On Wed, Jul 7, 2021 at 3:36 PM Helge Deller <deller@xxxxxx> wrote:
> * Arnd Bergmann <arnd@xxxxxxxx>:
> > My first thought was this was a result of -finline-functions getting
> > enabled by default in gcc-10, but enabling it on gcc-9 did not
> > help here. It's likely that the gcc behavior was fixed in the process
> > of enabling the more aggressive inliner by default though.
> >
> > I also tried building genhd.o for every architecture that has
> > gcc-9.4 support and did not find the problem anywhere except
> > on parisc.
> >
> > Using CONFIG_CC_OPTIMIZE_FOR_SIZE did solve the
> > problem for me (frame size down to 164 bytes), but I could not
> > pinpoint a specific -f option that fixes it for -O2. Maybe we can
> > simply change the defconfig here? 32-bit parisc systems are
> > probably memory limited enough that building a -Os kernel
> > is a good idea anyway. 64-bit parisc already builds with -Os
> > but does not see the warning with -O2 either.
>
> I agree that the simplest solution is to increase the default value for
> parisc here.
> On parisc we have a 32k stack (either 1x32k or 2x16k when using IRQ
> stacks). I increased the default value to 1280 in 2017, but as can be
> seen here this isn't sufficient. Either way, we have an active runtime
> check for stack overflows which has never triggered yet, so let's just
> remove the compiler warning by increasing the value to 2048. Patch is
> below.
>
> [PATCH] parisc: Increase gcc stack frame check to 2048 for 32- and 64-bit
>
> parisc uses much bigger frames than other architectures, so increase the
> stack frame check value to 2048 to avoid compiler warnings.
I think setting it to 2048 is rather excessive, and it would make you miss
other real bugs. What I suggested was to change the defconfig to use
CONFIG_CC_OPTIMIZE_FOR_SIZE instead.
The reasoning for the 1280 byte limit on parisc is that it needs a few extra
bytes for its larger stack frames, and 1024 for the other 32-bit architectures
is only there because anything smaller warns for a handful of functions
that are fine-tuned to need slightly less than that, when the call chain
is predictable and using less would impact performance.
I actually think we should reduce the warning limit for 64-bit architectures
to 1280 bytes as well, but that triggers a couple of warnings that still
need to be resolved first. In almost all cases, a kernel function needing
more than 512 bytes is an indication of either a bug in the kernel, or
(rarely) in the compiler.
Arnd
