Re: [dpdk-dev] [RFC] eal: make rte_rand() MT safe

[Bruce Richardson <bruce.richardson@intel.com>]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain; charset="UTF-8", Size: 1388 bytes --]

On Fri, Apr 05, 2019 at 03:51:39PM +0200, Mattias Rönnblom wrote:
> On 2019-04-05 15:45, Mattias Rönnblom wrote:
> > The rte_rand() documentation left it unspecified if the rte_rand() was
> > multi-thread safe or not, and the implementation (based on lrand48())
> > was not.
> > 
> > This commit makes rte_rand() safe to use from any lcore thread by
> > using lrand48_r() and per-lcore random state structs. Besides the
> > obvious improvement in terms of correctness (for concurrent users),
> > this also much improves rte_rand() performance, since the threads no
> > longer shares state. For the single-threaded case, this patch causes
> > ~10% rte_rand() performance degradation.
> > 
> 
> It's a little unclear to me, if lrand48_r() exists in FreeBSD or not. Could
> someone confirm?
>
Nothing shows up for me in the man pages for such a function on FreeBSD 12,
so I suspect they aren't available.
 
> Another question I have is in which section of the version.map file the new
> symbols should go. Experimental, or 19.05?
> 
I think it should be 19.05. Since the APIs have been around as inline
functions for some time now, I don't see the point of having them be
experimental for a time.

> The source interface is backward compatible, but the functions are no longer
> inline functions in the header file, and thus needs to go somewhere to be
> properly exported.

Regards,
/Bruce