[PATCH 2/2] interconnect: qcom: add MSM8x60 NoC driver

[Herman van Hazendonk <github.com@herrie.org>]

Add a Qualcomm interconnect driver for the MSM8x60 family modelling the
four NoC fabrics (APPSS, System, MMSS, DFAB) that connect masters and
slaves on these Scorpion-class SoCs. The driver implements the
interconnect-provider API to manage bandwidth between specific masters
and slaves via the RPM arbitration tables.

Each fabric carries:
  - A bus clock (managed via clk_bulk APIs) whose rate is the
    aggregated bandwidth divided by the fabric bus width, with a
    minimum floor of 384 MHz to prevent USB starvation observed when
    the fabric drops to the previously-used 266 MHz minimum.
  - An RPM arbitration buffer (arb / bwsum) that the RPM firmware
    consumes via its shared-memory protocol; commits go via the
    qcom-rpm driver's set_resource API.

msm8660_get_rpm() pins the supplier with device_link_add() before
reading drvdata so an unbind/rebind window cannot leave a stale
qcom_rpm pointer. clk_bulk_prepare_enable is paired with a
devm_add_action_or_reset cleanup so an EPROBE_DEFER from the RPM
lookup does not leak the prepare/enable refcount across retries. The
fabric rate cap uses min_t(u64,...) so a bandwidth request exceeding
4 GiB/s cannot wrap through u32 truncation into a near-zero clock
rate that would halt the interconnect.

Signed-off-by: Herman van Hazendonk <github.com@herrie.org>
---
 drivers/interconnect/qcom/Kconfig             |   14 +
 drivers/interconnect/qcom/Makefile            |    2 +
 drivers/interconnect/qcom/msm8660.c           | 1147 +++++++++++++++++
 .../dt-bindings/interconnect/qcom,msm8660.h   |    6 +-
 4 files changed, 1166 insertions(+), 3 deletions(-)
 create mode 100644 drivers/interconnect/qcom/msm8660.c

diff --git a/drivers/interconnect/qcom/Kconfig b/drivers/interconnect/qcom/Kconfig
index 786b4eda44b4..d0a8df7cc700 100644
--- a/drivers/interconnect/qcom/Kconfig
+++ b/drivers/interconnect/qcom/Kconfig
@@ -80,6 +80,20 @@ config INTERCONNECT_QCOM_MSM8953
 	  This is a driver for the Qualcomm Network-on-Chip on msm8953-based
 	  platforms.
 
+config INTERCONNECT_QCOM_MSM8660
+	# bool, not tristate: this provider must be registered at
+	# core_initcall to be ready before icc_init (subsys_initcall) walks
+	# devicetree consumers; a loadable module cannot satisfy that
+	# ordering. Build-in only.
+	bool "Qualcomm MSM8x60 interconnect driver"
+	depends on INTERCONNECT_QCOM=y
+	depends on MFD_QCOM_RPM
+	help
+	  This is a driver for the Qualcomm fabric-based bus interconnect
+	  on MSM8x60 family (MSM8260/MSM8660/APQ8060) platforms (e.g., HP TouchPad).
+	  The driver manages APPSS, System, and MMSS fabrics and sends
+	  per-port bandwidth arbitration requests to RPM firmware.
+
 config INTERCONNECT_QCOM_MSM8974
 	tristate "Qualcomm MSM8974 interconnect driver"
 	depends on INTERCONNECT_QCOM
diff --git a/drivers/interconnect/qcom/Makefile b/drivers/interconnect/qcom/Makefile
index cdf2c6c9fbf3..0c849c30a907 100644
--- a/drivers/interconnect/qcom/Makefile
+++ b/drivers/interconnect/qcom/Makefile
@@ -13,6 +13,7 @@ qnoc-msm8916-objs			:= msm8916.o
 qnoc-msm8937-objs			:= msm8937.o
 qnoc-msm8939-objs			:= msm8939.o
 qnoc-msm8953-objs			:= msm8953.o
+qnoc-msm8660-objs			:= msm8660.o
 qnoc-msm8974-objs			:= msm8974.o
 qnoc-msm8976-objs			:= msm8976.o
 qnoc-msm8996-objs			:= msm8996.o
@@ -58,6 +59,7 @@ obj-$(CONFIG_INTERCONNECT_QCOM_MSM8916) += qnoc-msm8916.o
 obj-$(CONFIG_INTERCONNECT_QCOM_MSM8937) += qnoc-msm8937.o
 obj-$(CONFIG_INTERCONNECT_QCOM_MSM8939) += qnoc-msm8939.o
 obj-$(CONFIG_INTERCONNECT_QCOM_MSM8953) += qnoc-msm8953.o
+obj-$(CONFIG_INTERCONNECT_QCOM_MSM8660) += qnoc-msm8660.o
 obj-$(CONFIG_INTERCONNECT_QCOM_MSM8974) += qnoc-msm8974.o
 obj-$(CONFIG_INTERCONNECT_QCOM_MSM8976) += qnoc-msm8976.o
 obj-$(CONFIG_INTERCONNECT_QCOM_MSM8996) += qnoc-msm8996.o
diff --git a/drivers/interconnect/qcom/msm8660.c b/drivers/interconnect/qcom/msm8660.c
new file mode 100644
index 000000000000..f4bb852eaa2b
--- /dev/null
+++ b/drivers/interconnect/qcom/msm8660.c
@@ -0,0 +1,1147 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Qualcomm MSM8x60 family (MSM8260/MSM8660/APQ8060) interconnect driver
+ *
+ * Copyright (c) 2026 Herman van Hazendonk <github.com@herrie.org>
+ *
+ * Based on msm8974.c by Brian Masney <masneyb@onstation.org>
+ * and legacy vendor kernel msm_bus_board_8660.c / msm_bus_fabric.c
+ * Copyright (c) 2010-2011, Code Aurora Forum. All rights reserved.
+ *
+ * MSM8x60 has a fabric-based bus architecture:
+ *
+ *              +------------------+
+ *              |   APPSS Fabric   |  (CPU, L2, Memory)
+ *              +--------+---------+
+ *                       |
+ *         +-------------+-------------+
+ *         |                           |
+ *  +------+------+            +-------+-------+
+ *  | MMSS Fabric |            | System Fabric |
+ *  | (Display,   |            | (Peripherals, |
+ *  |  Camera,    |            |  DMA, etc)    |
+ *  |  Video)     |            +-------+-------+
+ *  +-------------+                    |
+ *                           +---------+---------+
+ *                           |                   |
+ *                    +------+------+     +------+------+
+ *                    | System FPB  |     |  CPSS FPB   |
+ *                    | (RPM, PMIC) |     | (GSBI, USB) |
+ *                    +-------------+     +-------------+
+ *
+ * Each fabric has an RPM arbitration interface that programs per-port
+ * bandwidth and priority tier via MM_FABRIC_ARB / SYS_FABRIC_ARB /
+ * APPS_FABRIC_ARB registers.  The legacy vendor kernel sent these as packed
+ * u16 arrays (bwsum + arb) through msm_rpm_set().  This driver uses
+ * the mainline qcom_rpm_write() interface to do the same.
