Re: [PATCH] CDRW packet writing support for 2.6.7-bk13

[Christoph Hellwig <hch@infradead.org>]

On Mon, Jul 05, 2004 at 01:49:43AM +0200, Peter Osterlund wrote:
> For fsync_bdev(), which you say is not needed anyway. Also for the
> invalidate_bdev() call in pkt_remove_dev(). The loop driver is also
> calling invalidate_bdev() in its loop_clr_fd() function, so I guess
> that call is needed.

Only as long as it's reusing the same gendisk for different devices,
more on that below.

> > > +	for (i = 0; i < MAX_WRITERS; i++) {
> > > +		struct pktcdvd_device *pd = &pkt_devs[i];
> > > +		struct gendisk *disk = disks[i];
> > > +		disk->major = PACKET_MAJOR;
> > > +		disk->first_minor = i;
> > > +		disk->fops = &pktcdvd_ops;
> > > +		disk->flags = GENHD_FL_REMOVABLE;
> > > +		sprintf(disk->disk_name, "pktcdvd%d", i);
> > > +		sprintf(disk->devfs_name, "pktcdvd/%d", i);
> > > +		disk->private_data = pd;
> > > +		disk->queue = blk_alloc_queue(GFP_KERNEL);
> > > +		if (!disk->queue)
> > > +			goto out_mem3;
> > > +		add_disk(disk);
> > > +	}
> > 
> > Please allocate all these on demand only when you actually attach
> > a device.
> 
> But I need to open the device to be able to perform the ioctl to
> attach a device, and I can't open the device until add_disk() has been
> called. The loop device does the same thing.

And loops is broken the same way ;-)  This I need an blockdevice to actually
attach it scheme breaks in many ways, like the need to reset a gendisk
instead of allocating a new one, lots of warts in ->open and problems with
udev.  Just keep a single character device around for all administrative
work (like device mapper) - otoh if packet writing really moves to the
block layer all this is magically fixed..