[EMAIL PROTECTED] wrote: > Galen Shipman wrote: > >> I would like to be able to extend an existing registration such that >> the driver would take advantage of the fact that part of the extended >> registration is already registered, i.e. only the "new" memory would >> be pinned and made resident. > > Assuming you are referring to registration of virtual memory > from user space, this seems as a request for the hw drivers > to support overlapping memory regions in the sense that MR A > can overlap with MR B such that till both A and B are > unregistered the overlapped section is in place (pinned , > resident, mapped in the HCA MMU etc). >
Are you asking for the semantics of a reregister for virtual memory? If so, that would imply that it is the equivalent of deregistering and reregistering, except that the provider SHOULD optimize the process and minimize risk of giving up resources that it might not get back. That is different than simply extending an existing memory region, for example having some middleware register the widest range of the user's data space that has been seen. That would imply that you could expand an MR even while MWs were bound to it. >> although we would prefer that it wouldn't block if possible > > mmm. All the current memory registration verbs both user and > kernel are blocking, is it an issue for you? > If you need to do memory registrations in a context where blocking is not an option then you really need FMR work requests as in RDMAC and InfiniBand 1.2 verbs. _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
