Forwarded by Akihiro Shimura <shimura@pure.cpdc.canon.co.jp>
---------------- Original message follows ----------------
From: Akihiro Shimura <shimura@pure.cpdc.canon.co.jp>
To: Nagasaka Fumio <Nagasaka.Fumio@exc.epson.co.jp>
Cc: p1394@pwg.org
Date: Mon, 9 Mar 1998 18:41:46 +0900 (JST)
Subject: Re: P1394> unordered execution
--On Mon, 9 Mar 1998 10:14:15 +0900 Nagasaka Fumio <Nagasaka.Fumio@exc.epson.co.jp> wrote:
> I am curious why unordered execution is required for 'HPT'. > May be there are something I am missing.
The HPT(High Performance Transport) essentially requires unordered execution model in order to multiplex two independent sequences of operation requests (read request sequence and write request sequence) into single linked list of ORB's. In other words, the HPT provides full duplex communication path (independent flow for each direction) by using unordered execution model with one fetch agent (one login).
The unordered model under the basic task management model allows the target to unrestrictedly reorder the active tasks, and the responsibility for the assurance of data integrity is left with the initiator.
The SBP-2 describes following sentences as a note.
"NOTE - In multitasking operating system environments, independent execution threads may generate tasks that have ordering constraints within each thread but not with respect to other threads. If this is the case, an initiator may manage the constraints of each thread yet still keep the target substantially busy. This avoids the undesirable latencies that occur if the target is allowed to become idle before new ORB's are signaled."
The HPT uses this model. Essentially, the independent execution threads are read thread and write thread in the HPT. Each thread generates tasks that have ordering constraints within each thread but not with respect to another thread. So, the HPT provides in-order delivery in each direction while keeping each direction independent.
A certain operating system uses two I/O request queues for each direction to serve as full duplex. The requests in these I/O request queues are multiplexed into single linked list of ORB by message(ORB or status_block, but NOT data itself) flow control in the HPT. The ordering constraints are maintained in these two I/O request queues as usual. The status_block returned by the target indicates to which of the head of these I/O request queues the completed I/O request belongs. Since the message flow control does not require additional packet exchange with another peer in the HPT, the flow control does not have an overhead from the peer's latency. (The HPT also uses the same mechanism to multiplex multiple channels.)
There will be three places to put multiplexing information for independent data paths within on using SBP-2.
a) Data buffer b) ORB c) 1394 bus address
The protocol uses "a)" will be 1284.4 over SBP-2. (Single fetch agent, single execution agent, data flow control) The protocol uses "b)" will be HPT. (Single fetch agent, independent execution agents, message flow control) The protocol uses "c)" will be dual login or cross login. (Independent fetch agents, independent execution agents)
Since plain SBP-2 provides master controlled half duplex path, the profile (rev. 0.1d) may be between half duplex "plain SBP-2" and full duplex "SBP-2 dual or cross login" because it may require some sort of retries(re-ordering in the initiator) in case of asynchronous back channel data though it uses ordered model. (Three quarters duplex?)
-- Akihiro Shimura (shimura@pure.cpdc.canon.co.jp) Office Imaging Products Development Center 3 CANON INC.