[PATCH 4/5] io_uring: add support for dma pre-mapping

[Keith Busch]

On Wed, Jul 27, 2022 at 03:11:05PM +0100, Al Viro wrote:
> On Tue, Jul 26, 2022 at 10:38:13AM -0700, Keith Busch wrote:
> 
> > +	file = fget(map.fd);
> > +	if (!file)
> > +		return -EBADF;
> > +
> > +	if (S_ISBLK(file_inode(file)->i_mode))
> > +		bdev = I_BDEV(file->f_mapping->host);
> > +	else if (S_ISREG(file_inode(file)->i_mode))
> > +		bdev = file->f_inode->i_sb->s_bdev;
> > +	else
> > +		return -EOPNOTSUPP;
> > +
> > +	for (i = map.buf_start; i < map.buf_end; i++) {
> > +		struct io_mapped_ubuf *imu = ctx->user_bufs[i];
> > +		void *tag;
> > +
> > +		if (imu->dma_tag) {
> > +			ret = -EBUSY;
> > +			goto err;
> > +		}
> > +
> > +		tag = block_dma_map(bdev, imu->bvec, imu->nr_bvecs);
> > +		if (IS_ERR(tag)) {
> > +			ret = PTR_ERR(tag);
> > +			goto err;
> > +		}
> > +
> > +		imu->dma_tag = tag;
> > +		imu->dma_file = file;
> > +		imu->bdev = bdev;
> > +	}
> > +
> > +	fput(file);
> 
> This, BTW, is completely insane - what happens if you follow that
> with close(map.fd)?  A bunch of dangling struct file references?

This should have been tied to files registered with the io_uring instance
holding a reference, and cleaned up when the files are unregistered. I may be
missing some cases here, so I'll fix that up.

> I really don't understand what you are trying to do here

We want to register userspace addresses with the block_device just once. We can
skip costly per-IO setup this way.