On Wed, Jun 23, 2021 at 02:13:25PM +0100, Luca Boccassi wrote: > On Wed, 2021-06-23 at 12:57 +0100, Christoph Hellwig wrote: > > On Wed, Jun 23, 2021 at 12:58:58PM +0200, Matteo Croce wrote: > > > From: Matteo Croce <mcroce@xxxxxxxxxxxxx> > > > > > > On a very loaded system, if there are many events queued up from multiple > > > attach/detach cycles, it's impossible to match them up with the > > > LOOP_CONFIGURE or LOOP_SET_FD call, since we don't know where the position > > > of our own association in the queue is[1]. > > > Not even an empty uevent queue is a reliable indication that we already > > > received the uevent we were waiting for, since with multi-partition block > > > devices each partition's event is queued asynchronously and might be > > > delivered later. > > > > > > Increment the disk sequence number when setting or changing the backing > > > file, so the userspace knows which backing file generated the event: > > > > Instead of manually incrementing the sequence here, can we make loop > > generate the DISK_EVENT_MEDIA_CHANGE event on a backing device (aka > > media) change? > > Hi, > > This was answered in the v1 thread: > > https://lore.kernel.org/linux-fsdevel/20210315201331.GA2577561@xxxxxxxxxxxxxxxxxxxx/t/#m8a677028572e826352cbb1e19d1b9c1f3b6bff4b > > The fundamental issue is that we'd be back at trying to correlate > events to loopdev instances, which does not work reliably - hence this > patch series. With the new ioctl, we can get the id immediately and > without delay when we create the device, with no possible races. Then > we can handle events reliably, as we can correlate correctly in all > cases. I very much disagree with your reply there. The device now points to a different media. Both for the loop device, a floppy or a CD changer probably by some kind of user action. In the last cast it might even by done entirely locally through a script just like the loop device.
