freebsd-skq/sys/netgraph/NOTES

$FreeBSD$
Development ideas..

Archie's suggestions... :-)

 - There should be a new malloc type: M_NETGRAPH
	[DONE]
	- all mallocs/frees now changed to use this.. JRE
	- might further split them out some time.

 - Use MALLOC and FREE macros instead of direct function calls
	[DONE]
	- They allow conditional compilation which keeps
	  statistics & counters on various memory allocation
	  (or so it seems)
	   - Now tend to have NG_FREE_XX() macros. they
	     allow better debugging

 - In struct ng_mesg: at least make "header" into "hdr", if not
   getting rid of it altogether. It doesn't seem necessary and
   makes all my C code lines too long.

	- I understand.. one thought however.. consider..
	  if char data[0] were not legal,  so that data[1] needed to be
	  used instead, then the only way to get the size of the header
	  would be sizeof(msg.header) as sizeof(msg) would include the dummy
	  following bytes. this is a portability issue and I hope
	  it will be ported eventually :)

	- Baloney! you can use sizeof(msg) - 1 then.. or just
	  make it a macro, then its always portable:

	  #ifdef __GNU_C__
	    #define NG_MSG_HDR_SIZE	(sizeof(struct ng_message))
	  #else
	    #define NG_MSG_HDR_SIZE	(sizeof(struct ng_message) - 1)
	  #endif

	  - inertia rules :-b


 - Have a user level program to print out and manipulate nodes, etc.
	- [DONE]
		see ngctl, nghook

 - "Netgraph global" flags to turn on tracing, etc.

 - ngctl needs to be rewritten using libnetgraph. Also it needs a
   command to list all existing nodes (in case you don't know the
   name of what you're looking for).
	[DONE]

 - Need a way to get a list of ALL nodes.
	[DONE]
	- see NGM_LISTNODES

 - Enhance "netstat" to display all netgraph nodes -- or at least
   all netgraph socket nodes.
	[DONE]

 - BUG FIX: bind() on a socket should neither require nor allow a
   colon character at the end of the name. Note ngctl allows you
   to do it either way!
	[DONE] (I think)
	- bind on a control socket has been disabled
	  it was a bad idea.

 - Need to implement passing meta information through socket nodes
   using sendmsg() and recvmsg().

 - Stuff needing to be added to manual:

   - Awareness of SPL level, use ng_queue*() functions when necessary.
   - Malloc all memory with type M_NETGRAPH. -DONE
   - Write code so it can be an LKM or built into the kernel.. this means
     be careful with things like #ifdef INET.
   - All nodes assume that all data mbufs have the M_PKTHDR flag set!
     The ng_send_data() and related functions should have an
     #ifdef DIAGNOSTIC check to check this assumption for every mbuf.
     -DONE with INVARIANTS. Framework should test this more.
   - More generally, netgraph code should make liberal use of the
     #ifdef DIAGNOSTIC definition.
     -INVARIANTS.
   - Since data and messages are sent functionally, programmers need
     to watch out for infinite feedback loops. Should ng_base.c detect
     this automatically?
      - I've been thinking about this. each node could have a 'colour'
	which is set to the colour of the packet as you pass through.
	hitting a node already of your colour would abort. Each packet
	has another (incremented) colour.
	-new 'item' type can hold a hopcount...

NEW in 2001
All piggyback responses have gone away.
use the node ID in the return address field for quick response delivery.

Every node has a queue, plus there is a list of nodes that have queued work.
Extensive use of Mutexes. Getting in shape for SMP.

Messages and data are deliverd in a new form. An Item now has
all information needed to queue such a request and deliver it later, so 
it is now the basis of all data transfer since any transfer may need to
be queued.