man 2 write states:

 POSIX requires that a read(2) which can be proved to occur after a write() has returned returns the new data.

This means that in the msg_store, when we're reading a msg, we only need to make sure that write has really been called on the msg. We do not need to sync here. So, add a test, explicitly to hit this condition.

We were using both buffering in the fhc, and also delayed_write. This is dangerous, because we can think that we've written out from the fhc, but even though we're writing > the 64 kbyte default limit in delayed_write, we may not be writing a whole number of 64kb blocks, thus there may be something left in the delayed_write buffer, so remove that. Performance doesn't suffer because we're writing blocks of 1MB anyway!

Also, the current_offset/1 fun in fhc was returning the virtual address - i.e. where we would be if we flushed out the write buffer. We actually need both variants - getting the real raw offset of the handle, and getting the virtual offset.

Finally, as a result of all of this, the sync fun in msg_store now has no need to do any work unless there are outstanding transactions.

0 files changed, 0 insertions, 0 deletions