On 12/05/2015 21:54, Jason Gunthorpe wrote: > On Tue, May 12, 2015 at 09:50:51AM +0300, Haggai Eran wrote: >> Looking at the code though, I now notice that the other call site to >> cm_destroy_id, from within the error path of cm_process_work could now >> theoretically destroy an ID with existing references. Is that what you >> meant? > > No, but that is certainly a problem. > >> Since only listening CM IDs are now shared in RDMA CM, this should not >> happen in this patch-set, but perhaps the code can be changed to >> make > > I think you need to enforce those semantics.. > > Firstly, it looks to me like we, again, have two krefs, the one you > added and the 'ref_count' in the priv structure which is 99% of a > kref. So, again, don't do that. > > If you want to share listening CM IDs, then do exactly and only > that. Use the existing ref count scheme for keeping track of the > kfree/etc, The existing reference count scheme is for synchronization in cm_destroy_id. The function waits for active handlers to complete before destroying the id. Decrementing ref_count to zero doesn't cause the id to be freed. > and add some kind of sharable listen ref count. That's basically what the patch does. I can change it from a kref to a direct reference count if you prefer. > Early exit > from cm_destroy_id when the there are still people listening. > > That sounds like it keeps the basic rule of cm_destroy_id being > properly paired with the alloc, and allows listen sharing without the > confusion of what does multiple destroy mean. Won't you find it confusing if a call to ib_destroy_cm_id is successful but the id actually continues to live? I prefer the API to explicitly show that you are only decreasing the reference count and that the id might not be destroyed immediately. Regards, Haggai -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
