Re: [PATCH 3/5] riscv: Add the clone3 wrapper

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 29/05/23 11:04, Palmer Dabbelt wrote:
> On Fri, 03 Feb 2023 09:12:35 PST (-0800), adhemerval.zanella@linaro.org wrote:
>> It follows the internal signature:
>>
>>   extern int clone3 (struct clone_args *__cl_args, size_t __size,
>>  int (*__func) (void *__arg), void *__arg);
>>
>> Checked on riscv64-linux-gnu-rv64imafdc-lp64d.
> 
> Due to a bug we only had clone3 on rv64 until 59a4e0d5511b ("RISC-V: Include clone3() on rv32"), which landed in 5.15 and was then backported all the way to 5.4.
> 
>> ---
>>  sysdeps/unix/sysv/linux/riscv/clone3.S | 80 ++++++++++++++++++++++++++
>>  sysdeps/unix/sysv/linux/riscv/sysdep.h |  1 +
>>  2 files changed, 81 insertions(+)
>>  create mode 100644 sysdeps/unix/sysv/linux/riscv/clone3.S
>>
>> diff --git a/sysdeps/unix/sysv/linux/riscv/clone3.S b/sysdeps/unix/sysv/linux/riscv/clone3.S
>> new file mode 100644
>> index 0000000000..ee5780ee2f
>> --- /dev/null
>> +++ b/sysdeps/unix/sysv/linux/riscv/clone3.S
>> @@ -0,0 +1,80 @@
>> +/* The clone3 syscall wrapper.  Linux/RISC-V version.
>> +   Copyright (C) 2023 Free Software Foundation, Inc.
>> +
>> +   This file is part of the GNU C Library.
>> +
>> +   The GNU C Library is free software; you can redistribute it and/or
>> +   modify it under the terms of the GNU Lesser General Public
>> +   License as published by the Free Software Foundation; either
>> +   version 2.1 of the License, or (at your option) any later version.
>> +
>> +   The GNU C Library is distributed in the hope that it will be useful,
>> +   but WITHOUT ANY WARRANTY; without even the implied warranty of
>> +   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
>> +   Lesser General Public License for more details.
>> +
>> +   You should have received a copy of the GNU Lesser General Public
>> +   License along with the GNU C Library; if not, see
>> +   <https://www.gnu.org/licenses/>.  */
>> +
>> +#include <asm/errno.h>
>> +#include <sys/asm.h>
>> +#include <sysdep.h>
>> +
>> +/* The userland implementation is:
>> +   int clone3 (struct clone_args *cl_args, size_t size,
>> +               int (*func)(void *arg), void *arg);
>> +
>> +   the kernel entry is:
>> +   int clone3 (struct clone_args *cl_args, size_t size);
>> +
>> +   The parameters are passed in registers from userland:
>> +   a0: cl_args
>> +   a1: size
>> +   a2: func
>> +   a3: arg  */
>> +
>> +        .text
>> +ENTRY(__clone3)
>> +    /* Sanity check args.  */
>> +    beqz    a0, L(invalid)    /* No NULL cl_args pointer.  */
>> +    beqz    a2, L(invalid)  /* No NULL function pointer.  */
>> +
>> +    /* Do the system call, the kernel expects:
>> +       a7: system call number
>> +       a0: cl_args
>> +       a1: size  */
>> +    li    a7, __NR_clone3
> 
> IIUC this only builds with kernel headers 5.15 or newer on rv32.  I'm not sure if that's a problem: unless I'm missing something we support building for those systems now, but rv32 is pretty esoteric so I'm not sure anyone cares -- maybe someone uses 5.15 on rv32 (some rv64 hardware is still being released with 5.10-based vendor kernels), but 5.4 is pretty old for us.

The __NR_XXX macros are now provided by arch-syscall.h, which in turn is generated
by the kernel header. And we already have clone3 for rv32 since 3387c40a8bb.

> 
>> +    scall
>> +
>> +    bltz    a0, L(error)
>> +    beqz    a0, L(thread_start)
>> +
>> +    ret
>> +
>> +L(invalid):
>> +    li    a0, -EINVAL
>> +L(error):
>> +    tail    __syscall_error
>> +END (__clone3)
>> +
>> +ENTRY(__thread_start_clone3)
>> +L(thread_start):
>> +    /* Terminate call stack by noting ra is undefined.  Use a dummy
>> +       .cfi_label to force starting the FDE.  */
>> +    .cfi_label .Ldummy
>> +    cfi_undefined (ra)
>> +
>> +    /* Restore the arg for user's function and call the user's
>> +       function.  */
>> +    mv        a1, a2    /* Function pointer.  */
> 
> I think we don't need that mv, we can just call a2 directly?  Unless I'm missing some reason the thread main needs to see a pointer to itself?

We can, I will change it.

> 
>> +    mv        a0, a3    /* Argument pointer.  */
>> +    jalr        a1
>> +
>> +    /* Call exit with the function's return value.  */
>> +    li        a7, __NR_exit
>> +    scall
>> +END(__thread_start_clone3)
>> +
>> +libc_hidden_def (__clone3)
>> +weak_alias (__clone3, clone3)
>> diff --git a/sysdeps/unix/sysv/linux/riscv/sysdep.h b/sysdeps/unix/sysv/linux/riscv/sysdep.h
>> index 4af5fe5dbc..e96a930409 100644
>> --- a/sysdeps/unix/sysv/linux/riscv/sysdep.h
>> +++ b/sysdeps/unix/sysv/linux/riscv/sysdep.h
>> @@ -151,6 +151,7 @@
>>
>>  /* RV32 does not support the gettime VDSO syscalls.  */
>>  # endif
>> +# define HAVE_CLONE3_WRAPPER        1
>>
>>  /* List of system calls which are supported as vsyscalls (for RV32 and
>>     RV64).  */
> 
> That's just some minor comments though, so
> 
> Reviewed-by: Palmer Dabbelt <palmer@rivosinc.com>

Thanks.

> 
> Thanks! (and sorry I missed it for a bit)