On Fri, Aug 25, 2023 at 03:47:56PM +0200, Jan Kara wrote:

I can see the appeal of not having to introduce the new bdev_handle type
and just using struct file which unifies in-kernel and userspace block
device opens. But I can see downsides too - the last fput() happening from
task work makes me a bit nervous whether it will not break something
somewhere with exclusive bdev opens. Getting from struct file to bdev is
somewhat harder but I guess a helper like F_BDEV() would solve that just
fine.

So besides my last fput() worry about I think this could work and would be
probably a bit nicer than what I have. But before going and redoing the whole
series let me gather some more feedback so that we don't go back and forth.
Christoph, Christian, Jens, any opinion?

Redoing is not an issue - it can be done on top of your series just
as well.  Async behaviour of fput() might be, but...  need to look
through the actual users; for a lot of them it's perfectly fine.

FWIW, from a cursory look there appears to be a missing primitive: take
an opened bdev (or bdev_handle, with your variant, or opened file if we
go that way eventually) and claim it.

I mean, look at claim_swapfile() for example:
                p->bdev = blkdev_get_by_dev(inode->i_rdev,
                                   FMODE_READ | FMODE_WRITE | FMODE_EXCL, p);
                if (IS_ERR(p->bdev)) {
                        error = PTR_ERR(p->bdev);
                        p->bdev = NULL;
                        return error;
                }
                p->old_block_size = block_size(p->bdev);
                error = set_blocksize(p->bdev, PAGE_SIZE);
                if (error < 0)
                        return error;
we already have the file opened, and we keep it opened all the way until
the swapoff(2); here we have noticed that it's a block device and we
	* open the fucker again (by device number), this time claiming
it with our swap_info_struct as holder, to be closed at swapoff(2) time
(just before we close the file)
	* flip the block size to PAGE_SIZE, to be reverted at swapoff(2)
time That really looks like it ought to be
	* take the opened file, see that it's a block device
	* try to claim it with that holder
	* on success, flip the block size
with close_filp() in the swapoff(2) (or failure exit path in swapon(2))
doing what it would've done for an O_EXCL opened block device.
The only difference from O_EXCL userland open is that here we would
end up with holder pointing not to struct file in question, but to our
swap_info_struct.  It will do the right thing.

This extra open is entirely due to "well, we need to claim it and the
primitive that does that happens to be tied to opening"; feels rather
counter-intuitive.

For that matter, we could add an explicit "unclaim" primitive - might
be easier to follow.  That would add another example where that could
be used - in blkdev_bszset() we have an opened block device (it's an
ioctl, after all), we want to change block size and we *really* don't
want to have that happen under a mounted filesystem.  So if it's not
opened exclusive, we do a temporary exclusive open of own and act on
that instead.   Might as well go for a temporary claim...

BTW, what happens if two threads call ioctl(fd, BLKBSZSET, &n)
for the same descriptor that happens to have been opened O_EXCL?
Without O_EXCL they would've been unable to claim the sucker at the same
time - the holder we are using is the address of a function argument,
i.e. something that points to kernel stack of the caller.  Those would
conflict and we either get set_blocksize() calls fully serialized, or
one of the callers would eat -EBUSY.  Not so in "opened with O_EXCL"
case - they can very well overlap and IIRC set_blocksize() does *not*
expect that kind of crap...  It's all under CAP_SYS_ADMIN, so it's not
as if it was a meaningful security hole anyway, but it does look fishy.