On Mon June 5 2006 15:36, Roland Dreier wrote: > Boyd> I have a question about the intent of the mutex lock > Boyd> ib_uverbs_idr_mutex used in kernel interface from user > Boyd> libraries. It appears to be a lock on the idr linked lists > Boyd> but as a great many, if not all, of the ib_uverbs commands > Boyd> grab the mutex at the start of the function and hold it to > Boyd> the end, it acts to serialize all the library accesses to > Boyd> the kernel. Is this intended? > > Yes, when I first implemented things it made things a lot easier to > serialize things. For example holding the mutex during the entire > process of creating a QP prevents the associated CQs from being > destroyed in the middle of the operation. > > It does seem to be a scalability problem for some devices/workloads, > and Robert Walsh from qlogic was planning on looking at this > (replacing the mutex with a reference counting scheme). I don't know > if he's started on this or not. > > - R.
It is in the pipe. Sweet. Thanks, Boyd -- Boyd R. Faulkner Open Grid Computing, Inc. Phone: 512-343-9196 x109 Fax: 512-343-5450 _______________________________________________ openib-general mailing list [email protected] http://openib.org/mailman/listinfo/openib-general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
