On Thu, 11 Jul 2024 at 22:44, Jens Axboe <axboe@xxxxxxxxx> wrote:
>
> Note the statx change for atomic writes will cause a trivial conflict
> with a patch that went into the 6.9 kernel via the vfs tree later than
> what the 6.11 block tree was based on. Easy to resolve, only mentioning
> it for completeness sake.
It may be easy to resolve, but as I was resolving it I threw up in my
mouth a little.
That whole
struct inode *backing_inode;
...
backing_inode = d_backing_inode(path.dentry);
if (S_ISBLK(backing_inode->i_mode))
bdev_statx(backing_inode, stat, request_mask);
does not belong in the generic stat helper. The whole *point* of
bdev_statx() is to not need that kind of bdev-specific knowledge.
So I rewrote it to be
if (S_ISBLK(stat->mode))
bdev_statx(path, stat, request_mask);
instead, and that whole "backing_inode" stuff is now inside bdev_statx().
Because wasn't that the whole point of abstracting this out, so that
block devices could do their own thing?
And if the backing inode is a special block device, but the "front"
inode doesn't say S_IFBLK, something is very screwed up.
But hey, maybe I screwed something up. But randomly having that whole
"d_backing_inode()" logic in there seemed entirely against the
abstraction layering.
Linus
