Message ID | 20230910082911.3378782-6-guoren@kernel.org (mailing list archive) |
---|---|
State | Changes Requested |
Headers | show |
Series | riscv: Add Native/Paravirt qspinlock support | expand |
Context | Check | Description |
---|---|---|
conchuod/tree_selection | fail | Failed to apply to next/pending-fixes, riscv/for-next or riscv/master |
On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > From: Guo Ren <guoren@linux.alibaba.com> > > The requirements of qspinlock have been documented by commit: > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > atomics"). > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > doesn't satisfy the requirements of qspinlock above, it won't prevent > some riscv vendors from implementing a strong fwd guarantee LR/SC in > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > is the one. > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > queued_spinlock 0.5109/0.00 > ticket_spinlock 0.5814/0.00 > > perf futex/hash (+6.7%): > queued_spinlock 1444393 operations/sec (+- 0.09%) > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > perf futex/wake-parallel (+8.6%): > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > perf futex/requeue (+4.2%): > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > System Benchmarks (+6.4%) > queued_spinlock: > System Benchmarks Index Values BASELINE RESULT INDEX > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > Double-Precision Whetstone 55.0 182422.8 33167.8 > Execl Throughput 43.0 13116.6 3050.4 > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > Pipe Throughput 12440.0 23058600.5 18535.9 > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > Process Creation 126.0 12537.3 995.0 > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > System Call Overhead 15000.0 33308301.3 22205.5 > ======== > System Benchmarks Index Score 12426.1 > > ticket_spinlock: > System Benchmarks Index Values BASELINE RESULT INDEX > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > Double-Precision Whetstone 55.0 181921.0 33076.5 > Execl Throughput 43.0 12625.1 2936.1 > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > Pipe Throughput 12440.0 20594018.7 16554.7 > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > Process Creation 126.0 10798.4 857.0 > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > System Call Overhead 15000.0 30766778.4 20511.2 > ======== > System Benchmarks Index Score 11670.7 > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > cores platform than the ticket_lock. > > Signed-off-by: Guo Ren <guoren@kernel.org> > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > --- > arch/riscv/Kconfig | 16 ++++++++++++++++ > arch/riscv/include/asm/Kbuild | 3 ++- > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > 3 files changed, 35 insertions(+), 1 deletion(-) > create mode 100644 arch/riscv/include/asm/spinlock.h > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > index 2c346fe169c1..7f39bfc75744 100644 > --- a/arch/riscv/Kconfig > +++ b/arch/riscv/Kconfig > @@ -471,6 +471,22 @@ config NODES_SHIFT > Specify the maximum number of NUMA Nodes available on the target > system. Increases memory reserved to accommodate various tables. > > +choice > + prompt "RISC-V spinlock type" > + default RISCV_TICKET_SPINLOCKS > + > +config RISCV_TICKET_SPINLOCKS > + bool "Using ticket spinlock" > + > +config RISCV_QUEUED_SPINLOCKS > + bool "Using queued spinlock" > + depends on SMP && MMU > + select ARCH_USE_QUEUED_SPINLOCKS > + help > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > + Otherwise, stay at ticket-lock. > +endchoice > + > config RISCV_ALTERNATIVE > bool > depends on !XIP_KERNEL > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > index 504f8b7e72d4..a0dc85e4a754 100644 > --- a/arch/riscv/include/asm/Kbuild > +++ b/arch/riscv/include/asm/Kbuild > @@ -2,10 +2,11 @@ > generic-y += early_ioremap.h > generic-y += flat.h > generic-y += kvm_para.h > +generic-y += mcs_spinlock.h > generic-y += parport.h > -generic-y += spinlock.h IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) and include the asm-generic headers of mcs_spinlock and qspinlock. In this case, the qspinlock.h will provide the arch_spin_*() interfaces, which seems the oposite of the above description (ticket spinlocks being the standard). Shouldn't ticket-spinlock.h also get included here? (Also, I am probably missing something, as I dont' see the use of mcs_spinlock here.) > generic-y += spinlock_types.h > generic-y += qrwlock.h > generic-y += qrwlock_types.h > +generic-y += qspinlock.h > generic-y += user.h > generic-y += vmlinux.lds.h > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > new file mode 100644 > index 000000000000..c644a92d4548 > --- /dev/null > +++ b/arch/riscv/include/asm/spinlock.h > @@ -0,0 +1,17 @@ > +/* SPDX-License-Identifier: GPL-2.0 */ > + > +#ifndef __ASM_RISCV_SPINLOCK_H > +#define __ASM_RISCV_SPINLOCK_H > + > +#ifdef CONFIG_QUEUED_SPINLOCKS > +#define _Q_PENDING_LOOPS (1 << 9) > +#endif Any reason the above define couldn't be merged on the ifdef below? > + > +#ifdef CONFIG_QUEUED_SPINLOCKS > +#include <asm/qspinlock.h> > +#include <asm/qrwlock.h> > +#else > +#include <asm-generic/spinlock.h> > +#endif > + > +#endif /* __ASM_RISCV_SPINLOCK_H */ > -- > 2.36.1 > Thanks! Leo
On Thu, Sep 14, 2023 at 4:29 AM Leonardo Bras <leobras@redhat.com> wrote: > > On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > > From: Guo Ren <guoren@linux.alibaba.com> > > > > The requirements of qspinlock have been documented by commit: > > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > > atomics"). > > > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > > doesn't satisfy the requirements of qspinlock above, it won't prevent > > some riscv vendors from implementing a strong fwd guarantee LR/SC in > > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > > is the one. > > > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > > queued_spinlock 0.5109/0.00 > > ticket_spinlock 0.5814/0.00 > > > > perf futex/hash (+6.7%): > > queued_spinlock 1444393 operations/sec (+- 0.09%) > > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > > > perf futex/wake-parallel (+8.6%): > > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > > > perf futex/requeue (+4.2%): > > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > > > System Benchmarks (+6.4%) > > queued_spinlock: > > System Benchmarks Index Values BASELINE RESULT INDEX > > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > > Double-Precision Whetstone 55.0 182422.8 33167.8 > > Execl Throughput 43.0 13116.6 3050.4 > > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > > Pipe Throughput 12440.0 23058600.5 18535.9 > > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > > Process Creation 126.0 12537.3 995.0 > > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > > System Call Overhead 15000.0 33308301.3 22205.5 > > ======== > > System Benchmarks Index Score 12426.1 > > > > ticket_spinlock: > > System Benchmarks Index Values BASELINE RESULT INDEX > > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > > Double-Precision Whetstone 55.0 181921.0 33076.5 > > Execl Throughput 43.0 12625.1 2936.1 > > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > > Pipe Throughput 12440.0 20594018.7 16554.7 > > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > > Process