Re: [PATCH v3 05/11] x86/entry: Convert ret_from_fork to C

[Brian Gerst]

On Thu, Mar 4, 2021 at 2:16 PM Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>
> ret_from_fork is written in asm, slightly differently, for x86_32 and
> x86_64.  Convert it to C.
>
> This is a straight conversion without any particular cleverness.  As a
> further cleanup, the code that sets up the ret_from_fork argument registers
> could be adjusted to put the arguments in the correct registers.

An alternative would be to stash the function pointer and argument in
the pt_regs of the new kthread task, and test for PF_KTHREAD instead.
That would eliminate the need to pass those two values to the C
version of ret_from_fork().

> This will cause the ORC unwinder to find pt_regs even for kernel threads on
> x86_64.  This seems harmless.
>
> The 32-bit comment above the now-deleted schedule_tail_wrapper was
> obsolete: the encode_frame_pointer mechanism (see copy_thread()) solves the
> same problem more cleanly.
>
> Cc: Josh Poimboeuf <jpoimboe@xxxxxxxxxx>
> Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
> ---
>  arch/x86/entry/common.c          | 23 ++++++++++++++
>  arch/x86/entry/entry_32.S        | 51 +++++---------------------------
>  arch/x86/entry/entry_64.S        | 33 +++++----------------
>  arch/x86/include/asm/switch_to.h |  2 +-
>  arch/x86/kernel/process.c        |  2 +-
>  arch/x86/kernel/process_32.c     |  2 +-
>  arch/x86/kernel/unwind_orc.c     |  2 +-
>  7 files changed, 43 insertions(+), 72 deletions(-)
>
> diff --git a/arch/x86/entry/common.c b/arch/x86/entry/common.c
> index 95776f16c1cb..ef1c65938a6b 100644
> --- a/arch/x86/entry/common.c
> +++ b/arch/x86/entry/common.c
> @@ -214,6 +214,29 @@ SYSCALL_DEFINE0(ni_syscall)
>         return -ENOSYS;
>  }
>
> +void ret_from_fork(struct task_struct *prev,
> +                  int (*kernel_thread_fn)(void *),
> +                  void *kernel_thread_arg,
> +                  struct pt_regs *user_regs);
> +
> +__visible void noinstr ret_from_fork(struct task_struct *prev,
> +                                    int (*kernel_thread_fn)(void *),
> +                                    void *kernel_thread_arg,
> +                                    struct pt_regs *user_regs)
> +{
> +       instrumentation_begin();
> +
> +       schedule_tail(prev);
> +
> +       if (kernel_thread_fn) {

This should have an unlikely(), as kernel threads should be the rare case.

--
Brian Gerst