The blk files have blocks written to them as they come in. Each blk file and rev file reference the same blocks, but the order in which block data is written into the rev (undo) files is in order of activation, i.e. in block height order.
Thus, when the code decides to stop writing into blk000X.dat, it will flush this and rev000X.dat, and then move to X+1. The problem is that since blocks are activated later than they are stored in the block file, we will still be writing to rev X for awhile after this.
This code fixes this by simply finalizing the previous undo file each time we finish up a block file. This pretty much guarantees that we will never pre-allocate space and never flush for an undo file (which is happening right now).
Unfortunately for anyone with affected machines, there will be wasted drive space for each rev file if they did IBD before this fix. Probably affects all architectures.
Edit: note that this PR initially added a new critical section for undo pos and some other stuff, but I scratched that for this simpler solution, hence practicalswift’s comment below.
Fixes #17890.
