214 lines
		
	
	
		
			10 KiB
		
	
	
	
		
			Plaintext
		
	
	
	
			
		
		
	
	
			214 lines
		
	
	
		
			10 KiB
		
	
	
	
		
			Plaintext
		
	
	
	
| ****************************
 | |
| RDMA Transport (RTRS)
 | |
| ****************************
 | |
| 
 | |
| RTRS (RDMA Transport) is a reliable high speed transport library
 | |
| which provides support to establish optimal number of connections
 | |
| between client and server machines using RDMA (InfiniBand, RoCE, iWarp)
 | |
| transport. It is optimized to transfer (read/write) IO blocks.
 | |
| 
 | |
| In its core interface it follows the BIO semantics of providing the
 | |
| possibility to either write data from an sg list to the remote side
 | |
| or to request ("read") data transfer from the remote side into a given
 | |
| sg list.
 | |
| 
 | |
| RTRS provides I/O fail-over and load-balancing capabilities by using
 | |
| multipath I/O (see "add_path" and "mp_policy" configuration entries in
 | |
| Documentation/ABI/testing/sysfs-class-rtrs-client).
 | |
| 
 | |
| RTRS is used by the RNBD (RDMA Network Block Device) modules.
 | |
| 
 | |
| ==================
 | |
| Transport protocol
 | |
| ==================
 | |
| 
 | |
| Overview
 | |
| --------
 | |
| An established connection between a client and a server is called rtrs
 | |
| session. A session is associated with a set of memory chunks reserved on the
 | |
| server side for a given client for rdma transfer. A session
 | |
| consists of multiple paths, each representing a separate physical link
 | |
| between client and server. Those are used for load balancing and failover.
 | |
| Each path consists of as many connections (QPs) as there are cpus on
 | |
| the client.
 | |
| 
 | |
| When processing an incoming write or read request, rtrs client uses memory
 | |
| chunks reserved for him on the server side. Their number, size and addresses
 | |
| need to be exchanged between client and server during the connection
 | |
| establishment phase. Apart from the memory related information client needs to
 | |
| inform the server about the session name and identify each path and connection
 | |
| individually.
 | |
| 
 | |
| On an established session client sends to server write or read messages.
 | |
| Server uses immediate field to tell the client which request is being
 | |
| acknowledged and for errno. Client uses immediate field to tell the server
 | |
| which of the memory chunks has been accessed and at which offset the message
 | |
| can be found.
 | |
| 
 | |
| Module parameter always_invalidate is introduced for the security problem
 | |
| discussed in LPC RDMA MC 2019. When always_invalidate=Y, on the server side we
 | |
| invalidate each rdma buffer before we hand it over to RNBD server and
 | |
| then pass it to the block layer. A new rkey is generated and registered for the
 | |
| buffer after it returns back from the block layer and RNBD server.
 | |
| The new rkey is sent back to the client along with the IO result.
 | |
| The procedure is the default behaviour of the driver. This invalidation and
 | |
| registration on each IO causes performance drop of up to 20%. A user of the
 | |
| driver may choose to load the modules with this mechanism switched off
 | |
| (always_invalidate=N), if he understands and can take the risk of a malicious
 | |
| client being able to corrupt memory of a server it is connected to. This might
 | |
| be a reasonable option in a scenario where all the clients and all the servers
 | |
| are located within a secure datacenter.
 | |
| 
 | |
| 
 | |
| Connection establishment
 | |
| ------------------------
 | |
| 
 | |
| 1. Client starts establishing connections belonging to a path of a session one
 | |
| by one via attaching RTRS_MSG_CON_REQ messages to the rdma_connect requests.
 | |
| Those include uuid of the session and uuid of the path to be
 | |
| established. They are used by the server to find a persisting session/path or
 | |
| to create a new one when necessary. The message also contains the protocol
 | |
