I did not realize my last patch was so large. It looks like it was to large for linux-scsi, so here is the code broken up into a patchset which is what I should have done first. These patches were made against 2.6.15-rc2, but should apply against scsi-misc. For a open-iscsi svn users a patch is here: http://www.cs.wisc.edu/~michaelc/iscsi/open-iscsi/transport-class/v2/ Now on to the problem: The current iscsi class is not usable by HW iscsi drivers. This patch refactors the class code into a interface and lib so that we both can use the same class. SW iscsi currently allocates a scsi_host per session (roughly a scsi_host per I_T Nexus, or in other words I guess you could think of it as a scsi_host per target port (not really but I guess close enough for this discussion)). Some nice benefits to this were: - Finer Error handler. A host reset can map to a target reset (or iscsi session level recovery) and there is no waiting for every target to quiesce during timeouts. - Scanning is very simple and fast becuase a session is not serialized through the host and we do not need a workqueue. When we add a session, we just kick out to userspace and scan from there (the FC rport code could probbaly go to usersacpe too it looks like). - There are probably others but I am too tired. So for this transport class reshaping, I continued the scsi_host per session allocation for HW iscsi cards too. This may be similar to libata's scsi_host per port. When a HW iSCSI LLD is in it's driver->probe function they can call iscsi_alloc/add_iscsi_host(). This creates and adds a iscsi_host (similar to libatas's host_set) which is used similar to how the scsi_host is in a SPI driver. Then when a iSCSI LLD creates a session it can call iscsi_create_session() (iscsi_create_session() can also be called from a userspace event if like for open-iscsi discovery is done in userspace). iscsi_create_session, will then allocate a scsi_host and setup a session. We could also move the libata hostset concept to scsi-ml too and let the transport classes do whatever for the scsi_host usage? Just a thought. I added the setup and teardown code and a basic queucommand lib function. The next step would be to further libata'ify iscsi and add the iscsi error handler code. For the eh code though (and some of the queucommand lib fn), I think I would like to somehow break out some of the SAS task management code and scsi_error.c code so that we can share it. - : 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