+ */
+
+#include <dt-bindings/interconnect/qcom,msm8660.h>
+#include <dt-bindings/mfd/qcom-rpm.h>
+
+#include <linux/args.h>
+#include <linux/clk.h>
+#include <linux/device.h>
+#include <linux/interconnect-provider.h>
+#include <linux/io.h>
+#include <linux/mfd/qcom_rpm.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/of_platform.h>
+#include <linux/platform_device.h>
+#include <linux/slab.h>
+
+/* Internal node IDs - these map to the DT binding IDs plus fabric offset */
+enum {
+	/* APPSS Fabric nodes */
+	MSM8660_AFAB_MAS_AMPSS_M0 = 1,
+	MSM8660_AFAB_MAS_AMPSS_M1,
+	MSM8660_AFAB_SLV_EBI_CH0,
+	MSM8660_AFAB_SLV_AMPSS_L2,
+	MSM8660_AFAB_TO_MMSS,
+	MSM8660_AFAB_TO_SYSTEM,
+
+	/* System Fabric nodes */
+	MSM8660_SFAB_MAS_APPSS,
+	MSM8660_SFAB_MAS_SPS,
+	MSM8660_SFAB_MAS_ADM0_PORT0,
+	MSM8660_SFAB_MAS_ADM0_PORT1,
+	MSM8660_SFAB_MAS_ADM1_PORT0,
+	MSM8660_SFAB_MAS_ADM1_PORT1,
+	MSM8660_SFAB_MAS_LPASS_PROC,
+	MSM8660_SFAB_MAS_MSS_PROCI,
+	MSM8660_SFAB_MAS_MSS_PROCD,
+	MSM8660_SFAB_MAS_MSS_MDM_PORT0,
+	MSM8660_SFAB_MAS_LPASS,
+	MSM8660_SFAB_MAS_MMSS_FPB,
+	MSM8660_SFAB_MAS_ADM1_CI,
+	MSM8660_SFAB_MAS_ADM0_CI,
+	MSM8660_SFAB_MAS_MSS_MDM_PORT1,
+	MSM8660_SFAB_MAS_USB_HS,
+	MSM8660_SFAB_TO_APPSS,
+	MSM8660_SFAB_TO_SYSTEM_FPB,
+	MSM8660_SFAB_TO_CPSS_FPB,
+	MSM8660_SFAB_SLV_SPS,
+	MSM8660_SFAB_SLV_SYSTEM_IMEM,
+	MSM8660_SFAB_SLV_AMPSS,
+	MSM8660_SFAB_SLV_MSS,
+	MSM8660_SFAB_SLV_LPASS,
+	MSM8660_SFAB_SLV_MMSS_FPB,
+	MSM8660_SFAB_TO_DFAB,
+
+	/* Daytona Fabric nodes (DFAB) - peripheral bus */
+	MSM8660_DFAB_MAS_SDC1,
+	MSM8660_DFAB_MAS_SDC2,
+	MSM8660_DFAB_MAS_SDC3,
+	MSM8660_DFAB_MAS_SDC4,
+	MSM8660_DFAB_MAS_SDC5,
+	MSM8660_DFAB_MAS_ADM0_MASTER,
+	MSM8660_DFAB_MAS_ADM1_MASTER,
+	MSM8660_DFAB_TO_SFAB,
+	MSM8660_DFAB_SLV_SDC1,
+	MSM8660_DFAB_SLV_SDC2,
+	MSM8660_DFAB_SLV_SDC3,
+	MSM8660_DFAB_SLV_SDC4,
+	MSM8660_DFAB_SLV_SDC5,
+	MSM8660_DFAB_MAS_USB_HS,	/* USB HS DFAB voter */
+	MSM8660_DFAB_MAS_DSPS,		/* DSPS DFAB voter */
+
+	/* MMSS Fabric nodes */
+	MSM8660_MMFAB_MAS_MDP_PORT0,
+	MSM8660_MMFAB_MAS_MDP_PORT1,
+	MSM8660_MMFAB_MAS_ADM1_PORT0,
+	MSM8660_MMFAB_MAS_ROTATOR,
+	MSM8660_MMFAB_MAS_GRAPHICS_3D,
+	MSM8660_MMFAB_MAS_JPEG_DEC,
+	MSM8660_MMFAB_MAS_GRAPHICS_2D_CORE0,
+	MSM8660_MMFAB_MAS_VFE,
+	MSM8660_MMFAB_MAS_VPE,
+	MSM8660_MMFAB_MAS_JPEG_ENC,
+	MSM8660_MMFAB_MAS_GRAPHICS_2D_CORE1,
+	MSM8660_MMFAB_MAS_HD_CODEC_PORT0,
+	MSM8660_MMFAB_MAS_HD_CODEC_PORT1,
+	MSM8660_MMFAB_TO_APPSS,
+	MSM8660_MMFAB_SLV_SMI,
+	MSM8660_MMFAB_SLV_MM_IMEM,
+};
+
+#define to_msm8660_icc_provider(_provider) \
+	container_of(_provider, struct msm8660_icc_provider, provider)
+
+/*
+ * Minimum fabric clock rate to prevent bus starvation.
+ *
+ * When no consumers request bandwidth, the rate calculation yields 0,
+ * causing fabric clocks to drop to minimum. This creates bimodal
+ * performance: fast when other subsystems (like display) happen to
+ * request bandwidth, slow otherwise.
+ *
+ * 384 MHz keeps fabric fast during concurrent MDP display scanout
+ * and USB gadget traffic. legacy vendor kernel docs: "AXI bus frequency needs to be
+ * kept at maximum value while USB data transfers are happening."
+ * 266 MHz was insufficient - USB crashed during display activity.
+ */
+#define MSM8660_FABRIC_MIN_RATE		384000000UL	/* 384 MHz */
+
+/*
+ * Maximum RPM ARB buffer size across all fabrics.
+ * MM fabric is largest at 23 u32 words.
+ */
+#define MSM8660_MAX_RPM_BUF		23
+
+/*
+ * RPM fabric arbitration data format (from legacy vendor kernel msm_bus_fabric.c):
+ *
+ * Each u16 arb entry: bit 15 = tier (1=TIER1 high priority), bits 14-0 = BW
+ * Bandwidth is in 128KB units (bytes >> 17).
+ * Two u16 values are packed into each u32 RPM register word.
+ *
+ * Buffer layout: [bwsum pairs] [arb pairs]
+ * bwsum[slave_port] = total bandwidth to that slave
+ * arb[(tier-1)*nmasters + master_port] = per-master arbitration entry
+ */
+#define ARB_BWMASK		0x7FFF
+#define ARB_TIERMASK		0x8000
+#define ARB_TIER1		1
+#define ARB_TIER2		2
+
+static const struct clk_bulk_data msm8660_afab_clocks[] = {
+	{ .id = "bus" },
+	{ .id = "bus_a" },
+	{ .id = "ebi1" },
+	{ .id = "ebi1_a" },
+};
+
+static const struct clk_bulk_data msm8660_sfab_clocks[] = {
+	{ .id = "bus" },
+	{ .id = "bus_a" },
+};
+
+static const struct clk_bulk_data msm8660_mmfab_clocks[] = {
+	{ .id = "bus" },
+	{ .id = "bus_a" },
+	{ .id = "smi" },
+	{ .id = "smi_a" },
+};
+
+static const struct clk_bulk_data msm8660_dfab_clocks[] = {
+	{ .id = "bus" },
+	{ .id = "bus_a" },
+};
+
+/**
+ * struct msm8660_icc_node - MSM8660 specific interconnect nodes
+ * @name: the node name used in debugfs
+ * @id: a unique node identifier
+ * @links: an array of nodes where we can go next while traversing
+ * @num_links: the total number of @links
+ * @buswidth: width of the interconnect between a node and the bus (bytes)
+ * @rate: current bus clock rate in Hz
+ * @mas_port: master port index for RPM ARB (-1 if not a master)
+ * @slv_port: slave port index for RPM bwsum (-1 if not a slave)
+ * @mas_tier: master priority tier (ARB_TIER1 or ARB_TIER2, 0 if N/A)
+ */
+struct msm8660_icc_node {
+	unsigned char *name;
+	u16 id;
+#define MSM8660_ICC_MAX_LINKS	3
+	u16 links[MSM8660_ICC_MAX_LINKS];
+	u16 num_links;
+	u16 buswidth;
+	s8 mas_port;
+	s8 slv_port;
+	u8 mas_tier;
+};
+
+/**
+ * struct msm8660_icc_desc - Fabric descriptor
+ * @nodes: array of node pointers
+ * @num_nodes: number of nodes
+ * @bus_clks: clock definitions
+ * @num_clks: number of clocks
+ * @rpm_resource: QCOM_RPM_*_FABRIC_ARB constant, or -1 for no ARB
+ * @nmasters: number of master ports in this fabric (for ARB array sizing)
+ * @nslaves: number of slave ports in this fabric (for bwsum array sizing)
+ * @ntieredslaves: number of tiered slaves (ARB rows)
+ * @default_tiered_slave: 1-based index of default tiered slave for masters
+ * @rpm_buf_size: number of u32 words for RPM write
+ * @bus_width: representative fabric bus width in bytes, used as the
+ *             divisor for translating aggregate bytes/sec into a single
+ *             clock rate that drives the whole fabric
+ */
+struct msm8660_icc_desc {
+	struct msm8660_icc_node * const *nodes;
+	size_t num_nodes;
+	const struct clk_bulk_data *bus_clks;
+	size_t num_clks;
+	int rpm_resource;
+	u8 nmasters;
+	u8 nslaves;
+	u8 ntieredslaves;
+	u8 default_tiered_slave;
+	u8 rpm_buf_size;
+	u8 bus_width;
+};
+
+/**
+ * struct msm8660_icc_provider - MSM8660 specific interconnect provider
+ * @provider: generic interconnect provider
+ * @bus_clks: the clk_bulk_data table of bus clocks
+ * @num_clks: the total number of clk_bulk_data entries
+ * @rpm: RPM handle for fabric arbitration writes
+ * @desc: fabric descriptor with RPM metadata
+ * @arb: pre-allocated arbitration array (nmasters * ntieredslaves u16 entries)
+ * @bwsum: pre-allocated bandwidth sum array (nslaves u16 entries)
+ * @rpm_buf: pre-allocated RPM write buffer (rpm_buf_size u32 entries)
+ * @rate: last clock rate applied to the fabric bus_clks, used as the
+ *        single source of truth for whether the rate actually needs to
+ *        be reprogrammed (per-node caching would desync when different
+ *        masters update at different times)
+ */
+struct msm8660_icc_provider {
+	struct icc_provider provider;
+	struct clk_bulk_data *bus_clks;
+	int num_clks;
+	struct qcom_rpm *rpm;
+	const struct msm8660_icc_desc *desc;
+	u16 *arb;
+	u16 *bwsum;
+	u32 *rpm_buf;
+	u32 rate;
+};
+
+/*
+ * Node definitions with RPM port mapping.
