On Sun, Oct 22, 2017 at 09:32:00PM +0300, Sagi Grimberg wrote: > >> Currently, blk_mq_tagset_iter() iterate over initial hctx tags only. >> In case scheduler is used, it doesn't iterate the hctx scheduler tags >> and the static request aren't been updated. >> For example, while using NVMe over Fabrics RDMA host, this cause us not to >> reinit the scheduler requests and thus not re-register all the memory regions >> during the tagset re-initialization in the reconnect flow. > > I think this is a sign that we should cease from embedding memory > regions on the pre-allocated requests. Its too much resources > that we waste. In our case, tags are not really cheap given > that they take a physical HW resource (rdma memory region). > > I think we should switch (again) to a pool design instead. I guess its > time for a generic MR pool that will serve nvmf, xprt, srp, iser and > friends. Liks drivers/infiniband/core/mr_pool.c? :)
