Re: [PATCH v3 1/6] block: add disk sequence number

[Luca Boccassi]

On Wed, 2021-06-23 at 16:01 +0200, Hannes Reinecke wrote:
> On 6/23/21 3:51 PM, Lennart Poettering wrote:
> > On Mi, 23.06.21 15:10, Matteo Croce (mcroce@xxxxxxxxxxxxxxxxxxx) wrote:
> > 
> > > On Wed, Jun 23, 2021 at 1:49 PM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> > > > On Wed, Jun 23, 2021 at 12:58:53PM +0200, Matteo Croce wrote:
> > > > > +void inc_diskseq(struct gendisk *disk)
> > > > > +{
> > > > > +     static atomic64_t diskseq;
> > > > 
> > > > Please don't hide file scope variables in functions.
> > > > 
> > > 
> > > I just didn't want to clobber that file namespace, as that is the only
> > > point where it's used.
> > > 
> > > > Can you explain a little more why we need a global sequence count vs
> > > > a per-disk one here?
> > > 
> > > The point of the whole series is to have an unique sequence number for
> > > all the disks.
> > > Events can arrive to the userspace delayed or out-of-order, so this
> > > helps to correlate events to the disk.
> > > It might seem strange, but there isn't a way to do this yet, so I come
> > > up with a global, monotonically incrementing number.
> > 
> > To extend on this and given an example why the *global* sequence number
> > matters:
> > 
> > Consider you plug in a USB storage key, and it gets named
> > /dev/sda. You unplug it, the kernel structures for that device all
> > disappear. Then you plug in a different USB storage key, and since
> > it's the only one it will too be called /dev/sda.
> > 
> > With the global sequence number we can still distinguish these two
> > devices even though otherwise they can look pretty much identical. If
> > we had per-device counters then this would fall flat because the
> > counter would be flushed out when the device disappears and when a device
> > reappears under the same generic name we couldn't assign it a
> > different sequence number than before.
> > 
> > Thus: a global instead of local sequence number counter is absolutely
> > *key* for the problem this is supposed to solve
> > 
> Well ... except that you'll need to keep track of the numbers (otherwise 
> you wouldn't know if the numbers changed, right?).
> And if you keep track of the numbers you probably will have to implement 
> an uevent listener to get the events in time.
> But if you have an uevent listener you will also get the add/remove 
> events for these devices.
> And if you get add and remove events you can as well implement sequence 
> numbers in your application, seeing that you have all information 
> allowing you to do so.
> So why burden the kernel with it?
> 
> Cheers,
> 
> Hannes

Hi,

We need this so that we can reliably correlate events to instances of a
device. Events alone cannot solve this problem, because events _are_
the problem.

-- 
Kind regards,
Luca Boccassi

Attachment: signature.asc
Description: This is a digitally signed message part