Re: [RFC v4 linux-next 19/19] fs & block: remove bdev->bd_inode

[Al Viro <viro@zeniv.linux.org.uk>]

On Fri, Mar 22, 2024 at 03:43:47PM +0000, Al Viro wrote:
> On Fri, Mar 22, 2024 at 02:52:16PM +0800, Yu Kuai wrote:
> > Hi,
> > 
> > 在 2024/03/22 14:39, Al Viro 写道:
> > > On Fri, Mar 22, 2024 at 06:37:18AM +0000, Al Viro wrote:
> > > > On Thu, Mar 21, 2024 at 08:15:06PM +0800, Yu Kuai wrote:
> > > > 
> > > > > > blkdev_iomap_begin() etc. may be an arbitrary filesystem block device
> > > > > > inode. But why can't you use I_BDEV(inode->i_mapping->host) to get to the
> > > > > > block device instead of your file_bdev(inode->i_private)? I don't see any
> > > > > > advantage in stashing away that special bdev_file into inode->i_private but
> > > > > > perhaps I'm missing something...
> > > > > > 
> > > > > 
> > > > > Because we're goning to remove the 'block_device' from iomap and
> > > > > buffer_head, and replace it with a 'bdev_file'.
> > > > 
> > > > What of that?  file_inode(file)->f_mapping->host will give you bdevfs inode
> > > > just fine...
> > > 
> > > file->f_mapping->host, obviously - sorry.
> > > .
> > 
> > Yes, we already get bdev_inode this way, and use it in
> > blkdev_iomap_begin() and blkdev_get_block(), the problem is that if we
> > want to let iomap and buffer_head to use bdev_file for raw block fops as
> > well, we need a 'bdev_file' somehow.
> 
> Explain, please.  Why would anything care whether the file is bdevfs
> one or coming from devtmpfs/xfs/ext2/whatnot?

Yecchhh...  I see one possible reason, unfortunately, but I really doubt
that your approach is workable.  iomap is not a problem; nothing in
there will persist past the destruction of struct file you've used;
buffer_head, OTOH, is a problem.  They are, by their nature,
shared between various openers and we can't really withdraw them.

Why do we want ->b_bdev replaced with struct file * in the first place?
AFAICS, your patch tries to make it unique per opened bdev; that
makes the lifetime rules really convoluted, but that aside, what's
in that struct file that is not in struct block_device?

I don't see any point trying to shove that down into buffer_head, or,
Cthulhu forbid, bio.  Details, please...