Re: [PATCH v2 2/2] mmc: sdhci-npcm: Add NPCM SDHCI driver

[Adrian Hunter <adrian.hunter@intel.com>]

On 7/12/22 15:25, Andy Shevchenko wrote:
> On Wed, Dec 7, 2022 at 3:01 PM Tomer Maimon <tmaimon77@gmail.com> wrote:
>> On Mon, 5 Dec 2022 at 16:33, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>> On 5/12/22 16:17, Andy Shevchenko wrote:
>>>> On Mon, Dec 5, 2022 at 4:14 PM Andy Shevchenko
>>>> <andy.shevchenko@gmail.com> wrote:
>>>>> On Mon, Dec 5, 2022 at 3:41 PM Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>>> On 5/12/22 15:25, Andy Shevchenko wrote:
>>>>>>> On Mon, Dec 5, 2022 at 1:20 PM Tomer Maimon <tmaimon77@gmail.com> wrote:
> 
> ...
> 
>>>>>>> devm_ is problematic in your case.
>>>>>>> TL;DR: you need to use clk_get_optional() and clk_put().
>>>>>>
>>>>>> devm_ calls exactly those, so what is the issue?
>>>>>
>>>>> The issue is the error path or removal stage where it may or may be
>>>>> not problematic. To be on the safe side, the best approach is to make
>>>>> sure that allocated resources are being deallocated in the reversed
>>>>> order. That said, the
>>>>>
>>>>> 1. call non-devm_func()
>>>>> 2. call devm_func()
>>>>>
>>>>> is wrong strictly speaking.
>>>>
>>>> To elaborate more, the
>>>>
>>>> 1. call all devm_func()
>>>> 2. call only non-devm_func()
>>>>
>>>> is the correct order.
>>>
>>> 1. WRT pltfm_host->clk, that is what is happening
>>> 2. WRT other resources that is simply not always possible because not every resource is wrapped by devm_
>>> e.g. mmc_alloc_host() / mmc_free_host()
>> I little confused about what to decide, should I use only
>> non-devm_func because mmc_alloc_host() / mmc_free_host() is not
>> warrped with devm_?
> 
> It is up to you how to proceed. I pointed out the problem with your
> code which may or may not be fatal.
> 
> If you want to solve it, there are several approaches:
> 1) get rid of devm_ completely;
> 2) properly shuffle the ordering in ->probe(), so all devm_ calls are
> followed by non-devm_;
> 3) wrap non-devm_ cals to become managed (see
> devm_add_action_or_reset() approach);
> 4) fix SDHCI / MMC layer by providing necessary devm_ calls and/or fix
> sdhci_pltfm_register() to handle the clock.

I can take care of sdhci_pltfm when I next have some time.
Otherwise it looks OK to me, so I am acking it.

> 
> Personally, the list order is from the least, what I prefer, to the
> most (i.o.w. I would like to see rather 4) than 1) to be implemented).
> 
>>>> Hence in this case the driver can be worked around easily (by
>>>> shuffling the order in ->probe() to call devm_ first), but as I said
>>>> looking into implementation of the _unregister() I'm pretty sure that
>>>> clock management should be in sdhci-pltfm, rather than in all callers
>>>> who won't need the full customization.
>>>>
>>>> Hope this helps to understand my point.
>>>>
>>>>>>> Your ->remove() callback doesn't free resources in the reversed order
>>>>>>> which may or, by luck, may not be the case of all possible crashes,
>>>>>>> UAFs, races, etc during removal stage. All the same for error path in
>>>>>>> ->probe().
>>>>>
>>>>> I also pointed out above what would be the outcome of neglecting this rule.
> 
> ...
> 
>>>>>>>>> Why can't you use sdhci_pltfm_register()?
>>>>>>>> two things are missing in sdhci_pltfm_register
>>>>>>>> 1. clock.
>>>>>>>
>>>>>>> Taking into account the implementation of the corresponding
>>>>>>> _unregister() I would add the clock handling to the _register() one.
>>>>>>> Perhaps via a new member of the platform data that supplies the name
>>>>>>> and index of the clock and hence all clk_get_optional() / clk_put will
>>>>>>> be moved there.
>> Do you mean to add it to sdhci_pltfm_register function? if yes I
>> believe it will take some time to modify sdhci_pltfm_register
>> I prefer not to use sdhci_pltfm_register.
> 
> In the Linux kernel we are trying hard to avoid code duplication. Why
> do you need it to be open coded? (Yes, I heard you, but somebody
> should fix the issues with that funcion at some point, right?)
> 
>>>>>>>> 2. Adding SDHCI_CAN_DO_8BIT capability according the eMMC capabilities.
>>>>>>>
>>>>>>> All the same, why can't platform data be utilised for this?
>