Re: [PATCH net-next] net: ipa: always inline ipa_aggr_granularity_val()

From: Alex Elder
Date: Wed Aug 11 2021 - 08:06:06 EST


On 8/11/21 7:02 AM, Leon Romanovsky wrote:
> On Tue, Aug 10, 2021 at 11:02:13AM -0500, Alex Elder wrote:
>> It isn't required, but all callers of ipa_aggr_granularity_val()
>> pass a constant value (IPA_AGGR_GRANULARITY) as the usec argument.
>> Two of those callers are in ipa_validate_build(), with the result
>> being passed to BUILD_BUG_ON().
>>
>> Evidently the "sparc64-linux-gcc" compiler (at least) doesn't always
>> inline ipa_aggr_granularity_val(), so the result of the function is
>> not constant at compile time, and that leads to build errors.
>>
>> Define the function with the __always_inline attribute to avoid the
>> errors. And given that the function is inline, we can switch the
>> WARN_ON() there to be BUILD_BUG_ON().
>>
>> Fixes: 5bc5588466a1f ("net: ipa: use WARN_ON() rather than assertions")
>> Reported-by: kernel test robot <lkp@xxxxxxxxx>
>> Signed-off-by: Alex Elder <elder@xxxxxxxxxx>
>> ---
>>
>> David/Jakub, this fixes a bug in a commit in net-next/master. -Alex
>>
>> drivers/net/ipa/ipa_main.c | 4 ++--
>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/net/ipa/ipa_main.c b/drivers/net/ipa/ipa_main.c
>> index 25bbb456e0078..f90b3521e266b 100644
>> --- a/drivers/net/ipa/ipa_main.c
>> +++ b/drivers/net/ipa/ipa_main.c
>> @@ -255,9 +255,9 @@ ipa_hardware_config_qsb(struct ipa *ipa, const struct ipa_data *data)
>> * less than the number of timer ticks in the requested period. 0 is not
>> * a valid granularity value.
>> */
>> -static u32 ipa_aggr_granularity_val(u32 usec)
>> +static __always_inline u32 ipa_aggr_granularity_val(u32 usec)
>> {
>> - WARN_ON(!usec);
>> + BUILD_BUG_ON(!usec);
>
> So what exactly are you checking here if all callers pass same value?
> It is in-kernel API, declared as static inside one module. There is no
> need to protect from itself.

Yeah that's a good point. It can just as well be removed.
I think the check was added before I knew it was only going
to be used with a single constant value. That said, the
point was to check at runtime a required constraint.

I'll post version 2 that simply removes it. Thanks.

-Alex

>
> Thanks
>
>>
>> return DIV_ROUND_CLOSEST(usec * TIMER_FREQUENCY, USEC_PER_SEC) - 1;
>> }
>> --
>> 2.27.0
>>