Creation 126.0 10798.4 857.0 > > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > > System Call Overhead 15000.0 30766778.4 20511.2 > > ======== > > System Benchmarks Index Score 11670.7 > > > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > > cores platform than the ticket_lock. > > > > Signed-off-by: Guo Ren <guoren@kernel.org> > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > --- > > arch/riscv/Kconfig | 16 ++++++++++++++++ > > arch/riscv/include/asm/Kbuild | 3 ++- > > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > > 3 files changed, 35 insertions(+), 1 deletion(-) > > create mode 100644 arch/riscv/include/asm/spinlock.h > > > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > > index 2c346fe169c1..7f39bfc75744 100644 > > --- a/arch/riscv/Kconfig > > +++ b/arch/riscv/Kconfig > > @@ -471,6 +471,22 @@ config NODES_SHIFT > > Specify the maximum number of NUMA Nodes available on the target > > system. Increases memory reserved to accommodate various tables. > > > > +choice > > + prompt "RISC-V spinlock type" > > + default RISCV_TICKET_SPINLOCKS > > + > > +config RISCV_TICKET_SPINLOCKS > > + bool "Using ticket spinlock" > > + > > +config RISCV_QUEUED_SPINLOCKS > > + bool "Using queued spinlock" > > + depends on SMP && MMU > > + select ARCH_USE_QUEUED_SPINLOCKS > > + help > > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > > + Otherwise, stay at ticket-lock. > > +endchoice > > + > > config RISCV_ALTERNATIVE > > bool > > depends on !XIP_KERNEL > > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > > index 504f8b7e72d4..a0dc85e4a754 100644 > > --- a/arch/riscv/include/asm/Kbuild > > +++ b/arch/riscv/include/asm/Kbuild > > @@ -2,10 +2,11 @@ > > generic-y += early_ioremap.h > > generic-y += flat.h > > generic-y += kvm_para.h > > +generic-y += mcs_spinlock.h > > generic-y += parport.h > > -generic-y += spinlock.h > > IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) > and include the asm-generic headers of mcs_spinlock and qspinlock. > > In this case, the qspinlock.h will provide the arch_spin_*() interfaces, > which seems the oposite of the above description (ticket spinlocks being > the standard). > > Shouldn't ticket-spinlock.h also get included here? > (Also, I am probably missing something, as I dont' see the use of > mcs_spinlock here.) No, because asm-generic/spinlock.h: ... #include <asm-generic/ticket_spinlock.h> ... > > > generic-y += spinlock_types.h > > generic-y += qrwlock.h > > generic-y += qrwlock_types.h > > +generic-y += qspinlock.h > > generic-y += user.h > > generic-y += vmlinux.lds.h > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > new file mode 100644 > > index 000000000000..c644a92d4548 > > --- /dev/null > > +++ b/arch/riscv/include/asm/spinlock.h > > @@ -0,0 +1,17 @@ > > +/* SPDX-License-Identifier: GPL-2.0 */ > > + > > +#ifndef __ASM_RISCV_SPINLOCK_H > > +#define __ASM_RISCV_SPINLOCK_H > > + > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > +#define _Q_PENDING_LOOPS (1 << 9) > > +#endif > > Any reason the above define couldn't be merged on the ifdef below? Easy for the next patch to modify. See Waiman's comment: https://lore.kernel.org/linux-riscv/4cc7113a-0e4e-763a-cba2-7963bcd26c7a@redhat.com/ > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > index c644a92d4548..9eb3ad31e564 100644 > --- a/arch/riscv/include/asm/spinlock.h > +++ b/arch/riscv/include/asm/spinlock.h > @@ -7,11 +7,94 @@ > #define _Q_PENDING_LOOPS (1 << 9) > #endif > I see why you separated the _Q_PENDING_LOOPS out. > > > + > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > +#include <asm/qspinlock.h> > > +#include <asm/qrwlock.h> > > +#else > > +#include <asm-generic/spinlock.h> > > +#endif > > + > > +#endif /* __ASM_RISCV_SPINLOCK_H */ > > -- > > 2.36.1 > > > > Thanks! > Leo >
On Thu, Sep 14, 2023 at 12:46:56PM +0800, Guo Ren wrote: > On Thu, Sep 14, 2023 at 4:29 AM Leonardo Bras <leobras@redhat.com> wrote: > > > > On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > > > From: Guo Ren <guoren@linux.alibaba.com> > > > > > > The requirements of qspinlock have been documented by commit: > > > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > > > atomics"). > > > > > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > > > doesn't satisfy the requirements of qspinlock above, it won't prevent > > > some riscv vendors from implementing a strong fwd guarantee LR/SC in > > > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > > > is the one. > > > > > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > > > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > > > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > > > > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > > > queued_spinlock 0.5109/0.00 > > > ticket_spinlock 0.5814/0.00 > > > > > > perf futex/hash (+6.7%): > > > queued_spinlock 1444393 operations/sec (+- 0.09%) > > > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > > > > > perf futex/wake-parallel (+8.6%): > > > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > > > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > > > > > perf futex/requeue (+4.2%): > > > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > > > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > > > > > System Benchmarks (+6.4%) > > > queued_spinlock: > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > > > Double-Precision Whetstone 55.0 182422.8 33167.8 > > > Execl Throughput 43.0 13116.6 3050.4 > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > > > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > > > Pipe Throughput 12440.0 23058600.5 18535.9 > > > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > > > Process Creation 126.0 12537.3 995.0 > > > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > > > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > > > System Call Overhead 15000.0 33308301.3 22205.5 > > > ======== > > > System Benchmarks Index Score 12426.1 > > > > > > ticket_spinlock: > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > > > Double-Precision Whetstone 55.0 181921.0 33076.5 > > > Execl Throughput 43.0 12625.1 2936.1 > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > > > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > > > Pipe Throughput 12440.0 20594018.7 16554.7 > > > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > > > Process Creation 126.0 10798.4 857.0 > > > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > > > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > > > System Call Overhead 15000.0 30766778.4 20511.2 > > > ======== > > > System Benchmarks Index Score 11670.7 > > > > > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > > > cores platform than the ticket_lock. > > > > > > Signed-off-by: Guo Ren <guoren@kernel.org> > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > > --- > > > arch/riscv/Kconfig | 16 ++++++++++++++++ > > > arch/riscv/include/asm/Kbuild | 3 ++- > > > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > > > 3 files changed, 35 insertions(+), 1 deletion(-) > > > create mode 100644 arch/riscv/include/asm/spinlock.h > > > > > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > > > index 2c346fe169c1..7f39bfc75744 100644 > > > --- a/arch/riscv/Kconfig > > > +++ b/arch/riscv/Kconfig > > > @@ -471,6 +471,22 @@ config NODES_SHIFT > > > Specify the maximum number of NUMA Nodes available on the target > > > system. Increases memory reserved to accommodate various tables. > > > > > > +choice > > > + prompt "RISC-V spinlock type" > > > + default RISCV_TICKET_SPINLOCKS > > > + > > > +config RISCV_TICKET_SPINLOCKS > > > + bool "Using ticket spinlock" > > > + > > > +config RISCV_QUEUED_SPINLOCKS > > > + bool "Using queued spinlock" > > > + depends on SMP && MMU > > > + select ARCH_USE_QUEUED_SPINLOCKS > > > + help > > > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > > > + Otherwise, stay at ticket-lock. > > > +endchoice > > > + > > > config RISCV_ALTERNATIVE > > > bool > > > depends on !XIP_KERNEL > > > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > > > index 504f8b7e72d4..a0dc85e4a754 100644 > > > --- a/arch/riscv/include/asm/Kbuild > > > +++ b/arch/riscv/include/asm/Kbuild > > > @@ -2,10 +2,11 @@ > > > generic-y += early_ioremap.h > > > generic-y += flat.h > > > generic-y += kvm_para.h > > > +generic-y += mcs_spinlock.h > > > generic-y += parport.h > > > -generic-y += spinlock.h > > > > IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) > > and include the asm-generic headers of mcs_spinlock and qspinlock. > > > > In this case, the qspinlock.h will provide the arch_spin_*() interfaces, > > which seems the oposite of the above description (ticket spinlocks being > > the standard). > > > > Shouldn't ticket-spinlock.h also get included here? > > (Also, I am probably missing something, as I dont' see the use of > > mcs_spinlock here.) > No, because asm-generic/spinlock.h: > ... > #include <asm-generic/ticket_spinlock.h> > ... > But aren't you removing asm-generic/spinlock.h below ? -generic-y += spinlock.h > > > > > generic-y += spinlock_types.h > > > generic-y += qrwlock.h > > > generic-y += qrwlock_types.h > > > +generic-y += qspinlock.h > > > generic-y += user.h > > > generic-y += vmlinux.lds.h > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > new file mode 100644 > > > index 000000000000..c644a92d4548 > > > --- /dev/null > > > +++ b/arch/riscv/include/asm/spinlock.h > > > @@ -0,0 +1,17 @@ > > > +/* SPDX-License-Identifier: GPL-2.0 */ > > > + > > > +#ifndef __ASM_RISCV_SPINLOCK_H > > > +#define __ASM_RISCV_SPINLOCK_H > > > + > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > +#define _Q_PENDING_LOOPS (1 << 9) > > > +#endif > > > > Any reason the above define couldn't be merged on the ifdef below? > Easy for the next patch to modify. See Waiman's comment: > > https://lore.kernel.org/linux-riscv/4cc7113a-0e4e-763a-cba2-7963bcd26c7a@redhat.com/ > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > index c644a92d4548..9eb3ad31e564 100644 > > --- a/arch/riscv/include/asm/spinlock.h > > +++ b/arch/riscv/include/asm/spinlock.h > > @@ -7,11 +7,94 @@ > > #define _Q_PENDING_LOOPS (1 << 9) > > #endif > > > > I see why you separated the _Q_PENDING_LOOPS out. > I see, should be fine then. Thanks! Leo > > > > > > + > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > +#include <asm/qspinlock.h> > > > +#include <asm/qrwlock.h> > > > +#else > > > +#include <asm-generic/spinlock.h> > > > +#endif > > > + > > > +#endif /* __ASM_RISCV_SPINLOCK_H */ > > > -- > > > 2.36.1 > > > > > > > Thanks! > > Leo > > > > > -- > Best Regards > Guo Ren >
On Thu, Sep 14, 2023 at 5:43 PM Leonardo Bras <leobras@redhat.com> wrote: > > On Thu, Sep 14, 2023 at 12:46:56PM +0800, Guo Ren wrote: > > On Thu, Sep 14, 2023 at 4:29 AM Leonardo Bras <leobras@redhat.com> wrote: > > > > > > On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > > > > From: Guo Ren <guoren@linux.alibaba.com> > > > > > > > > The requirements of qspinlock have been documented by commit: > > > > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > > > > atomics"). > > > > > > > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > > > > doesn't satisfy the requirements of qspinlock above, it won't prevent > > > > some riscv vendors from implementing a strong fwd guarantee LR/SC in > > > > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > > > > is the one. > > > > > > > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > > > > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > > > > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > > > > > > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > > > > queued_spinlock 0.5109/0.00 > > > > ticket_spinlock 0.5814/0.00 > > > > > > > > perf futex/hash (+6.7%): > > > > queued_spinlock 1444393 operations/sec (+- 0.09%) > > > > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > > > > > > > perf futex/wake-parallel (+8.6%): > > > > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > > > > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > > > > > > > perf futex/requeue (+4.2%): > > > > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > > > > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > > > > > > > System Benchmarks (+6.4%) > > > > queued_spinlock: > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > > > > Double-Precision Whetstone 55.0 182422.8 33167.8 > > > > Execl Throughput 43.0 13116.6 3050.4 > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > > > > Pipe Throughput 12440.0 23058600.5 18535.9 > > > > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > > > > Process Creation 126.0 12537.3 995.0 > > > > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > > > > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > > > > System Call Overhead 15000.0 33308301.3 22205.5 > > > > ======== > > > > System Benchmarks Index Score 12426.1 > > > > > > > > ticket_spinlock: > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > > > > Double-Precision Whetstone 55.0 181921.0 33076.5 > > > > Execl Throughput 43.0 12625.1 2936.1 > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > > > > Pipe Throughput 12440.0 20594018.7 16554.7 > > > > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > > > > Process Creation 126.0 10798.4 857.0 > > > > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > > > > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > > > > System Call Overhead 15000.0 30766778.4 20511.2 > > > > ======== > > > > System Benchmarks Index Score 11670.7 > > > > > > > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > > > > cores platform than the ticket_lock. > > > > > > > > Signed-off-by: Guo Ren <guoren@kernel.org> > > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > > > --- > > > > arch/riscv/Kconfig | 16 ++++++++++++++++ > > > > arch/riscv/include/asm/Kbuild | 3 ++- > > > > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > > > > 3 files changed, 35 insertions(+), 1 deletion(-) > > > > create mode 100644 arch/riscv/include/asm/spinlock.h > > > > > > > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > > > > index 2c346fe169c1..7f39bfc75744 100644 > > > > --- a/arch/riscv/Kconfig > > > > +++ b/arch/riscv/Kconfig > > > > @@ -471,6 +471,22 @@ config NODES_SHIFT > > > > Specify the maximum number of NUMA Nodes available on the target > > > > system. Increases memory reserved to accommodate various tables. > > > > > > > > +choice > > > > + prompt "RISC-V spinlock type" > > > > + default RISCV_TICKET_SPINLOCKS > > > > + > > > > +config RISCV_TICKET_SPINLOCKS > > > > + bool "Using ticket spinlock" > > > > + > > > > +config