Re: [PATCH] Enable '-Werror' by default for all kernel builds

On 2021-09-08 12:41 a.m., Linus Torvalds wrote:

On Tue, Sep 7, 2021 at 8:52 PM Harry Wentland [off-list ref] wrote:

quoted

Attached patches fix these x86_64 ones reported by Nick:

Hmm.

You didn't seem to fix up the calling convention for print__xyz(),
which still take those xyz structs as pass-by-value.

Obviously it would be good to do things incrementally, so if that
attached patch was just [1/N] I won't complain..

You're right. I was focussed on the stack frame limit but fixed up
the rest as well now and sent the series out.

https://lkml.org/lkml/2021/9/8/933

quoted

I'm also seeing one more that might be more challenging to fix but is nearly at 1024:

drivers/gpu/drm/amd/amdgpu/../display/dc/dml/dcn21/display_mode_vba_21.c:3397:6: error: stack frame size of 1064 bytes in function 'dml21_ModeSupportAndSystemConfigurationFull' [-Werror,-Wframe-larger-than=]

Oh Gods, that function is truly something else..

Is there some reason why it's one humongous function, with the
occasional single-line comment?

Because it really looks to me like pretty much everywhere I see one of
those rare comments, I would go "this part should be a function of its
own", and then there would be one caller fuynction that just calls
each of those sub-functions one after the other.

Yeah, that's what I'm thinking as well. It would likely fix the stack
size, even without dynamically allocating the two structs you mention
below.

That would - I think - make the code easier to read, and then it would
also make it very obvious where it magically uses a lot of stack.

My suspicion is actually "nowhere". The stack use is just hugely
spread out, and the compiler has just kept accumulating more spill
variables on the frame with no single big reason.

Yes, I see a couple of local structures:

                Pipe myPipe;
                HostVM myHostVM;

but more than that I see several function calls that have basically 62
arguments. And I wish I was making that number up. I'm not. That
"CalculatePrefetchSchedule()" call literally has 62 arguments.

But *all* of the top-level loops in that function literally look like
they could - and should - be functions in their own right. Some of
them would be fairly complex even so (ie that code under the comment

        //Prefetch Check

would be quite the big function all of its own.

We have a coding style thing:

    Documentation/process/coding-style.rst

that says that you should strive to have functions that are "short and
sweet" and fit on one or two screenfuls of text.

That one function from hell is 1832 lines of code.

It really could be improved upon.

Absolutely.

The file comes with an (easy to miss) disclaimer that it is
"gcc-parseable HW gospel":
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/amd/display/dc/dml/dcn20/display_mode_vba_20.c?h=v5.14#n32

The display_mode_vba* stuff deals with bandwidth formulas from HW
designers. At some point in the past we attempted to convert them
to something more readable and elegant but would often run into
difficulties getting support from the right people when things
wouldn't work. Using the HW designer's code directly tends to
short circuit any arguments about SW correctness.

In short, I don't really like this code but it works. It helps
prevent black screens and underflows on the display.

We try to follow the coding-style.rst for the most part elsewhere,
though there are still plenty of areas where we can improve.

Harry

               Linus