On Wed, 2011-11-30 at 05:17 -0500, Christoph Hellwig wrote: > On Tue, Nov 29, 2011 at 11:27:13PM -0800, Nicholas A. Bellinger wrote: > > So I'm still undecided if pushing this completely out into fabric > > drivers in order to drop the current complexity of queue full with > > target-core is a better tradeoff for not.. > > > > This type of core internal queue-full handling is used by ib_srpt and > > other fabrics potentially using ACK_KREF, so I think we want to at least > > still handle generically for fabrics (if necessary) that need it atm. > > How exactly do they use it? > WRITEs signal a requeue of se_cmd with write-pending payload in another transport_processing_thread context after the initial process context returned -EAGAIN or -ENOMEM.. For READs + non GOOD status, use the same errnos to queue up attempts in transport_processing_thread context to get a hardware fabric descriptor resource. > Note that we really can't do better than the dumb busy waiting done now > in the core code. ... > The ressources generally are per-hba, and without > introducing a huge mess of methods for waiting and wakeups it's > impossible to get this right outside the drivers. Now that we use > workqueues for the submission and completion pathes blocing inside > ->queue_data_in and ->queue_status isn't a problem, and all these issues > can be fixed pretty trivially by adding a waitqueue in the driver. > Don't we still want to be able to enforce target_submit_cmd() write ordering on queue-full, instead of leaving this completely up being done by fabrics..? > > > + } > > > + > > > + return error; > > > } > > > EXPORT_SYMBOL(qla_tgt_xmit_response); > > > > > > > Note outgoing TFO->write_pending() -> qla_tgt_rdy_to_xfer() failures > > will also need to be addressed for this to work as expected for FCP > > WRITE. > > Indeed. But it's just another waiter on the waitqueue. What's more > important is finding out where it should be woken. > > Given that you wrote the initial code - what is the actual resource > the we run out of, and where does it get used/released? Running out of HW fabric descriptors for HW target mode fabrics. Anyways, I'm fine with it once write_pending for qla2xxx and ib_srpt can be sorted out as well. --nab -- To unsubscribe from this list: send the line "unsubscribe target-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
