All of lore.kernel.org
 help / color / mirror / Atom feed
From: "Radim Krčmář" <rkrcmar@ventanamicro.com>
To: "yunhui cui" <cuiyunhui@bytedance.com>
Cc: <masahiroy@kernel.org>, <nathan@kernel.org>,
	<nicolas.schier@linux.dev>, <dennis@kernel.org>, <tj@kernel.org>,
	<cl@gentwo.org>, <paul.walmsley@sifive.com>, <palmer@dabbelt.com>,
	<aou@eecs.berkeley.edu>, <alex@ghiti.fr>, <andybnac@gmail.com>,
	<bjorn@rivosinc.com>, <cyrilbur@tenstorrent.com>,
	<rostedt@goodmis.org>, <puranjay@kernel.org>,
	<ben.dooks@codethink.co.uk>, <zhangchunyan@iscas.ac.cn>,
	<ruanjinjie@huawei.com>, <jszhang@kernel.org>,
	<charlie@rivosinc.com>, <cleger@rivosinc.com>,
	<antonb@tenstorrent.com>, <ajones@ventanamicro.com>,
	<debug@rivosinc.com>, <haibo1.xu@intel.com>,
	<samuel.holland@sifive.com>, <linux-kbuild@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>, <linux-mm@kvack.org>,
	<linux-riscv@lists.infradead.org>,
	"linux-riscv" <linux-riscv-bounces@lists.infradead.org>,
	<wangziang.ok@bytedance.com>
Subject: [PATCH] RISC-V: store precomputed percpu_offset in the task struct
Date: Thu, 10 Jul 2025 18:40:02 +0200	[thread overview]
Message-ID: <DB8IV9GMLI6X.25TJF0WMVHZEM@ventanamicro.com> (raw)
In-Reply-To: <CAEEQ3wmxJ50PZHVpdexeyy1ELqKw+5mrb+8gRCA4KNj9zsrykA@mail.gmail.com>

2025-07-10T19:47:27+08:00, yunhui cui <cuiyunhui@bytedance.com>:
> On Thu, Jul 10, 2025 at 2:35 PM Radim Krčmář <rkrcmar@ventanamicro.com> wrote:
>> It would be to store the percpu offset in CSR_SCRATCH permanently, do
>> the early exception register shuffling with a percpu area storage, and
>> load the thread pointer from there as well.
>> That method would also eliminate writing CSR_SCRATCH on every exception
>> entry+exit, so maybe it makes sense to try it even if CSRs are slow...
>
> Based on the patch, optimizations for percpu offset have been added,
> with the following data:
> 6.989 7.046 6.976 6.986 7.001 7.017 7.007 7.064 7.008 7.039
> Geometric mean: 7.013248303
> Compared to reusing the scratch register, the performance has improved
> by approximately 0.7%.

Nice, thanks.  The CSR_SCRATCH accesses seem much slower than GPRs, and
possibly even slower than L1 hit -- we might gain more by storing the
precomputed offset in the task struct.

Can you check this patch as well?

(It should be compared against a variant of CSR_SCRATCH that uses the
 TASK_TI_PERCPU_OFFSET optimizations, but we can try to interpolate. :])

---8<---
RISC-V: store precomputed percpu_offset in the task struct

Exploring the memoization trade-off... hoping that __set_task_cpu covers
everything. :)

I didn't put any though into where the percpu_offset should live, and
the naive approach is to put it next to cpu.
This needs more work to not break build on other arches, because I
directly added RISC-V specific code to __set_task_cpu, to save time
figuring out where else it could be.
---
 arch/riscv/include/asm/asm.h         | 6 +-----
 arch/riscv/include/asm/percpu.h      | 8 ++++++++
 arch/riscv/include/asm/thread_info.h | 3 ++-
 arch/riscv/kernel/asm-offsets.c      | 1 +
 arch/riscv/kernel/smpboot.c          | 6 ++++++
 kernel/sched/sched.h                 | 1 +
 6 files changed, 19 insertions(+), 6 deletions(-)
 create mode 100644 arch/riscv/include/asm/percpu.h

diff --git a/arch/riscv/include/asm/asm.h b/arch/riscv/include/asm/asm.h
index a8a2af6dfe9d..2a6b831d9cdf 100644
--- a/arch/riscv/include/asm/asm.h
+++ b/arch/riscv/include/asm/asm.h
@@ -91,11 +91,7 @@
 #endif
 
