Skip to contentSkip to conversation

arm64: dts: qcom: glymur: use polling mode for SCMI transfers

PATCH In review Jesse Casco 3 participants 5 messages Aug 8 – Nov 10

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 arm64: dts: qcom: glymur: use polling mode for SCMI transfers

On glymur the CPUCP firmware writes correct SCMI replies into the shared memory but never rings the mailbox doorbell for them. Every transfer that waits for the completion interrupt therefore times out, scmi-cpufreq fails to probe with -110, and the machine ends up with no cpufreq at all -- all CPUs pinned at their boot OPP.

Measured on an ASUS Zenbook A16 (UX3607OA) through the SCMI RAW debugfs
interface: a Performance protocol (0x13) PROTOCOL_ATTRIBUTES request returns status 0 and reports three performance domains when the transport polls for completion, and never completes when it waits for the doorbell. The mailbox interrupt count stays at zero throughout, so the reply is being produced but never signalled.

Mark the transport arm,no-completion-irq so it polls. With this applied scmi-cpufreq probes, three performance domains appear, and cpu0/cpufreq/scaling_driver reads "scmi".

The property is documented in Documentation/devicetree/bindings/firmware/arm,scmi.yaml, is consumed by drivers/firmware/arm_scmi/driver.c, and is already used in-tree by renesas/r8a78000-ironhide.

Assisted-by Anthropic:Claude-Opus-5
---
 arch/arm64/boot/dts/qcom/glymur.dtsi | 6 ++++++
 1 file changed, 6 insertions(+)
arch/arm64/boot/dts/qcom/glymur.dtsi +6 -0
index f04058d2e..b9db81a19 100644
@@ -412,6 +412,12 @@ scmi {
 			mbox-names = "tx", "rx";
 			shmem = <&cpu_scp_lpri1>, <&cpu_scp_lpri0>;
 
+			/*
+			 * The CPUCP writes SCMI replies into the shmem but
+			 * never rings the completion doorbell, so poll.
+			 */
+			arm,no-completion-irq;
+
 			#address-cells = <1>;
 			#size-cells = <0>;
 

base-commit: 6b8c8af514d739d0335f5579b585e02babe8a727 --
2.55.0

#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Dangling pointer in xfer->async_done leads to stack Use-After-Free or NULL pointer dereference on late delayed response. --

--- Patch [1]: [PATCH] arm64: dts: qcom: glymur: use polling mode for SCMI transfers --- commit 7c8d066ac6f378514463d4f6eb7b970a95b13e88
Author: Jesse Casco <[email protected]>

arm64: dts: qcom: glymur: use polling mode for SCMI transfers

This commit marks the SCMI transport on qcom glymur to use polling mode. This works around the CPUCP firmware writing replies to shared memory but failing to ring the completion doorbell. Using arm,no-completion-irq ensures the scmi-cpufreq driver successfully probes.

Quoted 7 lines
arch/arm64/boot/dts/qcom/glymur.dtsi +0 -0
index f04058d2e0891..b9db81a19d186 100644
@@ -412,6 +412,12 @@ scmi {
 			mbox-names = "tx", "rx";
 			shmem = <&cpu_scp_lpri1>, <&cpu_scp_lpri0>;

arch/arm64/boot/dts/qcom/glymur.dtsi scmi { · line 423

Quoted 10 lines
arch/arm64/boot/dts/qcom/glymur.dtsi +6 -0
 
+			/*
+			 * The CPUCP writes SCMI replies into the shmem but
+			 * never rings the completion doorbell, so poll.
+			 */
+			arm,no-completion-irq;
+
 			#address-cells = <1>;
 			#size-cells = <0>;
 

[Severity: Critical]
This isn't a bug introduced by this patch, but could a late delayed response trigger a data race causing a use-after-free on the async_done stack frame or a NULL pointer dereference?

In drivers/firmware/arm_scmi/driver.c:do_xfer_with_response(), if a delayed response takes longer than SCMI_MAX_RESPONSE_TIMEOUT, wait_for_completion_timeout() times out.