+ *
+ * DEFINE_QNODE(_name, _id, _buswidth, _mas_port, _slv_port, _tier, links...)
+ *
+ * _mas_port: master port index for RPM ARB array (-1 if not a master)
+ * _slv_port: slave port index for RPM bwsum array (-1 if not a slave)
+ * _tier: master priority tier (ARB_TIER1=1, ARB_TIER2=2, 0 if N/A)
+ */
+#define DEFINE_QNODE(_name, _id, _buswidth, _mas, _slv, _tier, ...)	\
+	static struct msm8660_icc_node _name = {			\
+		.name = #_name,						\
+		.id = _id,						\
+		.buswidth = _buswidth,					\
+		.num_links = COUNT_ARGS(__VA_ARGS__),			\
+		.links = { __VA_ARGS__ },				\
+		.mas_port = _mas,					\
+		.slv_port = _slv,					\
+		.mas_tier = _tier,					\
+	}
+
+/*
+ * =========================================================================
+ * APPSS Fabric nodes
+ *
+ * 4 masters, 4 slaves, 2 tiered slaves
+ * Master ports: SMPSS_M0=0, SMPSS_M1=1, FAB_MMSS=2, FAB_SYSTEM=3
+ * Slave ports:  EBI_CH0=0, SMPSS_L2=1, MMSS_FAB=2, SYSTEM_FAB=3
+ * Default target: tiered slave 1 (EBI_CH0)
+ * =========================================================================
+ */
+DEFINE_QNODE(mas_ampss_m0, MSM8660_AFAB_MAS_AMPSS_M0, 8, 0, -1, ARB_TIER2,
+	     MSM8660_AFAB_SLV_EBI_CH0, MSM8660_AFAB_TO_MMSS, MSM8660_AFAB_TO_SYSTEM);
+DEFINE_QNODE(mas_ampss_m1, MSM8660_AFAB_MAS_AMPSS_M1, 8, 1, -1, ARB_TIER2,
+	     MSM8660_AFAB_SLV_EBI_CH0, MSM8660_AFAB_TO_MMSS, MSM8660_AFAB_TO_SYSTEM);
+DEFINE_QNODE(slv_ebi_ch0, MSM8660_AFAB_SLV_EBI_CH0, 8, -1, 0, 0);
+DEFINE_QNODE(slv_ampss_l2, MSM8660_AFAB_SLV_AMPSS_L2, 8, -1, 1, 0);
+/*
+ * Gateway nodes need links to both the cross-fabric gateway AND the memory
+ * slave to enable cross-fabric paths. Without link to EBI_CH0, path_find()
+ * can't route from MMSS/System fabric masters to main memory.
+ *
+ * AFAB_TO_MMSS doubles as AFAB master port 2 (the FAB_MMSS master). MDP
+ * scanout and GPU traffic enter AFAB through this gateway. Mark it
+ * ARB_TIER1 so display/multimedia traffic keeps priority over CPU L2
+ * misses inside the APPSS fabric — without this, MDP TIER1 priority
+ * earned in MMFAB is dropped at the AFAB boundary and MDP fetches lose
+ * arbitration to CPU traffic, producing PRIMARY_INTF_UDERRUN.
+ */
+DEFINE_QNODE(afab_to_mmss, MSM8660_AFAB_TO_MMSS, 8, 2, 2, ARB_TIER1,
+	     MSM8660_MMFAB_TO_APPSS, MSM8660_AFAB_SLV_EBI_CH0);
+DEFINE_QNODE(afab_to_system, MSM8660_AFAB_TO_SYSTEM, 8, 3, 3, ARB_TIER2,
+	     MSM8660_SFAB_TO_APPSS, MSM8660_AFAB_SLV_EBI_CH0);
+
+static struct msm8660_icc_node * const msm8660_afab_nodes[] = {
+	[AFAB_MAS_AMPSS_M0] = &mas_ampss_m0,
+	[AFAB_MAS_AMPSS_M1] = &mas_ampss_m1,
+	[AFAB_SLV_EBI_CH0] = &slv_ebi_ch0,
+	[AFAB_SLV_AMPSS_L2] = &slv_ampss_l2,
+	[AFAB_TO_MMSS] = &afab_to_mmss,
+	[AFAB_TO_SYSTEM] = &afab_to_system,
+};
+
+static const struct msm8660_icc_desc msm8660_afab = {
+	.nodes = msm8660_afab_nodes,
+	.num_nodes = ARRAY_SIZE(msm8660_afab_nodes),
+	.bus_clks = msm8660_afab_clocks,
+	.num_clks = ARRAY_SIZE(msm8660_afab_clocks),
+	.rpm_resource = QCOM_RPM_APPS_FABRIC_ARB,
+	.nmasters = 4,
+	.nslaves = 4,
+	.ntieredslaves = 2,
+	.default_tiered_slave = 1,	/* EBI_CH0 */
+	.rpm_buf_size = 6,
+	.bus_width = 8,			/* 64-bit APPSS fabric datapath */
+};
+
+/*
+ * =========================================================================
+ * System Fabric nodes
+ *
+ * 17 masters, 9 slaves, 2 tiered slaves
+ * Master ports: see enum msm_bus_8660_master_ports_type in legacy vendor kernel
+ * Slave ports:  APPSS_FAB=0, SPS=1, SYSTEM_IMEM=2, SMPSS=3, MSS=4,
+ *               LPASS=5, CPSS_FPB=6, SYSTEM_FPB=7, MMSS_FPB=8
+ * Default target: tiered slave 1 (APPSS gateway)
+ * =========================================================================
+ */
+DEFINE_QNODE(sfab_mas_appss, MSM8660_SFAB_MAS_APPSS, 8, 0, -1, ARB_TIER2,
+	     MSM8660_AFAB_TO_SYSTEM);
+DEFINE_QNODE(sfab_mas_sps, MSM8660_SFAB_MAS_SPS, 8, 1, -1, ARB_TIER2,
+	     MSM8660_SFAB_SLV_SPS);
+/*
+ * ADM DMA masters - route through SFAB_TO_APPSS to reach EBI memory.
+ * Path: ADM -> SFAB_TO_APPSS -> AFAB_TO_SYSTEM -> AFAB_SLV_EBI_CH0
+ * This enables proper EBI bandwidth voting for DMA operations.
