by Philipp Zabel

[permalink] [raw]

Subject: Re: [PATCH v2 2/4] reset: Instantiate reset GPIO controller for shared reset-gpios

On Di, 2024-01-09 at 11:59 +0100, Krzysztof Kozlowski wrote:
> On 08/01/2024 13:17, Philipp Zabel wrote:
> > On Fr, 2024-01-05 at 16:59 +0100, Krzysztof Kozlowski wrote:
> > Is that true?
>
> It should be true and my tests confirmed it.
>
> The code below looks like overwrites of_phandle_args so
> > that only one reset-gpio device is spawned for each gpio node.
>
> The code will iterate over list of of_node and of_phandle_args and
> compare them with __reset_gpios_args_match(). If all match: do not
> create new platform device.
>
> If they do not match, e.g. ACTIVE_LOW -> ACTIVE_HIGH, create new
> platform device. This will be the second device for the same GPIO.
> Probing of that device in reset-gpio driver will fail:
>
> [ 19.198775] reset-gpio reset-gpio.2.auto: error -EBUSY: Could not get
> reset gpios
>
> because GPIO is used by reset-gpio.1.auto already.

Thank you for the clarification.
I only understood later in the mail and didn't update this properly.

> > > + /* Not freed in normal path, persisent subsyst data */
> > > + rgpio_dev = kzalloc(sizeof(*rgpio_dev), GFP_KERNEL);
> >
> > Since this is persistent, instead of letting the reset-gpio driver call
> > of_parse_phandle_with_args() again, this could be passed in via
> > platform data. Is there a reason not to do that instead?
>
> We can pass it as read only platform data, but we cannot pass the
> ownership. This is associated with registered platform device, not with
> bound one device->driver.
>
> Imagine case:
> 1. modprobe reset-gpio,
> 2. Driver is bound to the device,
> 3. unbind (echo > unbind)
> 4. rmmod
> 5. goto 1

Keeping ownership on the list is fine, the reset-gpio driver makes its
own copy of of_phandle_args anyway. I was just wondering whether it
could make this copy from platform data instead of from the
of_parse_phandle_with_args() return value.

[...]
> >
> > > @@ -839,21 +960,50 @@ __of_reset_control_get(struct device_node *node, const char *id, int index,
> > > index, &args);
> > > if (ret == -EINVAL)
> > > return ERR_PTR(ret);
> > > - if (ret)
> > > - return optional ? NULL : ERR_PTR(ret);
> > > + if (ret) {
> > > + /*
> > > + * There can be only one reset-gpio for regular devices, so
> > > + * don't bother with GPIO index.
> > > + */
> >
> > I don't understand this comment. The GPIO index should be checked as
> > part of __reset_gpios_args_match(), or should it not?
>
> This and earlier comment are result of a bit hacky approach to the
> problem: how to find reset controllers for that GPIO?
>
> The point is that our reset gpio controller has only 1 reset, thus
> of_reset_n_cells=1. However args_count from of_parse_handle is >0, which
> later is compared in reset core:
>
> https://elixir.bootlin.com/linux/latest/source/drivers/reset/core.c#L859
>
> That part we need to match.
>
> I could make the reset-gpio driver to have of_reset_n_cells=2, but what
> would be the point? The rest of the cells are not really relevant,
> because you cannot refer to this reset gpio controller with any other
> arguments.
>
> To remind: my solution spawns one reset-gpio controller for one GPIO.

Thank you. I think we could also just make that check

if (WARN_ON(!rcdev->of_args && ...))

instead and skip the of_xlate call in that case (or implement of_xlate
in the reset-gpio driver to make this more explicit).

> >
> > > + ret = of_parse_phandle_with_args(node, "reset-gpios", "#gpio-cells",
> > > + 0, &args);
> > > + if (ret)
> > > + return optional ? NULL : ERR_PTR(ret);
> > >
> > > - mutex_lock(&reset_list_mutex);
> > > - rcdev = NULL;
> > > - list_for_each_entry(r, &reset_controller_list, list) {
> > > - if (args.np == r->of_node) {
> > > - rcdev = r;
> > > - break;
> > > - }
> > > + gpio_fallback = true;
> >
> > Is there a reason not just call __reset_add_reset_gpio_device() here?
> > With that, there should be no need to call __reset_find_rcdev() twice.
>
> Hm, could be, although not sure if code would be simpler.
>
> This entire function handles two cases:
> 1. Get normal reset controller ("resets" OF property),
> 2. If above fails, get reset-gpio controller ("reset-gpios" OF property)
>
> Therefore the entire solution is following approach:
> 1. of_parse_phandle(resets)
> 1b. error? Then of_parse_phandle(reset-gpios)
> 2. Find reset-controller based on any of above phandles.
> 3. error? Check if we created reset-gpios platform device. If not:
> create new reset-gpios platform device/
> 3b. Platform device could probe, so lookup again for reset controller or
> defer probe.
>
> What type of flow do you propose?

I propose to reorder after parsing the phandles: check/create the gpio
platform device right after parsing the gpio handle. Only then lock
reset_list_mutex look for the rcdev.

