Seeking advice on API return type inconsistency

[Tom Rini]

On Wed, Aug 13, 2025 at 01:41:21PM +0200, Michal Suchánek wrote:
> Hello,
> 
> On Wed, Aug 13, 2025 at 11:58:19AM +0100, Andrew Goodbody wrote:
> > On 12/08/2025 16:03, Tom Rini wrote:
> > > On Tue, Aug 12, 2025 at 03:46:56PM +0100, Andrew Goodbody wrote:
> > > > On 12/08/2025 15:33, Tom Rini wrote:
> > > > > On Tue, Aug 12, 2025 at 10:17:47AM +0100, Andrew Goodbody wrote:
> > > > > > On 11/08/2025 17:36, Quentin Schulz wrote:
> > > > > > > Hi Andrew,
> > > > > > > 
> > > > > > > On 8/11/25 5:24 PM, Andrew Goodbody wrote:
> > > > > > > > Hi,
> > > > > > > > 
> > > > > > > > I was wondering what people's thoughts were on API return types. In
> > > > > > > > particular there is this and other examples in include/clk-uclass.h
> > > > > > > > 
> > > > > > > > /**
> > > > > > > >     * get_rate() - Get current clock rate.
> > > > > > > >     * @clk:    The clock to query.
> > > > > > > >     *
> > > > > > > >     * This returns the current rate of a clock. If the clock is
> > > > > > > > disabled, it
> > > > > > > >     * returns the rate at which the clock would run if it was enabled. The
> > > > > > > >     * following pseudo-code should hold::
> > > > > > > >     *
> > > > > > > >     *   disable(clk)
> > > > > > > >     *   rate = get_rate(clk)
> > > > > > > >     *   enable(clk)
> > > > > > > >     *   assert(get_rate(clk) == rate)
> > > > > > > >     *
> > > > > > > >     * Return:
> > > > > > > >     * * The rate of @clk
> > > > > > > >     * * -%ENOSYS if this function is not implemented for @clk
> > > > > > > >     * * -%ENOENT if @clk->id is invalid. Prefer using an assert
> > > > > > > > instead, and doing
> > > > > > > >     *   this check in request().
> > > > > > > >     * * Another negative error value (such as %EIO or %ECOMM) if the
> > > > > > > > rate could
> > > > > > > >     *   not be determined due to a bus error.
> > > > > > > >     */
> > > > > > > > ulong get_rate(struct clk *clk);
> > > > > > > > 
> > > > > > > > 
> > > > > > > > get_rate is declared as returning a ulong but the description says
> > > > > > > > that it can return negative errors. A simple test of the return
> > > > > > > > value for being less than 0 will always fail so errors can go
> > > > > > > > undetected. Casting to a signed type seems less than ideal.
> > > > > > > > 
> > > > > > > > What is the best way to deal with this? Cast to a signed or update
> > > > > > > > the API to be signed or...?
> > > > > > > > 
> > > > > > > 
> > > > > > > Note that clk_get_rate() in the kernel has the same function signature
> > > > > > > so I would refrain from changing the type otherwise we'll have some
> > > > > > > "funny" bugs to handle considering it isn't that uncommon to import
> > > > > > > drivers almost as-is from the Linux kernel.
> > > > > > 
> > > > > > Ah yes. The difference being that the kernel does not seem to attempt to
> > > > > > push an error code through this API, you get a rate or you get 0.
> > > > > 
> > > > > How is the error code pushed? Or is it up to the caller to decide that 0
> > > > > means on a case by case basis?
> 
> Presumably it returns -EBUSY which is then implicitly converted to a large
> positive value. For pointers there is IS_ERR macro to decode that, and
> for integers there is another corresponding macro (I don't recall the
> name).
> 
> For memory this assumes that the last page that holds the 'error'
> pointers cannot be mapped, and there was already a bug in the kernel
> that this was not true.
> 
> For clock rates this assumes that rates close to the maximum value are
> invalid which is not guaranteed in any way whatsoever.
> 
> Consequently, this should not be used.
> 
> 0 is the only clock rate guaranteed to be invalid.
> 
> As far as I have seen in the u-boot code almost no caller checks the
> result, and those that do cannot do anything meaningful with it anyway,
> at most a message is printed. That can be done by the driver without
> passing around the error.
> 
> > > > In the Linux kernel almost no code checks the return of clk_get_rate at all.
> > > > Some code will check the value is sensible and 0 is obviously not sensible.
> > > > But pretty much the call to clk_get_rate is not expected to fail.
> > > 
> > > Perhaps getting lost in the specifics then, but perhaps for this case we
> > > should just do the same? But your question was more general, so another
> > > example might help.
> > 
> > I suspect that the answer is always going to that it depends on the
> > specifics of each case. The U-Boot implementation of clk_get_rate seems to
> > have become more complicated leading to the addition of returning error
> > codes. This leads to the question about what level of compatibility should
> > there be maintained with the kernel? That addition of returning error codes
> > in U-Boot already means that the API is no longer compatible with that of
> > the kernel.
> 
> I would suggest to abolish returning 'errors' as clock rates.

Yes, I think this is a case where we should change our internal API,
both for easier driver migrations from the linux kernel and also because
it seems our design here can't quite work as intended.

-- 
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20250813/0f4f5f5a/attachment.sig>