On Wed, Aug 5, 2020 at 6:07 PM David Howells [off-list ref] wrote:

Miklos Szeredi [off-list ref] wrote:

quoted

Shoun't we just make sure that the likelyhood of overruns is low

That's not necessarily easy.  To avoid overruns you need a bigger buffer.  The
buffer is preallocated from unswappable kernel space.  Yes, you can increase
the size of the buffer, but it eats out of your pipe bufferage limit.

Further, it's a *general* notifications queue, not just for a specific
purpose, but that means it might get connected to multiple sources, and doing
something like tearing down a container might generate enough notifications to
overrun the queue.

quoted

and if it happens, just reinitialize everthing from scratch (shouldn't be
*that* expensive).

If you then spend time reinitialising everything, you're increasing the
likelihood of racing with further events.  Further, there multiple expenses:
firstly, you have to tear down and discard all the data that you've spent time
setting up; secondly, it takes time doing all this; thirdly, it takes cpu
cycles away from applications.

The reason I put the event counters in there and made it so that fsinfo()
could read all the mounts in a subtree and their event counters in one go is
to make it faster for the user to find out what changed in the event that a
notification is lost.

That's just overdesigning it, IMO.

If the protocol is extensible (as you state) then the counters can be
added as needed.  And unless the above CPU cycle wastage is actually
observed in practice, the whole thing is unnecessary.

Thanks,
Miklos