|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: iSCSI: URL schemeI would like to propose the following requirement for any naming scheme. * Any naming scheme proposed MUST support multiple targets behind a single IP address * Any naming scheme proposed SHOULD support multiple targets behind a single port. This helps make traffic analysis easier. Consider the topology pictured below. An initiator is connected to a private network that has a gateway to the Internet. Similarly, the target is connected to a private network which is connected by another gateway to the Internet. Initiator - Private Network - Gateway 1 \ \ Internet / / Target - Private Network - Gateway 2 (bob) | Target (fred) How does the Initiator get to talk to the Target bob? ------------- Current model ------------- SCSI device identifier is a string in two parts: a hostname and a host-specific name (HSN). The hostname is either a domain name or an IPv4 or IPv6 address. Steps: 0) Configuration mechanism tells initiator to talk to SCSI device identifier "bob" 1) Do a DNS (or other name service) lookup on bob. Reply arrives with IP address of Gateway #2 2) TCP connect to gateway #2 on well known iSCSI port 3) As part of first login packet, send "Target: bob" 4) Gateway #2 resolves the name "bob" on the internal network and gets the private IP address of the Target Bob 5) Gateway #2 opens a TCP connection to Target Bob and passes all traffic between connections Third party command from bob to fred 0) Configuration tells initiator that bob should talk to SCSI device identifier "fred" 1) Initiator sends command to bob which has SCSI device identifier "fred" 2) Bob resolves fred's name and gets Fred's private Ip address 3) Bob connects to Fred and they do their thing Refinements ----------- Configuration mechanism (e.g. LDAP server) in step 0, along with returning the SCSI device identifier "bob", also returns the IP address of gateway #2. The DNS lookup in #1 is avoided. DNS lookup of bob actually returns gateway #1 (would work for iSCSI and HTTP/1.1 but unlikely to work for other protocols). Gateway #1 then looks up the name on the Internet and gets gateway #2... and so on... ---------------- IP address model ---------------- A target name is made of two parts: an IP address and a target ID. The target ID can be either a string or a fixed-length binary quantity. The target ID is not necessarily globally unique; it can be IP-address specific. Step 0) Initiator queries the configuration mechanisms and gets the IP address of gateway #2 and the target ID for bob. Step 1) Open TCP connection to IP address of gateway #2 on well-known iSCSI port Step 2) As part of first login packet, initiator send the target ID for bob Step 3) Gateway #2 opens connection to bob based on the target ID Step 4) Gateway #2 passes all info between connections Third party commands - passable way -------------------- Step 0) Configuartion mechanism tells initiator that bob should talk to fred. The configuration mechanism tells the initiator about the private IP address/target ID of Fred as seen by Bob (how does it find this out??) Step 1) Initiator sends third party command with Fred's IP address and target ID Step 2) Bob opens connection to Fred's IP address+port Third party commands - icky way -------------------- Step 0) Configuartion mechanism tells initiator that bob should talk to fred. The configuration mechanism tells the initiator about the IP address of gateway #2 and target ID of fred Step 1) Initiator sends third party command with IP address of gateway#2 and target ID of fred Step 2) Gateway #2 intercepts the third party command and rewrites it to have Fred's private IP address (based on the target ID) Step 3) Bob opens connection to Fred's private IP address on well known iSCSI port, sends target ID in the first iSCSI PDU, etc. or Step 2) Bob opens a connection to gateway #2's IP address and passes target ID Step 3) Bob gets redirected to Fred (or... gateway opens connection to Fred and passes all commands between Fred & Bob) In this approach, the gateway must be able to map a target ID to an IP address and a (possibly new) target ID. Note on using port numbers instead of target IDs ------------------------------------------------ The only thing that changes in the analysis above is that the target ID need not be sent on connection open, since the destination port number on the TCP connection identifies the target. In this specific scenario, it seems to me that using strings entirely is cleaner than passing IP addresses which are potentially non-sensical. -Costa
Home Last updated: Tue Sep 04 01:06:51 2001 6315 messages in chronological order |