Pageset2 pages changed!

Being challenged about your assumptions is a good thing.

Not long ago, Rafael expressed doubt about whether I could really be sure that the part of the image we save before the atomic copy ('pageset2') was really not changing while being saved. So I implemented a simple check, MD5 checksumming all the pages prior to starting to save them, and doing it again afterwards. As a result, I learnt that there are indeed instances where a small number of the pages' contents do change.

So now I'm on the hunt. The question, of course, is "Why?". At the moment, I'm tracing down the functions that add (directly or indirectly) pages to the LRU, to learn more about it's contents. I hope to then be able to see what is special about the pages that are being changed, so that I can either stop them being changed (freeze what's not being frozen at the moment) or use their characteristics to have them put in the atomically copied pages to start with.

All this takes time, however. And I don't have a lot of time at the moment, so it's taking more than usual.

Oh well.