On 2026-06-05 at 12:02:42, Ratheesh Kannoth (rkannoth@marvell.com) wrote:

Add NIX_LF_DONT_FREE_DFT_IDXS so the PF can send NIX LF free during hw
reinit or teardown without the AF freeing CN20K default NPC rule indexes
while the driver still owns that state (otx2_init_hw_resources and
otx2_free_hw_resources).

On CN20K, allocate default NPC rules from NIX LF alloc before
nix_interface_init, roll back with npc_cn20k_dft_rules_free on failure,
and free from NIX LF free when the new flag is not set. Tighten
rvu_mbox_handler_nix_lf_alloc error handling: use a single rc, propagate
qmem_alloc and other errors, and set -ENOMEM only when kcalloc fails
(remove the blanket -ENOMEM at the free_mem path).

Signed-off-by: Ratheesh Kannoth <redacted>

https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260605063245.3553861-1-rkannoth%40marvell.com says

quoted

 		intf = NIX_INTF_TYPE_SDP;

-	err = nix_interface_init(rvu, pcifunc, intf, nixlf, rsp,
-				 !!(req->flags & NIX_LF_LBK_BLK_SEL));
-	if (err)
-		goto free_mem;
+	if (is_cn20k(rvu->pdev)) {
+		rc = npc_cn20k_dft_rules_idx_get(rvu, pcifunc, &bcast, &mcast,
+						 &promisc, &ucast);
+		if (rc) {
+			rc = npc_cn20k_dft_rules_alloc(rvu, pcifunc);
+			if (rc)
+				goto free_mem;
+
+			rules_created = true;
+		}
+	}

Does the rules_created flag faithfully reflect whether allocation
actually happened?
npc_cn20k_dft_rules_alloc() returns 0 without doing any work when the
pcifunc is not CGX or LBK mapped (for example SDP):
if (!npc_is_cgx_or_lbk(rvu, pcifunc)) {
	dev_dbg(rvu->dev,
		"%s: dft rule allocation is only for cgx mapped device, pcifunc=%#x\n",
		__func__, pcifunc);
	return 0;
}
For such a pcifunc, npc_cn20k_dft_rules_idx_get() returns -ESRCH, the
if branch is taken, dft_rules_alloc() is a no-op returning 0, and
rules_created becomes true even though no rule was created.
The free_dft path then calls npc_cn20k_dft_rules_free(), which itself
early-returns on the same !npc_is_cgx_or_lbk predicate, so the safety
relies on that hidden coupling rather than an explicit signal.
Would it be clearer to have npc_cn20k_dft_rules_alloc() return a
distinct created/not-created indication, or to gate rules_created on
npc_is_cgx_or_lbk(rvu, pcifunc) at the call site?

quoted

+

dft rules are only for cgx/lbk interfaces. So for sdp, eventhough npc_cn20k_dft_rules_alloc() returns 0
and rules_created is set to true, failure case, it will call dft_rules_free(), but there is nothing
to free. So this will work.

quoted

+++ b/drivers/net/ethernet/marvell/octeontx2/af/rvu_npc.c

@@ -4085,12 +4095,10 @@ void rvu_npc_clear_ucast_entry(struct rvu *rvu, int pcifunc, int nixlf)

 	ucast_idx = npc_get_nixlf_mcam_index(mcam, pcifunc,
 					     nixlf, NIXLF_UCAST_ENTRY);
-	if (ucast_idx < 0) {
-		dev_err(rvu->dev,
-			"%s: Error to get ucast entry for pcifunc=%#x\n",
-			__func__, pcifunc);
+
+	/* In cn20k, default rules are freed before detach rsrc */
+	if (ucast_idx < 0)
 		return;
-	}

The comment says this silent return is for the CN20K case where default
rules are freed before detach rsrc, but the change drops the dev_err
unconditionally for every silicon that reaches rvu_npc_clear_ucast_entry()
via the generic rvu_detach_block() path, including CN10K and OcteonTX2.
On legacy silicon a negative ucast_idx was previously treated as
exceptional and worth logging; should this silent return be gated on
is_cn20k(rvu->pdev) so the dev_err is preserved for the other variants?

No. In cn10k, ucast entry is not dynamic, so it is treated as error. This is
a teardown path, so want to keep same code for cn10k and cn20k.