RISCV_QUEUED_SPINLOCKS > > > > + bool "Using queued spinlock" > > > > + depends on SMP && MMU > > > > + select ARCH_USE_QUEUED_SPINLOCKS > > > > + help > > > > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > > > > + Otherwise, stay at ticket-lock. > > > > +endchoice > > > > + > > > > config RISCV_ALTERNATIVE > > > > bool > > > > depends on !XIP_KERNEL > > > > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > > > > index 504f8b7e72d4..a0dc85e4a754 100644 > > > > --- a/arch/riscv/include/asm/Kbuild > > > > +++ b/arch/riscv/include/asm/Kbuild > > > > @@ -2,10 +2,11 @@ > > > > generic-y += early_ioremap.h > > > > generic-y += flat.h > > > > generic-y += kvm_para.h > > > > +generic-y += mcs_spinlock.h > > > > generic-y += parport.h > > > > -generic-y += spinlock.h > > > > > > IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) > > > and include the asm-generic headers of mcs_spinlock and qspinlock. > > > > > > In this case, the qspinlock.h will provide the arch_spin_*() interfaces, > > > which seems the oposite of the above description (ticket spinlocks being > > > the standard). > > > > > > Shouldn't ticket-spinlock.h also get included here? > > > (Also, I am probably missing something, as I dont' see the use of > > > mcs_spinlock here.) > > No, because asm-generic/spinlock.h: > > ... > > #include <asm-generic/ticket_spinlock.h> > > ... > > > > But aren't you removing asm-generic/spinlock.h below ? > -generic-y += spinlock.h Yes, current is: arch/riscv/include/asm/spinlock.h -> include/asm-generic/spinlock.h -> include/asm-generic/ticket_spinlock.h +#ifdef CONFIG_QUEUED_SPINLOCKS +#include <asm/qspinlock.h> +#include <asm/qrwlock.h> +#else +#include <asm-generic/spinlock.h> +#endif So, you want me: +#ifdef CONFIG_QUEUED_SPINLOCKS +#include <asm/qspinlock.h> +#else +#include <asm-generic/ticket_spinlock.h> +#endif +#include <asm/qrwlock.h> Right? > > > > > > > > generic-y += spinlock_types.h > > > > generic-y += qrwlock.h > > > > generic-y += qrwlock_types.h > > > > +generic-y += qspinlock.h > > > > generic-y += user.h > > > > generic-y += vmlinux.lds.h > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > new file mode 100644 > > > > index 000000000000..c644a92d4548 > > > > --- /dev/null > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > @@ -0,0 +1,17 @@ > > > > +/* SPDX-License-Identifier: GPL-2.0 */ > > > > + > > > > +#ifndef __ASM_RISCV_SPINLOCK_H > > > > +#define __ASM_RISCV_SPINLOCK_H > > > > + > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > +#define _Q_PENDING_LOOPS (1 << 9) > > > > +#endif > > > > > > Any reason the above define couldn't be merged on the ifdef below? > > Easy for the next patch to modify. See Waiman's comment: > > > > https://lore.kernel.org/linux-riscv/4cc7113a-0e4e-763a-cba2-7963bcd26c7a@redhat.com/ > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > index c644a92d4548..9eb3ad31e564 100644 > > > --- a/arch/riscv/include/asm/spinlock.h > > > +++ b/arch/riscv/include/asm/spinlock.h > > > @@ -7,11 +7,94 @@ > > > #define _Q_PENDING_LOOPS (1 << 9) > > > #endif > > > > > > > I see why you separated the _Q_PENDING_LOOPS out. > > > > I see, should be fine then. > > Thanks! > Leo > > > > > > > > > > + > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > +#include <asm/qspinlock.h> > > > > +#include <asm/qrwlock.h> > > > > +#else > > > > +#include <asm-generic/spinlock.h> > > > > +#endif > > > > + > > > > +#endif /* __ASM_RISCV_SPINLOCK_H */ > > > > -- > > > > 2.36.1 > > > > > > > > > > Thanks! > > > Leo > > > > > > > > > -- > > Best Regards > > Guo Ren > > >
On Fri, Sep 15, 2023 at 10:10:25AM +0800, Guo Ren wrote: > On Thu, Sep 14, 2023 at 5:43 PM Leonardo Bras <leobras@redhat.com> wrote: > > > > On Thu, Sep 14, 2023 at 12:46:56PM +0800, Guo Ren wrote: > > > On Thu, Sep 14, 2023 at 4:29 AM Leonardo Bras <leobras@redhat.com> wrote: > > > > > > > > On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > > > > > From: Guo Ren <guoren@linux.alibaba.com> > > > > > > > > > > The requirements of qspinlock have been documented by commit: > > > > > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > > > > > atomics"). > > > > > > > > > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > > > > > doesn't satisfy the requirements of qspinlock above, it won't prevent > > > > > some riscv vendors from implementing a strong fwd guarantee LR/SC in > > > > > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > > > > > is the one. > > > > > > > > > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > > > > > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > > > > > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > > > > > > > > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > > > > > queued_spinlock 0.5109/0.00 > > > > > ticket_spinlock 0.5814/0.00 > > > > > > > > > > perf futex/hash (+6.7%): > > > > > queued_spinlock 1444393 operations/sec (+- 0.09%) > > > > > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > > > > > > > > > perf futex/wake-parallel (+8.6%): > > > > > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > > > > > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > > > > > > > > > perf futex/requeue (+4.2%): > > > > > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > > > > > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > > > > > > > > > System Benchmarks (+6.4%) > > > > > queued_spinlock: > > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > > > > > Double-Precision Whetstone 55.0 182422.8 33167.8 > > > > > Execl Throughput 43.0 13116.6 3050.4 > > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > > > > > Pipe Throughput 12440.0 23058600.5 18535.9 > > > > > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > > > > > Process Creation 126.0 12537.3 995.0 > > > > > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > > > > > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > > > > > System Call Overhead 15000.0 33308301.3 22205.5 > > > > > ======== > > > > > System Benchmarks Index Score 12426.1 > > > > > > > > > > ticket_spinlock: > > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > > > > > Double-Precision Whetstone 55.0 181921.0 33076.5 > > > > > Execl Throughput 43.0 12625.1 2936.1 > > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > > > > > Pipe Throughput 12440.0 20594018.7 16554.7 > > > > > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > > > > > Process Creation 126.0 10798.4 857.0 > > > > > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > > > > > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > > > > > System Call Overhead 15000.0 30766778.4 20511.2 > > > > > ======== > > > > > System Benchmarks Index Score 11670.7 > > > > > > > > > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > > > > > cores platform than the ticket_lock. > > > > > > > > > > Signed-off-by: Guo Ren <guoren@kernel.org> > > > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > > > > --- > > > > > arch/riscv/Kconfig | 16 ++++++++++++++++ > > > > > arch/riscv/include/asm/Kbuild | 3 ++- > > > > > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > > > > > 3 files changed, 35 insertions(+), 1 deletion(-) > > > > > create mode 100644 arch/riscv/include/asm/spinlock.h > > > > > > > > > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > > > > > index 2c346fe169c1..7f39bfc75744 100644 > > > > > --- a/arch/riscv/Kconfig > > > > > +++ b/arch/riscv/Kconfig > > > > > @@ -471,6 +471,22 @@ config NODES_SHIFT > > > > > Specify the maximum number of NUMA Nodes available on the target > > > > > system. Increases memory reserved to accommodate various tables. > > > > > > > > > > +choice > > > > > + prompt "RISC-V spinlock type" > > > > > + default RISCV_TICKET_SPINLOCKS > > > > > + > > > > > +config RISCV_TICKET_SPINLOCKS > > > > > + bool "Using ticket spinlock" > > > > > + > > > > > +config RISCV_QUEUED_SPINLOCKS > > > > > + bool "Using queued spinlock" > > > > > + depends on SMP && MMU > > > > > + select ARCH_USE_QUEUED_SPINLOCKS > > > > > + help > > > > > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > > > > > + Otherwise, stay at ticket-lock. > > > > > +endchoice > > > > > + > > > > > config RISCV_ALTERNATIVE > > > > > bool > > > > > depends on !XIP_KERNEL > > > > > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > > > > > index 504f8b7e72d4..a0dc85e4a754 100644 > > > > > --- a/arch/riscv/include/asm/Kbuild > > > > > +++ b/arch/riscv/include/asm/Kbuild > > > > > @@ -2,10 +2,11 @@ > > > > > generic-y += early_ioremap.h > > > > > generic-y += flat.h > > > > > generic-y += kvm_para.h > > > > > +generic-y += mcs_spinlock.h > > > > > generic-y += parport.h > > > > > -generic-y += spinlock.h > > > > > > > > IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) > > > > and include the asm-generic headers of mcs_spinlock and qspinlock. > > > > > > > > In this case, the qspinlock.h will provide the arch_spin_*() interfaces, > > > > which seems the oposite of the above description (ticket spinlocks being > > > > the standard). > > > > > > > > Shouldn't ticket-spinlock.h also get included here? > > > > (Also, I am probably missing something, as I dont' see the use of > > > > mcs_spinlock here.) > > > No, because asm-generic/spinlock.h: > > > ... > > > #include <asm-generic/ticket_spinlock.h> > > > ... > > > > > > > But aren't you removing asm-generic/spinlock.h below ? > > -generic-y += spinlock.h > Yes, current is: > > arch/riscv/include/asm/spinlock.h -> include/asm-generic/spinlock.h -> > include/asm-generic/ticket_spinlock.h I did a little reading on how generic-y works (which I was unaware): "If an architecture uses a verbatim copy of a header from include/asm-generic then this is listed in the file arch/$(SRCARCH)/include/asm/Kbuild [...] During the prepare phase of the build a wrapper include file is generated in the directory [...]" Oh, so you are removing the asm-generic/spinlock.h because it's link was replaced by a new asm/spinlock.h. You add qspinlock.h to generic-y because it's new in riscv, and add mcs_spinlock.h because it's needed by qspinlock.h. Ok, it makes sense now. Sorry about this noise. I was unaware of how generic-y worked, and (wrongly) assumed it was about including headers automatically in the build. > > +#ifdef CONFIG_QUEUED_SPINLOCKS > +#include <asm/qspinlock.h> > +#include <asm/qrwlock.h> > +#else > +#include <asm-generic/spinlock.h> > +#endif > > So, you want me: > +#ifdef CONFIG_QUEUED_SPINLOCKS > +#include <asm/qspinlock.h> > +#else > +#include <asm-generic/ticket_spinlock.h> > +#endif > > +#include <asm/qrwlock.h> > > Right? No, I didn't mean that. I was just worried about the arch_spin_*() interfaces, but they should be fine. BTW, according to kernel doc on generic-y, shouldn't be a better idea to add 'ticket_spinlock.h' to generic-y, and include above as asm/ticket_spinlock.h? Or is generic-y reserved only for stuff which is indirectly included by other headers? Thanks! Leo > > > > > > > > > > > > generic-y += spinlock_types.h > > > > > generic-y += qrwlock.h > > > > > generic-y += qrwlock_types.h > > > > > +generic-y += qspinlock.h > > > > > generic-y += user.h > > > > > generic-y += vmlinux.lds.h > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > > new file mode 100644 > > > > > index 000000000000..c644a92d4548 > > > > > --- /dev/null > > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > > @@ -0,0 +1,17 @@ > > > > > +/* SPDX-License-Identifier: GPL-2.0 */ > > > > > + > > > > > +#ifndef __ASM_RISCV_SPINLOCK_H > > > > > +#define __ASM_RISCV_SPINLOCK_H > > > > > + > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > > +#define _Q_PENDING_LOOPS (1 << 9) > > > > > +#endif > > > > > > > > Any reason the above define couldn't be merged on the ifdef below? > > > Easy for the next patch to modify. See Waiman's comment: > > > > > > https://lore.kernel.org/linux-riscv/4cc7113a-0e4e-763a-cba2-7963bcd26c7a@redhat.com/ > > > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > index c644a92d4548..9eb3ad31e564 100644 > > > > --- a/arch/riscv/include/asm/spinlock.h > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > @@ -7,11 +7,94 @@ > > > > #define _Q_PENDING_LOOPS (1 << 9) > > > > #endif > > > > > > > > > > I see why you separated the _Q_PENDING_LOOPS out. > > > > > > > I see, should be fine then. > > > > Thanks! > > Leo > > > > > > > > > > > > > > + > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > > +#include <asm/qspinlock.h> > > > > > +#include <asm/qrwlock.h> > > > > > +#else > > > > > +#include <asm-generic/spinlock.h> > > > > > +#endif > > > > > + > > > > > +#endif /* __ASM_RISCV_SPINLOCK_H */ > > > > > -- > > > > > 2.36.1 > > > > > > > > > > > > > Thanks! > > > > Leo > > > > > > > > > > > > > -- > > > Best Regards > > > Guo Ren > > > > > > > > -- > Best Regards > Guo Ren >
On Fri, Sep 15, 2023 at 5:08 PM Leonardo Bras <leobras@redhat.com> wrote: > > On Fri, Sep 15, 2023 at 10:10:25AM +0800, Guo Ren wrote: > > On Thu, Sep 14, 2023 at 5:43 PM Leonardo Bras <leobras@redhat.com> wrote: > > > > > > On Thu, Sep 14, 2023 at 12:46:56PM +0800, Guo Ren wrote: > > > > On Thu, Sep 14, 2023 at 4:29 AM Leonardo Bras <leobras@redhat.com> wrote: > > > > > > > > > > On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > > > > > > From: Guo Ren <guoren@linux.alibaba.com> > > > > > > > > > > > > The requirements of qspinlock have been documented by commit: > > > > > > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > > > > > > atomics"). > > > > > > > > > > > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > > > > > > doesn't satisfy the requirements of qspinlock above, it won't prevent > > > > > > some riscv vendors from implementing a strong fwd guarantee LR/SC in > > > > > > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > > > > > > is the one. > > > > > > > > > > > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > > > > > > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > > > > > > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > > > > > > > > > > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > > > > > > queued_spinlock 0.5109/0.00 > > > > > > ticket_spinlock 0.5814/0.00 > > > > > > > > > > > > perf futex/hash (+6.7%): > > > > > > queued_spinlock 1444393 operations/sec (+- 0.09%) > > > > > > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > > > > > > > > > > > perf futex/wake-parallel (+8.6%): > > > > > > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > > > > > > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > > > > > > > > > > > perf futex/requeue (+4.2%): > > > > > > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > > > > > > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > > > > > > > > > > > System