On 06/27/13 01:04, Dave Chinner wrote:
From: Dave Chinner<dchinner@xxxxxxxxxx>
During review of the separate project quota inode patches, it bcame
obvious that the dquot log reservation calculation underestimated
the number dquots that can be modified in a transaction. This has
it's roots way back in the Irix quota implementation.
That is, when quotas were first implemented in XFS, it only
supported user and project quotas as Irix did not have group quotas.
Hence the worst case operation involving dquot modification was
calculated to involve 2 user dquots and 1 project dquot or 1 user
dequot and 2 project dquots. i.e. 3 dquots. This was determined back
in 1996, and has remained unchanged ever since.
However, back in 2001, the Linux XFS port dropped all support for
project quota and implmented group quotas over the top. This was
effectively done with a search-and-replace of project with group,
and as such the log reservation was not changed. However, with the
advent of group quotas, chmod and rename now could modify more than
3 dquots in a single transaction - both could modify 4 dquots. Hence
this log reservation has been wrong for a long time.
In 2005, project quotas were reintroduced into Linux, but they were
implemented to be mutually exclusive to group quotas, and so this
didn't add any new changes to the dquot log reservation. hence when
project quotas were in use, everything was still fine, just like
in the Irix days.
Now, with the addition of the separate project quota inode, group
and project quotas are no longer mutually exclusive, and hence
operations can now modify three dquots per inode where previously it
was only two. The worst case here is the rename transaction, which
can allocate/free space on two different directory inodes, and if
they have different uid/gid/prid configurations and are world
writeable, then rename can actually modify 6 different dquots now.
Further, the dquot log reservation doesn't take into account the
space used by the dquot log format structure that preceeds the dquot
that is logged, and hence further underestimates the worst case
log space required by dquots during a transaction.
Hence the worst case log reservation needs to be increased from 3 to
6, and it needs to take into account a log format header for each of
Signed-off-by: Dave Chinner<dchinner@xxxxxxxxxx>
Reviewed-by: Mark Tinguely <tinguely@xxxxxxx>