[ofw] RE: Release IBA resources in DISPATCH_LEVEL ?

[Fab Tillier]

Hi Leo,

> -----Original Message-----
> From: Leonid Keller [mailto:leonid at mellanox.co.il]
> Sent: Sunday, June 15, 2008 2:08 AM
>
>> The only reason waiting is needed is because the driver
>> implementation forces it.
> Generally, you are right, but it's a bit more
> complicated then that. There are dependencies between the commands,
> which can cause problems in fully asynchronous work with the card.

I disagree: the blocking semantics are independent of hardware limitations.

Dependencies between commands just affects whether you can issue the command right away, of if you have to wait for a previous command to complete.  You need that anyway as the hardware has a limit to the number of commands that it can handle concurrently, even with the synchronous mechanisms in use today.  With the synchronous model of the driver, you may block before the command is issued, and you will block while the command is being processed by the hardware.  The logic must already be there to resume command processing when a command completes, regardless of whether the command is blocking or asynchronous.

With an asynchronous model, the caller's request would always return STATUS_PENDING.  It would do this when either
  a). a command is in progress that is a prerequisiste for this command;
  b). there are too many commands in flight to issue this command;
  c). the command is issued to the hardware;

For cases a) and b), the command would be issued once the prerequisite command or first previous command completes, respectively.

In all cases, the client would be notified of the request completion when the command is complete.

I would actually suspect handling commands asynchronously would simplify the HCA driver.  All the logic for blocking and waking up the caller would be eliminated.  The HCA driver wouldn't have to care whether callers are at DISPATCH or lower, the processing would be the same always.

-Fab