| version and magic for compatibility, total number of connections per session
 | |
| (as many as cpus on the client), the id of the current connection and
 | |
| the reconnect counter, which is used to resolve the situations where
 | |
| client is trying to reconnect a path, while server is still destroying the old
 | |
| one.
 | |
| 
 | |
| 2. Server accepts the connection requests one by one and attaches
 | |
| RTRS_MSG_CONN_RSP messages to the rdma_accept. Apart from magic and
 | |
| protocol version, the messages include error code, queue depth supported by
 | |
| the server (number of memory chunks which are going to be allocated for that
 | |
| session) and the maximum size of one io, RTRS_MSG_NEW_RKEY_F flags is set
 | |
| when always_invalidate=Y.
 | |
| 
 | |
| 3. After all connections of a path are established client sends to server the
 | |
| RTRS_MSG_INFO_REQ message, containing the name of the session. This message
 | |
| requests the address information from the server.
 | |
| 
 | |
| 4. Server replies to the session info request message with RTRS_MSG_INFO_RSP,
 | |
| which contains the addresses and keys of the RDMA buffers allocated for that
 | |
| session.
 | |
| 
 | |
| 5. Session becomes connected after all paths to be established are connected
 | |
| (i.e. steps 1-4 finished for all paths requested for a session)
 | |
| 
 | |
| 6. Server and client exchange periodically heartbeat messages (empty rdma
 | |
| messages with an immediate field) which are used to detect a crash on remote
 | |
| side or network outage in an absence of IO.
 | |
| 
 | |
| 7. On any RDMA related error or in the case of a heartbeat timeout, the
 | |
| corresponding path is disconnected, all the inflight IO are failed over to a
 | |
| healthy path, if any, and the reconnect mechanism is triggered.
 | |
| 
 | |
| CLT                                     SRV
 | |
| *for each connection belonging to a path and for each path:
 | |
| RTRS_MSG_CON_REQ  ------------------->
 | |
|                    <------------------- RTRS_MSG_CON_RSP
 | |
| ...
 | |
| *after all connections are established:
 | |
| RTRS_MSG_INFO_REQ ------------------->
 | |
|                    <------------------- RTRS_MSG_INFO_RSP
 | |
| *heartbeat is started from both sides:
 | |
|                    -------------------> [RTRS_HB_MSG_IMM]
 | |
| [RTRS_HB_MSG_ACK] <-------------------
 | |
| [RTRS_HB_MSG_IMM] <-------------------
 | |
|                    -------------------> [RTRS_HB_MSG_ACK]
 | |
| 
 | |
| IO path
 | |
| -------
 | |
| 
 | |
| * Write (always_invalidate=N) *
 | |
| 
 | |
| 1. When processing a write request client selects one of the memory chunks
 | |
| on the server side and rdma writes there the user data, user header and the
 | |
| RTRS_MSG_RDMA_WRITE message. Apart from the type (write), the message only
 | |
| contains size of the user header. The client tells the server which chunk has
 | |
| been accessed and at what offset the RTRS_MSG_RDMA_WRITE can be found by
 | |
| using the IMM field.
 | |
| 
 | |
| 2. When confirming a write request server sends an "empty" rdma message with
 | |
| an immediate field. The 32 bit field is used to specify the outstanding
 | |
| inflight IO and for the error code.
 | |
| 
 | |
| CLT                                                          SRV
 | |
| usr_data + usr_hdr + rtrs_msg_rdma_write -----------------> [RTRS_IO_REQ_IMM]
 | |
| [RTRS_IO_RSP_IMM]                        <----------------- (id + errno)
 | |
| 
 | |
| * Write (always_invalidate=Y) *
 | |
| 
 | |
| 1. When processing a write request client selects one of the memory chunks
 | |
| on the server side and rdma writes there the user data, user header and the
 | |
| RTRS_MSG_RDMA_WRITE message. Apart from the type (write), the message only
 | |
| contains size of the user header. The client tells the server which chunk has
 | |
| been accessed and at what offset the RTRS_MSG_RDMA_WRITE can be found by
 | |
| using the IMM field, Server invalidate rkey associated to the memory chunks
