When calculating the space used for pruning, if a log chunk is
compressed, that size is used otherwise the uncompressed size is
used. This is intended to reach a steady state where 1/4 of the log
buffer is the uncompressed log chunk that is being written to and the
other 3/4 of the log buffer is compressed logs.
If we wait until there are no readers referencing the log chunk before
compressing it, we end up with 2 uncompressed logs (the one that was
just filled, that readers are still referencing, and the new one that
was allocated to fit the most recent log), which take up 1/2 of the
log buffer's allotted size and will thus cause prune to delete more
compressed logs than it should.
Instead, we should always compress the log chunks in FinishWriting()
such that the compressed size will always be used for log chunks other
than the one that is not actively written to.
Decompressed logs due to readers are ephemeral by their nature and
thus don't add to the log buffer size for pruning.
Test: observe that log buffers can be filled in the presence of a reader.
Change-Id: Ie21ccff032e41c4a0e51710cc435c5ab316563cb
SerializedFlushToState::PopNextUnreadLog() was calling
AddMinHeapEntry() to replenish the element that was just popped off of
the heap, however AddMinHeapEntry() also manages reference counts for
the buffers, and this resulting in the following scenario:
PopNextUnreadLog() returns a pointer referencing log buffer #1
AddMinHeapEntry() sees that all logs from buffer #1 has been read, so
it decrements the reference count
The caller of PopNextUnreadLog() uses the result which references
invalid memory.
This calls CheckForNewLogs() within HasUnreadLogs() instead of
requiring a separate call, which fixes an additional issue where
continuing from the loop in SerializedLogBuffer::FlushTo() may not
pick up subsequent logs in a given log buffer, since CheckForNewLogs()
wouldn't be called. This was exacerbated by the above change.
This adds a test to check the reference counts for this case and fixes
an argument mismatch in SerializedFlushToStateTest.
This adds the corpus that surfaced the issue.
Bug: 159753229
Bug: 159783005
Test: these unit tests, run fuzzer without error
Change-Id: Ib2636dfc14293b7e2cd00876b9def6e9dbbff4ce
In order of severity:
1) Add a CHECK() that a pointer is not nullptr, where the analyzer
believes this is possible.
2) Add `final` appropriately to functions called from constructors.
3) Add missing cloexec flags.
4) Add missing `noexcept` and other subtle performance warnings
Test: build with clang-tidy
Change-Id: Ifd9a1299a51027a47382926b2224748b5750d6cf
Initial commit for a SerializedLogBuffer. The intention here is for
the serialized data to be compressed (currently using zlib) to allow
for substantially longer logs in the same memory footprint.
Test: unit tests
Change-Id: I2528e4e1ff1cf3bc91130173a107f371f04d911a
Tom Cherry