+void icssm_prueth_sw_fdb_tbl_init(struct prueth *prueth)
+{
+	struct fdb_tbl *t = prueth->fdb_tbl;
+
+	t->index_a = (struct fdb_index_array_t *)((__force const void *)
+			prueth->mem[V2_1_FDB_TBL_LOC].va +
+			V2_1_FDB_TBL_OFFSET);

We have

+#define V2_1_FDB_TBL_LOC          PRUETH_MEM_SHARED_RAM

and existing code like:

void __iomem *sram_base = prueth->mem[PRUETH_MEM_SHARED_RAM].va;

so it seems like

t->index_a = sram_base + V2_1_FDB_TBL_OFFSET;

with no needs for any casts, since sram_base is a void * so can be
assigned to any pointer type.

And there are lots of cascading defines like:

/* 4 queue descriptors for port 0 (host receive). 32 bytes */
#define HOST_QUEUE_DESC_OFFSET          (HOST_QUEUE_SIZE_ADDR + 16)

/* table offset for queue size:
 * 3 ports * 4 Queues * 1 byte offset = 12 bytes
 */
#define HOST_QUEUE_SIZE_ADDR            (HOST_QUEUE_OFFSET_ADDR + 8)
/* table offset for queue:
 * 4 Queues * 2 byte offset = 8 bytes
 */
#define HOST_QUEUE_OFFSET_ADDR          (HOST_QUEUE_DESCRIPTOR_OFFSET_ADDR + 8)
/* table offset for Host queue descriptors:
 * 1 ports * 4 Queues * 2 byte offset = 8 bytes
 */
#define HOST_QUEUE_DESCRIPTOR_OFFSET_ADDR       (HOST_Q4_RX_CONTEXT_OFFSET + 8)

allowing code like:

	sram = sram_base + HOST_QUEUE_SIZE_ADDR;
	sram = sram_base + HOST_Q1_RX_CONTEXT_OFFSET;
	sram = sram_base + HOST_QUEUE_OFFSET_ADDR;
	sram = sram_base + HOST_QUEUE_DESCRIPTOR_OFFSET_ADDR;
	sram = sram_base + HOST_QUEUE_DESC_OFFSET;

+	t->mac_tbl_a = (struct fdb_mac_tbl_array_t *)((__force const void *)
+			t->index_a + FDB_INDEX_TBL_MAX_ENTRIES *
+			sizeof(struct fdb_index_tbl_entry_t));

So i think this could follow the same pattern, also allowing some of
these casts to be removed.

I just don't like casts, they suggest bad design.

+static u8 icssm_pru_lock_done(struct fdb_tbl *fdb_tbl)
+{
+	return readb((u8 __iomem *)&fdb_tbl->locks->pru_locks);

And maybe the __iomem attribute can be added to struct, either per
member, or at the top level? It is all iomem, so we want sparse to be
able to check all accesses.

+static int icssm_prueth_sw_fdb_spin_lock(struct fdb_tbl *fdb_tbl)
+{
+	u8 done;
+	int ret;
+
+	/* Take the host lock */
+	writeb(1, (u8 __iomem *)&fdb_tbl->locks->host_lock);
+
+	/* Wait for the PRUs to release their locks */
+	ret = read_poll_timeout(icssm_pru_lock_done, done, done == 0,
+				1, 10, false, fdb_tbl);
+	if (ret)
+		return -ETIMEDOUT;
+
+	return 0;

Documentation says:

 * Returns: 0 on success and -ETIMEDOUT upon a timeout.

So no need for the if statement.

+static s16
+icssm_prueth_sw_fdb_search(struct fdb_mac_tbl_array_t *mac_tbl,
+			   struct fdb_index_tbl_entry_t *bucket_info,
+			   const u8 *mac)
+{
+	u8 mac_tbl_idx = bucket_info->bucket_idx;
+	int i;
+
+	for (i = 0; i < bucket_info->bucket_entries; i++, mac_tbl_idx++) {
+		if (ether_addr_equal(mac,
+				     mac_tbl->mac_tbl_entry[mac_tbl_idx].mac))
+			return mac_tbl_idx;
+	}
+
+	return -ENODATA;

It is traditional to return errno in an int. But i don't see why a s16
cannot be used.

+icssm_prueth_sw_fdb_find_bucket_insert_point(struct fdb_tbl *fdb,
+					     struct fdb_index_tbl_entry_t
+					     *bkt_info,
+					     const u8 *mac, const u8 port)
+{
+	struct fdb_mac_tbl_array_t *mac_tbl = fdb->mac_tbl_a;
+	struct fdb_mac_tbl_entry_t *e;
+	u8 mac_tbl_idx;
+	int i, ret;
+	s8 cmp;
+
+	mac_tbl_idx = bkt_info->bucket_idx;
+
+	for (i = 0; i < bkt_info->bucket_entries; i++, mac_tbl_idx++) {
+		e = &mac_tbl->mac_tbl_entry[mac_tbl_idx];
+		cmp = memcmp(mac, e->mac, ETH_ALEN);
+		if (cmp < 0) {
+			return mac_tbl_idx;
+		} else if (cmp == 0) {
+			if (e->port != port) {
+				/* MAC is already in FDB, only port is
+				 * different. So just update the port.
+				 * Note: total_entries and bucket_entries
+				 * remain the same.
+				 */
+				ret = icssm_prueth_sw_fdb_spin_lock(fdb);
+				if (ret) {
+					pr_err("PRU lock timeout\n");
+					return -ETIMEDOUT;
+				}

icssm_prueth_sw_fdb_spin_lock() returns an errno. Don't replace it.

Also, pr_err() is bad practice and probably checkpatch is telling you
this. Ideally you want to indicate which device has an error, so you
should be using dev_err(), or maybe netdev_err().

+	if (left > 0) {
+		hash_prev =
+			icssm_prueth_sw_fdb_hash
+			(FDB_MAC_TBL_ENTRY(left - 1)->mac);
+	}

+		empty_slot_idx =
+			icssm_prueth_sw_fdb_check_empty_slot_right(mt, mti);

There are a couple of odd indentations like this. I wounder if it
makes sense to shorten the prefix? Do you really need all of
icssm_prueth_sw_fdb_ ?

	Andrew