Benchmarks (+6.4%) > > > > > > queued_spinlock: > > > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > > > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > > > > > > Double-Precision Whetstone 55.0 182422.8 33167.8 > > > > > > Execl Throughput 43.0 13116.6 3050.4 > > > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > > > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > > > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > > > > > > Pipe Throughput 12440.0 23058600.5 18535.9 > > > > > > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > > > > > > Process Creation 126.0 12537.3 995.0 > > > > > > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > > > > > > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > > > > > > System Call Overhead 15000.0 33308301.3 22205.5 > > > > > > ======== > > > > > > System Benchmarks Index Score 12426.1 > > > > > > > > > > > > ticket_spinlock: > > > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > > > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > > > > > > Double-Precision Whetstone 55.0 181921.0 33076.5 > > > > > > Execl Throughput 43.0 12625.1 2936.1 > > > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > > > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > > > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > > > > > > Pipe Throughput 12440.0 20594018.7 16554.7 > > > > > > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > > > > > > Process Creation 126.0 10798.4 857.0 > > > > > > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > > > > > > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > > > > > > System Call Overhead 15000.0 30766778.4 20511.2 > > > > > > ======== > > > > > > System Benchmarks Index Score 11670.7 > > > > > > > > > > > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > > > > > > cores platform than the ticket_lock. > > > > > > > > > > > > Signed-off-by: Guo Ren <guoren@kernel.org> > > > > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > > > > > --- > > > > > > arch/riscv/Kconfig | 16 ++++++++++++++++ > > > > > > arch/riscv/include/asm/Kbuild | 3 ++- > > > > > > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > > > > > > 3 files changed, 35 insertions(+), 1 deletion(-) > > > > > > create mode 100644 arch/riscv/include/asm/spinlock.h > > > > > > > > > > > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > > > > > > index 2c346fe169c1..7f39bfc75744 100644 > > > > > > --- a/arch/riscv/Kconfig > > > > > > +++ b/arch/riscv/Kconfig > > > > > > @@ -471,6 +471,22 @@ config NODES_SHIFT > > > > > > Specify the maximum number of NUMA Nodes available on the target > > > > > > system. Increases memory reserved to accommodate various tables. > > > > > > > > > > > > +choice > > > > > > + prompt "RISC-V spinlock type" > > > > > > + default RISCV_TICKET_SPINLOCKS > > > > > > + > > > > > > +config RISCV_TICKET_SPINLOCKS > > > > > > + bool "Using ticket spinlock" > > > > > > + > > > > > > +config RISCV_QUEUED_SPINLOCKS > > > > > > + bool "Using queued spinlock" > > > > > > + depends on SMP && MMU > > > > > > + select ARCH_USE_QUEUED_SPINLOCKS > > > > > > + help > > > > > > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > > > > > > + Otherwise, stay at ticket-lock. > > > > > > +endchoice > > > > > > + > > > > > > config RISCV_ALTERNATIVE > > > > > > bool > > > > > > depends on !XIP_KERNEL > > > > > > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > > > > > > index 504f8b7e72d4..a0dc85e4a754 100644 > > > > > > --- a/arch/riscv/include/asm/Kbuild > > > > > > +++ b/arch/riscv/include/asm/Kbuild > > > > > > @@ -2,10 +2,11 @@ > > > > > > generic-y += early_ioremap.h > > > > > > generic-y += flat.h > > > > > > generic-y += kvm_para.h > > > > > > +generic-y += mcs_spinlock.h > > > > > > generic-y += parport.h > > > > > > -generic-y += spinlock.h > > > > > > > > > > IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) > > > > > and include the asm-generic headers of mcs_spinlock and qspinlock. > > > > > > > > > > In this case, the qspinlock.h will provide the arch_spin_*() interfaces, > > > > > which seems the oposite of the above description (ticket spinlocks being > > > > > the standard). > > > > > > > > > > Shouldn't ticket-spinlock.h also get included here? > > > > > (Also, I am probably missing something, as I dont' see the use of > > > > > mcs_spinlock here.) > > > > No, because asm-generic/spinlock.h: > > > > ... > > > > #include <asm-generic/ticket_spinlock.h> > > > > ... > > > > > > > > > > But aren't you removing asm-generic/spinlock.h below ? > > > -generic-y += spinlock.h > > Yes, current is: > > > > arch/riscv/include/asm/spinlock.h -> include/asm-generic/spinlock.h -> > > include/asm-generic/ticket_spinlock.h > > I did a little reading on how generic-y works (which I was unaware): > > "If an architecture uses a verbatim copy of a header from > include/asm-generic then this is listed in the file > arch/$(SRCARCH)/include/asm/Kbuild [...] During the prepare phase of the > build a wrapper include file is generated in the directory [...]" > > Oh, so you are removing the asm-generic/spinlock.h because it's link > was replaced by a new asm/spinlock.h. > > You add qspinlock.h to generic-y because it's new in riscv, and add > mcs_spinlock.h because it's needed by qspinlock.h. > > Ok, it makes sense now. > > Sorry about this noise. > I was unaware of how generic-y worked, and (wrongly) > assumed it was about including headers automatically in the build. > > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > +#include <asm/qspinlock.h> > > +#include <asm/qrwlock.h> > > +#else > > +#include <asm-generic/spinlock.h> > > +#endif > > > > So, you want me: > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > +#include <asm/qspinlock.h> > > +#else > > +#include <asm-generic/ticket_spinlock.h> > > +#endif > > > > +#include <asm/qrwlock.h> > > > > Right? > > No, I didn't mean that. > I was just worried about the arch_spin_*() interfaces, but they should be > fine. > > BTW, according to kernel doc on generic-y, shouldn't be a better idea to > add 'ticket_spinlock.h' to generic-y, and include above as > asm/ticket_spinlock.h? > > Or is generic-y reserved only for stuff which is indirectly included by > other headers? It's okay to add generic-y for ticket_spinlock.h, and I'm okay with the following: +#ifdef CONFIG_QUEUED_SPINLOCKS +#include <asm/qspinlock.h> +#else +#include <asm/ticket_spinlock.h> +#endif +#include <asm/qrwlock.h> > > Thanks! > Leo > > > > > > > > > > > > > > > > > generic-y += spinlock_types.h > > > > > > generic-y += qrwlock.h > > > > > > generic-y += qrwlock_types.h > > > > > > +generic-y += qspinlock.h > > > > > > generic-y += user.h > > > > > > generic-y += vmlinux.lds.h > > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > > > new file mode 100644 > > > > > > index 000000000000..c644a92d4548 > > > > > > --- /dev/null > > > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > > > @@ -0,0 +1,17 @@ > > > > > > +/* SPDX-License-Identifier: GPL-2.0 */ > > > > > > + > > > > > > +#ifndef __ASM_RISCV_SPINLOCK_H > > > > > > +#define __ASM_RISCV_SPINLOCK_H > > > > > > + > > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > > > +#define _Q_PENDING_LOOPS (1 << 9) > > > > > > +#endif > > > > > > > > > > Any reason the above define