)
     return kBaseTableLen;
@@ -366,7 +369,7 @@
       block_files_(path),
       mask_(0),
       max_size_(0),
-      io_delay_(0),
+      up_ticks_(0),
       cache_type_(net::DISK_CACHE),
       uma_report_(0),
       user_flags_(0),
@@ -392,7 +395,7 @@
       block_files_(path),
       mask_(mask),
       max_size_(0),
-      io_delay_(0),
+      up_ticks_(0),
       cache_type_(net::DISK_CACHE),
       uma_report_(0),
       user_flags_(kMask),
@@ -975,7 +978,8 @@
 
 void BackendImpl::OnEntryDestroyEnd() {
   DecreaseNumRefs();
-  if (data_->header.num_bytes > max_size_ && !read_only_)
+  if (data_->header.num_bytes > max_size_ && !read_only_ &&
+      (up_ticks_ > kTrimDelay || user_flags_ & disk_cache::kNoRandom))
     eviction_.TrimCache(false);
 }
 
@@ -1190,6 +1194,8 @@
   CACHE_UMA(COUNTS, "ByteIORate", 0, byte_count_ / 1024);
   entry_count_ = 0;
   byte_count_ = 0;
+  if (up_ticks_ < kTrimDelay * 2)  // No need to count forever.

[gavinp, 2011/08/31 02:20:32]

Nit: Why not count forever?  Seems easier and less confusing.

[rvargas (doing something else), 2011/08/31 02:33:19]

Not easier because then I would have to either clamp the value or care about
integer overflows. To be fair, this is also a clamp, just at a lower value (so
few instructions saved). I could as easily change the constant to be kint32max
if you prefer it that way.

[gavinp, 2011/08/31 03:01:20]

I'm not convinced a clamp is necessary, because if I'm not missing something,
the overflow is about 2043 years, right?  Admittedly, since we use a signed int,
so after that we then would not evict cache entries for 2043 years + 5 minutes,
which is a long time, and so the disk would probably get full...

[rvargas (doing something else), 2011/08/31 18:06:35]

Convinced. Clamp gone.

+    up_ticks_++;
 
   if (!data_)
     first_timer_ = false;