| Message ID | 20250305-riscv_optimize_entry-v5-4-6507b5dff3ce@rivosinc.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | entry: Move ret_from_fork() to C and inline syscall_exit_to_user_mode() | expand |
| Context | Check | Description |
|---|---|---|
| bjorn/pre-ci_am | success | Success |
| bjorn/build-rv32-defconfig | success | build-rv32-defconfig |
| bjorn/build-rv64-clang-allmodconfig | success | build-rv64-clang-allmodconfig |
| bjorn/build-rv64-gcc-allmodconfig | success | build-rv64-gcc-allmodconfig |
| bjorn/build-rv64-nommu-k210-defconfig | success | build-rv64-nommu-k210-defconfig |
| bjorn/build-rv64-nommu-k210-virt | success | build-rv64-nommu-k210-virt |
| bjorn/checkpatch | success | checkpatch |
| bjorn/dtb-warn-rv64 | success | dtb-warn-rv64 |
| bjorn/header-inline | success | header-inline |
| bjorn/kdoc | success | kdoc |
| bjorn/module-param | success | module-param |
| bjorn/verify-fixes | success | verify-fixes |
| bjorn/verify-signedoff | success | verify-signedoff |
On Wed, Mar 05 2025 at 16:43, Charlie Jenkins wrote: > Architectures using the generic entry code can be optimized by having > syscall_exit_to_user_mode inlined. That's a pretty handwavy claim. What's the actual benefit in numbers? Thanks, tglx
On Sat, Mar 08, 2025 at 01:37:30PM +0100, Thomas Gleixner wrote: > On Wed, Mar 05 2025 at 16:43, Charlie Jenkins wrote: > > Architectures using the generic entry code can be optimized by having > > syscall_exit_to_user_mode inlined. > > That's a pretty handwavy claim. What's the actual benefit in numbers? I put the numbers in the cover letter! Here is the data I gathered: 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] - Charlie > > Thanks, > > tglx
On Sat, Mar 08 2025 at 16:08, Charlie Jenkins wrote: > On Sat, Mar 08, 2025 at 01:37:30PM +0100, Thomas Gleixner wrote: >> On Wed, Mar 05 2025 at 16:43, Charlie Jenkins wrote: >> > Architectures using the generic entry code can be optimized by having >> > syscall_exit_to_user_mode inlined. >> >> That's a pretty handwavy claim. What's the actual benefit in numbers? > > I put the numbers in the cover letter! Here is the data I gathered: They should be in the change log as the cover letter is not part of what ends up in git.
On Sun, Mar 09, 2025 at 09:23:20AM +0100, Thomas Gleixner wrote: > On Sat, Mar 08 2025 at 16:08, Charlie Jenkins wrote: > > On Sat, Mar 08, 2025 at 01:37:30PM +0100, Thomas Gleixner wrote: > >> On Wed, Mar 05 2025 at 16:43, Charlie Jenkins wrote: > >> > Architectures using the generic entry code can be optimized by having > >> > syscall_exit_to_user_mode inlined. > >> > >> That's a pretty handwavy claim. What's the actual benefit in numbers? > > > > I put the numbers in the cover letter! Here is the data I gathered: > > They should be in the change log as the cover letter is not part of what > ends up in git. Yes good point. I will add this data here in the next version of the patch. - Charlie
Hi Charlie, On 06/03/2025 01:43, Charlie Jenkins wrote: > Architectures using the generic entry code can be optimized by having > syscall_exit_to_user_mode inlined. > > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> > --- > include/linux/entry-common.h | 43 ++++++++++++++++++++++++++++++++++++-- > kernel/entry/common.c | 49 +------------------------------------------- > 2 files changed, 42 insertions(+), 50 deletions(-) > > diff --git a/include/linux/entry-common.h b/include/linux/entry-common.h > index fc61d0205c97084acc89c8e45e088946f5e6d9b2..f94f3fdf15fc0091223cc9f7b823970302e67312 100644 > --- a/include/linux/entry-common.h > +++ b/include/linux/entry-common.h > @@ -14,6 +14,7 @@ > #include <linux/kmsan.h> > > #include <asm/entry-common.h> > +#include <asm/syscall.h> > > /* > * Define dummy _TIF work flags if not defined by the architecture or for > @@ -366,6 +367,15 @@ static __always_inline void exit_to_user_mode(void) > lockdep_hardirqs_on(CALLER_ADDR0); > } > > +/** > + * syscall_exit_work - Handle work before returning to user mode > + * @regs: Pointer to current pt_regs > + * @work: Current thread syscall work > + * > + * Do one-time syscall specific work. > + */ > +void syscall_exit_work(struct pt_regs *regs, unsigned long work); > + > /** > * syscall_exit_to_user_mode_work - Handle work before returning to user mode > * @regs: Pointer to currents pt_regs > @@ -379,7 +389,30 @@ static __always_inline void exit_to_user_mode(void) > * make the final state transitions. Interrupts must stay disabled between > * return from this function and the invocation of exit_to_user_mode(). > */ > -void syscall_exit_to_user_mode_work(struct pt_regs *regs); > +static __always_inline void syscall_exit_to_user_mode_work(struct pt_regs *regs) > +{ > + unsigned long work = READ_ONCE(current_thread_info()->syscall_work); > + unsigned long nr = syscall_get_nr(current, regs); > + > + CT_WARN_ON(ct_state() != CT_STATE_KERNEL); > + > + if (IS_ENABLED(CONFIG_PROVE_LOCKING)) { > + if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr)) > + local_irq_enable(); > + } > + > + rseq_syscall(regs); > + > + /* > + * Do one-time syscall specific work. If these work items are > + * enabled, we want to run them exactly once per syscall exit with > + * interrupts enabled. > + */ > + if (unlikely(work & SYSCALL_WORK_EXIT)) > + syscall_exit_work(regs, work); > + local_irq_disable_exit_to_user(); > + exit_to_user_mode_prepare(regs); > +} > > /** > * syscall_exit_to_user_mode - Handle work before returning to user mode > @@ -410,7 +443,13 @@ void syscall_exit_to_user_mode_work(struct pt_regs *regs); > * exit_to_user_mode(). This function is preferred unless there is a > * compelling architectural reason to use the separate functions. > */ > -void syscall_exit_to_user_mode(struct pt_regs *regs); > +static __always_inline void syscall_exit_to_user_mode(struct pt_regs *regs) > +{ > + instrumentation_begin(); > + syscall_exit_to_user_mode_work(regs); > + instrumentation_end(); > + exit_to_user_mode(); > +} > > /** > * irqentry_enter_from_user_mode - Establish state before invoking the irq handler > diff --git a/kernel/entry/common.c b/kernel/entry/common.c > index e33691d5adf7aab4af54cf2bf8e5ef5bd6ad1424..f55e421fb196dd5f9d4e34dd85ae096c774cf879 100644 > --- a/kernel/entry/common.c > +++ b/kernel/entry/common.c > @@ -146,7 +146,7 @@ static inline bool report_single_step(unsigned long work) > return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP; > } > > -static void syscall_exit_work(struct pt_regs *regs, unsigned long work) > +void syscall_exit_work(struct pt_regs *regs, unsigned long work) > { > bool step; > > @@ -173,53 +173,6 @@ static void syscall_exit_work(struct pt_regs *regs, unsigned long work) > ptrace_report_syscall_exit(regs, step); > } > > -/* > - * Syscall specific exit to user mode preparation. Runs with interrupts > - * enabled. > - */ > -static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs) > -{ > - unsigned long work = READ_ONCE(current_thread_info()->syscall_work); > - unsigned long nr = syscall_get_nr(current, regs); > - > - CT_WARN_ON(ct_state() != CT_STATE_KERNEL); > - > - if (IS_ENABLED(CONFIG_PROVE_LOCKING)) { > - if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr)) > - local_irq_enable(); > - } > - > - rseq_syscall(regs); > - > - /* > - * Do one-time syscall specific work. If these work items are > - * enabled, we want to run them exactly once per syscall exit with > - * interrupts enabled. > - */ > - if (unlikely(work & SYSCALL_WORK_EXIT)) > - syscall_exit_work(regs, work); > -} > - > -static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs) > -{ > - syscall_exit_to_user_mode_prepare(regs); > - local_irq_disable_exit_to_user(); > - exit_to_user_mode_prepare(regs); > -} > - > -void syscall_exit_to_user_mode_work(struct pt_regs *regs) > -{ > - __syscall_exit_to_user_mode_work(regs); > -} > - > -__visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs) > -{ > - instrumentation_begin(); > - __syscall_exit_to_user_mode_work(regs); > - instrumentation_end(); > - exit_to_user_mode(); > -} > - > noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs) > { > enter_from_user_mode(regs); > Reviewed-by: Alexandre Ghiti <alexghiti@rivosinc.com> Thanks, Alex
