Re: [PATCH 1/4] interconnect: icc-clk: support N:1 master-to-slave topology

[Konrad Dybcio <konrad.dybcio@oss.qualcomm.com>]

On 10/7/26 8:50 AM, Kathiravan Thirumoorthy wrote:
> icc_clk_register() assumes each icc_clk_data entry describes a unique
> master/slave pair and creates both nodes for every entry. When several
> masters share the same slave_id, the shared slave node is created and
> added again on each later entry. icc_node_create() returns the existing
> node, so icc_node_add() hits WARN_ON(node->provider):

[...]

> -	return clk_set_rate(qn->clk, icc_units_to_bps(src->peak_bw));
> +	/*
> +	 * Multiple master nodes can share the same underlying clock (N:1
> +	 * topology, e.g. several CPU clusters voting on one L3 clock).
> +	 * Aggregate the peak bandwidth across all of them before setting
> +	 * the rate, otherwise the last caller wins and can undervote what
> +	 * another master already asked for.
> +	 */
> +	list_for_each_entry(node, &provider->nodes, node_list) {
> +		struct icc_clk_node *n = node->data;
> +
> +		if (n && clk_is_match(n->clk, qn->clk))
> +			rate = max(rate, icc_units_to_bps(node->peak_bw));
> +	}

GPT has the following concern:

[P2] Recompute the shared clock rate when a master drops its vote  
drivers/interconnect/icc-clk.c:36-42

The zero-bandwidth path returns before the new shared-clock aggregation.
If one master had set the shared clock to a high rate and then drops its
vote while another master still votes, the clock remains at the old,
higher rate: the remaining master’s icc_clk_set() is not necessarily
called. Recompute the maximum vote from all nodes sharing the clock on
this path, and apply that rate when other masters still have votes.

Konrad