| Message ID | 20220628081707.1997728-2-guoren@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | riscv: Add qspinlock support with combo style | expand |
On 6/28/22 04:17, guoren@kernel.org wrote: > From: Guo Ren <guoren@linux.alibaba.com> > > Remove unnecessary atomic_read in arch_spin_value_unlocked(lock), > because the value has been in lock. This patch could prevent > arch_spin_value_unlocked contend spin_lock data again. > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > Signed-off-by: Guo Ren <guoren@kernel.org> > Cc: Peter Zijlstra (Intel) <peterz@infradead.org> > Cc: Arnd Bergmann <arnd@arndb.de> > Cc: Palmer Dabbelt <palmer@rivosinc.com> > --- > include/asm-generic/spinlock.h | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > index fdfebcb050f4..f1e4fa100f5a 100644 > --- a/include/asm-generic/spinlock.h > +++ b/include/asm-generic/spinlock.h > @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > > static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > { > - return !arch_spin_is_locked(&lock); > + u32 val = lock.counter; > + > + return ((val >> 16) == (val & 0xffff)); > } > > #include <asm/qrwlock.h> lockref.c is the only current user of arch_spin_value_unlocked(). This change is probably OK with this particular use case. Do you have any performance data about the improvement due to this change? You may have to document that we have to revisit that if another use case shows up. Cheers, Longman
On Wed, Jun 29, 2022 at 2:06 AM Waiman Long <longman@redhat.com> wrote: > > On 6/28/22 04:17, guoren@kernel.org wrote: > > From: Guo Ren <guoren@linux.alibaba.com> > > > > Remove unnecessary atomic_read in arch_spin_value_unlocked(lock), > > because the value has been in lock. This patch could prevent > > arch_spin_value_unlocked contend spin_lock data again. > > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > Signed-off-by: Guo Ren <guoren@kernel.org> > > Cc: Peter Zijlstra (Intel) <peterz@infradead.org> > > Cc: Arnd Bergmann <arnd@arndb.de> > > Cc: Palmer Dabbelt <palmer@rivosinc.com> > > --- > > include/asm-generic/spinlock.h | 4 +++- > > 1 file changed, 3 insertions(+), 1 deletion(-) > > > > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > > index fdfebcb050f4..f1e4fa100f5a 100644 > > --- a/include/asm-generic/spinlock.h > > +++ b/include/asm-generic/spinlock.h > > @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > > > > static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > > { > > - return !arch_spin_is_locked(&lock); > > + u32 val = lock.counter; > > + > > + return ((val >> 16) == (val & 0xffff)); > > } > > > > #include <asm/qrwlock.h> > > lockref.c is the only current user of arch_spin_value_unlocked(). This > change is probably OK with this particular use case. Do you have any > performance data about the improvement due to this change? I don't have performance data and I just check the asm code, previous version has an additional unnecessary atomic_read. About this point, we've talked before, but I & palmer missed that point when we pick peter's patch again. https://lore.kernel.org/linux-riscv/YHbmXXvuG442ZDfN@hirez.programming.kicks-ass.net/ ---- > > +static __always_inline int ticket_value_unlocked(arch_spinlock_t lock) > > +{ > > + return !ticket_is_locked(&lock); > Are you sure to let ticket_is_locked->atomic_read(lock) again, the > lock has contained all information? > > return lock.tickets.owner == lock.tickets.next; Yeah, I wrote then the wrong way around. Couldn't be bothered to go back when I figured it out. --- It's just a small typo. > > You may have to document that we have to revisit that if another use > case shows up. > > Cheers, > Longman > -- Best Regards Guo Ren ML: https://lore.kernel.org/linux-csky/
From: guoren@kernel.org > Sent: 28 June 2022 09:17 > > From: Guo Ren <guoren@linux.alibaba.com> > > Remove unnecessary atomic_read in arch_spin_value_unlocked(lock), > because the value has been in lock. This patch could prevent > arch_spin_value_unlocked contend spin_lock data again. I'm confused (as usual). Isn't atomic_read() pretty much free? .. > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > index fdfebcb050f4..f1e4fa100f5a 100644 > --- a/include/asm-generic/spinlock.h > +++ b/include/asm-generic/spinlock.h > @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > > static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > { > - return !arch_spin_is_locked(&lock); > + u32 val = lock.counter; > + > + return ((val >> 16) == (val & 0xffff)); That almost certainly needs a READ_ONCE(). The result is also inherently stale. So the uses must be pretty limited. David > } > > #include <asm/qrwlock.h> > -- > 2.36.1 - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On Wed, Jun 29, 2022 at 4:27 PM David Laight <David.Laight@aculab.com> wrote: > > From: guoren@kernel.org > > Sent: 28 June 2022 09:17 > > > > From: Guo Ren <guoren@linux.alibaba.com> > > > > Remove unnecessary atomic_read in arch_spin_value_unlocked(lock), > > because the value has been in lock. This patch could prevent > > arch_spin_value_unlocked contend spin_lock data again. > > I'm confused (as usual). > Isn't atomic_read() pretty much free? When a cache line is shared with multi-harts, not as free as you think. Preventing touching contended data is the basic principle. atomic_read in alpha is heavy, It could be a potential user of ticket-lock. > > .. > > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > > index fdfebcb050f4..f1e4fa100f5a 100644 > > --- a/include/asm-generic/spinlock.h > > +++ b/include/asm-generic/spinlock.h > > @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > > > > static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > > { > > - return !arch_spin_is_locked(&lock); > > + u32 val = lock.counter; > > + > > + return ((val >> 16) == (val & 0xffff)); > > That almost certainly needs a READ_ONCE(). > > The result is also inherently stale. > So the uses must be pretty limited. The previous read_once could get 64bit, use the API to check the 32bit atomic data part. > > David > > > } > > > > #include <asm/qrwlock.h> > > -- > > 2.36.1 > > - > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > Registration No: 1397386 (Wales) >
