Robert Love wrote:
This series represents the continuation of the RFC archived here: http://marc.info/?l=linux-scsi&m=126463466126962&w=2. The sysfs organization has taken shape and object allocations make more sense than the previous RFC. The patches have undergone my own developer testing, which has mostly focused on NPIV port creation and deletion and some I/O. This series creates a Fibre Channel subsystem that has an improved sysfs device tree (the port's view of the fabric) and begins to abstract itself from SCSI. This RFC is to solicit any feedback before I continue to refine these patches. I've created two diagrams to help people understand the change. The first shows the current sysfs device layout and the second shows the allocation scheme for libfc and fcoe's primary data structures. http://open-fcoe.org/rwlove/fc_sysfs.jpg http://open-fcoe.org/rwlove/libfc_alloc.jpg These patches apply to scsi-misc + the recently submitted fcoe patches, which is the same as the fcoe-next tree on open-fcoe.org. Overview of N_Port creation/login (for fcoe.ko): 1) First, fc.ko is installed and then scsi_transport_fcp.ko is installed. scsi_transport_fcp.ko registers itself (SCSI initiator) as a FC4 with fc.ko. 2) When a fcoe port is "created" from 'echo ethX > /sys/module/fcoe/parameters/create' a fcport is allocated and added to sysfs. All of the port's physical properties are added as attributes. <PCI root>/fcport_0 3) At this point libfc/libfcoe/fcoe needs an lport and an fcoe_port. The lport and fcoe_ports are private data of a fcvport so a fcvport is allocated, but not added to sysfs. 4) libfcoe.ko discovers FCFs and allocates/adds them as fcfabrics to sysfs as they are discovered. [ I think that we may want to revisit the idea of having a fcfport between the fcport and fcfabric. Right now I'm creating a fcfabric for each FCF that is discovered at the FIP phase, but that doesn't make much sense since multiple FCFs on the same fabric can be discovered. Adding a fcfport for each FCF discovered might be more correct. ] <PCI root>/fcport_0/fcfabric_0 5) libfc.ko will FLOGI into the fabric and upon the FLOGI ACC the fcvport will be added to sysfs. <PCI root>/fcport_0/fcfabric_0/fcvport_0 5.1) The FC layer will call fc4_init_add in the fc4_template. If SCSI is the FC4, fcp_init_add (scsi_transport_fcp) will be called. This routine calls scsi_host_alloc/scsi_add_host. There is also a callback down to the LLD so that it can setup any SCSI-FCP private data. [ I'm not sure how non-FCoE HBAs will do this. I imagine they will just create all three sysfs devices (fcport, fcfabric and fcvport) on the FLOGI ACC since they do not rely on a discovery protocol. ] <PCI root>/fcport_0/fcfabric_0/fcvport_0/fcpinit_X/hostX 6) libfc will start discovery and as remote ports are discovered libfc will notify FC to create and add fcrports to sysfs. <PCI root>/fcport_0/fcfabric_0/fcvport_0/fcpinit_X/hostX /fcrport_0 6.1) When the fcrports are created the FC layer will call fc4_targ_add in the fc4_template. If SCSI is the FC4, fcp_targ_add (scsi_transport_fcp) will be called. This routine will allocate a fcptarg device, add it to sysfs and notify SCSI to scan it. <PCI root>/fcport_0/fcfabric_0/fcvport_0/fcpinit_X/hostX /fcrport_0/fcptargX:0-0 The current series shortcomings are: 1) The FC4 interface needs work to become more FC4 agnostic. (end of include/fc/fc.h) 2) Point-to-Point mode needs to be addressed 3) BSG needs to be tested and STGT should become another FC4 that is registered with FC. Since libfc/fcoe doesn't plug into STGT yet this will be a challenge. 4) user space updates - For the fcoe-utils package, libhbalinux (user space HBA API library) needs to be updated to read the new layout before heavy testing of the new layout can be done. 5) Miscellaneous problems like error handling, fcfabric deletion cleaning up vports, using attribute containers / scsi transport style for SCSI-FCP devices, warnings, etc... 6) FCoE attributes... Other thoughts: I would love to collaborate with someone who maintains a HBA that supports native FC since I'm not aware of the specific needs of traditional HBAs. I would hope that this new layout could be added to the kernel (when ready) and the existing FC Transport deprecated as I'm not sure how to get all HBA maintainers to switch their drivers at the same time. I think that only adding FCoE attributes to this layout would be an incentive to get LLDs to migrate. I've tried to keep this covermail brief, so there are plenty of details left out. Please ask if you have a question. Thanks, //Rob
<snip> Hi Rob, I'm running the patches and they work OK (very lightly used). I'd prefer we kept the naming in /sys/class shorter and keep underscores after "fc" like we currently have in fc_remote_port. Maybe you could continue to use the existing names where they apply, after getting the other transport code converted. Here are my suggested name changes: fcvport becomes fc_vport fcvport/fcvport_0 becomes fc_vport/0 fcpinit_4 becomes fcp_init4 fcptarg-X:0-1 becomes fcp_targ/1 fcrport:xxx-0 becomes fc_rport/0 (or fc_remote_ports/xxx eventually) fcfabric/fcfabric_0 becomes fc_fabric/0 fcport/fcport_0 becomes fc_port/0 When we have local FCP target ports (as opposed to remote ones) how will we represent that? You could include that in the diagram. When we see remote initiators, how do we show them? Maybe we just don't create anything under fcport_N in that case. I may have more comments later after I look through the code a bit further. I'm modifying fcc to handle the new trees. Minor issues: I noticed that fnic no longer compiles and checkpatch spots some issues, but I'm sure those will be fixed later. Regards, Joe -- 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
