Re: [PATCH] soc: qcom: ice: Avoid probe deferring for un-supported ICE feature

[Manivannan Sadhasivam]

On Mon, Feb 02, 2026 at 02:02:28PM +0530, Sumit Garg wrote:
> On Fri, Jan 30, 2026 at 01:36:28PM +0100, Konrad Dybcio wrote:
> > On 1/30/26 1:29 PM, Sumit Garg wrote:
> > > On Fri, Jan 30, 2026 at 12:55:53PM +0100, Konrad Dybcio wrote:
> > >> On 1/30/26 12:36 PM, Sumit Garg wrote:
> > >>> On Fri, Jan 30, 2026 at 10:59:18AM +0100, Konrad Dybcio wrote:
> > >>>> On 1/30/26 10:52 AM, Sumit Garg wrote:
> > >>>>> On Fri, Jan 30, 2026 at 10:34:26AM +0100, Konrad Dybcio wrote:
> > >>>>>> On 1/30/26 10:11 AM, Sumit Garg wrote:
> > >>>>>>> From: Sumit Garg <sumit.garg@xxxxxxxxxxxxxxxx>
> > >>
> > >> [...]
> > >>
> > >>>>> Since qcom,ufs depends on qcom,ice via a phandle, so isn't the probe
> > >>>>> orderering automatically taken care off? Or that isn't the case here?
> > >>>>
> > >>>> No, that's guaranteed by devlink only with certain properties.
> > >>>
> > >>> Okay I see. The other alternate solution I can come up is following to
> > >>> keep the deferred probing intact. Let me know your thoughts on this:
> > >>>
> > >>> diff --git a/drivers/soc/qcom/ice.c b/drivers/soc/qcom/ice.c
> > >>> index ab9586b8caf5..76bf9f94fbaf 100644
> > >>> --- a/drivers/soc/qcom/ice.c
> > >>> +++ b/drivers/soc/qcom/ice.c
> > >>> @@ -559,7 +559,7 @@ static struct qcom_ice *qcom_ice_create(struct device *dev,
> > >>>
> > >>>         if (!qcom_scm_ice_available()) {
> > >>>                 dev_warn(dev, "ICE SCM interface not found\n");
> > >>> -               return NULL;
> > >>> +               return ERR_PTR(-ENODEV);
> > >>>         }
> > >>>
> > >>>         engine = devm_kzalloc(dev, sizeof(*engine), GFP_KERNEL);
> > >>> @@ -648,11 +648,14 @@ static struct qcom_ice *of_qcom_ice_get(struct device *dev)
> > >>>         }
> > >>>
> > >>>         ice = platform_get_drvdata(pdev);
> > >>> -       if (!ice) {
> > >>> +       if (IS_ERR_OR_NULL(ice)) {
> > >>>                 dev_err(dev, "Cannot get ice instance from %s\n",
> > >>>                         dev_name(&pdev->dev));
> > >>>                 platform_device_put(pdev);
> > >>> -               return NULL;
> > >>> +               if (PTR_ERR(ice) == -ENODEV)
> > >>> +                       return NULL;
> > >>> +               else
> > >>> +                       return ERR_PTR(-EPROBE_DEFER);
> > >>>         }
> > >>>
> > >>>         link = device_link_add(dev, &pdev->dev, DL_FLAG_AUTOREMOVE_SUPPLIER);
> > >>> @@ -726,7 +729,7 @@ static int qcom_ice_probe(struct platform_device *pdev)
> > >>>         }
> > >>>
> > >>>         engine = qcom_ice_create(&pdev->dev, base);
> > >>> -       if (IS_ERR(engine))
> > >>> +       if (IS_ERR(engine) && PTR_ERR(engine) != -ENODEV)
> > >>>                 return PTR_ERR(engine);
> > >>>
> > >>>         platform_set_drvdata(pdev, engine);
> > >>
> > >> This looks more robust. Although the UFS and MMC drivers today check
> > >> for EOPNOTSUPP, so perhaps throwing that would be even better
> > > 
> > > Sure, I can use that error code instead.
> > > 
> > >>
> > >>>