+ */
+DEFINE_QNODE(sfab_mas_adm0_port0, MSM8660_SFAB_MAS_ADM0_PORT0, 8, 2, -1, ARB_TIER2,
+	     MSM8660_SFAB_TO_APPSS);
+DEFINE_QNODE(sfab_mas_adm0_port1, MSM8660_SFAB_MAS_ADM0_PORT1, 8, 3, -1, ARB_TIER2,
+	     MSM8660_SFAB_TO_APPSS);
+DEFINE_QNODE(sfab_mas_adm1_port0, MSM8660_SFAB_MAS_ADM1_PORT0, 8, 4, -1, ARB_TIER2,
+	     MSM8660_SFAB_TO_APPSS);
+DEFINE_QNODE(sfab_mas_adm1_port1, MSM8660_SFAB_MAS_ADM1_PORT1, 8, 5, -1, ARB_TIER2,
+	     MSM8660_SFAB_TO_APPSS);
+DEFINE_QNODE(sfab_mas_lpass_proc, MSM8660_SFAB_MAS_LPASS_PROC, 8, 6, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_mss_proci, MSM8660_SFAB_MAS_MSS_PROCI, 8, 7, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_mss_procd, MSM8660_SFAB_MAS_MSS_PROCD, 8, 8, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_mss_mdm_port0, MSM8660_SFAB_MAS_MSS_MDM_PORT0, 8, 9, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_lpass, MSM8660_SFAB_MAS_LPASS, 8, 10, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_mmss_fpb, MSM8660_SFAB_MAS_MMSS_FPB, 8, 13, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_adm1_ci, MSM8660_SFAB_MAS_ADM1_CI, 8, 14, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_adm0_ci, MSM8660_SFAB_MAS_ADM0_CI, 8, 15, -1, ARB_TIER2);
+DEFINE_QNODE(sfab_mas_mss_mdm_port1, MSM8660_SFAB_MAS_MSS_MDM_PORT1, 8, 16, -1, ARB_TIER2);
+/* USB HS has no dedicated master port in legacy vendor kernel SFAB - bandwidth voting only */
+DEFINE_QNODE(sfab_mas_usb_hs, MSM8660_SFAB_MAS_USB_HS, 8, -1, -1, 0,
+	     MSM8660_SFAB_TO_APPSS);
+DEFINE_QNODE(sfab_to_appss, MSM8660_SFAB_TO_APPSS, 8, -1, 0, 0,
+	     MSM8660_AFAB_TO_SYSTEM);
+DEFINE_QNODE(sfab_to_system_fpb, MSM8660_SFAB_TO_SYSTEM_FPB, 4, -1, 7, 0);
+DEFINE_QNODE(sfab_to_cpss_fpb, MSM8660_SFAB_TO_CPSS_FPB, 4, -1, 6, 0);
+DEFINE_QNODE(sfab_slv_sps, MSM8660_SFAB_SLV_SPS, 8, -1, 1, 0);
+DEFINE_QNODE(sfab_slv_system_imem, MSM8660_SFAB_SLV_SYSTEM_IMEM, 8, -1, 2, 0);
+DEFINE_QNODE(sfab_slv_ampss, MSM8660_SFAB_SLV_AMPSS, 8, -1, 3, 0);
+DEFINE_QNODE(sfab_slv_mss, MSM8660_SFAB_SLV_MSS, 8, -1, 4, 0);
+DEFINE_QNODE(sfab_slv_lpass, MSM8660_SFAB_SLV_LPASS, 8, -1, 5, 0);
+DEFINE_QNODE(sfab_slv_mmss_fpb, MSM8660_SFAB_SLV_MMSS_FPB, 8, -1, 8, 0);
+/*
+ * Gateway to DFAB: links to DFAB_TO_SFAB for path traversal.
+ * Also links to SFAB_TO_APPSS to enable DFAB->SFAB->AFAB->memory paths.
+ * No slave port in legacy vendor kernel SFAB config (DFAB is separate fabric).
+ */
+DEFINE_QNODE(sfab_to_dfab, MSM8660_SFAB_TO_DFAB, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB, MSM8660_SFAB_TO_APPSS);
+
+static struct msm8660_icc_node * const msm8660_sfab_nodes[] = {
+	[SFAB_MAS_APPSS] = &sfab_mas_appss,
+	[SFAB_MAS_SPS] = &sfab_mas_sps,
+	[SFAB_MAS_ADM0_PORT0] = &sfab_mas_adm0_port0,
+	[SFAB_MAS_ADM0_PORT1] = &sfab_mas_adm0_port1,
+	[SFAB_MAS_ADM1_PORT0] = &sfab_mas_adm1_port0,
+	[SFAB_MAS_ADM1_PORT1] = &sfab_mas_adm1_port1,
+	[SFAB_MAS_LPASS_PROC] = &sfab_mas_lpass_proc,
+	[SFAB_MAS_MSS_PROCI] = &sfab_mas_mss_proci,
+	[SFAB_MAS_MSS_PROCD] = &sfab_mas_mss_procd,
+	[SFAB_MAS_MSS_MDM_PORT0] = &sfab_mas_mss_mdm_port0,
+	[SFAB_MAS_LPASS] = &sfab_mas_lpass,
+	[SFAB_MAS_MMSS_FPB] = &sfab_mas_mmss_fpb,
+	[SFAB_MAS_ADM1_CI] = &sfab_mas_adm1_ci,
+	[SFAB_MAS_ADM0_CI] = &sfab_mas_adm0_ci,
+	[SFAB_MAS_MSS_MDM_PORT1] = &sfab_mas_mss_mdm_port1,
+	[SFAB_MAS_USB_HS] = &sfab_mas_usb_hs,
+	[SFAB_TO_APPSS] = &sfab_to_appss,
+	[SFAB_TO_SYSTEM_FPB] = &sfab_to_system_fpb,
+	[SFAB_TO_CPSS_FPB] = &sfab_to_cpss_fpb,
+	[SFAB_SLV_SPS] = &sfab_slv_sps,
+	[SFAB_SLV_SYSTEM_IMEM] = &sfab_slv_system_imem,
+	[SFAB_SLV_AMPSS] = &sfab_slv_ampss,
+	[SFAB_SLV_MSS] = &sfab_slv_mss,
+	[SFAB_SLV_LPASS] = &sfab_slv_lpass,
+	[SFAB_SLV_MMSS_FPB] = &sfab_slv_mmss_fpb,
+	[SFAB_TO_DFAB] = &sfab_to_dfab,
+};
+
+static const struct msm8660_icc_desc msm8660_sfab = {
+	.nodes = msm8660_sfab_nodes,
+	.num_nodes = ARRAY_SIZE(msm8660_sfab_nodes),
+	.bus_clks = msm8660_sfab_clocks,
+	.num_clks = ARRAY_SIZE(msm8660_sfab_clocks),
+	.rpm_resource = QCOM_RPM_SYS_FABRIC_ARB,
+	.nmasters = 17,
+	.nslaves = 9,
+	.ntieredslaves = 2,
+	.default_tiered_slave = 1,	/* APPSS gateway */
+	.rpm_buf_size = 22,
+	.bus_width = 8,			/* 64-bit System fabric datapath */
+};
+
+/*
+ * =========================================================================
+ * MMSS Fabric nodes - Multimedia subsystem (MDP, camera, video, GPU)
+ *
+ * 14 masters, 4 slaves, 3 tiered slaves
+ * Master ports: MDP0=0, MDP1=1, ADM1=2, ROT=3, 3D=4, JPEG_DEC=5,
+ *               2D_CORE0=6, VFE=7, VPE=8, JPEG_ENC=9, 2D_CORE1=10,
+ *               (APPS_FAB=11), HD_CODEC0=12, HD_CODEC1=13
+ * Slave ports:  SMI=0, APPSS_FAB=1, (APPSS_FAB_1=2), MM_IMEM=3
+ * Tiered slaves: SMI=1, APPSS_FAB=2, MM_IMEM=3
+ * Default target: tiered slave 2 (APPSS gateway -> main memory)
+ *
+ * MDP ports get TIER1 (high priority) for guaranteed display refresh.
+ * All other masters get TIER2 (default priority).
