On 30.08.2024 5:19 PM, Dan Carpenter wrote:
> On Fri, Aug 30, 2024 at 12:03:20PM +0200, Konrad Dybcio wrote:
>> On 30.08.2024 10:08 AM, Krzysztof Kozlowski wrote:
>>> On 29/08/2024 14:48, Yu Jiaoliang wrote:
>>>> Error handling in probe() can be a bit simpler with dev_err_probe().
>>>>
>>>> Signed-off-by: Yu Jiaoliang <yujiaoliang@xxxxxxxx>
>>>> ---
>>>> drivers/soc/qcom/qcom-pbs.c | 7 +++----
>>>> 1 file changed, 3 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/drivers/soc/qcom/qcom-pbs.c b/drivers/soc/qcom/qcom-pbs.c
>>>> index 77a70d3d0d0b..ab9de12ec901 100644
>>>> --- a/drivers/soc/qcom/qcom-pbs.c
>>>> +++ b/drivers/soc/qcom/qcom-pbs.c
>>>> @@ -201,10 +201,9 @@ static int qcom_pbs_probe(struct platform_device *pdev)
>>>> }
>>>>
>>>> ret = device_property_read_u32(pbs->dev, "reg", &val);
>>>> - if (ret < 0) {
>>>> - dev_err(pbs->dev, "Couldn't find reg, ret = %d\n", ret);
>>>> - return ret;
>>>> - }
>>>> + if (ret < 0)
>>>> + return dev_err_probe(pbs->dev, ret, "Couldn't find reg\n");
>>>
>>> This cannot defer, so not much benefits. And you ignore other place in
>>> the probe()... That's like a weird pattern with all your patches change
>>> something irrelevant, but leave other places unchanged.
>>>
>>> That's pointless and churn.
>>
>> Hm, that's a good point.. Maybe the static checker folks could come up
>> with a way that would find functions that call something that can defer
>> at one point or another and suggest (not) using dev_err_probe with W=1/2..
>> (although that is probably not going to be very high prio given all the
>> other static checker issues we're still yet to resolve)
>>
>> Unless we have something like that already? +CC Dan
>
> I believe these patches are from people writing their own Coccinelle scripts to
> do the conversions. There aren't any published scripts which care one way or
> the other.
>
> device_property_read_u32() can't return -EPROBE_DEFER. It's documented in a
> comment. So this is just a question of preferred style. There isn't a kernel
> wide preferred style on this. Some maintainers prefer to not use dev_err_probe()
> where it "doesn't make sense because ret isn't -EPROBE_DEFER". Other maintainers
> use it all the time even for error code literals like:
> return dev_err_probe(pbs->dev, -EINVAL, "invalid input.\n");
> Because "it's cleaner, more uniform and only takes one line". I think Julia
> said she didn't want to get involved with this debate and I definitely don't.
Personally, I don't mind either.. so longer as it's consistent within
the file
>
> There are a few things which we could do:
>
> 1) Returning -EPROBE_DEFER to an ioctl or something besides a probe()
> This is a bug right? -EPROBE_DEFER is basically kernel internal for probe()
> functions. It tried to write this but it was complicated so I gave up.
Maybe call_tree.pl can somehow be used with an if name[-5:] == "probe"
or something along those lines..
>
> 2) Printing an error message for -EPROBE_DEFER warnings
> I've written this check and I can test it tonight.
>
> 3) Not propagating the -EPROBE_DEFER returns
> This shouldn't be too hard to write.
>
> Let me add a KTODO in case anyone wants to do this before I get around to it.
>
> KTODO: write Smatch EPROBE_DEFER warnings
Thanks!
Konrad
