True number of device openers in 2.5

True number of device openers in 2.5

[Andrey Borzenkov]

For single-partition devices it is just bd_openers but I am more intersted in 
multi-partition case. Here I need to know about sum of opens for all 
partitions including the whole device. It appears that for a container 
bd_openers is incremented for every open of itself and for the _first_ open 
of partition and bd_part_count is incremented for every (including first) 
open of partition. So bdev->bd_contains->bd_openers + 
bdev->bd_contains->bd_part_count really gives an access (number of opened 
partitions but this is unknown otherwise).

Thank you

-andrey

Re: True number of device openers in 2.5

[viro]

On Thu, Jun 12, 2003 at 07:37:53PM +0400, Andrey Borzenkov wrote:
> For single-partition devices it is just bd_openers but I am more intersted in 
> multi-partition case. Here I need to know about sum of opens for all 
> partitions including the whole device. It appears that for a container 
> bd_openers is incremented for every open of itself and for the _first_ open 
> of partition and bd_part_count is incremented for every (including first) 
> open of partition. So bdev->bd_contains->bd_openers + 
> bdev->bd_contains->bd_part_count really gives an access (number of opened 
> partitions but this is unknown otherwise).

There is no way to get anything useful out of that.  _Any_ kernel code
that relies on such counters to tell if there's somebody else who might've
accessed the thing does not account for the fact that fork() and dup()
do not go beyond the struct file.

IOW, any place that does
	if (number of openers == 1)
		do something that breaks if IO is going on
is FUBAR.  Variant with
	if (number of openers == 0)
		block ->open()
		do something ...
will work, but it means that we are not triggering it from ioctl() on that
device.  And here we only care about zero/non-zero, so ->bd_openers on
the entire disk will do just fine.

Re: True number of device openers in 2.5

[Andrey Borzenkov]

On Thursday 12 June 2003 20:37, viro@parcelfarce.linux.theplanet.co.uk wrote:
> On Thu, Jun 12, 2003 at 07:37:53PM +0400, Andrey Borzenkov wrote:
> > For single-partition devices it is just bd_openers but I am more
> > intersted in multi-partition case. Here I need to know about sum of opens
> > for all partitions including the whole device. It appears that for a
> > container bd_openers is incremented for every open of itself and for the
> > _first_ open of partition and bd_part_count is incremented for every
> > (including first) open of partition. So bdev->bd_contains->bd_openers +
> > bdev->bd_contains->bd_part_count really gives an access (number of opened
> > partitions but this is unknown otherwise).
>
> There is no way to get anything useful out of that.  _Any_ kernel code
> that relies on such counters to tell if there's somebody else who might've
> accessed the thing does not account for the fact that fork() and dup()
> do not go beyond the struct file.
>
> IOW, any place that does
> 	if (number of openers == 1)
> 		do something that breaks if IO is going on
> is FUBAR.  Variant with

Hmm ... this is exactly what EJECT ioctls for some removables do. ide-floppy 
and cdrom for sure. sd does not do it because it does not actually support 
atomic EJECT ioctl so far. I am not interested in anything more than that.

> 	if (number of openers == 0)
> 		block ->open()
> 		do something ...
> will work, but it means that we are not triggering it from ioctl() on that
> device.  And here we only care about zero/non-zero, so ->bd_openers on
> the entire disk will do just fine.

My concern is EJECT implemetation. So far EJECT was handled by individual 
drivers that used private usage counters to decide when drive was free. Now 
CDEJECT is moved into drivers/block/scsi_ioctl.c which means that at least 
for IDE and SCSI removables device driver is never entered here. I am looking 
for reliable method to detect busy device on block layer - and yes, the code 
is triggered from ioctl (mostly).

regards

-andrey