> Not necessarily. If the block that is written contains some sort
> of checksum as the last piece of information that is written in
> the block, the validity of the block can be easily checked. Thus
> if the whole block (together with the checksum) is written and then
> the machine crashes, the block will still contain valid data. And
> if the machine crashes before the checksum is written, the block
> will be invalid and will be reported as such when the file is
> accessed.
this breaks things like 'DMA 64K worth of file pages directly into
user-space', which Linux does very well currently.
but no, soft updates aka 'delayed updates' work differently, and it's a
nice concept. Read the paper ;) I havent found any negative thing, other
than the cost of double-updating filesystem metadata. (but since this
method _really_ gives an always correct filesystem, it can be considered
an 'i can live with it')
another nice thing about soft updates is that the higher level side is
totally asynchronous, just like ext2fs is currently. So nothing will
_ever_ sleep on a write. (other than getting free memory that is, or
explicit fsync() or mount)
-- mingo