Similar to commit 221a164035fd ("entry: Move
syscall_enter_from_user_mode() to header file"), move
syscall_exit_to_user_mode() to the header file as well.

Testing was done with the byte-unixbench [1] syscall benchmark (which
calls getpid) and QEMU. On riscv I measured a 7.09246% improvement, on
x86 a 2.98843% improvement, on loongarch a 6.07954% improvement, and on
s390 a 11.1328% improvement.

The Intel bot also reported "kernel test robot noticed a 1.9%
improvement of stress-ng.seek.ops_per_sec" [2]

Since this is on QEMU, I know these numbers are not perfect, but they
show a trend of general improvement across all architectures that use
the generic entry code.

[1] https://github.com/kdlucas/byte-unixbench
[2] https://lore.kernel.org/linux-riscv/202502051555.85ae6844-lkp@intel.com/

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
Changes in v6:
- Update commit message of patch 4 to contain performance numbers
- Link to v5: https://lore.kernel.org/r/20250305-riscv_optimize_entry-v5-0-6507b5dff3ce@rivosinc.com

Changes in v5:
- Rebase on 6.14-rc5
- Link to v4: https://lore.kernel.org/r/20250127-riscv_optimize_entry-v4-0-868cf7702dc9@rivosinc.com

Changes in v4:
- I had messed up warning for ct_state() on rebase, correct that issue
- Link to v3: https://lore.kernel.org/r/20250124-riscv_optimize_entry-v3-0-869f36b9e43b@rivosinc.com

Changes in v3:
- Fixup comment to properly reflect args (Alex)
- Fix prototypes for loongarch (Huacai)
- Link to v2: https://lore.kernel.org/r/20250123-riscv_optimize_entry-v2-0-7c259492d508@rivosinc.com

Changes in v2:
- Fixup compilation issues for loongarch
- Fixup compilation issues with CONFIG_CONTEXT_TRACKING_USER
- Link to v1: https://lore.kernel.org/r/20250122-riscv_optimize_entry-v1-0-4ee95559cfd0@rivosinc.com

---
Charlie Jenkins (4):
      riscv: entry: Convert ret_from_fork() to C
      riscv: entry: Split ret_from_fork() into user and kernel
      LoongArch: entry: Migrate ret_from_fork() to C
      entry: Inline syscall_exit_to_user_mode()

 arch/loongarch/include/asm/asm-prototypes.h |  8 +++++
 arch/loongarch/kernel/entry.S               | 22 ++++++-------
 arch/loongarch/kernel/process.c             | 33 +++++++++++++++----
 arch/riscv/include/asm/asm-prototypes.h     |  2 ++
 arch/riscv/kernel/entry.S                   | 20 +++++++-----
 arch/riscv/kernel/process.c                 | 21 +++++++++++--
 include/linux/entry-common.h                | 43 +++++++++++++++++++++++--
 kernel/entry/common.c                       | 49 +----------------------------
 8 files changed, 119 insertions(+), 79 deletions(-)
---
base-commit: 7eb172143d5508b4da468ed59ee857c6e5e01da6
change-id: 20240402-riscv_optimize_entry-583843420325
-- 
- Charlie

On Thu, Mar 20 2025 at 10:29, Charlie Jenkins wrote:
> Similar to commit 221a164035fd ("entry: Move
> syscall_enter_from_user_mode() to header file"), move
> syscall_exit_to_user_mode() to the header file as well.
>
> Testing was done with the byte-unixbench [1] syscall benchmark (which
> calls getpid) and QEMU. On riscv I measured a 7.09246% improvement, on
> x86 a 2.98843% improvement, on loongarch a 6.07954% improvement, and on
> s390 a 11.1328% improvement.
>
> The Intel bot also reported "kernel test robot noticed a 1.9%
> improvement of stress-ng.seek.ops_per_sec" [2]
>
> Since this is on QEMU, I know these numbers are not perfect, but they
> show a trend of general improvement across all architectures that use
> the generic entry code.

This looks sane now. I've bookmarked it as post-merge window material.

Thanks,

        tglx

On Fri, Mar 21, 2025 at 10:22:36PM +0100, Thomas Gleixner wrote:
> On Thu, Mar 20 2025 at 10:29, Charlie Jenkins wrote:
> > Similar to commit 221a164035fd ("entry: Move
> > syscall_enter_from_user_mode() to header file"), move
> > syscall_exit_to_user_mode() to the header file as well.
> >
> > Testing was done with the byte-unixbench [1] syscall benchmark (which
> > calls getpid) and QEMU. On riscv I measured a 7.09246% improvement, on
> > x86 a 2.98843% improvement, on loongarch a 6.07954% improvement, and on
> > s390 a 11.1328% improvement.
> >
> > The Intel bot also reported "kernel test robot noticed a 1.9%
> > improvement of stress-ng.seek.ops_per_sec" [2]
> >
> > Since this is on QEMU, I know these numbers are not perfect, but they
> > show a trend of general improvement across all architectures that use
> > the generic entry code.
> 
> This looks sane now. I've bookmarked it as post-merge window material.

Has this been applied somewhere?

- Charlie

> 
> Thanks,
> 
>         tglx

On Mon, Apr 28 2025 at 14:54, Charlie Jenkins wrote:
> On Fri, Mar 21, 2025 at 10:22:36PM +0100, Thomas Gleixner wrote:
>> 
>> This looks sane now. I've bookmarked it as post-merge window material.
>
> Has this been applied somewhere?

Thanks for the nudge. It indeed fell through the cracks. Tending to it
now.

Thanks,

        tglx