On Wed, Jun 12, 2019 at 4:32 PM Ira Weiny <ira.weiny@xxxxxxxxx> wrote: > > On Wed, Jun 12, 2019 at 03:54:19PM -0700, Dan Williams wrote: > > On Wed, Jun 12, 2019 at 3:12 PM Ira Weiny <ira.weiny@xxxxxxxxx> wrote: > > > > > > On Wed, Jun 12, 2019 at 04:14:21PM -0300, Jason Gunthorpe wrote: > > > > On Wed, Jun 12, 2019 at 02:09:07PM +0200, Jan Kara wrote: > > > > > On Wed 12-06-19 08:47:21, Jason Gunthorpe wrote: > > > > > > On Wed, Jun 12, 2019 at 12:29:17PM +0200, Jan Kara wrote: > > > > > > > > > > > > > > > The main objection to the current ODP & DAX solution is that very > > > > > > > > > little HW can actually implement it, having the alternative still > > > > > > > > > require HW support doesn't seem like progress. > > > > > > > > > > > > > > > > > > I think we will eventually start seein some HW be able to do this > > > > > > > > > invalidation, but it won't be universal, and I'd rather leave it > > > > > > > > > optional, for recovery from truely catastrophic errors (ie my DAX is > > > > > > > > > on fire, I need to unplug it). > > > > > > > > > > > > > > > > Agreed. I think software wise there is not much some of the devices can do > > > > > > > > with such an "invalidate". > > > > > > > > > > > > > > So out of curiosity: What does RDMA driver do when userspace just closes > > > > > > > the file pointing to RDMA object? It has to handle that somehow by aborting > > > > > > > everything that's going on... And I wanted similar behavior here. > > > > > > > > > > > > It aborts *everything* connected to that file descriptor. Destroying > > > > > > everything avoids creating inconsistencies that destroying a subset > > > > > > would create. > > > > > > > > > > > > What has been talked about for lease break is not destroying anything > > > > > > but very selectively saying that one memory region linked to the GUP > > > > > > is no longer functional. > > > > > > > > > > OK, so what I had in mind was that if RDMA app doesn't play by the rules > > > > > and closes the file with existing pins (and thus layout lease) we would > > > > > force it to abort everything. Yes, it is disruptive but then the app didn't > > > > > obey the rule that it has to maintain file lease while holding pins. Thus > > > > > such situation should never happen unless the app is malicious / buggy. > > > > > > > > We do have the infrastructure to completely revoke the entire > > > > *content* of a FD (this is called device disassociate). It is > > > > basically close without the app doing close. But again it only works > > > > with some drivers. However, this is more likely something a driver > > > > could support without a HW change though. > > > > > > > > It is quite destructive as it forcibly kills everything RDMA related > > > > the process(es) are doing, but it is less violent than SIGKILL, and > > > > there is perhaps a way for the app to recover from this, if it is > > > > coded for it. > > > > > > I don't think many are... I think most would effectively be "killed" if this > > > happened to them. > > > > > > > > > > > My preference would be to avoid this scenario, but if it is really > > > > necessary, we could probably build it with some work. > > > > > > > > The only case we use it today is forced HW hot unplug, so it is rarely > > > > used and only for an 'emergency' like use case. > > > > > > I'd really like to avoid this as well. I think it will be very confusing for > > > RDMA apps to have their context suddenly be invalid. I think if we have a way > > > for admins to ID who is pinning a file the admin can take more appropriate > > > action on those processes. Up to and including killing the process. > > > > Can RDMA context invalidation, "device disassociate", be inflicted on > > a process from the outside? Identifying the pid of a pin holder only > > leaves SIGKILL of the entire process as the remediation for revoking a > > pin, and I assume admins would use the finer grained invalidation > > where it was available. > > No not in the way you are describing it. As Jason said you can hotplug the > device which is "from the outside" but this would affect all users of that > device. > > Effectively, we would need a way for an admin to close a specific file > descriptor (or set of fds) which point to that file. AFAIK there is no way to > do that at all, is there? You can certainly give the lease holder the option to close the file voluntarily via the siginfo_t that can be attached to a lease break signal. But it's not really "close" you want as much as a finer grained disassociate. All that said you could require the lease taker opt-in to SIGKILL via F_SETSIG before marking the lease "exclusive". That effectively precludes failing truncate, but it's something we can enforce today and work on finer grained / less drastic escalations over time for something that should "never" happen.
