From: Herman van Hazendonk <github.com@herrie.org>
To: Conor Dooley <conor+dt@kernel.org>,
	devicetree@vger.kernel.org,
	Georgi Djakov <djakov@kernel.org>,
	Krzysztof Kozlowski <krzk+dt@kernel.org>,
	linux-arm-msm@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	linux-pm@vger.kernel.org,
	Rob Herring <robh@kernel.org>
Subject: [PATCH v2 0/2] interconnect: qcom: add MSM8x60 NoC driver
Date: Sun, 31 May 2026 06:09:05 +0200
Message-ID: <cover.1780197411.git.github.com@herrie.org>
In-Reply-To: <cover.1780148149.git.github.com@herrie.org>
References: <cover.1780148149.git.github.com@herrie.org>
Precedence: bulk
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Message-ID: <20260531040905.ZAdI1TQarTaVwyV5T3xfExjsPFLtfkVkMq5lR39VWUA@z>

Hi all,

Self-review (with Sashiko AI assist) caught five real issues in v1's
NoC driver before the maintainer review cycle reached them, so
re-rolling promptly. v1:

  https://lore.kernel.org/linux-arm-msm/cover.1780148149.git.github.com@herrie.org/

v2 changes (driver patch only; the binding header is unchanged from v1):

  - MMFAB master port collision: mmfab_to_appss claimed port 2, but
    that port was already owned by mmfab_mas_adm1_port0. Per the
    fabric's port table the APPS_FAB gateway lives at port 11; fix
    the DEFINE_QNODE so ADM1 DMA's arbitration vote on MMFAB is no
    longer overwritten by the gateway.

  - msm8660_rpm_commit() double-converted bandwidth units: it called
    icc_units_to_bps() (which already returns bytes/s) and then
    immediately divided by 8 "bits -> bytes", asking RPM for 1/8 of
    the bandwidth that consumers had requested. Drop the spurious
    divide.

  - Fabric rate cache hoisted to provider scope and divided by the
    fabric's own bus width rather than the triggering node's local
    buswidth. A fabric has one shared hardware clock; using each
    master's local buswidth makes the requested clk rate oscillate
    depending on which master called icc_set_bw() last, and caching
    the result on the node lets a subsequent update skip clk_set_rate
    even though a different node has already moved the hardware. The
    new layout adds desc->bus_width per fabric (AFAB/SFAB/DFAB=8,
    MMFAB=16) and qp->rate as the single source of truth.

  - msm8660_get_rpm() rewritten:
      * returns ERR_PTR(-EPROBE_DEFER) when the RPM phandle resolves
        but the device hasn't probed yet, instead of returning NULL
        and silently disabling ARB forever;
      * adds a device_link to the RPM device so the devres-managed
        qcom_rpm struct can't be freed out from under us if the RPM
        driver is unbound at runtime (previously a use-after-free
        risk because we kept a pointer past put_device()).
    Probe propagates IS_ERR(qp->rpm) up so the framework retries.

  - devm_clk_bulk_get_optional() now distinguishes -EPROBE_DEFER
    (propagated) from real "no clocks" (continue without scaling).
    Previously every error was swallowed, which permanently disabled
    clock scaling if the icc driver happened to probe before the
    clock provider - which it usually does, given the core_initcall
    ordering.

No functional change intended on the working paths; this is purely
correctness work.

Thanks,
Herman

Herman van Hazendonk (2):
  dt-bindings: interconnect: qcom: add msm8660 fabric IDs
  interconnect: qcom: add MSM8x60 NoC driver

 drivers/interconnect/qcom/Kconfig             |   10 +
 drivers/interconnect/qcom/Makefile            |    2 +
 drivers/interconnect/qcom/msm8660.c           | 1069 +++++++++++++++++
 .../dt-bindings/interconnect/qcom,msm8660.h   |  156 +++
 4 files changed, 1237 insertions(+)
 create mode 100644 drivers/interconnect/qcom/msm8660.c
 create mode 100644 include/dt-bindings/interconnect/qcom,msm8660.h

-- 
2.43.0