1. of_parse_phandle(resets)
1b. error? Then of_parse_phandle(reset-gpios)
2b. gpio? Then check if we created reset-gpios platform device. If not:
create new reset-gpios platform device/, defer if probe failed
3. Lock reset_list_mutex, find reset-controller based on any of above
phandles.

>
> >
> > > }
> > >
> > > + mutex_lock(&reset_list_mutex);
> > > + rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
> >
> > This gets called with args as parsed. If there is a match, this will
> > overwrite args (in the gpio_fallback case) and return NULL.
>
> Overwrite not complete. It will only overwrite args_count and return a
> valid rcdev.
> I do not see overwriting in case of returning NULL.

Sorry, I meant to write

"This gets called with args as parsed. If there is a match, this will
overwrite args (in the gpio_fallback case) _or_ return NULL."

at least at the end, when I understood the following.

> >
> > > +
> > > if (!rcdev) {

So in this non-NULL branch there was no overwriting.

> > > - rstc = ERR_PTR(-EPROBE_DEFER);
> > > - goto out;
> > > + if (gpio_fallback) {
> > > + /*
> > > + * Registering reset-gpio device might cause immediate
> > > + * bind, thus taking reset_list_mutex lock via
> > > + * reset_controller_register().
> > > + */
> > > + mutex_unlock(&reset_list_mutex);
> > > + ret = __reset_add_reset_gpio_device(node, &args);
> >
> > So this will also be called with args as parsed.
> >
> > > + mutex_lock(&reset_list_mutex);
> > > + if (ret) {
> > > + rstc = ERR_PTR(ret);
> > > + goto out;
> > > + }
> > > + /*
> > > + * Success: reset-gpio could probe immediately, so
> > > + * re-check the lookup.
> > > + */
> > > + rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
> >
> > And this will again be called with args as parsed and overwrite args
> > again.>
> > > + if (!rcdev) {
> > > + rstc = ERR_PTR(-EPROBE_DEFER);
> > > + goto out;
> > > + }
> > > + /* Success, rcdev is valid thus do not bail out */
> > > + } else {
> > > + rstc = ERR_PTR(-EPROBE_DEFER);
> > > + goto out;
> > > + }
> > > }
> >
> > So at this point args is overwritten in the gpio_fallback case. I would
> > find it much clearer to just overwrite args here and make the first
> > parameter to __reset_find_rcdev() const.
>
> I think I get your point. Overwriting happens after we store the
> original of_args, but the code is indeed not intuitive. I think I can
> move it further, as you suggested.

Now I think we can skip the overwriting altogether and just adapt the
following of_reset_n_cells check ad of_xlate call as mentioned above.

regards
Philipp

2024-01-11 09:48:28

by Krzysztof Kozlowski

[permalink] [raw]

Subject: Re: [PATCH v2 2/4] reset: Instantiate reset GPIO controller for shared reset-gpios

On 09/01/2024 12:58, Philipp Zabel wrote:
>>>> + /* Not freed in normal path, persisent subsyst data */
>>>> + rgpio_dev = kzalloc(sizeof(*rgpio_dev), GFP_KERNEL);
>>>
>>> Since this is persistent, instead of letting the reset-gpio driver call
>>> of_parse_phandle_with_args() again, this could be passed in via
>>> platform data. Is there a reason not to do that instead?
>>
>> We can pass it as read only platform data, but we cannot pass the
>> ownership. This is associated with registered platform device, not with
>> bound one device->driver.
>>
>> Imagine case:
>> 1. modprobe reset-gpio,
>> 2. Driver is bound to the device,
>> 3. unbind (echo > unbind)
>> 4. rmmod
>> 5. goto 1
>
> Keeping ownership on the list is fine, the reset-gpio driver makes its
> own copy of of_phandle_args anyway. I was just wondering whether it
> could make this copy from platform data instead of from the
> of_parse_phandle_with_args() return value.

Looks like it could. This could save us few lines of code in
reset-gpio.c. I'll try it.

>
> [...]
>>>
>>>> @@ -839,21 +960,50 @@ __of_reset_control_get(struct device_node *node, const char *id, int index,
>>>> index, &args);
>>>> if (ret == -EINVAL)
>>>> return ERR_PTR(ret);
>>>> - if (ret)
>>>> - return optional ? NULL : ERR_PTR(ret);
>>>> + if (ret) {
>>>> + /*
>>>> + * There can be only one reset-gpio for regular devices, so
>>>> + * don't bother with GPIO index.
>>>> + */
>>>
>>> I don't understand this comment. The GPIO index should be checked as
>>> part of __reset_gpios_args_match(), or should it not?
>>
>> This and earlier comment are result of a bit hacky approach to the
>> problem: how to find reset controllers for that GPIO?
>>
>> The point is that our reset gpio controller has only 1 reset, thus
>> of_reset_n_cells=1. However args_count from of_parse_handle is >0, which
>> later is compared in reset core:
>>
>> https://elixir.bootlin.com/linux/latest/source/drivers/reset/core.c#L859
>>
>> That part we need to match.
>>
>> I could make the reset-gpio driver to have of_reset_n_cells=2, but what
>> would be the point? The rest of the cells are not really relevant,
>> because you cannot refer to this reset gpio controller with any other
>> arguments.
>>
>> To remind: my solution spawns one reset-gpio controller for one GPIO.
>
> Thank you. I think we could also just make that check
>
> if (WARN_ON(!rcdev->of_args && ...))
>
> instead and skip the of_xlate call in that case (or implement of_xlate
> in the reset-gpio driver to make this more explicit).

