mount time (was Re: Fwd: Hello -- NAND/NOR Flash design for GTA02)

[Werner Almesberger]

Neng-Yu Tu (Tony Tu) wrote:
> Well, we might need to provide more chocolate to burst his idea.

Hmm, that one had been my only trip to Taipei/China so far where I
actually gained weight ...

Anyway, there are indeed many interesting ideas to look into, once you
start thinking of it. Here's what I've found so far:

- make the partition we have to scan smaller
  - distinct partitions (Andy)
  - initramfs
  - overlay (Joachim, OpenWRT uses squashfs+jffs2)

- as a variation of the above, separate read-only data from read-write
  data - only the latter needs scanning (that's from OpenWRT again)

- do not do a linear scan of everything
  - checkpointing (yaffs2)
  - better data structures (LogFS)

- accelerate the speed at which we do the linear scan
  - make NAND access faster (hwecc, DMA, readahead)
  - skip blocks we know can't contain meta-data (*)

- access the file system before the scan completes (*)
  - order files during FS creation
  - use pointers to "quick access" files

(*) not sure to what extent, if at all, mechanisms for these exist
    yet.

Many of these concepts could also be combined.

Anyway, there's more to examine, and I need to gather a bit more
background data as well. I rather like the idea behind LogFS, because
it seems to have the smallest number of "hidden penalties", but it
also has the disadvantage of being quite unproven and not even part
of the mainline kernel yet.

- Werner