This is an attempt to fix issues with dir seeking in a filesystem
undergoing changes. The problem with the previous dstart-relative
position encoding is that if we deleted/created new entries outside of
our current directory, we didn't if they were inside or outside of the
current directory, so we couldn't always update our position correctly.
Instead of using a dstart-relative position, this solution crams both
the mid and rid into a single 31-bit integer. Things get a bit tight
here, so we use the current block_size as a heuristic for how many
possible rids we can ever have in a single mdir. The idea is the larger
the rid encoding needs to be, the smaller the mid encoding needs to be,
and we should, _roughly_, approach the same encoding limitation we would
have to dstart-relative position anyways.
Making some assumptions about the maximum possible number of rids in a
block gives us at most ~block_size/8 rids per mdir.
So for 4096 byte blocks (note the exact encoding is dynamic):
sbbbbbbb bbbbbbbb bbbbbbbr rrrrrrrr
^'-----------+----------''----+---'
'------------|----------------|----- sign bit (used for errors)
'----------------|----- 22-bit metadata bid
'----- 9-bit metadata rid
Note this introduced as new, significant limitation on the number of
total mdirs in the system. Normally I would be against this solution for
that reason, however if we adopt this encoding elsewhere in the system it
may improve some RAM cost and in general simplify things by being able to
store any mid in a single integer. More work needs to be done here...
This approach needs some fleshing out and has its own issues (the
offset returned by tell quickly becomes out of date if the filesystem
is modified, but is that really a problem?), but it improves over the
previous implementation by making tell always correct at that moment.
Christopher Haster
edd12e1f93