Resending the the correct linux-riscv address.
In the RISC-V 32-bit glibc port [1] the siginfo_t struct in the kernel
doesn't line up with the struct in glibc. In glibc world the _sifields
union is 8 byte alligned (although I can't figure out why) while in the
kernel wordl the _sifields union is 4 bytes alligned. This results in
information being lost in the waitid syscall.
This doesn't seem to be a great fix, but it is somewhat similar to what
x32 does (which has 64-bit time_t like RV32) and I can't figure out why
the two allignments are different.
1: https://github.com/alistair23/glibc/commits/alistair/rv32.next
Alistair Francis (2):
uapi/asm-generic: Allow defining a custom __SIGINFO struct
riscv/include/uapi: Define a custom __SIGINFO struct for RV32
arch/riscv/include/uapi/asm/siginfo.h | 32 +++++++++++++++++++++++++++
include/uapi/asm-generic/siginfo.h | 32 ++++++++++++++-------------
2 files changed, 49 insertions(+), 15 deletions(-)
create mode 100644 arch/riscv/include/uapi/asm/siginfo.h
--
2.22.0
The glibc implementation of siginfo_t results in an allignment of 8 bytes
for the union _sifields on RV32. The kernel has an allignment of 4 bytes
for the _sifields union. This results in information being lost when
glibc parses the siginfo_t struct.
To fix the issue add a pad variable to the struct to avoid allignment
mismatches.
Signed-off-by: Alistair Francis <[email protected]>
---
arch/riscv/include/uapi/asm/siginfo.h | 32 +++++++++++++++++++++++++++
1 file changed, 32 insertions(+)
create mode 100644 arch/riscv/include/uapi/asm/siginfo.h
diff --git a/arch/riscv/include/uapi/asm/siginfo.h b/arch/riscv/include/uapi/asm/siginfo.h
new file mode 100644
index 000000000000..0854ad97bf44
--- /dev/null
+++ b/arch/riscv/include/uapi/asm/siginfo.h
@@ -0,0 +1,32 @@
+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
+#ifndef _ASM_RISCV_SIGINFO_H
+#define _ASM_RISCV_SIGINFO_H
+
+/* Add a pad element for RISC-V 32-bit. We need this as the
+ * _sifields union is 8 byte allgined in usperace.
+ */
+#if __riscv_xlen == 32
+#ifndef __ARCH_HAS_SWAPPED_SIGINFO
+#define __SIGINFO \
+struct { \
+ int si_signo; \
+ int si_errno; \
+ int si_code; \
+ int pad; \
+ union __sifields _sifields; \
+}
+#else
+#define __SIGINFO \
+struct { \
+ int si_signo; \
+ int si_code; \
+ int si_errno; \
+ int pad; \
+ union __sifields _sifields; \
+}
+#endif /* __ARCH_HAS_SWAPPED_SIGINFO */
+#endif
+
+#include <asm-generic/siginfo.h>
+
+#endif /* _ASM_RISCV_SIGINFO_H */
--
2.22.0
On Jul 02 2019, Alistair Francis <[email protected]> wrote:
> In the RISC-V 32-bit glibc port [1] the siginfo_t struct in the kernel
> doesn't line up with the struct in glibc. In glibc world the _sifields
> union is 8 byte alligned (although I can't figure out why)
Try ptype/o in gdb.
Andreas.
--
Andreas Schwab, SUSE Labs, [email protected]
GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE 1748 E4D4 88E3 0EEA B9D7
"And now for something completely different."
On Wed, Jul 3, 2019 at 12:08 AM Andreas Schwab <[email protected]> wrote:
>
> On Jul 02 2019, Alistair Francis <[email protected]> wrote:
>
> > In the RISC-V 32-bit glibc port [1] the siginfo_t struct in the kernel
> > doesn't line up with the struct in glibc. In glibc world the _sifields
> > union is 8 byte alligned (although I can't figure out why)
>
> Try ptype/o in gdb.
That's a useful tip, I'll be sure to use that next time.
Alistair
>
> Andreas.
>
> --
> Andreas Schwab, SUSE Labs, [email protected]
> GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE 1748 E4D4 88E3 0EEA B9D7
> "And now for something completely different."