+ * =========================================================================
+ */
+DEFINE_QNODE(mmfab_mas_mdp_port0, MSM8660_MMFAB_MAS_MDP_PORT0, 16, 0, -1, ARB_TIER1,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_mdp_port1, MSM8660_MMFAB_MAS_MDP_PORT1, 16, 1, -1, ARB_TIER1,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_adm1_port0, MSM8660_MMFAB_MAS_ADM1_PORT0, 8, 2, -1, ARB_TIER2);
+DEFINE_QNODE(mmfab_mas_rotator, MSM8660_MMFAB_MAS_ROTATOR, 16, 3, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_graphics_3d, MSM8660_MMFAB_MAS_GRAPHICS_3D, 16, 4, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_jpeg_dec, MSM8660_MMFAB_MAS_JPEG_DEC, 16, 5, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_graphics_2d_core0, MSM8660_MMFAB_MAS_GRAPHICS_2D_CORE0, 16,
+	     6, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_vfe, MSM8660_MMFAB_MAS_VFE, 16, 7, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_vpe, MSM8660_MMFAB_MAS_VPE, 16, 8, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_jpeg_enc, MSM8660_MMFAB_MAS_JPEG_ENC, 16, 9, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_graphics_2d_core1, MSM8660_MMFAB_MAS_GRAPHICS_2D_CORE1, 16,
+	     10, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_hd_codec_port0, MSM8660_MMFAB_MAS_HD_CODEC_PORT0, 16,
+	     12, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+DEFINE_QNODE(mmfab_mas_hd_codec_port1, MSM8660_MMFAB_MAS_HD_CODEC_PORT1, 16,
+	     13, -1, ARB_TIER2,
+	     MSM8660_MMFAB_SLV_SMI, MSM8660_MMFAB_TO_APPSS);
+/*
+ * Gateway from APPSS into MMSS: slave (port 1) for outbound traffic
+ * leaving MMSS, AND master (port 2) for inbound traffic arriving from
+ * APPSS (e.g. CPU memremap_wc accesses to SMI BOs). Without the master
+ * port and the forward links into MMFAB slaves (SMI / MM_IMEM), the
+ * ICC path-finder has no route from AMPSS_M0 to MMFAB_SLV_SMI; the
+ * cross-fabric gateway only worked for outbound traffic (MMSS masters
+ * reaching APPSS slaves like EBI).
+ *
+ * ARB_TIER1 keeps AMPSS->SMI traffic high-priority within MMFAB so
+ * CPU mmap reads/writes to SMI BOs don't get starved by MDP scanout.
+ */
+DEFINE_QNODE(mmfab_to_appss, MSM8660_MMFAB_TO_APPSS, 8, 11, 1, ARB_TIER1,
+	     MSM8660_AFAB_TO_MMSS,
+	     MSM8660_MMFAB_SLV_SMI,
+	     MSM8660_MMFAB_SLV_MM_IMEM);
+DEFINE_QNODE(mmfab_slv_smi, MSM8660_MMFAB_SLV_SMI, 16, -1, 0, 0);
+DEFINE_QNODE(mmfab_slv_mm_imem, MSM8660_MMFAB_SLV_MM_IMEM, 8, -1, 3, 0);
+
+static struct msm8660_icc_node * const msm8660_mmfab_nodes[] = {
+	[MMFAB_MAS_MDP_PORT0] = &mmfab_mas_mdp_port0,
+	[MMFAB_MAS_MDP_PORT1] = &mmfab_mas_mdp_port1,
+	[MMFAB_MAS_ADM1_PORT0] = &mmfab_mas_adm1_port0,
+	[MMFAB_MAS_ROTATOR] = &mmfab_mas_rotator,
+	[MMFAB_MAS_GRAPHICS_3D] = &mmfab_mas_graphics_3d,
+	[MMFAB_MAS_JPEG_DEC] = &mmfab_mas_jpeg_dec,
+	[MMFAB_MAS_GRAPHICS_2D_CORE0] = &mmfab_mas_graphics_2d_core0,
+	[MMFAB_MAS_VFE] = &mmfab_mas_vfe,
+	[MMFAB_MAS_VPE] = &mmfab_mas_vpe,
+	[MMFAB_MAS_JPEG_ENC] = &mmfab_mas_jpeg_enc,
+	[MMFAB_MAS_GRAPHICS_2D_CORE1] = &mmfab_mas_graphics_2d_core1,
+	[MMFAB_MAS_HD_CODEC_PORT0] = &mmfab_mas_hd_codec_port0,
+	[MMFAB_MAS_HD_CODEC_PORT1] = &mmfab_mas_hd_codec_port1,
+	[MMFAB_TO_APPSS] = &mmfab_to_appss,
+	[MMFAB_SLV_SMI] = &mmfab_slv_smi,
+	[MMFAB_SLV_MM_IMEM] = &mmfab_slv_mm_imem,
+};
+
+static const struct msm8660_icc_desc msm8660_mmfab = {
+	.nodes = msm8660_mmfab_nodes,
+	.num_nodes = ARRAY_SIZE(msm8660_mmfab_nodes),
+	.bus_clks = msm8660_mmfab_clocks,
+	.num_clks = ARRAY_SIZE(msm8660_mmfab_clocks),
+	.rpm_resource = QCOM_RPM_MM_FABRIC_ARB,
+	.nmasters = 14,
+	.nslaves = 4,
+	.ntieredslaves = 3,
+	.default_tiered_slave = 2,	/* APPSS gateway */
+	.rpm_buf_size = 23,
+	.bus_width = 16,		/* 128-bit Multimedia fabric datapath */
+};
+
+/*
+ * =========================================================================
+ * Daytona Fabric (DFAB) nodes - peripheral bus for SDCC and ADM DMA
+ *
+ * DFAB connects slower peripherals to SFAB via the DFAB_TO_SFAB gateway.
+ * SDCC controllers (eMMC, SD card) connect here.
+ *
+ * No RPM ARB for DFAB - it's a simple peripheral bus with clock-only control.
+ *
+ * USB HS is included as a DFAB voter for compatibility with the legacy
+ * legacy vendor kernel clock voting mechanism.
+ * =========================================================================
+ */
+DEFINE_QNODE(dfab_mas_sdc1, MSM8660_DFAB_MAS_SDC1, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_mas_sdc2, MSM8660_DFAB_MAS_SDC2, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_mas_sdc3, MSM8660_DFAB_MAS_SDC3, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_mas_sdc4, MSM8660_DFAB_MAS_SDC4, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_mas_sdc5, MSM8660_DFAB_MAS_SDC5, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_mas_adm0_master, MSM8660_DFAB_MAS_ADM0_MASTER, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_mas_adm1_master, MSM8660_DFAB_MAS_ADM1_MASTER, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+DEFINE_QNODE(dfab_to_sfab, MSM8660_DFAB_TO_SFAB, 8, -1, -1, 0,
+	     MSM8660_SFAB_TO_DFAB);
+DEFINE_QNODE(dfab_slv_sdc1, MSM8660_DFAB_SLV_SDC1, 8, -1, -1, 0);
+DEFINE_QNODE(dfab_slv_sdc2, MSM8660_DFAB_SLV_SDC2, 8, -1, -1, 0);
+DEFINE_QNODE(dfab_slv_sdc3, MSM8660_DFAB_SLV_SDC3, 8, -1, -1, 0);
+DEFINE_QNODE(dfab_slv_sdc4, MSM8660_DFAB_SLV_SDC4, 8, -1, -1, 0);
+DEFINE_QNODE(dfab_slv_sdc5, MSM8660_DFAB_SLV_SDC5, 8, -1, -1, 0);
+/* USB HS DFAB voter - keeps DFAB clock stable during USB activity */
+DEFINE_QNODE(dfab_mas_usb_hs, MSM8660_DFAB_MAS_USB_HS, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+/* DSPS DFAB voter - keeps DFAB clock stable during sensor activity */
+DEFINE_QNODE(dfab_mas_dsps, MSM8660_DFAB_MAS_DSPS, 8, -1, -1, 0,
+	     MSM8660_DFAB_TO_SFAB);
+
+static struct msm8660_icc_node * const msm8660_dfab_nodes[] = {
+	[DFAB_MAS_SDC1] = &dfab_mas_sdc1,
+	[DFAB_MAS_SDC2] = &dfab_mas_sdc2,
+	[DFAB_MAS_SDC3] = &dfab_mas_sdc3,
+	[DFAB_MAS_SDC4] = &dfab_mas_sdc4,
+	[DFAB_MAS_SDC5] = &dfab_mas_sdc5,
+	[DFAB_MAS_ADM0_MASTER] = &dfab_mas_adm0_master,
+	[DFAB_MAS_ADM1_MASTER] = &dfab_mas_adm1_master,
+	[DFAB_TO_SFAB] = &dfab_to_sfab,
+	[DFAB_SLV_SDC1] = &dfab_slv_sdc1,
+	[DFAB_SLV_SDC2] = &dfab_slv_sdc2,
+	[DFAB_SLV_SDC3] = &dfab_slv_sdc3,
+	[DFAB_SLV_SDC4] = &dfab_slv_sdc4,
+	[DFAB_SLV_SDC5] = &dfab_slv_sdc5,
+	[DFAB_MAS_USB_HS] = &dfab_mas_usb_hs,
+	[DFAB_MAS_DSPS] = &dfab_mas_dsps,
+};
+
+static const struct msm8660_icc_desc msm8660_dfab = {
+	.nodes = msm8660_dfab_nodes,
+	.num_nodes = ARRAY_SIZE(msm8660_dfab_nodes),
+	.bus_clks = msm8660_dfab_clocks,
+	.num_clks = ARRAY_SIZE(msm8660_dfab_clocks),
+	.rpm_resource = -1,		/* No RPM ARB for DFAB */
+	.bus_width = 8,			/* 64-bit Daytona fabric datapath */
+};
+
+/*
+ * Pack bwsum[] and arb[] arrays into the u32 RPM buffer.
