Fix for SGWC and SMF round robin selection

[kbarlee]

Previously, after searching for a SGWU/UPF designed to serve a particular TAC/APN/cell_ID and not finding a node that was suitable, SGWC and SMF would run this line to select a node in round robin fashion:

return next ? next : ogs_list_first(&ogs_pfcp_self()->peer_list);

However this did not check that node next was PFCP associated.
#555

The code now runs as follows:

• search for matching SGWU/UPF nodes as before
• if none are found, enter RR mode
• in RR mode, find next PFCP associated node in list and use it
• if there are no PFCP associated nodes, print error message (and select first from list)

[acetcom]

Hi @kbarlee

When I first designed it, I didn't make a distinction between RR mode and the rest.

For example, let's say you have 10 UPFs. At this time, I wanted to be able to select 3 UPFs as Round-Robin manner when the APN of 3 UPFs are internet.

Stabilization is in progress now. After that is done, let me look at this part again. Of course, you can also request a Pull Request again. If it fits my original design, I will merge it.

Thank you so much!
Sukchan

[kbarlee]

Hey @acetcom

You don't see the issue for the scenario you described. The code is executed along the following lines:

1. Search through node list to see if any are dedicated to support the TAC
   -- round robin between nodes that support the TAC && are PFCP associated
   -- if there are none, move to 2

2. Search through node list to see if any are dedicated to support the e_cell_ID
   -- round robin between nodes that support the e_cell_ID && are PFCP associated
   -- if there are none, move to 3

3. Search through node list to see if any are dedicated to support the nr_cell_ID
   -- round robin between nodes that support the nr_cell_ID && are PFCP associated
   -- if there are none, move to 4

4. Search through node list to see if any are dedicated to support the APN
   -- round robin between nodes that support the APN && are PFCP associated
   -- if there are none, move to 5

5. Round robin full list. As described above, this line does not check that the node selected in RR is PFCP associated

return next ? next : ogs_list_first(&ogs_pfcp_self()->peer_list);

So for the situation you describe where there are 3 UPFs that support APN 'internet', it will happily round robin between them. If one of the 3 is offline, it will round robin between the other two. This fix is for the situations where

• no TAC/e_cell_ID/nr_cell_ID/APN are specified in the list, and it defaults to full list RR
• none of the nodes support the TAC/e_cell_ID/nr_cell_ID/APN that the UE is connecting with, and it defaults to full list RR
• or no nodes that do support the TAC/e_cell_ID/nr_cell_ID/APN are currently PFCP associated, and it defaults to full list RR

(I had already implemented a fix for this here, but it didnt make it into the current master)

open5gs/src/smf/context.c

Lines 687 to 743 in bcd02b1

	if (ogs_pfcp_self()->upf_selection_mode == UPF_SELECT_RR) {
	/* Select UPF (PFCP) with round-robin manner */
	ogs_debug("Select UPF by RR");
	/*
	- starting from list position of last used UPF, search down list for next PFCP associated UPF
	- if PFCP associated UPF found
	- use it
	- if no associated UPF found, keep searching
	- if bottom of list reached, reset search to top of list
	- if completed full cyclic search of list and still not found a UPF, use default (first)
	*/
	int connect=0, numreset=0;
	char startUPFIP[OGS_ADDRSTRLEN];
	OGS_ADDR(&ogs_pfcp_self()->node->addr, startUPFIP);
	//search UPF list, find next UPF that is associated
	while (!connect)
	{
	// (when end of UPF (PFCP) list reached, reset search to top of UPF list)
	if(ogs_pfcp_self()->node == NULL){
	ogs_pfcp_self()->node = ogs_list_first(&ogs_pfcp_self()->n4_list);
	numreset = 1;
	}
	// search UPF (PFCP) list, find next UPF that is associated
	while(ogs_pfcp_self()->node && !connect) {
	// if cyclic list check complete and still not found a UPF, break
	OGS_ADDR(&ogs_pfcp_self()->node->addr, buf);
	if (numreset == 1 && !strcmp(buf,startUPFIP)) {
	break;
	}
	// has UPF <x> associated over PFCP?
	if (OGS_FSM_CHECK( &ogs_pfcp_self()->node->sm, smf_pfcp_state_associated) ){
	// then use it
	connect = 1;
	} else {
	// else check next UPF in list
	OGS_ADDR(&ogs_pfcp_self()->node->addr, buf);
	ogs_debug("UPF on IP[%s] is not PFCP associated",
	buf);
	ogs_pfcp_self()->node = ogs_list_next(ogs_pfcp_self()->node);
	}
	}
	// after checking from the top of the list and not finding a suitable UPF
	if (!connect && numreset == 1 ){
	// default to first in list
	ogs_error("No UPF (PFCP) is currently PFCP associated");
	ogs_error("Defaulting to first UPF (PFCP) in smf.yaml list");
	ogs_pfcp_self()->node = ogs_list_first(&ogs_pfcp_self()->n4_list);
	break;
	}
	}
	// list UPF used
	OGS_ADDR(&ogs_pfcp_self()->node->addr, buf);
	ogs_debug("UE using UPF on IP[%s]",
	buf);

Hope that makes sense!

Kenny

[acetcom]

@kbarlee I misunderstood your original code. Now I have merged your code. And also, The coding indentation was slightly modified. 228dd34

That's really nice work!

Thank you so much for your effort!
Sukchan