In D12132#254430, @jmg wrote:

So, you'll still suffer terribly with this, as random_harvest_fast is not PCPU, so you'll have the cache line bouncing around.

IMO, as we feed in the mbuf (and not the mbuf cluster) we are not getting great entropy from the current ethernet harvester. One solution would be to turn off all ethernet harvesting, and enable INTR_ENTROPY on all the ethernet drivers. This will provide timing data for each interrupt, which is good for a couple bits of entropy. This will automatically scale down the rate as most high end ethernet cards do interrupt coalescing.

We don't need to collect THAT much entropy from the system. We should assume we have a good seed (and an uncompromised seed). If it does get compromised, it's more about a time to recovery. We do not need to be "collecting" thousands of bits of entropy every second, even just 10 bit per second not including RDRAND, (we poll RDRAND a minimum of 10 times per second, it increases 2x+𝛆 per bytes read from /dev/random, and it provides 64-bits EACH poll, or 640 bits per second), would give a recover time of under 30 seconds.

Looks good to me, adding a few other committers just in case.

Sorry for the delay. Looks good to me now, thanks!

LGTM, approved.

Need more work.

(fixed a few bugs, tested with 2+TB ARC by quakelee at gmail.com).

In D12076#251688, @feld wrote:

If you meant a FreeBSD bug report, no -- this was something I discovered and resolved, but did not open a bug report for the FreeBSD project.

Sorry didn't act on this earlier.

There are multiple blank space issues. For consistency, please add space after all punctuation, unless they are immediately followed by a closing tag.

There are multiple blank space issues. For consistency, please add space after all punctuation, unless they are immediately followed by a closing tag.

Looks good to me.

Please see comments inline.

Thanks for the fixes!

Looks good overall, please fix the few minor nits.

In open_binary_fd: when using buffer size for strl* and snprintf,
always use >= instead of > to avoid truncation.