Re: [PATCH] rust: io: always inline methods leading to build_assert

["Alexandre Courbot" <acourbot@nvidia.com>]

On Fri Nov 28, 2025 at 9:27 AM JST, Alexandre Courbot wrote:
> On Thu Nov 27, 2025 at 11:53 PM JST, Daniel Almeida wrote:
>>
>>
>>> On 27 Nov 2025, at 10:30, Alexandre Courbot <acourbot@nvidia.com> wrote:
>>> 
>>> `build_assert` relies on the compiler to optimize out its error path,
>>> lest build fails with the dreaded error:
>>> 
>>>    ERROR: modpost: "rust_build_error" [drivers/gpu/nova-core/nova_core.ko] undefined!
>>> 
>>> It has been observed that very trivial code performing I/O accesses
>>> (sometimes even using an immediate value) would seemingly randomly fail
>>> with this error whenever `CLIPPY=1` was set. Removing the CLIPPY option
>>> makes the error go away, but that's obviously not a great workaround.
>>> Clippy appears to influence the way the compiler optimizes things,
>>> making it on occasion generate a method where we would need it to inline
>>> in order to satisfy a `build_assert`.
>>> 
>>> Fix this by instructing the compiler to always inline the methods
>>> leading to `build_assert`. This stronger directive is effective even
>>> when `CLIPPY=1` is specified, which gets rid of this error.
>>> 
>>> Signed-off-by: Alexandre Courbot <acourbot@nvidia.com>
>>> ---
>>> This is the same fix as for another build error triggered by the use of
>>> `build_assert` [1], which signals that all callers of this macro should
>>> all be tagged with `#[inline(always)]`, as inlining is a requirement for
>>> `build_assert` to perform properly anyway.
>>> 
>>> [1] https://lore.kernel.org/all/DEEUYUOAEZU3.1J1HM2YQ10EX1@nvidia.com/
>>> ---
>>> rust/kernel/io.rs | 9 ++++++---
>>> 1 file changed, 6 insertions(+), 3 deletions(-)
>>> 
>>> diff --git a/rust/kernel/io.rs b/rust/kernel/io.rs
>>> index 98e8b84e68d1..f161ec8056ce 100644
>>> --- a/rust/kernel/io.rs
>>> +++ b/rust/kernel/io.rs
>>> @@ -142,7 +142,8 @@ macro_rules! define_read {
>>>         /// Bound checks are performed on compile time, hence if the offset is not known at compile
>>>         /// time, the build will fail.
>>>         $(#[$attr])*
>>> -        #[inline]
>>> +        // Always inline so the error path of `io_addr_assert` is optimized out.
>>> +        #[inline(always)]
>>>         pub fn $name(&self, offset: usize) -> $type_name {
>>>             let addr = self.io_addr_assert::<$type_name>(offset);
>>> 
>>> @@ -171,7 +172,8 @@ macro_rules! define_write {
>>>         /// Bound checks are performed on compile time, hence if the offset is not known at compile
>>>         /// time, the build will fail.
>>>         $(#[$attr])*
>>> -        #[inline]
>>> +        // Always inline so the error path of `io_addr_assert` is optimized out.
>>> +        #[inline(always)]
>>>         pub fn $name(&self, value: $type_name, offset: usize) {
>>>             let addr = self.io_addr_assert::<$type_name>(offset);
>>> 
>>> @@ -239,7 +241,8 @@ fn io_addr<U>(&self, offset: usize) -> Result<usize> {
>>>         self.addr().checked_add(offset).ok_or(EINVAL)
>>>     }
>>> 
>>> -    #[inline]
>>> +    // Always inline so the error path of `build_assert!` is optimized out.
>>> +    #[inline(always)]
>>>     fn io_addr_assert<U>(&self, offset: usize) -> usize {
>>>         build_assert!(Self::offset_valid::<U>(offset, SIZE));
>>> 
>>> 
>>> ---
>>> base-commit: ea34511aaf755349999a1067b2984a541bee1492
>>> change-id: 20251127-io-build-assert-3579a5bfb81c
>>> 
>>> Best regards,
>>> -- 
>>> Alexandre Courbot <acourbot@nvidia.com>
>>> 
>>> 
>>
>> Reviewed-by: Daniel Almeida <daniel.almeida@collabora.com>
>>
>> I also faced this with genmask, by the way, i.e.: using genmask with an
>> in-bounds constant would trigger a build error. Very confusingly, this would be
>> randomly solved by moving the genmask invocation around in the code.
>>
>> I wonder if the same fix is needed for it as well?
>
> What you described is exactly the symptoms I was experimenting (notably
> the "moving around sometimes fixes it" thing) so yeah, I think this also
> applies to `bit_*` and `genmask_*` (and anything that invokes
> `build_assert`, really).

... anything that invoke `builds_assert` with one of its arguments as
parameter, that is. There are uses in e.g. `locked_by.rs` that validate
valid properties rather than function arguments, and they are not
affected by this.