Hifn driver in SMP (was Re: GELI - disk encryption GEOM class committed.)

Sam Leffler sam at errno.com
Wed Aug 10 05:56:20 GMT 2005


Mike Tancsa wrote:
> At 12:20 AM 09/08/2005, Sam Leffler wrote:
> 
>> I much prefer cryptotest for exercising the hardware.  If you increase 
>> the number of concurrent threads (-t I think) you can really load the 
>> device.
>>
>> I wouldn't be surprised if there were an smp locking bug in the crypto 
>> code
> 
> 
> 
> Found the program.  Under RELENG_4, its quite happy doing
> 
> cryptotest -t 4 -p -a aes 100000
> 92.781 sec,  200000    aes crypts,      16 bytes,    34490 byte/sec,     
> 0.3 Mb/sec
> dispatch->invoke: avg   3599 ns : min   1955 ns : max   55315 ns [800048 
> samps]
>     invoke->done: avg  56574 ns : min  32406 ns : max  158121 ns [800048 
> samps]
>         done->cb: avg   3707 ns : min   1677 ns : max   57270 ns [800048 
> samps]
>        cb->finis: avg   5286 ns : min   2235 ns : max   63695 ns [800048 
> samps]
> 
> ... continuously.  But I am pretty sure this is an IRQ routing issue and 
> not a HiFn specific thing.

FWIW -p significantly slows down performance.  Try something like -t 64 
or -t 128.

	Sam


More information about the freebsd-current mailing list