+ *
+ * Two u16 values are packed per u32 word: lower 16 bits first, upper 16 next.
+ * Layout: [bwsum pairs] then [arb pairs], handling odd boundaries.
+ *
+ * This matches the legacy vendor kernel msm_bus_fabric_rpm_commit() packing algorithm.
+ */
+static void msm8660_pack_rpm_data(const u16 *bwsum, int nslaves,
+				  const u16 *arb, int arb_size,
+				  u32 *buf)
+{
+	int i, index = 0;
+
+	/* Pack bwsum pairs */
+	for (i = 0; i + 1 < nslaves; i += 2) {
+		buf[index] = ((u32)bwsum[i + 1] << 16) | bwsum[i];
+		index++;
+	}
+
+	/*
+	 * Handle boundary between bwsum and arb for odd nslaves. When the
+	 * fabric has no master ports (arb_size == 0) the arb[0] access would
+	 * read out of bounds; pad the lone bwsum into the low half of the
+	 * word instead.
+	 */
+	if (nslaves & 1) {
+		if (arb_size > 0) {
+			buf[index] = ((u32)arb[0] << 16) | bwsum[i];
+			i = 1;
+		} else {
+			buf[index] = bwsum[i];
+			i = 0;
+		}
+		index++;
+	} else {
+		i = 0;
+	}
+
+	/* Pack arb pairs */
+	for (; i + 1 < arb_size; i += 2) {
+		buf[index] = ((u32)arb[i + 1] << 16) | arb[i];
+		index++;
+	}
+
+	/* Handle odd arb entry at end */
+	if (i < arb_size) {
+		buf[index] = arb[i];
+		index++;
+	}
+}
+
+/*
+ * Send fabric arbitration data to RPM.
+ *
+ * Iterates over all ICC nodes in the provider, builds bwsum/arb arrays
+ * from their aggregated bandwidth, and sends the packed data to RPM.
+ */
+static void msm8660_rpm_commit(struct msm8660_icc_provider *qp)
+{
+	const struct msm8660_icc_desc *desc = qp->desc;
+	struct icc_provider *provider = &qp->provider;
+	int nm = desc->nmasters;
+	int ns = desc->nslaves;
+	int nts = desc->ntieredslaves;
+	int arb_size = nm * nts;
+	int def_ts = desc->default_tiered_slave;
+	struct icc_node *n;
+	int ret;
+
+	memset(qp->bwsum, 0, ns * sizeof(u16));
+	memset(qp->arb, 0, arb_size * sizeof(u16));
+	memset(qp->rpm_buf, 0, desc->rpm_buf_size * sizeof(u32));
+
+	list_for_each_entry(n, &provider->nodes, node_list) {
+		struct msm8660_icc_node *qn = n->data;
+		u64 bw_bytes;
+		u16 bw_128k;
+
+		/* Use max of avg and peak bandwidth, convert to 128KB units */
+		bw_bytes = max(icc_units_to_bps(n->avg_bw),
+			       icc_units_to_bps(n->peak_bw));
+		bw_128k = (u16)min_t(u64, bw_bytes >> 17, ARB_BWMASK);
+
+		/* Set arb entry for masters at their default tiered slave */
+		if (qn->mas_port >= 0 && qn->mas_port < nm && def_ts > 0) {
+			int idx = (def_ts - 1) * nm + qn->mas_port;
+			u8 tier;
+
+			if (idx < arb_size) {
+				tier = bw_128k ? qn->mas_tier : ARB_TIER2;
+				qp->arb[idx] = (tier == ARB_TIER1 ? ARB_TIERMASK : 0)
+					       | (bw_128k & ARB_BWMASK);
+			}
+		}
+
+		/* Set bwsum for slaves */
+		if (qn->slv_port >= 0 && qn->slv_port < ns)
+			qp->bwsum[qn->slv_port] = bw_128k;
+	}
+
+	msm8660_pack_rpm_data(qp->bwsum, ns, qp->arb, arb_size, qp->rpm_buf);
+
+	ret = qcom_rpm_write(qp->rpm, QCOM_RPM_ACTIVE_STATE,
+			     desc->rpm_resource, qp->rpm_buf,
+			     desc->rpm_buf_size);
+	if (ret)
+		dev_err_ratelimited(provider->dev,
+				    "RPM fabric ARB write failed: %d\n", ret);
+}
+
+/*
+ * u64-safe replacement for icc_std_aggregate(): the standard helper sums
+ * average bandwidth into a u32, which can wrap around when summed across
+ * many high-bandwidth nodes. We accumulate in u64 internally and saturate
+ * back to U32_MAX on overflow rather than wrapping silently to a small
+ * value that would underclock the fabric.
+ */
+static int msm8660_icc_aggregate(struct icc_node *node, u32 tag,
+				 u32 avg_bw, u32 peak_bw,
+				 u32 *agg_avg, u32 *agg_peak)
+{
+	u64 new_avg = (u64)*agg_avg + avg_bw;
+
+	*agg_avg = (new_avg > U32_MAX) ? U32_MAX : (u32)new_avg;
+	*agg_peak = max(*agg_peak, peak_bw);
+	return 0;
+}
+
+static int msm8660_icc_set(struct icc_node *src, struct icc_node *dst)
+{
+	struct msm8660_icc_node *src_qn;
+	struct msm8660_icc_provider *qp;
+	u64 sum_bw, max_peak_bw, rate;
+	u32 agg_avg = 0, agg_peak = 0;
+	struct icc_provider *provider;
+	struct icc_node *n;
+	int ret, i;
+
+	src_qn = src->data;
+	provider = src->provider;
+	qp = to_msm8660_icc_provider(provider);
+
+	list_for_each_entry(n, &provider->nodes, node_list)
+		provider->aggregate(n, 0, n->avg_bw, n->peak_bw,
+				    &agg_avg, &agg_peak);
+
+	sum_bw = icc_units_to_bps(agg_avg);
+	max_peak_bw = icc_units_to_bps(agg_peak);
+
+	/*
+	 * Divide by the *fabric* bus width, not src_qn->buswidth: every
+	 * master on a given fabric shares the same hardware clock, so the
+	 * required clock rate is a single function of total bandwidth and
+	 * the fabric's bus width. Picking the bus width of whichever node
+	 * happened to trigger this update would make the rate oscillate
+	 * depending on which master called icc_set_bw() last.
+	 */
+	rate = max(sum_bw, max_peak_bw);
+	do_div(rate, qp->desc->bus_width);
+	/* Apply minimum floor to prevent bus starvation */
+	rate = max_t(u64, rate, MSM8660_FABRIC_MIN_RATE);
+	/*
+	 * Cap at INT_MAX in u64 space; min_t(u32, ...) would cast the u64
+	 * down to u32 BEFORE comparing, so a request above 4 GiB/s could
+	 * wrap to a near-zero value below INT_MAX and pass through, then
+	 * be programmed as a near-zero clock rate that effectively halts
+	 * the interconnect fabric.
