Re: [regression 4.20-rc1, bisected] cpufreq fails to register on db820c

[Srinivas Kandagatla]

Hi Niklas,

thanks for reporting this..

I found the issue, DT support seems to be totally broken in NVMEM.
Will send this fix once you can confirm it works for you!

Can you try this patch!
---------------------->cut<--------------------------

Author: Srinivas Kandagatla <srinivas.kandagatla@xxxxxxxxxx>
Date:   Tue Nov 6 14:32:30 2018 +0000

    nvmem: core: fix of_nvmem_cell_get()

    NVMEM DT support seems to be totally broken after
    e888d445ac33 ("nvmem: resolve cells from DT at registration time")
    Fix this!

    Fixes: e888d445ac33 ("nvmem: resolve cells from DT at registration 
time")
    Reported-by: Niklas Cassel <niklas.cassel@xxxxxxxxxx>
    Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@xxxxxxxxxx>

diff --git a/drivers/nvmem/core.c b/drivers/nvmem/core.c
index 9b18ce90f907..7ae8072f56ae 100644
--- a/drivers/nvmem/core.c
+++ b/drivers/nvmem/core.c
@@ -44,6 +44,7 @@ struct nvmem_cell {
        int                     bytes;
        int                     bit_offset;
        int                     nbits;
+       struct device_node      *np;
        struct nvmem_device     *nvmem;
        struct list_head        node;
 };
@@ -530,6 +531,7 @@ static int nvmem_add_cells_from_of(struct 
nvmem_device *nvmem)
                        return -ENOMEM;

                cell->nvmem = nvmem;
+               cell->np = child;
                cell->offset = be32_to_cpup(addr++);
                cell->bytes = be32_to_cpup(addr);
                cell->name = kasprintf(GFP_KERNEL, "%pOFn", child);
@@ -960,14 +962,13 @@ nvmem_cell_get_from_lookup(struct device *dev, 
const char *con_id)

 #if IS_ENABLED(CONFIG_OF)
 static struct nvmem_cell *
-nvmem_find_cell_by_index(struct nvmem_device *nvmem, int index)
+nvmem_find_cell_by_node(struct nvmem_device *nvmem, struct device_node *np)
 {
        struct nvmem_cell *cell = NULL;
-       int i = 0;

        mutex_lock(&nvmem_mutex);
        list_for_each_entry(cell, &nvmem->cells, node) {
-               if (index == i++)
+               if (np == cell->np)
                        break;
        }
        mutex_unlock(&nvmem_mutex);
@@ -1001,7 +1002,6 @@ struct nvmem_cell *of_nvmem_cell_get(struct 
device_node *np, const char *id)
        cell_np = of_parse_phandle(np, "nvmem-cells", index);
        if (!cell_np)
                return ERR_PTR(-EINVAL);
-
        nvmem_np = of_get_next_parent(cell_np);
        if (!nvmem_np)
                return ERR_PTR(-EINVAL);
@@ -1011,7 +1011,7 @@ struct nvmem_cell *of_nvmem_cell_get(struct 
device_node *np, const char *id)
        if (IS_ERR(nvmem))
                return ERR_CAST(nvmem);

-       cell = nvmem_find_cell_by_index(nvmem, index);
+       cell = nvmem_find_cell_by_node(nvmem, cell_np);
        if (!cell) {
                __nvmem_device_put(nvmem);
                return ERR_PTR(-ENOENT);

---------------------->cut<--------------------------
On 06/11/18 01:29, Niklas Cassel wrote:
> Hello Bartosz,
>
> I've run into a regression on v4.20-rc1, where cpufreq on dragonboard 820c
> fails to register.
>
> I've bisected the issue to:
> e888d445ac33 ("nvmem: resolve cells from DT at registration time")
>
> If I revert that patch, cpufreq starts working on dragonboard 820c again.
>
>
> Before your patch:
> [    5.052985] qcom_cpufreq_kryo_probe:110 speedbin_nvmem: ffff8000d4d51e00
> [    5.053133] qcom_cpufreq_kryo_probe:120 speedbin: ffff8000d4d51d00
> [    5.058895] qcom_cpufreq_kryo_probe:121 *speedbin: 0
>
>
> After your patch:
> [    4.791664] qcom_cpufreq_kryo_probe:110 speedbin_nvmem: ffff8000d472ed00
> [    4.791815] qcom_cpufreq_kryo_probe:120 speedbin: ffff8000d451b300
> [    4.797575] qcom_cpufreq_kryo_probe:121 *speedbin: 12
>
>
> The prints are debug prints that I've added locally:
>
> --- a/drivers/cpufreq/qcom-cpufreq-kryo.c
> +++ b/drivers/cpufreq/qcom-cpufreq-kryo.c
> @@ -107,6 +107,7 @@ static int qcom_cpufreq_kryo_probe(struct platform_device *pdev)
>          }
>   
>          speedbin_nvmem = of_nvmem_cell_get(np, NULL);
> +       pr_err("%s:%d speedbin_nvmem: %lx\n", __func__, __LINE__, speedbin_nvmem);
>          of_node_put(np);
>          if (IS_ERR(speedbin_nvmem)) {
>                  if (PTR_ERR(speedbin_nvmem) != -EPROBE_DEFER)
> @@ -116,6 +117,8 @@ static int qcom_cpufreq_kryo_probe(struct platform_device *pdev)
>          }
>   
>          speedbin = nvmem_cell_read(speedbin_nvmem, &len);
> +       pr_err("%s:%d speedbin: %lx\n", __func__, __LINE__, speedbin);
> +       pr_err("%s:%d *speedbin: %u\n", __func__, __LINE__, (unsigned int)(*speedbin));
>          nvmem_cell_put(speedbin_nvmem);
>          if (IS_ERR(speedbin))
>                  return PTR_ERR(speedbin);
>
>
> So it appears that the read from nvmem_cell_read()
> gets another value after your commit.
>
> This value is then given to dev_pm_opp_set_supported_hw(),
> however, since this new value (12) from nvmem_cell_read()
> doesn't have any matching opp-supported-hw in device tree,
> cpufreq fails to register.
>
> This error only happens when we read this new incorrect value
> (12 instead of 0).
> Any idea why nvmem_cell_read() returns a new value for the
> same efuse?
>
>
> Kind regards,
> Niklas