Ack

>
>>>
>>>> + ret = of_parse_phandle_with_args(node, "reset-gpios", "#gpio-cells",
>>>> + 0, &args);
>>>> + if (ret)
>>>> + return optional ? NULL : ERR_PTR(ret);
>>>>
>>>> - mutex_lock(&reset_list_mutex);
>>>> - rcdev = NULL;
>>>> - list_for_each_entry(r, &reset_controller_list, list) {
>>>> - if (args.np == r->of_node) {
>>>> - rcdev = r;
>>>> - break;
>>>> - }
>>>> + gpio_fallback = true;
>>>
>>> Is there a reason not just call __reset_add_reset_gpio_device() here?
>>> With that, there should be no need to call __reset_find_rcdev() twice.
>>
>> Hm, could be, although not sure if code would be simpler.
>>
>> This entire function handles two cases:
>> 1. Get normal reset controller ("resets" OF property),
>> 2. If above fails, get reset-gpio controller ("reset-gpios" OF property)
>>
>> Therefore the entire solution is following approach:
>> 1. of_parse_phandle(resets)
>> 1b. error? Then of_parse_phandle(reset-gpios)
>> 2. Find reset-controller based on any of above phandles.
>> 3. error? Check if we created reset-gpios platform device. If not:
>> create new reset-gpios platform device/
>> 3b. Platform device could probe, so lookup again for reset controller or
>> defer probe.
>>
>> What type of flow do you propose?
>
> I propose to reorder after parsing the phandles: check/create the gpio
> platform device right after parsing the gpio handle. Only then lock
> reset_list_mutex look for the rcdev.
>
> 1. of_parse_phandle(resets)
> 1b. error? Then of_parse_phandle(reset-gpios)
> 2b. gpio? Then check if we created reset-gpios platform device. If not:
> create new reset-gpios platform device/, defer if probe failed
> 3. Lock reset_list_mutex, find reset-controller based on any of above
> phandles.

Could work, let me try. I have impression this was my first approach
which resulted in a bit more complicated code, but I don't remember the
details now.

>
>>
>>>
>>>> }
>>>>
>>>> + mutex_lock(&reset_list_mutex);
>>>> + rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
>>>
>>> This gets called with args as parsed. If there is a match, this will
>>> overwrite args (in the gpio_fallback case) and return NULL.
>>
>> Overwrite not complete. It will only overwrite args_count and return a
>> valid rcdev.
>> I do not see overwriting in case of returning NULL.
>
> Sorry, I meant to write
>
> "This gets called with args as parsed. If there is a match, this will
> overwrite args (in the gpio_fallback case) _or_ return NULL."
>
> at least at the end, when I understood the following.
>
>>>
>>>> +
>>>> if (!rcdev) {
>
> So in this non-NULL branch there was no overwriting.
>
>>>> - rstc = ERR_PTR(-EPROBE_DEFER);
>>>> - goto out;
>>>> + if (gpio_fallback) {
>>>> + /*
>>>> + * Registering reset-gpio device might cause immediate
>>>> + * bind, thus taking reset_list_mutex lock via
>>>> + * reset_controller_register().
>>>> + */
>>>> + mutex_unlock(&reset_list_mutex);
>>>> + ret = __reset_add_reset_gpio_device(node, &args);
>>>
>>> So this will also be called with args as parsed.
>>>
>>>> + mutex_lock(&reset_list_mutex);
>>>> + if (ret) {
>>>> + rstc = ERR_PTR(ret);
>>>> + goto out;
>>>> + }
>>>> + /*
>>>> + * Success: reset-gpio could probe immediately, so
>>>> + * re-check the lookup.
>>>> + */
>>>> + rcdev = __reset_find_rcdev(&args, gpio_fallback, NULL);
>>>
>>> And this will again be called with args as parsed and overwrite args
>>> again.>
>>>> + if (!rcdev) {
>>>> + rstc = ERR_PTR(-EPROBE_DEFER);
>>>> + goto out;
>>>> + }
>>>> + /* Success, rcdev is valid thus do not bail out */
>>>> + } else {
>>>> + rstc = ERR_PTR(-EPROBE_DEFER);
>>>> + goto out;
>>>> + }
>>>> }
>>>
>>> So at this point args is overwritten in the gpio_fallback case. I would
>>> find it much clearer to just overwrite args here and make the first
>>> parameter to __reset_find_rcdev() const.
>>
>> I think I get your point. Overwriting happens after we store the
>> original of_args, but the code is indeed not intuitive. I think I can
>> move it further, as you suggested.
>
> Now I think we can skip the overwriting altogether and just adapt the
> following of_reset_n_cells check ad of_xlate call as mentioned above.

Yep!

Best regards,
Krzysztof