On Jul 03 2019, Alistair Francis <[email protected]> wrote:
> On Wed, Jul 3, 2019 at 12:08 AM Andreas Schwab <[email protected]> wrote:
>>
>> On Jul 02 2019, Alistair Francis <[email protected]> wrote:
>>
>> > In the RISC-V 32-bit glibc port [1] the siginfo_t struct in the kernel
>> > doesn't line up with the struct in glibc. In glibc world the _sifields
>> > union is 8 byte alligned (although I can't figure out why)
>>
>> Try ptype/o in gdb.
>
> That's a useful tip, I'll be sure to use that next time.
It was a serious note. If the structs don't line up then there is a
mismatch in types that cannot be solved by adding spurious padding. You
need to fix the types instead.
Andreas.
--
Andreas Schwab, SUSE Labs, [email protected]
GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE 1748 E4D4 88E3 0EEA B9D7
"And now for something completely different."
On Thu, Jul 4, 2019 at 9:20 AM Andreas Schwab <[email protected]> wrote:
>
> On Jul 03 2019, Alistair Francis <[email protected]> wrote:
>
> > On Wed, Jul 3, 2019 at 12:08 AM Andreas Schwab <[email protected]> wrote:
> >>
> >> On Jul 02 2019, Alistair Francis <[email protected]> wrote:
> >>
> >> > In the RISC-V 32-bit glibc port [1] the siginfo_t struct in the kernel
> >> > doesn't line up with the struct in glibc. In glibc world the _sifields
> >> > union is 8 byte alligned (although I can't figure out why)
> >>
> >> Try ptype/o in gdb.
> >
> > That's a useful tip, I'll be sure to use that next time.
>
> It was a serious note. If the structs don't line up then there is a
> mismatch in types that cannot be solved by adding spurious padding. You
> need to fix the types instead.
Would it be an option to align all the basic typedefs (off_t, time_t,
clock_t, ...)
between glibc and kernel then, and just use the existing
sysdeps/unix/sysv/linux/generic/bits/typesizes.h after all to avoid
surprises like this?
As of v2 the functional difference is
-#define __INO_T_TYPE __ULONGWORD_TYPE
+#define __INO_T_TYPE __UQUAD_TYPE
#define __INO64_T_TYPE __UQUAD_TYPE
#define __MODE_T_TYPE __U32_TYPE
-#define __NLINK_T_TYPE __U32_TYPE
-#define __OFF_T_TYPE __SLONGWORD_TYPE
+#define __NLINK_T_TYPE __UQUAD_TYPE
+#define __OFF_T_TYPE __SQUAD_TYPE
#define __OFF64_T_TYPE __SQUAD_TYPE
-#define __RLIM_T_TYPE __ULONGWORD_TYPE
+#define __RLIM_T_TYPE __UQUAD_TYPE
#define __RLIM64_T_TYPE __UQUAD_TYPE
-#define __BLKCNT_T_TYPE __SLONGWORD_TYPE
+#define __BLKCNT_T_TYPE __SQUAD_TYPE
#define __BLKCNT64_T_TYPE __SQUAD_TYPE
-#define __FSBLKCNT_T_TYPE __ULONGWORD_TYPE
+#define __FSBLKCNT_T_TYPE __UQUAD_TYPE
#define __FSBLKCNT64_T_TYPE __UQUAD_TYPE
-#define __FSFILCNT_T_TYPE __ULONGWORD_TYPE
+#define __FSFILCNT_T_TYPE __UQUAD_TYPE
#define __FSFILCNT64_T_TYPE __UQUAD_TYPE
-#define __FSWORD_T_TYPE __SWORD_TYPE
+#define __FSWORD_T_TYPE __SQUAD_TYPE
-#define __CLOCK_T_TYPE __SLONGWORD_TYPE
-#define __TIME_T_TYPE __SLONGWORD_TYPE
+#define __CLOCK_T_TYPE __SQUAD_TYPE
+#define __TIME_T_TYPE __SQUAD_TYPE
#define __USECONDS_T_TYPE __U32_TYPE
-#define __SUSECONDS_T_TYPE __SLONGWORD_TYPE
+#define __SUSECONDS_T_TYPE __SQUAD_TYPE
-#define __BLKSIZE_T_TYPE __S32_TYPE
+#define __BLKSIZE_T_TYPE __SQUAD_TYPE
#define __FSID_T_TYPE struct { int __val[2]; }
#define __SSIZE_T_TYPE __SWORD_TYPE
-#define __SYSCALL_SLONG_TYPE __SLONGWORD_TYPE
-#define __SYSCALL_ULONG_TYPE __ULONGWORD_TYPE
-#define __CPU_MASK_TYPE __ULONGWORD_TYPE
+#define __SYSCALL_SLONG_TYPE __SQUAD_TYPE
+#define __SYSCALL_ULONG_TYPE __UQUAD_TYPE
+#define __CPU_MASK_TYPE __UQUAD_TYPE
-#ifdef __LP64__
# define __RLIM_T_MATCHES_RLIM64_T 1
-#else
-# define __RLIM_T_MATCHES_RLIM64_T 0
-#endif
+#define __ASSUME_TIME64_SYSCALLS 1
+#define __ASSUME_RLIM64_SYSCALLS 1
Since the sysdeps/unix/sysv/linux/generic/bits/typesizes.h definitions
generally match the kernel, anything diverging from that has the potential
of breaking it, so the difference should probably be kept to the absolute
minimum.
