Messages in this thread | | | Date | Thu, 15 Mar 2018 15:46:06 +0200 | From | "Michael S. Tsirkin" <> | Subject | Re: [PATCH v5] drivers/misc: vm_gen_counter: initial driver implementation |
| |
On Thu, Mar 15, 2018 at 02:19:59PM +0100, Greg KH wrote: > On Wed, Mar 14, 2018 at 09:25:17PM +0200, Michael S. Tsirkin wrote: > > On Wed, Mar 14, 2018 at 07:25:36PM +0100, Greg KH wrote: > > > On Tue, Mar 13, 2018 at 07:40:51PM +0200, Michael S. Tsirkin wrote: > > > > I think it's a good idea to use sysfs for this. However, > > > > there are a couple of missing interfaces here: > > > > > > > > 1. Userspace needs a way to know when this value changes. > > > > I see no change notifications here and that does not seem right. > > > > > > How can these change? > > > > It's a hardware register. It changes when hardware feels like it :) > > Does the hardware notify the kernel that it changes? Or does the kernel > have to "poll" for it?
Yes - it sends an ACPI interrupt notification.
> > In particular, it changes whenever VM is migrated or snapshotted. > > So very rarely. And userspace always knows about those events already, > right?
Not at all.
> > > > 2. Userspace needs to be able to read these without > > > > system calls. > > > > > > Ick, what? Why not? > > > > > > > Pls add mmap support to the raw format. > > > > > > For a single integer? Why do you need mmap for this? What is so > > > "performant" that needs to touch a sysfs file? > > > > (Phys address is not guaranteed to be page-aligned so you will > > > > probably want an offset attribute for that as well). > > > > > > Ick ick ick, that's why it's good to just stick with a sysfs file. > > > > > > Have you tested just how long this takes to see if the open/read/close > > > is really the bottleneck, or if the io on reading the value is the > > > bottleneck? > > > > > > thanks, > > > > > > greg k-h > > > > Well an application needs to check this value basically after > > every database transaction. > > "every"? That's horrid, why would you write a database that has to do > an ACPI i/o call for every transaction? That's a sure way to write a > very slow database :(
Absolutely. That's why we might want to do an mmap and then get the ID from memory. An alternative is poll support so userspace can get notified about changes. Opens a bigger window during which you are doing duplicate work, but maybe that's not such a big issue.
> > So I'm pretty sure it's a performance sensitive path. > > Given that this api is not present today, why is this even needed? Who > wants/needs it so badly that it has to be tuned in ways like this? > > If it is _really_ performant critical, just make it a new syscall :)
Maybe you are right and it is a premature optimization. Let's put it out there without mmap support and see what happens - but then we definitely need poll support.
> > But yes, I > > didn't profile any apps since they > > are yet to be written to use this interface. > > Then what database are you talking about? What apps need/want this > thing? > > thanks, > > greg k-h
Anything that runs within a VM that is snapshoted is at risk of sending duplicate transactions when it's restored and time rolls back to a random point in the past.
If you want the application to have ability to detect these, then VM gen ID offers a way to do it: id=read_id() do_work() new_id=read_id() if (new_id != id) find_and_handle_duplicate_work()
-- MST
| |