dhcpcd-discuss

On 03/03/2019 13:42, Donald Sharp wrote:
> Hi Roy -
> To generate these routes I used FRRouting (https://frrouting.org).
> The code can be found here: https://github.com/frrouting/frr.git .
> Documentation to build on NetBSD is here:
> http://docs.frrouting.org/projects/dev-guide/en/latest/building-frr-for-netbsd7.html
> .  Essentially I used a new utility daemon to generate the routes via
> cli: http://docs.frrouting.org/en/latest/sharp.html .  This can be a
> bit complicated to setup, if you are interested in setting it up and
> run into issues feel free to reach out to me, I'll get you any help
> you may need here.

I'll look into that next week maybe.

> On to your questions about dhcpcd: I did see the rbtree()
> functionality when I went searching for a replacement, but it is not
> immediately available on linux..  I knew the rb_tree code would drop
> right in as that we use it in our project.  Personally I don't have a
> dog in this fight( hence the question in the original email ).  If you
> want to use something else please do so, no skin off my back.  You'll
> just need to point me at it if you want me to do the work :)
>
> Size of original data structures:
> (gdb) p sizeof(struct dhcpcd_ctx)
> $1 = 600
> (gdb) p sizeof(struct rt)
> $2 = 160
> (gdb)
>
> 1 million routes with original dhcpcd is ~19mb of main memory
>
> Size w/ rb_tree:
> (gdb) p sizeof(struct dhcpcd_ctx)
> $1 = 904
> (gdb) p sizeof(struct rt)
> $2 = 176(gdb)
>
> With 1 million routes with my code it's ~180mb of main memory.  This
> implies I've introduced a memory leak, Looks like I'll need to look at
> it this evening to see if I can figure out where I've gone wrong.  As
> that an extra 16 bytes of memory per `struct rt` shouldn't be causing
> this many problems.

I've implemented rbtree(3) into dhcpcd in a new branch here:
https://roy.marples.name/git/dhcpcd.git/commit/?h=rbtree&id=d1301b94bce1ab4f70ae4ef6222513bf93456b68
The commit message describes why it's not ready for production use.
It seems to work fine on NetBSD - you'll need the patch here if you try:
https://mail-index.netbsd.org/tech-userlevel/2019/03/03/msg011804.html

This saves ~5k on the binary size according to sizeof which is awesome!
Memory usage (with normal use) actually appears to be slightly lower by 
100 bytes or so - I suspect my patch fixed a slight bug with the free 
route list to save on mallocs.

I'll look into adding some portability gunk so this can compile on Linux 
at least, again, sometime next week. Probably before I look at the 
frrouting.

Roy