With all due respect Ted, you dreamed everything you described above 🙂 It doesn't do any of that except the checkpoint at the start, and that's just to flush out as many dirty pages from the buffer pool as it can, so as to reduce the amount of REDO recovery to do during the restore sequence.

[Edit: btw, a checkpoint doesn't take locks either - see my blog post for details: How do checkpoints work and what gets logged. Whatever differing descriptions you may read elsewhere are unfortunately incorrect - this description is based on the code itself.]

Take a look at my backup article I referenced a few posts above which explains exactly what happens.

It's worked the same way since 7.0 up to 2008.

Thanks