On Wed 05-02-20 10:50:16, Andreas Dilger wrote: > On Feb 5, 2020, at 3:01 AM, Jan Kara <jack@xxxxxxx> wrote: > > > > Implement proper creation of new directory entries in htree directories > > in ext2fs_link(). So far we just cleared EXT2_INDEX_FL and treated > > directory as unindexed however this results in mismatched checksums if > > metadata checksums are in use because checksums are placed in different > > places depending on htree node type. > > I'm definitely not agains this, as I believe it will also speed up e2fsck > for cases where a lot of entries are inserted into lost+found (sometimes > many millions of files). Currently e2fsck linearly scans the whole dir > for each insert, rather than saving the offset of the last entry. I have > a patch to fix that, but it needed several API changes and got bogged down > in performance testing and never made it out to the list. > > One potential risk is that if a directory is corrupted in some way, then > the htree index cannot always be trusted to do inserts during e2fsck, so > it might still have to fall back to clearing the flag and doing a linear > insertion. Well, but pass 2 checks internal consistency of all directories so if htree in lost+found is not usable, it will be detected and handled. So in pass 3 we should have usable lost+found (possibly without htree) to use. Also note that I have not implemented conversion of directory into indexed one so in the normal case lost+found will be just an ordinary linear directory. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
