Re: [PATCH v4 5/6] soc: qcom: Add LLCC support for multi channel DDR

From: Konrad Dybcio
Date: Wed Jun 28 2023 - 07:13:41 EST


On 28.06.2023 10:52, Komal Bajaj wrote:
>
>
> On 6/23/2023 8:28 PM, Konrad Dybcio wrote:
>> On 23.06.2023 16:18, Komal Bajaj wrote:
>>> Add LLCC support for multi channel DDR configuration
>>> based on a feature register. Reading DDR channel
>>> confiuration uses nvmem framework, so select the
>>> dependency in Kconfig. Without this, there will be
>>> errors while building the driver with COMPILE_TEST only.
>>>
>>> Signed-off-by: Komal Bajaj <quic_kbajaj@xxxxxxxxxxx>
>>> ---
>>>   drivers/soc/qcom/Kconfig     |  2 ++
>>>   drivers/soc/qcom/llcc-qcom.c | 33 ++++++++++++++++++++++++++++++---
>>>   2 files changed, 32 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/soc/qcom/Kconfig b/drivers/soc/qcom/Kconfig
>>> index a491718f8064..cc9ad41c63aa 100644
>>> --- a/drivers/soc/qcom/Kconfig
>>> +++ b/drivers/soc/qcom/Kconfig
>>> @@ -64,6 +64,8 @@ config QCOM_LLCC
>>>       tristate "Qualcomm Technologies, Inc. LLCC driver"
>>>       depends on ARCH_QCOM || COMPILE_TEST
>>>       select REGMAP_MMIO
>>> +    select NVMEM
>>> +    select QCOM_SCM
>>>       help
>>>         Qualcomm Technologies, Inc. platform specific
>>>         Last Level Cache Controller(LLCC) driver for platforms such as,
>>> diff --git a/drivers/soc/qcom/llcc-qcom.c b/drivers/soc/qcom/llcc-qcom.c
>>> index 6cf373da5df9..3c29612da1c5 100644
>>> --- a/drivers/soc/qcom/llcc-qcom.c
>>> +++ b/drivers/soc/qcom/llcc-qcom.c
>>> @@ -12,6 +12,7 @@
>>>   #include <linux/kernel.h>
>>>   #include <linux/module.h>
>>>   #include <linux/mutex.h>
>>> +#include <linux/nvmem-consumer.h>
>>>   #include <linux/of.h>
>>>   #include <linux/of_device.h>
>>>   #include <linux/regmap.h>
>>> @@ -943,6 +944,19 @@ static int qcom_llcc_cfg_program(struct platform_device *pdev,
>>>       return ret;
>>>   }
>>>   +static int qcom_llcc_get_cfg_index(struct platform_device *pdev, u8 *cfg_index)
>>> +{
>>> +    int ret;
>>> +
>>> +    ret = nvmem_cell_read_u8(&pdev->dev, "multi-chan-ddr", cfg_index);
>>> +    if (ret == -ENOENT) {
>>> +        *cfg_index = 0;
>>> +        return 0;
>>> +    }
>>> +
>>> +    return ret;
>>> +}
>>> +
>>>   static int qcom_llcc_remove(struct platform_device *pdev)
>>>   {
>>>       /* Set the global pointer to a error code to avoid referencing it */
>>> @@ -975,11 +989,13 @@ static int qcom_llcc_probe(struct platform_device *pdev)
>>>       struct device *dev = &pdev->dev;
>>>       int ret, i;
>>>       struct platform_device *llcc_edac;
>>> -    const struct qcom_llcc_config *cfg;
>>> +    const struct qcom_llcc_config *cfg, *entry;
>>>       const struct llcc_slice_config *llcc_cfg;
>>>       u32 sz;
>>> +    u8 cfg_index;
>>>       u32 version;
>>>       struct regmap *regmap;
>>> +    u32 num_entries = 0;
>>>         drv_data = devm_kzalloc(dev, sizeof(*drv_data), GFP_KERNEL);
>>>       if (!drv_data) {
>>> @@ -1040,8 +1056,19 @@ static int qcom_llcc_probe(struct platform_device *pdev)
>>>         drv_data->version = version;
>>>   -    llcc_cfg = cfg[0]->sct_data;
>>> -    sz = cfg[0]->size;
>>> +    ret = qcom_llcc_get_cfg_index(pdev, &cfg_index);
>>> +    if (ret)
>>> +        goto err;
>>> +
>>
>>> +    for (entry = cfg; entry->sct_data; entry++, num_entries++)
>>> +        ;
>>> +    if (cfg_index >= num_entries || cfg_index < 0) {
>> cfg_index is an unsigned variable, it can never be < 0
>
> Okay, will remove this condition.
>
>>
>>> +        ret = -EINVAL;
>>> +        goto err;
>>> +    }
>>> +
>> if (cfg_index >= entry->size)? With that, you can also keep the config
>> entries non-0-terminated in the previous patch, saving a whole lot of RAM.
>
> entry->size represents the size of sct table whereas num_entries represents the number
> of sct tables that we can have. And by this check we are validating the value read from the
> fuse register. Am I understanding your comment correctly?
Oh you're right.

I still see room for improvement, though.

For example, you duplicate assigning need_llcc_cfg, reg_offset
and edac_reg_offset. You can add a new struct, like "sct_config" and add
a pointer to sct_config[] & the length of this array to qcom_llcc_config.

Konrad

>
>>
>> Konrad
>>> +    llcc_cfg = cfg[cfg_index].sct_data;
>>> +    sz = cfg[cfg_index].size;
>>>         for (i = 0; i < sz; i++)
>>>           if (llcc_cfg[i].slice_id > drv_data->max_slices)
>