The function sets xfer->async_done = NULL without any synchronization and returns, destroying the async_response stack frame:

drivers/firmware/arm_scmi/driver.c:do_xfer_with_response() {
    ...
	if (!ret) {
		if (!wait_for_completion_timeout(xfer->async_done, timeout)) {
			dev_err(ph->dev,
				"timed out in delayed resp(caller: %pS)\n",
				(void *)_RET_IP_);
			ret = -ETIMEDOUT;
		} else if (xfer->hdr.status) {
    ...
	}

	xfer->async_done = NULL;
    ...
}

However, the transfer remains in the pending hash table until the caller invokes xfer_put().

If the delayed response interrupt fires in this window, scmi_handle_response() can successfully acquire the xfer and unconditionally access the unprotected async_done pointer:

drivers/firmware/arm_scmi/driver.c:scmi_handle_response() {
    ...
	if (xfer->hdr.type == MSG_TYPE_DELAYED_RESP) {
		scmi_clear_channel(info, cinfo);
		complete(xfer->async_done);
		scmi_inc_count(info->dbg, DELAYED_RESPONSE_OK);
	} else {
    ...
}

Could this lead to stack memory corruption if the async_done stack frame is already destroyed, or a kernel panic if async_done is read after being set to NULL?

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Jesse Casco
Quoted 5 lines · Jesse Casco

On glymur the CPUCP firmware writes correct SCMI replies into the shared memory but never rings the mailbox doorbell for them. Every transfer that waits for the completion interrupt therefore times out, scmi-cpufreq fails to probe with -110, and the machine ends up with no cpufreq at all -- all CPUs pinned at their boot OPP.

I'm not seeing this issue, but I heard others hit it, possibly due to some timing bug. If you set your kernel config to build all drivers in (i.e. change =m to =y), do you still face this?

Konrad

#3 lore raw

Tested. Everything SCMI-related was already =y except the mailbox controller itself, qcom_cpucp_mbox, which was =m. Built that in, pulled arm,no-completion-irq back out...No CMD timeout, no -110, cpufreq comes up fine. Looks like exactly the timing bug you mentioned.

Keeping arm,no-completion-irq for daily use since it doesn't depend on build config, but happy to test more if useful.

Jesse

Quoted 13 lines · Konrad Dybcio
Quoted 5 lines · Jesse Casco

On glymur the CPUCP firmware writes correct SCMI replies into the shared memory but never rings the mailbox doorbell for them. Every transfer that waits for the completion interrupt therefore times out, scmi-cpufreq fails to probe with -110, and the machine ends up with no cpufreq at all -- all CPUs pinned at their boot OPP.

I'm not seeing this issue, but I heard others hit it, possibly due to some timing bug. If you set your kernel config to build all drivers in (i.e. change =m to =y), do you still face this?

Konrad

Signature · 1 line
Jesse Casco
#4 lore raw

Hi Konrad, Sibi,

I tracked down the timing issue on the ASUS Zenbook A16 (7.4-rc1 based, arm,no-completion-irq removed). It depends on whether PSCI cpuidle is already running when SCMI sends its first PERF (0x13) message:

- CPUCP_MBOX=y: first PERF message at 0.37 s, before cpuidle starts. PERF works for the whole boot, cpuidle included.
- CPUCP_MBOX=m (arm64 defconfig): first PERF message at 1.1-1.4 s, after cpuidle starts. No PERF reply ever gets a doorbell, so scmi-cpufreq and scmi-perf-domain fail with -110. BASE replies still ring.
- CPUCP_MBOX=m with cpuidle.off=1: PERF works.

In the failing case the firmware still writes the PERF replies (the same request through SCMI RAW message_poll returns status 0), but RX_MBOX_STAT never gets a bit, with EN=0x3 and MAP=~0, the same as in the working case. So it looks like the CPUCP decides at the first PERF exchange whether to ring for PERF, rather than a mailbox driver bug.

Please consider the DT patch withdrawn. Sibi, is this a known CPUCP firmware behavior? In the meantime I could send an arm64 defconfig change building QCOM_CPUCP_MBOX in, so defconfig kernels get cpufreq on this machine, if that's acceptable.

Jesse