+	 */
+	rate = min_t(u64, rate, INT_MAX);
+
+	if (qp->rate != rate) {
+		for (i = 0; i < qp->num_clks; i++) {
+			ret = clk_set_rate(qp->bus_clks[i].clk, rate);
+			if (ret) {
+				dev_err(provider->dev,
+					"%s clk_set_rate(%llu) error: %d\n",
+					qp->bus_clks[i].id, rate, ret);
+				/*
+				 * Bail without updating qp->rate so the next
+				 * icc_set call will retry the rate change
+				 * rather than treating it as cached-applied.
+				 */
+				return ret;
+			}
+		}
+		qp->rate = rate;
+	}
+
+	/* Send RPM fabric arbitration if available */
+	if (qp->rpm && qp->desc->rpm_resource >= 0)
+		msm8660_rpm_commit(qp);
+
+	return 0;
+}
+
+static int msm8660_get_bw(struct icc_node *node, u32 *avg, u32 *peak)
+{
+	*avg = 0;
+	*peak = 0;
+	return 0;
+}
+
+/*
+ * devm cleanup paired with clk_bulk_prepare_enable() in probe. Registered
+ * via devm_add_action_or_reset() so any subsequent probe error path
+ * (including -EPROBE_DEFER from msm8660_get_rpm()) reliably releases the
+ * prepare/enable reference rather than leaking it across the retry.
+ */
+static void msm8660_icc_clk_release(void *data)
+{
+	struct msm8660_icc_provider *qp = data;
+
+	clk_bulk_disable_unprepare(qp->num_clks, qp->bus_clks);
+}
+
+/*
+ * Look up the RPM that owns fabric arbitration writes.
+ *
+ * Returns NULL if the DT does not have a "qcom,rpm" phandle (in which
+ * case the caller silently drops RPM ARB and runs the fabric purely
+ * via clk_set_rate).
+ *
+ * Returns ERR_PTR(-EPROBE_DEFER) if the RPM device exists in DT but
+ * its driver has not finished probing yet, or if device_link_add()
+ * fails. The caller is expected to propagate this so the interconnect
+ * driver gets retried once the RPM is ready.
+ *
+ * On success returns the qcom_rpm handle and pins the RPM device
+ * lifetime to ours via a consumer-supplier device link, so the
+ * devres-allocated qcom_rpm cannot be freed while we still hold a
+ * pointer to it.
+ */
+static struct qcom_rpm *msm8660_get_rpm(struct device *dev)
+{
+	struct device_node *rpm_np;
+	struct platform_device *rpm_pdev;
+	struct device_link *link;
+	struct qcom_rpm *rpm;
+
+	rpm_np = of_parse_phandle(dev->of_node, "qcom,rpm", 0);
+	if (!rpm_np) {
+		dev_dbg(dev, "no qcom,rpm phandle, RPM ARB disabled\n");
+		return NULL;
+	}
+
+	rpm_pdev = of_find_device_by_node(rpm_np);
+	of_node_put(rpm_np);
+	if (!rpm_pdev) {
+		dev_dbg(dev, "RPM device not found yet, deferring probe\n");
+		return ERR_PTR(-EPROBE_DEFER);
+	}
+
+	/*
+	 * Pin the supplier BEFORE reading its drvdata. The device link
+	 * (MANAGED, the default state) prevents the RPM driver from being
+	 * unbound while we hold the link, which closes the window where a
+	 * concurrent unbind+rebind could free the qcom_rpm pointer between
+	 * dev_get_drvdata() and the link being established. If the link
+	 * cannot be added (e.g. supplier is in the process of being
+	 * removed) we defer and retry.
+	 */
+	link = device_link_add(dev, &rpm_pdev->dev,
+			       DL_FLAG_AUTOREMOVE_CONSUMER);
+	put_device(&rpm_pdev->dev);
+	if (!link) {
+		dev_warn(dev, "failed to add device link to RPM, deferring\n");
+		return ERR_PTR(-EPROBE_DEFER);
+	}
+
+	/*
+	 * Safe to read drvdata now: the device link pins the supplier so
+	 * it cannot be unbound until our consumer (this interconnect
+	 * provider) is unbound first.
+	 */
+	rpm = dev_get_drvdata(&rpm_pdev->dev);
+	if (!rpm) {
+		dev_dbg(dev, "RPM not ready, deferring probe\n");
+		device_link_remove(dev, &rpm_pdev->dev);
+		return ERR_PTR(-EPROBE_DEFER);
+	}
+
+	return rpm;
+}
+
+static int msm8660_icc_probe(struct platform_device *pdev)
+{
+	const struct msm8660_icc_desc *desc;
+	struct msm8660_icc_node * const *qnodes;
+	struct msm8660_icc_provider *qp;
+	struct device *dev = &pdev->dev;
+	struct icc_onecell_data *data;
+	struct icc_provider *provider;
+	struct icc_node *node;
+	size_t num_nodes, i;
+	int ret;
+
+	desc = of_device_get_match_data(dev);
+	if (!desc)
+		return -EINVAL;
+
+	qnodes = desc->nodes;
+	num_nodes = desc->num_nodes;
+
+	qp = devm_kzalloc(dev, sizeof(*qp), GFP_KERNEL);
+	if (!qp)
+		return -ENOMEM;
+
+	data = devm_kzalloc(dev, struct_size(data, nodes, num_nodes),
+			    GFP_KERNEL);
+	if (!data)
+		return -ENOMEM;
+	data->num_nodes = num_nodes;
+
+	qp->bus_clks = devm_kmemdup(dev, desc->bus_clks,
+				    desc->num_clks * sizeof(*desc->bus_clks),
+				    GFP_KERNEL);
+	if (!qp->bus_clks)
+		return -ENOMEM;
+
+	qp->num_clks = desc->num_clks;
+
+	/*
+	 * MSM8660 fabric clocks are managed by RPM firmware and may not be
+	 * available in mainline Linux yet. Once the clock provider exists,
+	 * we want to honour it; until then we run without per-fabric clock
+	 * scaling. The crucial part is that -EPROBE_DEFER means "the
+	 * provider exists but hasn't probed yet" and MUST be propagated so
+	 * we get retried; only other errors (genuine -ENOENT, etc.) get
+	 * downgraded to "no clocks, continue".
+	 */
+	ret = devm_clk_bulk_get_optional(dev, qp->num_clks, qp->bus_clks);
+	if (ret == -EPROBE_DEFER)
+		return ret;
+	if (ret) {
+		dev_warn(dev, "Failed to get bus clocks: %d (continuing without clock scaling)\n",
+			 ret);
+		qp->num_clks = 0;
+	}
+
+	if (qp->num_clks) {
+		ret = clk_bulk_prepare_enable(qp->num_clks, qp->bus_clks);
+		if (ret) {
+			dev_warn(dev, "Failed to enable bus clocks: %d\n", ret);
+			qp->num_clks = 0;
+		} else {
+			/*
+			 * Register the cleanup right after a successful
+			 * prepare_enable so any later -EPROBE_DEFER or other
+			 * probe error path (e.g. msm8660_get_rpm failing
+			 * with -EPROBE_DEFER below) does not leak a clock
+			 * prepare/enable reference across the retry.
+			 */
+			ret = devm_add_action_or_reset(dev,
+				msm8660_icc_clk_release, qp);
+			if (ret)
+				return ret;
+		}
+	}
+
+	/* Set up RPM fabric arbitration */
+	qp->desc = desc;
+	if (desc->rpm_resource >= 0) {
+		qp->rpm = msm8660_get_rpm(dev);
+		if (IS_ERR(qp->rpm))
+			return PTR_ERR(qp->rpm);
+		if (qp->rpm) {
+			int arb_size = desc->nmasters * desc->ntieredslaves;
+
+			qp->bwsum = devm_kcalloc(dev, desc->nslaves,
+						 sizeof(u16), GFP_KERNEL);
+			qp->arb = devm_kcalloc(dev, arb_size,
+					       sizeof(u16), GFP_KERNEL);
+			qp->rpm_buf = devm_kcalloc(dev, desc->rpm_buf_size,
+						   sizeof(u32), GFP_KERNEL);
+			if (!qp->bwsum || !qp->arb || !qp->rpm_buf) {
+				dev_warn(dev, "RPM buffer alloc failed, ARB disabled\n");
+				qp->rpm = NULL;
+			} else {
+				int rc;
+
+				dev_info(dev, "RPM fabric ARB enabled (%d masters, %d slaves, %d tiered)\n",
+					 desc->nmasters, desc->nslaves,
+					 desc->ntieredslaves);
+
+				/*
+				 * One-shot sleep-context vote of zero bandwidth.