diff --git a/include/linux/entry-common.h b/include/linux/entry-common.h index fc61d0205c97084acc89c8e45e088946f5e6d9b2..f94f3fdf15fc0091223cc9f7b823970302e67312 100644 --- a/include/linux/entry-common.h +++ b/include/linux/entry-common.h @@ -14,6 +14,7 @@ #include <linux/kmsan.h> #include <asm/entry-common.h> +#include <asm/syscall.h> /* * Define dummy _TIF work flags if not defined by the architecture or for @@ -366,6 +367,15 @@ static __always_inline void exit_to_user_mode(void) lockdep_hardirqs_on(CALLER_ADDR0); } +/** + * syscall_exit_work - Handle work before returning to user mode + * @regs: Pointer to current pt_regs + * @work: Current thread syscall work + * + * Do one-time syscall specific work. + */ +void syscall_exit_work(struct pt_regs *regs, unsigned long work); + /** * syscall_exit_to_user_mode_work - Handle work before returning to user mode * @regs: Pointer to currents pt_regs @@ -379,7 +389,30 @@ static __always_inline void exit_to_user_mode(void) * make the final state transitions. Interrupts must stay disabled between * return from this function and the invocation of exit_to_user_mode(). */ -void syscall_exit_to_user_mode_work(struct pt_regs *regs); +static __always_inline void syscall_exit_to_user_mode_work(struct pt_regs *regs) +{ + unsigned long work = READ_ONCE(current_thread_info()->syscall_work); + unsigned long nr = syscall_get_nr(current, regs); + + CT_WARN_ON(ct_state() != CT_STATE_KERNEL); + + if (IS_ENABLED(CONFIG_PROVE_LOCKING)) { + if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr)) + local_irq_enable(); + } + + rseq_syscall(regs); + + /* + * Do one-time syscall specific work. If these work items are + * enabled, we want to run them exactly once per syscall exit with + * interrupts enabled. + */ + if (unlikely(work & SYSCALL_WORK_EXIT)) + syscall_exit_work(regs, work); + local_irq_disable_exit_to_user(); + exit_to_user_mode_prepare(regs); +} /** * syscall_exit_to_user_mode - Handle work before returning to user mode @@ -410,7 +443,13 @@ void syscall_exit_to_user_mode_work(struct pt_regs *regs); * exit_to_user_mode(). This function is preferred unless there is a * compelling architectural reason to use the separate functions. */ -void syscall_exit_to_user_mode(struct pt_regs *regs); +static __always_inline void syscall_exit_to_user_mode(struct pt_regs *regs) +{ + instrumentation_begin(); + syscall_exit_to_user_mode_work(regs); + instrumentation_end(); + exit_to_user_mode(); +} /** * irqentry_enter_from_user_mode - Establish state before invoking the irq handler diff --git a/kernel/entry/common.c b/kernel/entry/common.c index e33691d5adf7aab4af54cf2bf8e5ef5bd6ad1424..f55e421fb196dd5f9d4e34dd85ae096c774cf879 100644 --- a/kernel/entry/common.c +++ b/kernel/entry/common.c @@ -146,7 +146,7 @@ static inline bool report_single_step(unsigned long work) return work & SYSCALL_WORK_SYSCALL_EXIT_TRAP; } -static void syscall_exit_work(struct pt_regs *regs, unsigned long work) +void syscall_exit_work(struct pt_regs *regs, unsigned long work) { bool step; @@ -173,53 +173,6 @@ static void syscall_exit_work(struct pt_regs *regs, unsigned long work) ptrace_report_syscall_exit(regs, step); } -/* - * Syscall specific exit to user mode preparation. Runs with interrupts - * enabled. - */ -static void syscall_exit_to_user_mode_prepare(struct pt_regs *regs) -{ - unsigned long work = READ_ONCE(current_thread_info()->syscall_work); - unsigned long nr = syscall_get_nr(current, regs); - - CT_WARN_ON(ct_state() != CT_STATE_KERNEL); - - if (IS_ENABLED(CONFIG_PROVE_LOCKING)) { - if (WARN(irqs_disabled(), "syscall %lu left IRQs disabled", nr)) - local_irq_enable(); - } - - rseq_syscall(regs); - - /* - * Do one-time syscall specific work. If these work items are - * enabled, we want to run them exactly once per syscall exit with - * interrupts enabled. - */ - if (unlikely(work & SYSCALL_WORK_EXIT)) - syscall_exit_work(regs, work); -} - -static __always_inline void __syscall_exit_to_user_mode_work(struct pt_regs *regs) -{ - syscall_exit_to_user_mode_prepare(regs); - local_irq_disable_exit_to_user(); - exit_to_user_mode_prepare(regs); -} - -void syscall_exit_to_user_mode_work(struct pt_regs *regs) -{ - __syscall_exit_to_user_mode_work(regs); -} - -__visible noinstr void syscall_exit_to_user_mode(struct pt_regs *regs) -{ - instrumentation_begin(); - __syscall_exit_to_user_mode_work(regs); - instrumentation_end(); - exit_to_user_mode(); -} - noinstr void irqentry_enter_from_user_mode(struct pt_regs *regs) { enter_from_user_mode(regs);
Architectures using the generic entry code can be optimized by having syscall_exit_to_user_mode inlined. Signed-off-by: Charlie Jenkins <charlie@rivosinc.com> --- include/linux/entry-common.h | 43 ++++++++++++++++++++++++++++++++++++-- kernel/entry/common.c | 49 +------------------------------------------- 2 files changed, 42 insertions(+), 50 deletions(-)