 | |
| first, when it finishes, pass the IO to RNBD server module.
 | |
| 
 | |
| 2. When confirming a write request server sends an "empty" rdma message with
 | |
| an immediate field. The 32 bit field is used to specify the outstanding
 | |
| inflight IO and for the error code. The new rkey is sent back using
 | |
| SEND_WITH_IMM WR, client When it recived new rkey message, it validates
 | |
| the message and finished IO after update rkey for the rbuffer, then post
 | |
| back the recv buffer for later use.
 | |
| 
 | |
| CLT                                                          SRV
 | |
| usr_data + usr_hdr + rtrs_msg_rdma_write -----------------> [RTRS_IO_REQ_IMM]
 | |
| [RTRS_MSG_RKEY_RSP]                     <----------------- (RTRS_MSG_RKEY_RSP)
 | |
| [RTRS_IO_RSP_IMM]                        <----------------- (id + errno)
 | |
| 
 | |
| 
 | |
| * Read (always_invalidate=N)*
 | |
| 
 | |
| 1. When processing a read request client selects one of the memory chunks
 | |
| on the server side and rdma writes there the user header and the
 | |
| RTRS_MSG_RDMA_READ message. This message contains the type (read), size of
 | |
| the user header, flags (specifying if memory invalidation is necessary) and the
 | |
| list of addresses along with keys for the data to be read into.
 | |
| 
 | |
| 2. When confirming a read request server transfers the requested data first,
 | |
| attaches an invalidation message if requested and finally an "empty" rdma
 | |
| message with an immediate field. The 32 bit field is used to specify the
 | |
| outstanding inflight IO and the error code.
 | |
| 
 | |
| CLT                                           SRV
 | |
| usr_hdr + rtrs_msg_rdma_read --------------> [RTRS_IO_REQ_IMM]
 | |
| [RTRS_IO_RSP_IMM]            <-------------- usr_data + (id + errno)
 | |
| or in case client requested invalidation:
 | |
| [RTRS_IO_RSP_IMM_W_INV]      <-------------- usr_data + (INV) + (id + errno)
 | |
| 
 | |
| * Read (always_invalidate=Y)*
 | |
| 
 | |
| 1. When processing a read request client selects one of the memory chunks
 | |
| on the server side and rdma writes there the user header and the
 | |
| RTRS_MSG_RDMA_READ message. This message contains the type (read), size of
 | |
| the user header, flags (specifying if memory invalidation is necessary) and the
 | |
| list of addresses along with keys for the data to be read into.
 | |
| Server invalidate rkey associated to the memory chunks first, when it finishes,
 | |
| passes the IO to RNBD server module.
 | |
| 
 | |
| 2. When confirming a read request server transfers the requested data first,
 | |
| attaches an invalidation message if requested and finally an "empty" rdma
 | |
| message with an immediate field. The 32 bit field is used to specify the
 | |
| outstanding inflight IO and the error code. The new rkey is sent back using
 | |
| SEND_WITH_IMM WR, client When it recived new rkey message, it validates
 | |
| the message and finished IO after update rkey for the rbuffer, then post
 | |
| back the recv buffer for later use.
 | |
| 
 | |
| CLT                                           SRV
 | |
| usr_hdr + rtrs_msg_rdma_read --------------> [RTRS_IO_REQ_IMM]
 | |
| [RTRS_IO_RSP_IMM]            <-------------- usr_data + (id + errno)
 | |
| [RTRS_MSG_RKEY_RSP]	     <----------------- (RTRS_MSG_RKEY_RSP)
 | |
| or in case client requested invalidation:
 | |
| [RTRS_IO_RSP_IMM_W_INV]      <-------------- usr_data + (INV) + (id + errno)
 | |
| =========================================
 | |
| Contributors List(in alphabetical order)
 | |
| =========================================
 | |
| Danil Kipnis <danil.kipnis@profitbricks.com>
 | |
| Fabian Holler <mail@fholler.de>
 | |
| Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
 | |
| Jack Wang <jinpu.wang@profitbricks.com>
 | |
| Kleber Souza <kleber.souza@profitbricks.com>
 | |
| Lutz Pogrell <lutz.pogrell@cloud.ionos.com>
 | |
| Milind Dumbare <Milind.dumbare@gmail.com>
 | |
| Roman Penyaev <roman.penyaev@profitbricks.com>
 |