On Tue, Jun 28, 2022 at 04:17:03AM -0400, guoren@kernel.org wrote: > From: Guo Ren <guoren@linux.alibaba.com> > > Remove unnecessary atomic_read in arch_spin_value_unlocked(lock), > because the value has been in lock. This patch could prevent > arch_spin_value_unlocked contend spin_lock data again. > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > Signed-off-by: Guo Ren <guoren@kernel.org> > Cc: Peter Zijlstra (Intel) <peterz@infradead.org> > Cc: Arnd Bergmann <arnd@arndb.de> > Cc: Palmer Dabbelt <palmer@rivosinc.com> > --- > include/asm-generic/spinlock.h | 4 +++- > 1 file changed, 3 insertions(+), 1 deletion(-) > > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > index fdfebcb050f4..f1e4fa100f5a 100644 > --- a/include/asm-generic/spinlock.h > +++ b/include/asm-generic/spinlock.h > @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > > static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > { > - return !arch_spin_is_locked(&lock); > + u32 val = lock.counter; > + > + return ((val >> 16) == (val & 0xffff)); > } Wouldn't the right thing be to flip arch_spin_is_locked() and arch_spin_value_is_unlocked() ? diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h index fdfebcb050f4..63ab4da262f2 100644 --- a/include/asm-generic/spinlock.h +++ b/include/asm-generic/spinlock.h @@ -68,23 +68,25 @@ static __always_inline void arch_spin_unlock(arch_spinlock_t *lock) smp_store_release(ptr, (u16)val + 1); } -static __always_inline int arch_spin_is_locked(arch_spinlock_t *lock) +static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) { u32 val = atomic_read(lock); - return ((val >> 16) != (val & 0xffff)); + return (s16)((val >> 16) - (val & 0xffff)) > 1; } -static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) +static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) { - u32 val = atomic_read(lock); + u32 val = lock.counter; - return (s16)((val >> 16) - (val & 0xffff)) > 1; + return ((val >> 16) == (val & 0xffff)); } -static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) +static __always_inline int arch_spin_is_locked(arch_spinlock_t *lock) { - return !arch_spin_is_locked(&lock); + arch_spinlock_t val = READ_ONCE(*lock); + + return !arch_spin_value_unlocked(val); } #include <asm/qrwlock.h>
On Mon, Jul 4, 2022 at 5:52 PM Peter Zijlstra <peterz@infradead.org> wrote: > > On Tue, Jun 28, 2022 at 04:17:03AM -0400, guoren@kernel.org wrote: > > From: Guo Ren <guoren@linux.alibaba.com> > > > > Remove unnecessary atomic_read in arch_spin_value_unlocked(lock), > > because the value has been in lock. This patch could prevent > > arch_spin_value_unlocked contend spin_lock data again. > > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > Signed-off-by: Guo Ren <guoren@kernel.org> > > Cc: Peter Zijlstra (Intel) <peterz@infradead.org> > > Cc: Arnd Bergmann <arnd@arndb.de> > > Cc: Palmer Dabbelt <palmer@rivosinc.com> > > --- > > include/asm-generic/spinlock.h | 4 +++- > > 1 file changed, 3 insertions(+), 1 deletion(-) > > > > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > > index fdfebcb050f4..f1e4fa100f5a 100644 > > --- a/include/asm-generic/spinlock.h > > +++ b/include/asm-generic/spinlock.h > > @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > > > > static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > > { > > - return !arch_spin_is_locked(&lock); > > + u32 val = lock.counter; > > + > > + return ((val >> 16) == (val & 0xffff)); > > } > > Wouldn't the right thing be to flip arch_spin_is_locked() and > arch_spin_value_is_unlocked() ? Okay, I agree with your patch. Next version, I would take the below code. > > > diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h > index fdfebcb050f4..63ab4da262f2 100644 > --- a/include/asm-generic/spinlock.h > +++ b/include/asm-generic/spinlock.h > @@ -68,23 +68,25 @@ static __always_inline void arch_spin_unlock(arch_spinlock_t *lock) > smp_store_release(ptr, (u16)val + 1); > } > > -static __always_inline int arch_spin_is_locked(arch_spinlock_t *lock) > +static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > { > u32 val = atomic_read(lock); > > - return ((val >> 16) != (val & 0xffff)); > + return (s16)((val >> 16) - (val & 0xffff)) > 1; > } > > -static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) > +static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > { > - u32 val = atomic_read(lock); > + u32 val = lock.counter; > > - return (s16)((val >> 16) - (val & 0xffff)) > 1; > + return ((val >> 16) == (val & 0xffff)); > } > > -static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) > +static __always_inline int arch_spin_is_locked(arch_spinlock_t *lock) > { > - return !arch_spin_is_locked(&lock); > + arch_spinlock_t val = READ_ONCE(*lock); > + > + return !arch_spin_value_unlocked(val); > } > > #include <asm/qrwlock.h> >
diff --git a/include/asm-generic/spinlock.h b/include/asm-generic/spinlock.h index fdfebcb050f4..f1e4fa100f5a 100644 --- a/include/asm-generic/spinlock.h +++ b/include/asm-generic/spinlock.h @@ -84,7 +84,9 @@ static __always_inline int arch_spin_is_contended(arch_spinlock_t *lock) static __always_inline int arch_spin_value_unlocked(arch_spinlock_t lock) { - return !arch_spin_is_locked(&lock); + u32 val = lock.counter; + + return ((val >> 16) == (val & 0xffff)); } #include <asm/qrwlock.h>