Re: A question on Zero Copy

[csapuntz@cisco.com]

> Does the iSCSI layer want:
> 
> A) Plain Zero copy, where the upper layer (iSCSI) asks
>    to read the next available "message" from the wire
>    into a buffer passed to the transport by iSCSI? 
> 
> <OR>
> 
> B) A directed Zero Copy, where the upper layer (iSCSI) asks
>    to read a particular request to a specific buffer?
> 

Randall,

We would ideally have directed Zero Copy.

It has to do with the block interface in most OSes. Clients of the
block interface (the file system, the virtual memory manager) supply a
buffer to the block subsystem. The result of any READ operation
MUST appear in the supplied buffer. There is no opportunity for
the block interface to do A. Thus, B is required.
 
Actually, I lied a bit. If the buffer is in virtual memory, you can do
a cute hack called page flipping. It's a hack because it's not general.
If the READ data landed in physical memory with the same alignment as
the buffer given to the block interface, then we can often just change
the VM mapping to point to the new buffer. 

It's a hack because it requires the storage buffers and network
buffers to be interchangeable. To be workable in many cases, it
requires either very large transfers or page aligned buffers. It does
not readily generalize to where the client of the block interface
wants to specify a scatter list for the I/O READ.

Anyway, the crux of the issue is the block interface. It may be better
to modify the block interface in many OSes to be more modern and
accept a data buffer back from the storage driver. Or the current
interface may be better for applications that wish to keep simple
data structures and logic by allowing them to control where the
results of I/O operations lands.

Of course, the aim of iSCSI is not revolutionize the storage stack but
slip in under SCSI and block interface in a transparent fashion. So,
directed Zero copy it is!

Cheers,
-Costa