Re: swapon vs savecore dilemma

Pawel Worach wrote:
> Is fsck really that memory heavy so that it needs swap?

Yes, if you have a huge FS.

The problem is that the checking of the CG bitmaps during an fsck
require that you have all the bitmaps in core, and then linearly
traverse the entire directory structure to identify which bits
need to be cleared (to indicate that the block was deallocated
prior to the crash, without the bitmap being successfully written
out).

Because a block allocation for any file can be written (effectively)
anywhere on the disk, and there is no guarantee of cylinder group
locality, this basically means that you have to hold all the
bitmaps in memory simultaneously ...or you have to make multiple
passes over the directory structure, with the number of passes being
equal to the total number of CG's divided by the number of bitmaps
you can keep in core simultaneously.

This is hideously expensive, and it's never been implemented: you
are assumed to have enough memory (or memory and swap) available
to hold all the bitmaps in memory simultaneously.  If you have a
multiterabyte FS, passing over it 9 times instead of once with
swapping would be extremely dissatisfying: presumably you have
all that data for a reason, and need it back online as fast as
possible.

My suggestion (which has been my suggestion all along) is to add
two date stamped CG bitmap bitmaps somewhere (my favorite place
for this is to steal space at the front of inode 1, which is used
only rarely, since people don't use the whiteout feature, and
which can be made compatible with whiteouts, in any case).  Then
you don't let more than some small number be dirty simultaneously,
without flushing some of them out (you would need an additional
soft dependency to implement this).

If you did this, then you could guarantee a smaller set of data
to be simultaneously dirty, even for an arbitrarily large FS.  You
just load only those bitmaps that are marked dirty in the bitmap
logs, and do a single pass through the full directory structure.


> Wouldn't fsck -> mount -> savecore -> swapon be a more appropriate order?

If you had small enough disks, large enough RAM, or could limit
the number of CG bitmaps you had to simultaneously examine, then
yes.  Otherwise, no.

-- Terry