+				 * Without an explicit SLEEP_STATE write, RPM has no
+				 * fabric bandwidth target for deep-sleep and may
+				 * keep the active vote applied indefinitely,
+				 * preventing DDR from dropping its rate when CPUs
+				 * power-collapse. The buffer is devm_kcalloc'd so
+				 * it is all-zero at this point — written before
+				 * any consumer can drive an active vote that would
+				 * dirty it.
+				 *
+				 * msm8660_rpm_commit() writes ACTIVE_STATE only;
+				 * SLEEP_STATE remains zero for the provider's
+				 * lifetime, so this vote does not need refreshing.
+				 */
+				rc = qcom_rpm_write(qp->rpm,
+						    QCOM_RPM_SLEEP_STATE,
+						    desc->rpm_resource,
+						    qp->rpm_buf,
+						    desc->rpm_buf_size);
+				if (rc)
+					dev_warn(dev, "RPM fabric sleep vote failed: %d\n",
+						 rc);
+			}
+		}
+	}
+
+	provider = &qp->provider;
+	provider->dev = dev;
+	provider->set = msm8660_icc_set;
+	provider->aggregate = msm8660_icc_aggregate;
+	provider->xlate = of_icc_xlate_onecell;
+	provider->data = data;
+	provider->get_bw = msm8660_get_bw;
+
+	icc_provider_init(provider);
+
+	for (i = 0; i < num_nodes; i++) {
+		size_t j;
+
+		if (!qnodes[i])
+			continue;
+
+		node = icc_node_create(qnodes[i]->id);
+		if (IS_ERR(node)) {
+			ret = PTR_ERR(node);
+			goto err_remove_nodes;
+		}
+
+		node->name = qnodes[i]->name;
+		node->data = qnodes[i];
+		icc_node_add(node, provider);
+
+		dev_dbg(dev, "registered node %s\n", node->name);
+
+		/* populate links */
+		for (j = 0; j < qnodes[i]->num_links; j++)
+			icc_link_create(node, qnodes[i]->links[j]);
+
+		data->nodes[i] = node;
+	}
+
+	ret = icc_provider_register(provider);
+	if (ret)
+		goto err_remove_nodes;
+
+	platform_set_drvdata(pdev, qp);
+
+	dev_info(dev, "MSM8660 interconnect provider registered\n");
+
+	return 0;
+
+err_remove_nodes:
+	icc_nodes_remove(provider);
+	/*
+	 * Do NOT call clk_bulk_disable_unprepare() here: the devm cleanup
+	 * action registered after clk_bulk_prepare_enable() will run
+	 * automatically when probe returns an error and devres unwinds.
+	 * Calling it manually would double-unprepare and corrupt the
+	 * clock-framework refcount.
+	 */
+	return ret;
+}
+
+static void msm8660_icc_remove(struct platform_device *pdev)
+{
+	struct msm8660_icc_provider *qp = platform_get_drvdata(pdev);
+
+	icc_provider_deregister(&qp->provider);
+	icc_nodes_remove(&qp->provider);
+	/* clk cleanup happens via devm_add_action_or_reset on remove. */
+}
+
+static const struct of_device_id msm8660_noc_of_match[] = {
+	{ .compatible = "qcom,msm8660-apps-fabric", .data = &msm8660_afab },
+	{ .compatible = "qcom,msm8660-system-fabric", .data = &msm8660_sfab },
+	{ .compatible = "qcom,msm8660-mmss-fabric", .data = &msm8660_mmfab },
+	{ .compatible = "qcom,msm8660-daytona-fabric", .data = &msm8660_dfab },
+	{ },
+};
+MODULE_DEVICE_TABLE(of, msm8660_noc_of_match);
+
+static struct platform_driver msm8660_noc_driver = {
+	.probe = msm8660_icc_probe,
+	.remove = msm8660_icc_remove,
+	.driver = {
+		.name = "qnoc-msm8660",
+		.of_match_table = msm8660_noc_of_match,
+		.sync_state = icc_sync_state,
+	},
+};
+/*
+ * Register the NOC provider at core_initcall, matching the mainline pattern
+ * used by newer Qualcomm SoCs (sm8450, glymur, qdu1000, sc8280xp, sm8750).
+ *
+ * Why not module_platform_driver (device_initcall)?  drivers/Makefile lists
+ * drivers/interconnect/ at position 189, *after* every ICC-consumer subdir
+ * (clk/ @40, soc/ @46, gpu/ @68, base/mfd/ @76, spmi/ @89, usb/ @106,
+ * i2c/ @116, mmc/ @133, remoteproc/ @158).  Within a single initcall level
+ * execution order = link order, so a device_initcall registration here runs
+ * *after* every consumer has already tried to probe.  Mainline relies on
+ * deferred-probe retry to recover from that, but in this tree some consumer
+ * (apcs-msm8660 + cpufreq cascade suspected) fails to recover within
+ * deferred_probe_timeout=5 and boot dies at the Tux splash with no rootfs.
+ * Empirically confirmed 2026-05-29 with module_platform_driver (commits
+ * 99275d8a8ae9 + ca35c591854c, reverted).
+ *
+ * icc_provider_register does not require icc_init to have run first --
+ * the framework's locks are statically DEFINE_MUTEX'd -- so registering
+ * the provider at core_initcall (before icc_init at subsys_initcall) is
+ * safe, same as mainline sm8450 etc.
+ */
+static int __init msm8660_noc_driver_init(void)
+{
+	return platform_driver_register(&msm8660_noc_driver);
+}
+core_initcall(msm8660_noc_driver_init);
+
+/*
+ * No module_exit: Kconfig is bool, the driver is built-in only, and
+ * unbind/unload paths are not exercised. core_initcall + module_exit
+ * mix badly anyway (you cannot unload something registered earlier
+ * than module_init level).
+ */
+
+MODULE_DESCRIPTION("Qualcomm MSM8x60 interconnect driver");
+MODULE_LICENSE("GPL v2");
diff --git a/include/dt-bindings/interconnect/qcom,msm8660.h b/include/dt-bindings/interconnect/qcom,msm8660.h
index c9ce3f5a5276..bac9fc423da2 100644
--- a/include/dt-bindings/interconnect/qcom,msm8660.h
+++ b/include/dt-bindings/interconnect/qcom,msm8660.h
@@ -4,7 +4,7 @@
  *
  * Copyright (c) 2026 Herman van Hazendonk <github.com@herrie.org>
  *
- * Based on webOS kernel msm_bus_board_8660.c
+ * Based on legacy vendor kernel msm_bus_board_8660.c
  * Copyright (c) 2010-2011, Code Aurora Forum. All rights reserved.
  */
 
@@ -79,11 +79,11 @@
  * Daytona Fabric (DFAB) - Peripheral bus
  *
  * DFAB connects slower peripherals (SDCC, ADM DMA) to the system fabric.
- * The webOS kernel managed DFAB bandwidth via voter clocks (dfab_sdc*_clk,
+ * The legacy vendor kernel managed DFAB bandwidth via voter clocks (dfab_sdc*_clk,
  * dfab_usb_hs_clk). In mainline, this is handled by the interconnect framework.
  *
  * USB HS is included as a DFAB voter for compatibility with the legacy clock
- * voting mechanism. The webOS kernel comment said: "if usb link is in sps
+ * voting mechanism. The legacy vendor kernel comment said: "if usb link is in sps
  * there is no need for usb pclk as daytona fabric clock will be used instead".
  * This keeps DFAB clock stable when USB is active.
  */
-- 
2.43.0