couldn't be merged on the ifdef below? > > > > Easy for the next patch to modify. See Waiman's comment: > > > > > > > > https://lore.kernel.org/linux-riscv/4cc7113a-0e4e-763a-cba2-7963bcd26c7a@redhat.com/ > > > > > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > > index c644a92d4548..9eb3ad31e564 100644 > > > > > --- a/arch/riscv/include/asm/spinlock.h > > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > > @@ -7,11 +7,94 @@ > > > > > #define _Q_PENDING_LOOPS (1 << 9) > > > > > #endif > > > > > > > > > > > > > I see why you separated the _Q_PENDING_LOOPS out. > > > > > > > > > > I see, should be fine then. > > > > > > Thanks! > > > Leo > > > > > > > > > > > > > > > > > > + > > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > > > +#include <asm/qspinlock.h> > > > > > > +#include <asm/qrwlock.h> > > > > > > +#else > > > > > > +#include <asm-generic/spinlock.h> > > > > > > +#endif > > > > > > + > > > > > > +#endif /* __ASM_RISCV_SPINLOCK_H */ > > > > > > -- > > > > > > 2.36.1 > > > > > > > > > > > > > > > > Thanks! > > > > > Leo > > > > > > > > > > > > > > > > > -- > > > > Best Regards > > > > Guo Ren > > > > > > > > > > > > > -- > > Best Regards > > Guo Ren > > >
On Sun, Sep 17, 2023 at 11:02:47PM +0800, Guo Ren wrote: > On Fri, Sep 15, 2023 at 5:08 PM Leonardo Bras <leobras@redhat.com> wrote: > > > > On Fri, Sep 15, 2023 at 10:10:25AM +0800, Guo Ren wrote: > > > On Thu, Sep 14, 2023 at 5:43 PM Leonardo Bras <leobras@redhat.com> wrote: > > > > > > > > On Thu, Sep 14, 2023 at 12:46:56PM +0800, Guo Ren wrote: > > > > > On Thu, Sep 14, 2023 at 4:29 AM Leonardo Bras <leobras@redhat.com> wrote: > > > > > > > > > > > > On Sun, Sep 10, 2023 at 04:28:59AM -0400, guoren@kernel.org wrote: > > > > > > > From: Guo Ren <guoren@linux.alibaba.com> > > > > > > > > > > > > > > The requirements of qspinlock have been documented by commit: > > > > > > > a8ad07e5240c ("asm-generic: qspinlock: Indicate the use of mixed-size > > > > > > > atomics"). > > > > > > > > > > > > > > Although RISC-V ISA gives out a weaker forward guarantee LR/SC, which > > > > > > > doesn't satisfy the requirements of qspinlock above, it won't prevent > > > > > > > some riscv vendors from implementing a strong fwd guarantee LR/SC in > > > > > > > microarchitecture to match xchg_tail requirement. T-HEAD C9xx processor > > > > > > > is the one. > > > > > > > > > > > > > > We've tested the patch on SOPHGO sg2042 & th1520 and passed the stress > > > > > > > test on Fedora & Ubuntu & OpenEuler ... Here is the performance > > > > > > > comparison between qspinlock and ticket_lock on sg2042 (64 cores): > > > > > > > > > > > > > > sysbench test=threads threads=32 yields=100 lock=8 (+13.8%): > > > > > > > queued_spinlock 0.5109/0.00 > > > > > > > ticket_spinlock 0.5814/0.00 > > > > > > > > > > > > > > perf futex/hash (+6.7%): > > > > > > > queued_spinlock 1444393 operations/sec (+- 0.09%) > > > > > > > ticket_spinlock 1353215 operations/sec (+- 0.15%) > > > > > > > > > > > > > > perf futex/wake-parallel (+8.6%): > > > > > > > queued_spinlock (waking 1/64 threads) in 0.0253 ms (+-2.90%) > > > > > > > ticket_spinlock (waking 1/64 threads) in 0.0275 ms (+-3.12%) > > > > > > > > > > > > > > perf futex/requeue (+4.2%): > > > > > > > queued_spinlock Requeued 64 of 64 threads in 0.0785 ms (+-0.55%) > > > > > > > ticket_spinlock Requeued 64 of 64 threads in 0.0818 ms (+-4.12%) > > > > > > > > > > > > > > System Benchmarks (+6.4%) > > > > > > > queued_spinlock: > > > > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > > > > Dhrystone 2 using register variables 116700.0 628613745.4 53865.8 > > > > > > > Double-Precision Whetstone 55.0 182422.8 33167.8 > > > > > > > Execl Throughput 43.0 13116.6 3050.4 > > > > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 7762306.2 19601.8 > > > > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3417556.8 20649.9 > > > > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7427995.7 12806.9 > > > > > > > Pipe Throughput 12440.0 23058600.5 18535.9 > > > > > > > Pipe-based Context Switching 4000.0 2835617.7 7089.0 > > > > > > > Process Creation 126.0 12537.3 995.0 > > > > > > > Shell Scripts (1 concurrent) 42.4 57057.4 13456.9 > > > > > > > Shell Scripts (8 concurrent) 6.0 7367.1 12278.5 > > > > > > > System Call Overhead 15000.0 33308301.3 22205.5 > > > > > > > ======== > > > > > > > System Benchmarks Index Score 12426.1 > > > > > > > > > > > > > > ticket_spinlock: > > > > > > > System Benchmarks Index Values BASELINE RESULT INDEX > > > > > > > Dhrystone 2 using register variables 116700.0 626541701.9 53688.2 > > > > > > > Double-Precision Whetstone 55.0 181921.0 33076.5 > > > > > > > Execl Throughput 43.0 12625.1 2936.1 > > > > > > > File Copy 1024 bufsize 2000 maxblocks 3960.0 6553792.9 16550.0 > > > > > > > File Copy 256 bufsize 500 maxblocks 1655.0 3189231.6 19270.3 > > > > > > > File Copy 4096 bufsize 8000 maxblocks 5800.0 7221277.0 12450.5 > > > > > > > Pipe Throughput 12440.0 20594018.7 16554.7 > > > > > > > Pipe-based Context Switching 4000.0 2571117.7 6427.8 > > > > > > > Process Creation 126.0 10798.4 857.0 > > > > > > > Shell Scripts (1 concurrent) 42.4 57227.5 13497.1 > > > > > > > Shell Scripts (8 concurrent) 6.0 7329.2 12215.3 > > > > > > > System Call Overhead 15000.0 30766778.4 20511.2 > > > > > > > ======== > > > > > > > System Benchmarks Index Score 11670.7 > > > > > > > > > > > > > > The qspinlock has a significant improvement on SOPHGO SG2042 64 > > > > > > > cores platform than the ticket_lock. > > > > > > > > > > > > > > Signed-off-by: Guo Ren <guoren@kernel.org> > > > > > > > Signed-off-by: Guo Ren <guoren@linux.alibaba.com> > > > > > > > --- > > > > > > > arch/riscv/Kconfig | 16 ++++++++++++++++ > > > > > > > arch/riscv/include/asm/Kbuild | 3 ++- > > > > > > > arch/riscv/include/asm/spinlock.h | 17 +++++++++++++++++ > > > > > > > 3 files changed, 35 insertions(+), 1 deletion(-) > > > > > > > create mode 100644 arch/riscv/include/asm/spinlock.h > > > > > > > > > > > > > > diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig > > > > > > > index 2c346fe169c1..7f39bfc75744 100644 > > > > > > > --- a/arch/riscv/Kconfig > > > > > > > +++ b/arch/riscv/Kconfig > > > > > > > @@ -471,6 +471,22 @@ config NODES_SHIFT > > > > > > > Specify the maximum number of NUMA Nodes available on the target > > > > > > > system. Increases memory reserved to accommodate various tables. > > > > > > > > > > > > > > +choice > > > > > > > + prompt "RISC-V spinlock type" > > > > > > > + default RISCV_TICKET_SPINLOCKS > > > > > > > + > > > > > > > +config RISCV_TICKET_SPINLOCKS > > > > > > > + bool "Using ticket spinlock" > > > > > > > + > > > > > > > +config RISCV_QUEUED_SPINLOCKS > > > > > > > + bool "Using queued spinlock" > > > > > > > + depends on SMP && MMU > > > > > > > + select ARCH_USE_QUEUED_SPINLOCKS > > > > > > > + help > > > > > > > + Make sure your micro arch LL/SC has a strong forward progress guarantee. > > > > > > > + Otherwise, stay at ticket-lock. > > > > > > > +endchoice > > > > > > > + > > > > > > > config RISCV_ALTERNATIVE > > > > > > > bool > > > > > > > depends on !XIP_KERNEL > > > > > > > diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild > > > > > > > index 504f8b7e72d4..a0dc85e4a754 100644 > > > > > > > --- a/arch/riscv/include/asm/Kbuild > > > > > > > +++ b/arch/riscv/include/asm/Kbuild > > > > > > > @@ -2,10 +2,11 @@ > > > > > > > generic-y += early_ioremap.h > > > > > > > generic-y += flat.h > > > > > > > generic-y += kvm_para.h > > > > > > > +generic-y += mcs_spinlock.h > > > > > > > generic-y += parport.h > > > > > > > -generic-y += spinlock.h > > > > > > > > > > > > IIUC here you take the asm-generic/spinlock.h (which defines arch_spin_*()) > > > > > > and include the asm-generic headers of mcs_spinlock and qspinlock. > > > > > > > > > > > > In this case, the qspinlock.h will provide the arch_spin_*() interfaces, > > > > > > which seems the oposite of the above description (ticket spinlocks being > > > > > > the standard). > > > > > > > > > > > > Shouldn't ticket-spinlock.h also get included here? > > > > > > (Also, I am probably missing something, as I dont' see the use of > > > > > > mcs_spinlock here.) > > > > > No, because asm-generic/spinlock.h: > > > > > ... > > > > > #include <asm-generic/ticket_spinlock.h> > > > > > ... > > > > > > > > > > > > > But aren't you removing asm-generic/spinlock.h below ? > > > > -generic-y += spinlock.h > > > Yes, current is: > > > > > > arch/riscv/include/asm/spinlock.h -> include/asm-generic/spinlock.h -> > > > include/asm-generic/ticket_spinlock.h > > > > I did a little reading on how generic-y works (which I was unaware): > > > > "If an architecture uses a verbatim copy of a header from > > include/asm-generic then this is listed in the file > > arch/$(SRCARCH)/include/asm/Kbuild [...] During the prepare phase of the > > build a wrapper include file is generated in the directory [...]" > > > > Oh, so you are removing the asm-generic/spinlock.h because it's link > > was replaced by a new asm/spinlock.h. > > > > You add qspinlock.h to generic-y because it's new in riscv, and add > > mcs_spinlock.h because it's needed by qspinlock.h. > > > > Ok, it makes sense now. > > > > Sorry about this noise. > > I was unaware of how generic-y worked, and (wrongly) > > assumed it was about including headers automatically in the build. > > > > > > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > +#include <asm/qspinlock.h> > > > +#include <asm/qrwlock.h> > > > +#else > > > +#include <asm-generic/spinlock.h> > > > +#endif > > > > > > So, you want me: > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > +#include <asm/qspinlock.h> > > > +#else > > > +#include <asm-generic/ticket_spinlock.h> > > > +#endif > > > > > > +#include <asm/qrwlock.h> > > > > > > Right? > > > > No, I didn't mean that. > > I was just worried about the arch_spin_*() interfaces, but they should be > > fine. > > > > BTW, according to kernel doc on generic-y, shouldn't be a better idea to > > add 'ticket_spinlock.h' to generic-y, and include above as > > asm/ticket_spinlock.h? > > > > Or is generic-y reserved only for stuff which is indirectly included by > > other headers? > It's okay to add generic-y for ticket_spinlock.h, and I'm okay with > the following: > > +#ifdef CONFIG_QUEUED_SPINLOCKS > +#include <asm/qspinlock.h> > +#else > +#include <asm/ticket_spinlock.h> > +#endif > > +#include <asm/qrwlock.h> It does look more intuitive, so I am glad it works for you :) FWIW: Reviewed-by: Leonardo Bras <leobras@redhat.com> Thanks! Leo > > > > > Thanks! > > Leo > > > > > > > > > > > > > > > > > > > > > > generic-y += spinlock_types.h > > > > > > > generic-y += qrwlock.h > > > > > > > generic-y += qrwlock_types.h > > > > > > > +generic-y += qspinlock.h > > > > > > > generic-y += user.h > > > > > > > generic-y += vmlinux.lds.h > > > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > > > > new file mode 100644 > > > > > > > index 000000000000..c644a92d4548 > > > > > > > --- /dev/null > > > > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > > > > @@ -0,0 +1,17 @@ > > > > > > > +/* SPDX-License-Identifier: GPL-2.0 */ > > > > > > > + > > > > > > > +#ifndef __ASM_RISCV_SPINLOCK_H > > > > > > > +#define __ASM_RISCV_SPINLOCK_H > > > > > > > + > > > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > > > > +#define _Q_PENDING_LOOPS (1 << 9) > > > > > > > +#endif > > > > > > > > > > > > Any reason the above define couldn't be merged on the ifdef below? > > > > > Easy for the next patch to modify. See Waiman's comment: > > > > > > > > > > https://lore.kernel.org/linux-riscv/4cc7113a-0e4e-763a-cba2-7963bcd26c7a@redhat.com/ > > > > > > > > > > > diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h > > > > > > index c644a92d4548..9eb3ad31e564 100644 > > > > > > --- a/arch/riscv/include/asm/spinlock.h > > > > > > +++ b/arch/riscv/include/asm/spinlock.h > > > > > > @@ -7,11 +7,94 @@ > > > > > > #define _Q_PENDING_LOOPS (1 << 9) > > > > > > #endif > > > > > > > > > > > > > > > > I see why you separated the _Q_PENDING_LOOPS out. > > > > > > > > > > > > > I see, should be fine then. > > > > > > > > Thanks! > > > > Leo > > > > > > > > > > > > > > > > > > > > > > + > > > > > > > +#ifdef CONFIG_QUEUED_SPINLOCKS > > > > > > > +#include <asm/qspinlock.h> > > > > > > > +#include <asm/qrwlock.h> > > > > > > > +#else > > > > > > > +#include <asm-generic/spinlock.h> > > > > > > > +#endif > > > > > > > + > > > > > > > +#endif /* __ASM_RISCV_SPINLOCK_H */ > > > > > > > -- > > > > > > > 2.36.1 > > > > > > > > > > > > > > > > > > > Thanks! > > > > > > Leo > > > > > > > > > > > > > > > > > > > > > -- > > > > > Best Regards > > > > > Guo Ren > > > > > > > > > > > > > > > > > > -- > > > Best Regards > > > Guo Ren > > > > > > > > -- > Best Regards > Guo Ren >
diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig index 2c346fe169c1..7f39bfc75744 100644 --- a/arch/riscv/Kconfig +++ b/arch/riscv/Kconfig @@ -471,6 +471,22 @@ config NODES_SHIFT Specify the maximum number of NUMA Nodes available on the target system. Increases memory reserved to accommodate various tables. +choice + prompt "RISC-V spinlock type" + default RISCV_TICKET_SPINLOCKS + +config RISCV_TICKET_SPINLOCKS + bool "Using ticket spinlock" + +config RISCV_QUEUED_SPINLOCKS + bool "Using queued spinlock" + depends on SMP && MMU + select ARCH_USE_QUEUED_SPINLOCKS + help + Make sure your micro arch LL/SC has a strong forward progress guarantee. + Otherwise, stay at ticket-lock. +endchoice + config RISCV_ALTERNATIVE bool depends on !XIP_KERNEL diff --git a/arch/riscv/include/asm/Kbuild b/arch/riscv/include/asm/Kbuild index 504f8b7e72d4..a0dc85e4a754 100644 --- a/arch/riscv/include/asm/Kbuild +++ b/arch/riscv/include/asm/Kbuild @@ -2,10 +2,11 @@ generic-y += early_ioremap.h generic-y += flat.h generic-y += kvm_para.h +generic-y += mcs_spinlock.h generic-y += parport.h -generic-y += spinlock.h generic-y += spinlock_types.h generic-y += qrwlock.h generic-y += qrwlock_types.h +generic-y += qspinlock.h generic-y += user.h generic-y += vmlinux.lds.h diff --git a/arch/riscv/include/asm/spinlock.h b/arch/riscv/include/asm/spinlock.h new file mode 100644 index 000000000000..c644a92d4548 --- /dev/null +++ b/arch/riscv/include/asm/spinlock.h @@ -0,0 +1,17 @@ +/* SPDX-License-Identifier: GPL-2.0 */ + +#ifndef __ASM_RISCV_SPINLOCK_H +#define __ASM_RISCV_SPINLOCK_H + +#ifdef CONFIG_QUEUED_SPINLOCKS +#define _Q_PENDING_LOOPS (1 << 9) +#endif + +#ifdef CONFIG_QUEUED_SPINLOCKS +#include <asm/qspinlock.h> +#include <asm/qrwlock.h> +#else +#include <asm-generic/spinlock.h> +#endif + +#endif /* __ASM_RISCV_SPINLOCK_H */