Re: Hifn crypto driver: does it work for anyone?
Thor Lancelot Simon <[email protected]>
| Newsgroups | gmane.os.netbsd.current,gmane.os.netbsd.devel.kernel,gmane.os.netbsd.devel.network,gmane.os.netbsd.devel.crypto |
|---|---|
| Message-ID | <[email protected]> |
[ I've added tech-crypto, which should have been there in the first place. ] On Sun, Oct 16, 2005 at 03:51:42PM -0400, Gilles Roy wrote: > > I've used the NetBSD hifn driver with Soekris VPN1401 (actually a hifn > 7955 chip inside) and the driver worked fine under heavy loads for short > periods of time (benchmarking). > > Of course, only encryption works. The driver has an off by twelve bug > when it reads the result of any hash operation (it adds 12 bytes to a > pointer before the callback, and adds twelve again inside of the > callback). I sent a patch that fixes this to the list a few months ago. I missed the patch the first time. Applying it causes my FAST_IPSEC kernel in the machine with the 7955 to correctly handle AES/MD5 ESP packets... for about a minute after boot (a few tens of kilobytes of traffic). So it looks to be correct, and it's an improvement; I have committed it. However, that same machine still displays the symptom where the whole crypto subsystem grinds to a halt after it's run for a minute or so (and a few tens of kilobytes of traffic via ipsec, plus a few megabytes via OpenSSH's use of /dev/crypto have flowed through). I'm rebuilding it with options KTRACE so at least I can see what error code, exactly, the /dev/crypto operations are returning. When this happens, IPsec traffic stops too. Sigh. Thanks for the patch! Thor