Re: [PATCH v2 09/10] target/i386: implement 32-bit SYSENTER for linux-user

[Richard Henderson]

On 6/20/23 17:16, Paolo Bonzini wrote:
> TCG reports the SEP feature (SYSENTER/SYSEXIT) in user mode emulation,
> but does not plumb it into the linux-user run loop.  Split the helper into
> system emulation and user-mode emulation cases and implement the latter.
>
> SYSENTER does not have the best design for a kernel-mode entry
> instruction, and therefore Linux always makes it return to the
> vsyscall page.  Because QEMU does not provide the_contents_  of
> the vsyscall page, the instructions executed after SYSEXIT have
> to be emulated by hand until the first RET.
>
> Some corner cases, such as restarting the system call after the
> system call has rewritten the SYSENTER instruction, are not emulated
> correctly.  On Linux, the system call restart uses the SYSENTER
> call in the vsyscall page, while on QEMU it uses the emulated
> program's instruction.
>
> Signed-off-by: Paolo Bonzini<pbonzini@redhat.com>
> ---
>   linux-user/i386/cpu_loop.c          | 51 +++++++++++++++++++++++++++--
>   target/i386/cpu.c                   |  9 ++++-
>   target/i386/cpu.h                   |  1 +
>   target/i386/helper.h                |  2 +-
>   target/i386/tcg/seg_helper.c        | 33 -------------------
>   target/i386/tcg/sysemu/seg_helper.c | 33 +++++++++++++++++++
>   target/i386/tcg/translate.c         |  2 +-
>   target/i386/tcg/user/seg_helper.c   | 16 +++++++++
>   8 files changed, 109 insertions(+), 38 deletions(-)

I'm not keen on this.

This belongs with the rest of the vdso (see patches posted years ago; committing binary 
blobs rejected, still waiting on a decent way to invoke cross-compilers to build them).

Further, this shouldn't ever be reachable, because AT_SYSINFO won't be present to give the 
guest libc the location of the vdso routine to call.


r~