netstat: convert routing statistics to a sysctl

[kevans]

Summary

Exporting the relevant pcpustat is trivial, so let's do that. Use it
for netstat operation to avoid having to dig around in mem(4) for live
kernel statistics. Based on callers of kresolve_list(), this is the
last live path using mem(4) that could be functional today.

Tested both with netstat -rs and netstat -rs -M.

[zlei]

sys/net/route.c

77

This line seems wrong to me. The following line 80~81 looks good.

[kevans]

sys/net/route.c

77

Whoops, forgot to clear out a stale version of the patch on the machine I send them from. Will update in a few hours- this line disappears, the two below remain.

[kevans]

Fix the diff; locally, I split this up into

Author: Kyle Evans <[e-mail]>
Date:   Mon Oct 21 22:44:59 2024 -0500

    net: route: convert routing statistics to a sysctl
    
    Exporting the relevant pcpustat is trivial, so let's do that.  We will
    use it in a near-future change in netstat to avoid having to dig around
    in mem(4) for live kernel statistics.
    
    Differential Revision:  https://reviews.freebsd.org/D47231

Author: Kyle Evans <[e-mail]>
Date:   Tue Oct 22 08:33:35 2024 -0500

    netstat: switch to using the sysctl-exported stats for live stats
    
    Now that we export the relevant stats via the net.route.stats sysctl,
    switch to using that to avoid having to dig around in mem(4) for live
    kernel statistics.  Based on callers of kresolve_list(), this is the
    last live path using mem(4) that could be functional today.
    
    Tested both with `netstat -rs` and `netstat -rs -M`.
    
    Differential Revision:  https://reviews.freebsd.org/D47231

I'll commit the first one then wait a week or so to commit the second.

[zlei]

Looks good to me.

In D47231#1077100, @kevans wrote:

...

I'll commit the first one then wait a week or so to commit the second.

Why wait a week for the second commit ? Have it in main so it could be get tested as early as possible.

sys/net/route.c

80

A nit here. I'd prefer no period for the description of sysctl vars.

[kevans]

In D47231#1077112, @zlei wrote:

Looks good to me.

In D47231#1077100, @kevans wrote:

...

I'll commit the first one then wait a week or so to commit the second.

Why wait a week for the second commit ? Have it in main so it could be get tested as early as possible.

Just to be nicer to folks wanting to bisect and use netstat -rs for a little bit; you can still -M /dev/mem until we strip setgid in another change.

Also, I can probably get away with landing a small/quick sysctl during stabweek but probably better to be conservatie about its use.

[markj]

I wonder if we want to add some spare fields to struct rtstat before making it part of the user ABI?

sys/net/route.c

80

+1

[kevans]

In D47231#1077231, @markj wrote:

I wonder if we want to add some spare fields to struct rtstat before making it part of the user ABI?

Sure; maybe drop a uint64_t[3] in at the end to round up to a cool 14? Additions are fairly far and few between.

[markj]

In D47231#1077402, @kevans wrote:

In D47231#1077231, @markj wrote:

I wonder if we want to add some spare fields to struct rtstat before making it part of the user ABI?

Sure; maybe drop a uint64_t[3] in at the end to round up to a cool 14? Additions are fairly far and few between.

I'd make it an even 16 but it's up to you. :)