On Sun, Apr 07, 2024 at 06:11:19AM +0100, Al Viro wrote: > On Sun, Apr 07, 2024 at 05:57:58AM +0100, Al Viro wrote: > > > PS: in grow_dev_folio() we probably want > > struct address_space *mapping = bdev->bd_inode->i_mapping; > > instead of > > struct inode *inode = bdev->bd_inode; > > as one of the preliminary chunks. > > FWIW, it really looks like address_space (== page cache of block device, > > not an unreasonably candidate for primitive) and block size (well, > > logarithm thereof) cover the majority of what remains, with device > > size possibly being (remote) third... > > Incidentally, how painful would it be to switch __bread_gfp() and __bread() > to passing *logarithm* of block size instead of block size? And possibly > supply the same to clean_bdev_aliases()... > > That would reduce fs/buffer.c uses to just "give me the address_space of > that block device"... ... and from what I've seen in your series, it very much looks like after that we could replace ->bd_inode with ->bd_mapping, turning your bdev_mapping() into an inline and (hopefully) leaving the few remaining uses of bdev_inode() outside of block/bdev.c _not_ on hot paths. If nothing else, it would make it much easier to grep for remaining odd stuff. Might trim the btrfs parts of the series, at that - a lot of that seems to be "how do we propagate opened file instead of just bdev, so that we could get to its ->f_mapping deep in call chain"... Again, all of that is only if __bread...() conversion to log(size) is feasible without a massive PITA - there might be dragons...
