On Thu, Nov 1, 2018 at 11:32 PM Changbin Du <changbin.du@xxxxxxxxx> wrote:
>
> On Thu, Nov 01, 2018 at 12:32:48PM +0900, Masahiro Yamada wrote:
> > Hi Arnd,
> >
> >
> > On Thu, Nov 1, 2018 at 6:42 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:
> > >
> > > On 10/31/18, Randy Dunlap <rdunlap@xxxxxxxxxxxxx> wrote:
> > > > On 10/30/18 8:59 PM, Stephen Rothwell wrote:
> > > >> Hi all,
> > > >>
> > > >> Please do not add any v4.21/v5.1 code to your linux-next included trees
> > > >> until after the merge window closes.
> > > >>
> > > >> Changes since 20181030:
> > > >>
> > > >
> > > >
> > > > on i386:
> > > >
> > > > ld: drivers/virt/vboxguest/vboxguest_core.o: in function
> > > > `vbg_ioctl_hgcm_call':
> > > > vboxguest_core.c:(.text+0x212b): undefined reference to `vbg_hgcm_call32'
> > > >
> > > >
> > > > Full randconfig file is attached.
> > > >
> > > > Reported-by: Randy Dunlap <rdunlap@xxxxxxxxxxxxx>
> > >
> > > I ran into the same thing, and lots of others like it; they all go back to
> > > the -Og compiler option introduced in commit 94c7dfd01652 ("kernel
> > > hacking: support building kernel with -Og optimization level").
> > >
> > >
> > > I have patches for most of these, but I'd prefer if we could leave
> > > that patch out of 4.20 for now, and have it in linux-next for merging
> > > into the next merge window, to give us more time to fix up all the
> > > regressions. This particular instance is fixed by the patch below.
> >
> >
> > Hmm, I was considering the -Og support for the current MW,
> > but it might be better to postpone it for 4.21-rc1...
> >
> >
> >
> > > Other symptoms caused by the change are tons of 'section mismatch'
> > > warnings for functions that are no longer automatically inlined, and
> > > sprintf format string overflow warnings when gcc fails to propagate
> > > an upper bound on an integer argument that is printed into a fixed
> > > length string.
> > >
> > > Arnd
> > > ---
> > > From 8172e650c6797b9cc64c6f7b1c9e5433156290cc Mon Sep 17 00:00:00 2001
> > > From: Arnd Bergmann <arnd@xxxxxxxx>
> > > Date: Wed, 31 Oct 2018 16:24:35 +0100
> > > Subject: [PATCH] vbox: fix link error with 'gcc -Og'
> > >
> > > With the new CONFIG_CC_OPTIMIZE_FOR_DEBUGGING option, we get a link
> > > error in the vboxguest driver, when that fails to optimize out the
> > > call to the compat handler:
> > >
> > > drivers/virt/vboxguest/vboxguest_core.o: In function `vbg_ioctl_hgcm_call':
> > > vboxguest_core.c:(.text+0x1f6e): undefined reference to `vbg_hgcm_call32'
> > >
> > > Another compile-time check documents better what we want and avoids
> > > the error.
> > >
> > > Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> > >
> > > diff --git a/drivers/virt/vboxguest/vboxguest_core.c
> > > b/drivers/virt/vboxguest/vboxguest_core.c
> > > index 3093655c7b92..1475ed5ffcde 100644
> > > --- a/drivers/virt/vboxguest/vboxguest_core.c
> > > +++ b/drivers/virt/vboxguest/vboxguest_core.c
> > > @@ -1312,7 +1312,7 @@ static int vbg_ioctl_hgcm_call(struct vbg_dev *gdev,
> > > return -EINVAL;
> > > }
> > >
> > > - if (f32bit)
> > > + if (IS_ENABLED(CONFIG_COMPAT) && f32bit)
> > > ret = vbg_hgcm_call32(gdev, client_id,
> > > call->function, call->timeout_ms,
> > > VBG_IOCTL_HGCM_CALL_PARMS32(call),
> >
> > Right, but adding the CONFIG_COMPAT guard to both the caller and the callee
> > looks a bit redundant...
> >
> >
> > I think most cases causing this kind of problem
> > rely on the assumption that 'static' functions are inlined.
> >
> > We can force the compiler to do so
> > by replacing 'static' with 'static __always_inline'
> >
> Declaring it as normal 'inline' is enough. The inline functions are
> still inlined by gcc. Thanks.
How about clang?
For clang, -Og might be equivalent to -O1 at this moment, but I am not sure.
In my understanding, Clang does not inline functions marked with 'static inline'
for -Og (or -O1) optimization level.
Theoretically, 'inline' keyword is a just hint for the compiler, after all.
--
Best Regards
Masahiro Yamada