I think these ones are wrong and will cause bugs similar to the clock_t
issue if they are used with kernel interfaces:
__NLINK_T_TYPE, __FSWORD_T_TYPE, __CLOCK_T_TYPE,
__BLKSIZE_T_TYPE, __SYSCALL_ULONG_TYPE,
__SYSCALL_SLONG_TYPE, __CPU_MASK_TYPE
These are fine as long as they are only used in user space and to
wrap kernel syscalls, but I think most of them can end up being
passed to the kernel, so it seems safer not to have rv32 diverge
without a good reason.
The remaining ones (__INO_T_TYPE, __OFF_T_TYPE, __BLKCNT_T_TYPE,
__FSBLKCNT_T_TYPE, __FSFILCNT_T_TYPE, __TIME_T_TYPE) all
follow the pattern where the kernel has an old 32-bit type and a new
64-bit type, but the kernel tries not to expose the 32-bit interfaces
to user space on new architectures and only provide the 64-bit
replacements, but there are a couple of interfaces that never got
replaced, typically in driver and file system ioctls.
Since glibc already has code to deal with the 64-bit types and that
is well tested, it would seem safer to me to just #undef the old
types completely rather than defining them to 64-bit, which would
make them incompatible with the kernel's types.
Arnd
On Thu, Jul 4, 2019 at 2:19 AM Arnd Bergmann <[email protected]> wrote:
>
> On Thu, Jul 4, 2019 at 9:20 AM Andreas Schwab <[email protected]> wrote:
> >
> > On Jul 03 2019, Alistair Francis <[email protected]> wrote:
> >
> > > On Wed, Jul 3, 2019 at 12:08 AM Andreas Schwab <[email protected]> wrote:
> > >>
> > >> On Jul 02 2019, Alistair Francis <[email protected]> wrote:
> > >>
> > >> > In the RISC-V 32-bit glibc port [1] the siginfo_t struct in the kernel
> > >> > doesn't line up with the struct in glibc. In glibc world the _sifields
> > >> > union is 8 byte alligned (although I can't figure out why)
> > >>
> > >> Try ptype/o in gdb.
> > >
> > > That's a useful tip, I'll be sure to use that next time.
> >
> > It was a serious note. If the structs don't line up then there is a
> > mismatch in types that cannot be solved by adding spurious padding. You
> > need to fix the types instead.
>
> Would it be an option to align all the basic typedefs (off_t, time_t,
> clock_t, ...)
> between glibc and kernel then, and just use the existing
> sysdeps/unix/sysv/linux/generic/bits/typesizes.h after all to avoid
> surprises like this?
