Re: [PATCH vfs.all 22/26] block: stash a bdev_file to read/write raw blcok_device

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

On Sat, Apr 06, 2024 at 08:42:06PM +0100, Al Viro wrote:
> On Sat, Apr 06, 2024 at 05:09:26PM +0800, Yu Kuai wrote:
> > From: Yu Kuai <yukuai3@huawei.com>
> > 
> > So that iomap and bffer_head can convert to use bdev_file in following
> > patches.
> 
> Let me see if I got it straight.  You introduce dummy struct file instances
> (no methods, nothing).  The *ONLY* purpose they serve is to correspond to
> opened instances of struct bdev.  No other use is possible.
> 
> You shove them into ->i_private of bdevfs inodes.  Lifetime rules are...
> odd.
> 
> In bdev_open() you arrange for such beast to be present.  You never
> return it anywhere, they only get accessed via ->i_private, exposing
> it at least to fs/buffer.c.  Reference to those suckers get stored
> (without grabbing refcount) into buffer_head instances.
> 
> And all of that is for... what, exactly?

Put another way, what's the endgame here?  Are you going to try and
propagate those beasts down into bio_alloc()?  Because if you do not,
you need to keep struct block_device * around anyway.

We use ->b_bdev for several things:
	* passing to bio_alloc() (quite a few places)
	* %pg in debugging printks
	* (rare) passing to write_boundary_block().
	* (twice) passing to clean_bdev_aliases().
	* (once) passing to __find_get_block().
	* one irregular use as a key in lookup_bh_lru()

IDGI...