On Sat, 2011-05-28 at 13:31 -0500, James Bottomley wrote: > On Fri, 2011-05-27 at 17:52 -0700, Nicholas A. Bellinger wrote: > > On Fri, 2011-05-27 at 18:23 -0500, James Bottomley wrote: > > > On Thu, 2011-05-26 at 17:07 -0700, Nicholas A. Bellinger wrote: > > > > On Fri, 2011-05-27 at 08:47 +0900, FUJITA Tomonori wrote: > > > > > On Thu, 26 May 2011 16:28:10 -0700 > > > > > "Nicholas A. Bellinger" <nab@xxxxxxxxxxxxxxx> wrote: > > > > > > > > > > > As we have discussed at length over the years, the split needs to be all > > > > > > userspace or all kernelspace, and when implementations start doing > > > > > > things in-between they quickly get painful to debug, maintain and > > > > > > extend. I have no interest in trying to evolve this further when LIO > > > > > > > > > > Sorry, I disagree. As I explained, once user space passes established nexuses > > > > > to kernel, kernel handles all. I don't think it's painful. > > > > > > > > > > > > > > > > > > Then we are going to have to agree to disgree on this for an individual > > > > target endpoint context and being able to manage (via configfs) a > > > > complete set of iscsi-target features with native python library code. > > > > > > > > As for moving the mainline iscsi-target efforts to a more complex > > > > default direction is something that we (speaking as LIO maintainer and > > > > on behalf of RisingTide userspace) do not have an interest for an > > > > initial merge. We owe our users a complete set of functional and stable > > > > kernel and userspace library+shell, and not an untested design with > > > > undetermined time-frame for deployment. > > > > > > OK, so I understand the commercial imperatives. However, when a trusted > > > reviewer raises issues, and I check and find myself agreeing, I need > > > them addressed to make forward progress. > > > > > > > It's more than that. It's about the future maintainability of the > > target kernel code. Doing this reset to push the entire iSCSI login > > into userspace for drivers/target/iscsi/ is technically the wrong > > direction. > > You've stated this many times, but never actually given an explanation > of why. > I have been laying out the technical reasons against this, but these have apparently been snipped from my last response. :( > > We already have a userspace target, and going down this > > direction (again) for iscsi-target is not making forward progress. > > This isn't the correct way to look at it. The choice isn't all in > userspace or all in kernel space. For performance, the key idea is to > have all the fast paths in kernel space. However, login is hardly a > fast path. It's done once at start of day, then occasionally on > connection reset. > > Moving some code to userspace doesn't mean all has to move. > We already completely drive configuration from userspace, and have a python object library for doing this all of this so the application level developer does not have to care way the underyling configfs layout functions or actually looks like. This is modern kernel/user split that makes the most sense for iscsi-target mode and has taken years of effort to achieve, and is what I have committed to support to Linus (CC'ed) Trying to change that model by pushing iSCSI login state machine into userspace is an acedemic project, and going to produce poor results compared to what we already have. > > > The issue is simple: > > > > > > * We can put all the auth mechanisms in the kernel, so we need a > > > userspace upcall anyway > > > * Since we have to have an upcall, it should be the default path > > > for everything (so it gets well tested). > > > > > > Just saying "everything has to be in the kernel because mixed > > > user/kernel code is too complex" doesn't fly because we already have a > > > growing list of counter examples, some of which were cited. > > > > > > > I am not saying that everything has to be in the kernel. I am saying > > that the main iSCSI login state machines that *do* *not* effect > > authentication payloads of the iSCSI login process need to be in kernel > > for a kernel-level iscsi-target stack in order to properly support the > > real-time management of the /sys/kernel/config/target/iscsi/ control > > plane. > > So what I hear you say is that the login process has to be in-kernel > because otherwise we can't control it from configfs. > > I don't think this to be true: you can still have a configfs based > control plane and just do user space upcalls for the pieces that aren't > in-kernel. I agree with the goal of single control plane, but if > configfs is causing us to be inflexible, I think configfs, rather than > the design, needs to change. > Again, it's about consistency with the userspace library and the ability to access all of the underlying iscsi-target functionality from an object library for application developers, integrators, and test script writers. I understand from your high-level perspective it sounds like pushing iSCSI login to userspace does not make any different to how this work from the persective of rtslib, but it very much does so. > > Being able to support the proper shutdown of all outstanding I/O across > > N:M mapped iSCSI network portals <-> iSCSI TargetName > > +TargetPortalGroupTag endpoints is already not trivial, and wanting to > > move existing iscsi_target_login_thread() logic to (multi-threaded..?) > > userspace where we now have to sychronize with everything going on in > > the kernel is where there is a serious technical problem with what you > > are suggesting that is now required for an initial iscsi-target merge. > > I really don't see this. The broad abstract state machine looks like > > > |Target Configuration and initialisation| > | > | +----------------------------- (relogin) > V V | > |Login (userspace)| ----> |Operation (Kernel)| -> |Termination| > > That's nicely separable and doesn't really involve state > synchronisation. Nothing really says that Termination has to be in > userspace (or kernel space). > Wrong, this is a overly simplfied view of how things actually work. It involves alot of state sychronization. Working from iscsi_target_nego.c:iscsi_target_locate_portal(), a login attempt has to: *) Locate TargetName (tiqn) via iscsit_get_tiqn_for_login() *) Locate TargetPortalGroupTag (conn->tpg) via iscsit_get_tpg_from_np() *) Locate NodeACL (se_node_acl) via core_tpg_check_initiator_node_acl() The three structures that are located reference underlying struct config_group objects at tiqn->tiqn_wwn->wwn_group, tpg->tpg_se_tpg->tpg_group and se_node_acl->acl_group respectively. We have to consider that any of these objects may be removed / changed via configfs at any time. That all means sychronization of kernel objects if the complete login statemachine is living in userspace. But that's not even the whole story.. We also have to deal with iSCSI session reinstatement during the login process in iscsi_target_login.c:iscsi_check_for_session_reinstatement() which needs to: *) Walk the list of active sessions on the endpoint to find a match *) If an active session is found, force reinstatement via iscsit_stop_session(), and once that is complete call iscsit_close_session() to release the session. The forcing reinstatement involves waiting for all outstanding I/O to complete for the active nexus. Again, this all means sychronization of kernel objects if the complete login state machine is living in userspace. This already is by no means trivial code, and pushing this into userspace unnecessary complicates everything that is already has to be done, with no tangible upside benefit. --nab -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
