Re:Re: [PATCH v4] powerpc:85xx: Add missing of_node_put() in sgy_cst1000

From: Liang He
Date: Fri Jun 17 2022 - 04:18:06 EST





At 2022-06-17 14:53:13, "Christophe Leroy" <christophe.leroy@xxxxxxxxxx> wrote:
>
>
>Le 17/06/2022 à 08:45, Liang He a écrit :
>>
>>
>>
>> At 2022-06-17 14:28:56, "Christophe Leroy" <christophe.leroy@xxxxxxxxxx> wrote:
>>>
>>>
>>> Le 17/06/2022 à 08:08, Liang He a écrit :
>>>> In gpio_halt_probe(), of_find_matching_node() will return a node
>>>> pointer with refcount incremented. We should use of_node_put() in
>>>> fail path or when it is not used anymore.
>>>>
>>>> Signed-off-by: Liang He <windhl@xxxxxxx>
>>>> ---
>>>> changelog:
>>>> v4: reuse exist 'err' and use a simple code style, advised by CJ
>>>> v3: use local 'child_node' advised by Michael.
>>>> v2: use goto-label patch style advised by Christophe Leroy.
>>>> v1: add of_node_put() before each exit.
>>>>
>>>> arch/powerpc/platforms/85xx/sgy_cts1000.c | 35 ++++++++++++++---------
>>>> 1 file changed, 22 insertions(+), 13 deletions(-)
>>>>
>>>> diff --git a/arch/powerpc/platforms/85xx/sgy_cts1000.c b/arch/powerpc/platforms/85xx/sgy_cts1000.c
>>>> index 98ae64075193..e4588943fe7e 100644
>>>> --- a/arch/powerpc/platforms/85xx/sgy_cts1000.c
>>>> +++ b/arch/powerpc/platforms/85xx/sgy_cts1000.c
>>>> @@ -71,6 +71,7 @@ static int gpio_halt_probe(struct platform_device *pdev)
>>>> {
>>>> enum of_gpio_flags flags;
>>>> struct device_node *node = pdev->dev.of_node;
>>>> + struct device_node *child_node;
>>>> int gpio, err, irq;
>>>> int trigger;
>>>>
>>>> @@ -78,26 +79,29 @@ static int gpio_halt_probe(struct platform_device *pdev)
>>>> return -ENODEV;
>>>>
>>>> /* If there's no matching child, this isn't really an error */
>>>> - halt_node = of_find_matching_node(node, child_match);
>>>> - if (!halt_node)
>>>> + child_node = of_find_matching_node(node, child_match);
>>>> + if (!child_node)
>>>> return 0;
>>>>
>>>> /* Technically we could just read the first one, but punish
>>>> * DT writers for invalid form. */
>>>> - if (of_gpio_count(halt_node) != 1)
>>>> - return -EINVAL;
>>>> + if (of_gpio_count(child_node) != 1) {
>>>> + err = -EINVAL;
>>>> + goto err_put;
>>>> + }
>>>>
>>>> /* Get the gpio number relative to the dynamic base. */
>>>> - gpio = of_get_gpio_flags(halt_node, 0, &flags);
>>>> - if (!gpio_is_valid(gpio))
>>>> - return -EINVAL;
>>>> + gpio = of_get_gpio_flags(child_node, 0, &flags);
>>>> + if (!gpio_is_valid(gpio)) {
>>>> + err = -EINVAL;
>>>> + gotot err_put;
>>>
>>> Did you test the build ?
>>
>> Sorry for this fault.
>>
>> In fact, I am still finding an efficient way to building different arch source code as I only have x86-64.
>>
>> Now I am try using QEMU.
>>
>> Anyway, sorry for this fault.
>
>You can find cross compilers for most architectures for x86-64 here :
>https://mirrors.edge.kernel.org/pub/tools/crosstool/
>
>Christophe

Hi, Christophe and Conor.

Sorry to trouble you again.

Now I only know how to quickly identify the refcounting bugs, but I cannot efficiently give a build test.

For example, I use the cross compilers 'powerpc-linux-gnu-gcc' to compile 'arch/powerpc/platforms/85xx/sgy_cts1000.c' with -fsyntax-only flag.
But I meet too many header file missing errors. Even if I add some 'include' pathes, e.g., ./arch/powerpc/include, ./include,
there are still too many other errors.

So if there is any efficient way to check my patch code to avoid 'gotot' error again.

Thanks again, Christophe and Conor.

Liang