>
> As of v2 the functional difference is
>
> -#define __INO_T_TYPE __ULONGWORD_TYPE
> +#define __INO_T_TYPE __UQUAD_TYPE
> #define __INO64_T_TYPE __UQUAD_TYPE
> #define __MODE_T_TYPE __U32_TYPE
> -#define __NLINK_T_TYPE __U32_TYPE
> -#define __OFF_T_TYPE __SLONGWORD_TYPE
> +#define __NLINK_T_TYPE __UQUAD_TYPE
> +#define __OFF_T_TYPE __SQUAD_TYPE
> #define __OFF64_T_TYPE __SQUAD_TYPE
> -#define __RLIM_T_TYPE __ULONGWORD_TYPE
> +#define __RLIM_T_TYPE __UQUAD_TYPE
> #define __RLIM64_T_TYPE __UQUAD_TYPE
> -#define __BLKCNT_T_TYPE __SLONGWORD_TYPE
> +#define __BLKCNT_T_TYPE __SQUAD_TYPE
> #define __BLKCNT64_T_TYPE __SQUAD_TYPE
> -#define __FSBLKCNT_T_TYPE __ULONGWORD_TYPE
> +#define __FSBLKCNT_T_TYPE __UQUAD_TYPE
> #define __FSBLKCNT64_T_TYPE __UQUAD_TYPE
> -#define __FSFILCNT_T_TYPE __ULONGWORD_TYPE
> +#define __FSFILCNT_T_TYPE __UQUAD_TYPE
> #define __FSFILCNT64_T_TYPE __UQUAD_TYPE
> -#define __FSWORD_T_TYPE __SWORD_TYPE
> +#define __FSWORD_T_TYPE __SQUAD_TYPE
> -#define __CLOCK_T_TYPE __SLONGWORD_TYPE
> -#define __TIME_T_TYPE __SLONGWORD_TYPE
> +#define __CLOCK_T_TYPE __SQUAD_TYPE
> +#define __TIME_T_TYPE __SQUAD_TYPE
> #define __USECONDS_T_TYPE __U32_TYPE
> -#define __SUSECONDS_T_TYPE __SLONGWORD_TYPE
> +#define __SUSECONDS_T_TYPE __SQUAD_TYPE
> -#define __BLKSIZE_T_TYPE __S32_TYPE
> +#define __BLKSIZE_T_TYPE __SQUAD_TYPE
> #define __FSID_T_TYPE struct { int __val[2]; }
> #define __SSIZE_T_TYPE __SWORD_TYPE
> -#define __SYSCALL_SLONG_TYPE __SLONGWORD_TYPE
> -#define __SYSCALL_ULONG_TYPE __ULONGWORD_TYPE
> -#define __CPU_MASK_TYPE __ULONGWORD_TYPE
> +#define __SYSCALL_SLONG_TYPE __SQUAD_TYPE
> +#define __SYSCALL_ULONG_TYPE __UQUAD_TYPE
> +#define __CPU_MASK_TYPE __UQUAD_TYPE
>
> -#ifdef __LP64__
> # define __RLIM_T_MATCHES_RLIM64_T 1
> -#else
> -# define __RLIM_T_MATCHES_RLIM64_T 0
> -#endif
>
> +#define __ASSUME_TIME64_SYSCALLS 1
> +#define __ASSUME_RLIM64_SYSCALLS 1
>
> Since the sysdeps/unix/sysv/linux/generic/bits/typesizes.h definitions
> generally match the kernel, anything diverging from that has the potential
> of breaking it, so the difference should probably be kept to the absolute
> minimum.
>
> I think these ones are wrong and will cause bugs similar to the clock_t
> issue if they are used with kernel interfaces:
> __NLINK_T_TYPE, __FSWORD_T_TYPE, __CLOCK_T_TYPE,
> __BLKSIZE_T_TYPE, __SYSCALL_ULONG_TYPE,
> __SYSCALL_SLONG_TYPE, __CPU_MASK_TYPE
>
> These are fine as long as they are only used in user space and to
> wrap kernel syscalls, but I think most of them can end up being
> passed to the kernel, so it seems safer not to have rv32 diverge
> without a good reason.
>
> The remaining ones (__INO_T_TYPE, __OFF_T_TYPE, __BLKCNT_T_TYPE,
> __FSBLKCNT_T_TYPE, __FSFILCNT_T_TYPE, __TIME_T_TYPE) all
> follow the pattern where the kernel has an old 32-bit type and a new
> 64-bit type, but the kernel tries not to expose the 32-bit interfaces
> to user space on new architectures and only provide the 64-bit
> replacements, but there are a couple of interfaces that never got
> replaced, typically in driver and file system ioctls.
>
> Since glibc already has code to deal with the 64-bit types and that
> is well tested, it would seem safer to me to just #undef the old
> types completely rather than defining them to 64-bit, which would
> make them incompatible with the kernel's types.
#undef-ing these results in build failures unfortunately, it seems
like they are required.
I'm sending a v3 RFC to the glibc list right now. We can continue the
discussion there.
Alistair
>
> Arnd