> > >>>> In this case though, I think it could make sense to add it to the
> > >>>> "suppliers" list in drivers/of/property.c.
> > >>>>
> > >>>> I don't know if vendors adding their custom properties there is a
> > >>>> pattern that +Rob will be happy about though..
> > >>>
> > >>> Not sure if that's a shorter path as I would like to see fix for this
> > >>> issue backported as well.
> > >>>
> > >>> Aside, not sure how much stable ICE feature itself is as I got following
> > >>> crash with QLI boot firmware on Kodiak:
> > >>>
> > >>> [    5.172970] SError Interrupt on CPU6, code 0x00000000be000000 -- SError
> > >>> [    5.172986] CPU: 6 UID: 0 PID: 241 Comm: (udev-worker) Tainted: G   M                6.19.0-rc5-next-20260115-gc1a0fee87a05 #9 PREEMPT 
> > >>> [    5.172996] Tainted: [M]=MACHINE_CHECK
> > >>> [    5.172999] Hardware name: Qualcomm Technologies, Inc. Robotics RB3gen2 (DT)
> > >>> [    5.173003] pstate: 80400005 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
> > >>> [    5.173010] pc : qcom_ice_create.part.0+0x6c/0x24c [qcom_ice]
> > >>> [    5.173024] lr : qcom_ice_create.part.0+0xe4/0x24c [qcom_ice]
> > >>
> > >> Could you please decode the pc value with ./scripts/faddr2line?
> > >> My compiler produces different output
> > >>
> > >> ./scripts/faddr2line vmlinux(or path to .ko) <symbol_name>
> > >>
> > > 
> > > This points at:
> > > 
> > > static bool qcom_ice_check_supported(struct qcom_ice *ice)
> > > {
> > > -->   u32 regval = qcom_ice_readl(ice, QCOM_ICE_REG_VERSION);
> > > 
> > > $ ./scripts/faddr2line ./drivers/soc/qcom/qcom_ice.ko qcom_ice_create.part.0+0x6c/0x24c
> > > qcom_ice_create.part.0+0x6c/0x24c:
> > > readl at /home/sumit/build/upstream/linux/./include/asm-generic/io.h:232 (discriminator 1)
> > > (inlined by) qcom_ice_check_supported at /home/sumit/build/upstream/linux/drivers/soc/qcom/ice.c:118 (discriminator 1)
> > > (inlined by) qcom_ice_create at /home/sumit/build/upstream/linux/drivers/soc/qcom/ice.c:587 (discriminator 1)
> > > 
> > > To me it looks like an issue related to access control policy. Note that
> > > it's the Gunyah based stack only where this issue is seen.
> > 
> > Please first check if that's not just a lack of power, because apparently
> > we never guaranteed that would be present, see:
> > 
> > https://lore.kernel.org/linux-arm-msm/20260123-qcom_ice_power_and_clk_vote-v1-0-e9059776f85c@xxxxxxxxxxxxxxxx/
> > 
> > https://lore.kernel.org/linux-arm-msm/20260123-enable-ufs-ice-clock-scaling-v3-0-d0d8532abd98@xxxxxxxxxxxxxxxx/
> 
> Thanks for these references. TBH, if there are some known issues with
> ICE then it shouldn't be enabled in the common defconfig. BTW, it looks
> like Mani is already looking into proper fix for this issue as per
> offline communication.
> 

The existing ICE based platform driver probe and of_qcom_ice_get() are quite
racy. I'm working on a patch that removes registering the ICE driver as a
platform driver and moves qcom_ice_create() to of_qcom_ice_get() with proper
reference counting.

Will post the patch soon.

- Mani

-- 
மணிவண்ணன் சதாசிவம்