tech-kern archive

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]

Re: patch: multiple-request, async-request support for /dev/crypto

I have to say it looks great - excellent for high-performance
application designs with single-threaded event-driven models.
Multi-threaded apps would benefit too.

I may be a tiny bit biased having written part of it ;-).

Darran Hunt.

On 27/03/2008, at 7:13 PM, Thor Lancelot Simon wrote:

The attached patch adds ioctls for multiple request submission and
retrieval in a single system call, and for asynchronous operation via
select()/poll() to /dev/crypto.  It was written for a slightly older
NetBSD-current and I've hastily adapted it to Andrew's recent file
descriptor allocation changes, which I hope I got right -- it's well
tested before that adaptation but untested with it (caveat emptor).

With this patch and a slightly clever multithreaded application that
batches requests when pushing them to /dev/crypto we can do 84,000
trivial asymmetric operations/sec (32 bit modular math ops) to a
rather old crypto accellerator card, with a Core 2 Duo as the host.
I think it's worth the added complexity.  Even simpler applications
can benefit -- a lot -- from async operation and retrieving multiple
requests at once when poll() fires.

There is some duplicated code here in the multiple-request ioctls that
could be shared with the single-request ioctls. It'd be a moderate pain
to clean up and I'd prefer to do that after commit so I can get this
in the tree while I have time to focus on it.  Also, there is a large
comment here describing the new ioctls and parameters which text I'll
reproduce in or move to the manual page.

I will revise the openssl engine to work as efficiently as possible with
the new ioctls added here once I sort out some issues about updating
OpenSSL itself in our tree and feeding changes back to the OpenSSL project
in an effective way.


Thor Lancelot Simon
Coyote Point Systems, Inc.                      <>
Millerton, NY, USA<oc-multi.diff>

Home | Main Index | Thread Index | Old Index