Re: [PATCH v5] driver core: enforce device_lock for driver_match_device()

["Danilo Krummrich" <dakr@kernel.org>]

On Wed Jan 21, 2026 at 2:03 PM CET, Robin Murphy wrote:
> On 2026-01-21 11:02 am, Danilo Krummrich wrote:
>> On Wed Jan 21, 2026 at 11:40 AM CET, Danilo Krummrich wrote:
>>> So, the problem is that in the callstack of the arm-smmu driver's (a platform
>>> driver) probe() function, the QCOM specific code (through arm_smmu_impl_init())
>>> registers another platform driver. Since we are still in probe() of arm-smmu the
>>> call to platform_driver_register() happens with the device lock of the arm-smmu
>>> platform device held.
>>>
>>> platform_driver_register() eventually results in driver_attach() which iterates
>>> over all the devices of a bus. Since the device we are probing and the driver we
>>> are registering are for the same bus (i.e. the platform bus) it can now happen
>>> that by chance that we also match the exact same device that is currently probed
>>> again. And since we take the device lock for matching now, we actually take the
>>> same lock twice.
>>>
>>> Now, we could avoid this by not matching bound devices, but we check this
>>> through dev->driver while holding the device lock, so that doesn't help.
>>>
>>> But on the other hand, I don't see any reason why a driver would call
>>> platform_driver_register() from probe() in the first place. I think drivers
>>> should not do that and instead just register the driver through a normal
>>> initcall.
>>>
>>> (If, however, it turns out that registering drivers from probe() is something we
>>> really need for some reason, it is probably best to drop the patch and don't
>>> make any guarantees about whether match() is called with the device lock held or
>>> not.
>>>
>>> Consequently, driver_override must be protected with a separate lock (which
>>> would be the cleaner solution in any case).)
>> 
>> I assume that this should resolve the problem (unless there are more drivers
>> that register drivers in probe()):
>> 
>> diff --git a/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c b/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
>> index 573085349df3..9bb793efc35f 100644
>> --- a/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
>> +++ b/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
>> @@ -774,10 +774,6 @@ struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu)
>>   {
>>          const struct device_node *np = smmu->dev->of_node;
>>          const struct of_device_id *match;
>> -       static u8 tbu_registered;
>> -
>> -       if (!tbu_registered++)
>> -               platform_driver_register(&qcom_smmu_tbu_driver);
>> 
>>   #ifdef CONFIG_ACPI
>>          if (np == NULL) {
>> @@ -802,3 +798,5 @@ struct arm_smmu_device *qcom_smmu_impl_init(struct arm_smmu_device *smmu)
>> 
>>          return smmu;
>>   }
>> +
>> +builtin_platform_driver(qcom_smmu_tbu_driver);
>> 
>> @qcom maintainers: I'm aware of commit 0b4eeee2876f ("iommu/arm-smmu-qcom:
>> Register the TBU driver in qcom_smmu_impl_init"), but I think the above patch
>> should work fine as it is still *not only* registered when
>> CONFIG_ARM_SMMU_QCOM_DEBUG?
>
> In principle there should be nothing wrong with registering the driver 
> unconditionally - that existing tbu_registered logic looks racy in the 
> face of async_probe anyway - however I don't think the *_platform_driver 
> macros will work here, as this all gets combined into arm_smmu.ko 
> wherein ending up with multiple module_init declarations breaks the build.
>
> (Please do double-check all the build permutations of ARM_SMMU, 
> ARM_SMMU_QCOM and ARM_SMMU_QCOM_DEBUG)

Indeed, I accounted for this in the final patch I sent out, thanks!