Hi,
On Thu, May 20, 2021 at 03:43:37AM -0700, Joe Richey wrote:
> From: Joe Richey <[email protected]>
>
> The BIT(n) macro is used in the kernel as an alias for (1 << n).
> However, it is not defined in the UAPI headers, which means that any
> UAPI header files must be careful not to use it, or else the user
> will get a linker error.
Beware that the common definition of BIT() (in include/vdso/bits.h) is:
| #define BIT(nr) (UL(1) << (nr))
That UL() can be important if `nr` is ever greater than bits per int.
> For example, compiling the following program:
>
> #include <sys/auxv.h>
> #include <asm/hwcap2.h>
>
> // Detect if FSGSBASE instructions are enabled
> int main() {
> unsigned long val = getauxval(AT_HWCAP2);
> return !(val & HWCAP2_FSGSBASE);
> }
>
> Results in the following likner error:
>
> /usr/bin/ld: /tmp/cceFpAdR.o: in function `main':
> gs.c:(.text+0x21): undefined reference to `BIT'
>
> This patch series changes all UAPI uses of BIT() to just be open-coded.
In include/uapi/linux/const.h we have an equivaleint _BITUL() macro,
which I think should be used in preference of open-coding this (and is
already used in a number of uapi headers).
> However, there really should be a check for this in checkpatch.pl
> Currently, the script actually _encourages_ users to use the BIT macro
> even if adding things to UAPI.
I think having something that suggests s/BIT()/_BITUL()/ under uapi
would be good.
Thanks,
Mark.
>
> Running `rg "BIT\(" **/uapi/**` shows no more usage of BIT() in any
> UAPI headers. Tested by building a basic kernel. Changes are trivial.
>
> Joe Richey (6):
> x86/elf: Don't use BIT() macro in UAPI headers
> KVM: X86: Don't use BIT() macro in UAPI headers
> drivers: firmware: psci: Don't use BIT() macro in UAPI headers
> uacce: Don't use BIT() macro in UAPI headers
> media: vicodec: Don't use BIT() macro in UAPI headers
> tools headers UAPI: Sync pkt_sched.h with the kernel sources
>
> arch/x86/include/uapi/asm/hwcap2.h | 2 +-
> include/uapi/linux/kvm.h | 4 +-
> include/uapi/linux/psci.h | 2 +-
> include/uapi/linux/v4l2-controls.h | 22 ++---
> include/uapi/misc/uacce/uacce.h | 2 +-
> tools/include/uapi/linux/kvm.h | 4 +-
> tools/include/uapi/linux/pkt_sched.h | 122 ++++++++++++++++++++++++---
> 7 files changed, 130 insertions(+), 28 deletions(-)
>
> --
> 2.31.1
>
On Thu, May 20, 2021 at 4:11 AM Mark Rutland <[email protected]> wrote:
>
> Hi,
>
> On Thu, May 20, 2021 at 03:43:37AM -0700, Joe Richey wrote:
> > From: Joe Richey <[email protected]>
> >
> > The BIT(n) macro is used in the kernel as an alias for (1 << n).
> > However, it is not defined in the UAPI headers, which means that any
> > UAPI header files must be careful not to use it, or else the user
> > will get a linker error.
>
> Beware that the common definition of BIT() (in include/vdso/bits.h) is:
>
> | #define BIT(nr) (UL(1) << (nr))
>
> That UL() can be important if `nr` is ever greater than bits per int.
>
> > For example, compiling the following program:
> >
> > #include <sys/auxv.h>
> > #include <asm/hwcap2.h>
> >
> > // Detect if FSGSBASE instructions are enabled
> > int main() {
> > unsigned long val = getauxval(AT_HWCAP2);
> > return !(val & HWCAP2_FSGSBASE);
> > }
> >
> > Results in the following likner error:
> >
> > /usr/bin/ld: /tmp/cceFpAdR.o: in function `main':
> > gs.c:(.text+0x21): undefined reference to `BIT'
> >
> > This patch series changes all UAPI uses of BIT() to just be open-coded.
>
> In include/uapi/linux/const.h we have an equivaleint _BITUL() macro,
> which I think should be used in preference of open-coding this (and is
> already used in a number of uapi headers).
That's a good idea. I mostly just did the open-coding for
consistency with the surrounding code, but aside from kvm.h
there aren't really many usages to change, so I can do it.
For kvm.h, I think it might be best to leave it open-coded
and move the entire file to _BITUL() at once.
>
> > However, there really should be a check for this in checkpatch.pl
> > Currently, the script actually _encourages_ users to use the BIT macro
> > even if adding things to UAPI.
>
> I think having something that suggests s/BIT()/_BITUL()/ under uapi
> would be good.
I'll just change the script to recommend _BITUL() (instead
of BIT()) if the code path contains "uapi".
>
> Thanks,
> Mark.
>
> >
> > Running `rg "BIT\(" **/uapi/**` shows no more usage of BIT() in any
> > UAPI headers. Tested by building a basic kernel. Changes are trivial.
> >
> > Joe Richey (6):
> > x86/elf: Don't use BIT() macro in UAPI headers
> > KVM: X86: Don't use BIT() macro in UAPI headers
> > drivers: firmware: psci: Don't use BIT() macro in UAPI headers
> > uacce: Don't use BIT() macro in UAPI headers
> > media: vicodec: Don't use BIT() macro in UAPI headers
> > tools headers UAPI: Sync pkt_sched.h with the kernel sources
> >
> > arch/x86/include/uapi/asm/hwcap2.h | 2 +-
> > include/uapi/linux/kvm.h | 4 +-
> > include/uapi/linux/psci.h | 2 +-
> > include/uapi/linux/v4l2-controls.h | 22 ++---
> > include/uapi/misc/uacce/uacce.h | 2 +-
> > tools/include/uapi/linux/kvm.h | 4 +-
> > tools/include/uapi/linux/pkt_sched.h | 122 ++++++++++++++++++++++++---
> > 7 files changed, 130 insertions(+), 28 deletions(-)
> >
> > --
> > 2.31.1
> >