 .macro asm_per_cpu dst sym tmp
-	REG_L \tmp, TASK_TI_CPU_NUM(tp)
-	slli  \tmp, \tmp, PER_CPU_OFFSET_SHIFT
-	la    \dst, __per_cpu_offset
-	add   \dst, \dst, \tmp
-	REG_L \tmp, 0(\dst)
+	REG_L \tmp, TASK_TI_PERCPU_OFFSET(tp)
 	la    \dst, \sym
 	add   \dst, \dst, \tmp
 .endm
diff --git a/arch/riscv/include/asm/percpu.h b/arch/riscv/include/asm/percpu.h
new file mode 100644
index 000000000000..c37a0fce6ebc
--- /dev/null
+++ b/arch/riscv/include/asm/percpu.h
@@ -0,0 +1,8 @@
+#ifndef __ASM_PERCPU_H
+#define __ASM_PERCPU_H
+
+#define __my_cpu_offset (current_thread_info()->percpu_offset)
+
+#include <asm-generic/percpu.h>
+
+#endif
diff --git a/arch/riscv/include/asm/thread_info.h b/arch/riscv/include/asm/thread_info.h
index f5916a70879a..da776b7a1d02 100644
--- a/arch/riscv/include/asm/thread_info.h
+++ b/arch/riscv/include/asm/thread_info.h
@@ -60,8 +60,9 @@ struct thread_info {
 	 */
 	long			kernel_sp;	/* Kernel stack pointer */
 	long			user_sp;	/* User stack pointer */
-	int			cpu;
+	int			cpu;		// TODO: could be packed better
 	unsigned long		syscall_work;	/* SYSCALL_WORK_ flags */
+	unsigned long		percpu_offset;	// XXX: randomly placed here
 #ifdef CONFIG_SHADOW_CALL_STACK
 	void			*scs_base;
 	void			*scs_sp;
diff --git a/arch/riscv/kernel/asm-offsets.c b/arch/riscv/kernel/asm-offsets.c
index 6e8c0d6feae9..9c7bb4d7e3b3 100644
--- a/arch/riscv/kernel/asm-offsets.c
+++ b/arch/riscv/kernel/asm-offsets.c
@@ -50,6 +50,7 @@ void asm_offsets(void)
 #endif
 
 	OFFSET(TASK_TI_CPU_NUM, task_struct, thread_info.cpu);
+	OFFSET(TASK_TI_PERCPU_OFFSET, task_struct, thread_info.percpu_offset);
 	OFFSET(TASK_THREAD_F0,  task_struct, thread.fstate.f[0]);
 	OFFSET(TASK_THREAD_F1,  task_struct, thread.fstate.f[1]);
 	OFFSET(TASK_THREAD_F2,  task_struct, thread.fstate.f[2]);
diff --git a/arch/riscv/kernel/smpboot.c b/arch/riscv/kernel/smpboot.c
index 601a321e0f17..3c09c8f3e30c 100644
--- a/arch/riscv/kernel/smpboot.c
+++ b/arch/riscv/kernel/smpboot.c
@@ -41,6 +41,11 @@
 
 static DECLARE_COMPLETION(cpu_running);
 
+void __init smp_prepare_boot_cpu(void)
+{
+	current_thread_info()->percpu_offset = per_cpu_offset(smp_processor_id());
+}
+
 void __init smp_prepare_cpus(unsigned int max_cpus)
 {
 	int cpuid;
@@ -183,6 +188,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 {
 	int ret = 0;
 	tidle->thread_info.cpu = cpu;
+	tidle->thread_info.percpu_offset = per_cpu_offset(cpu);
 
 	ret = start_secondary_cpu(cpu, tidle);
 	if (!ret) {
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 475bb5998295..2180a85b1403 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -2199,6 +2199,7 @@ static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
 	 */
 	smp_wmb();
 	WRITE_ONCE(task_thread_info(p)->cpu, cpu);
+	WRITE_ONCE(task_thread_info(p)->percpu_offset, per_cpu_offset(cpu));
 	p->wake_cpu = cpu;
 #endif
 }

WARNING: multiple messages have this Message-ID (diff)
From: "Radim Krčmář" <rkrcmar@ventanamicro.com>
To: "yunhui cui" <cuiyunhui@bytedance.com>
Cc: <masahiroy@kernel.org>, <nathan@kernel.org>,
	<nicolas.schier@linux.dev>, <dennis@kernel.org>, <tj@kernel.org>,
	<cl@gentwo.org>, <paul.walmsley@sifive.com>, <palmer@dabbelt.com>,
	<aou@eecs.berkeley.edu>, <alex@ghiti.fr>, <andybnac@gmail.com>,
	<bjorn@rivosinc.com>, <cyrilbur@tenstorrent.com>,
	<rostedt@goodmis.org>, <puranjay@kernel.org>,
	<ben.dooks@codethink.co.uk>, <zhangchunyan@iscas.ac.cn>,
	<ruanjinjie@huawei.com>, <jszhang@kernel.org>,
	<charlie@rivosinc.com>, <cleger@rivosinc.com>,
	<antonb@tenstorrent.com>, <ajones@ventanamicro.com>,
	<debug@rivosinc.com>, <haibo1.xu@intel.com>,
	<samuel.holland@sifive.com>, <linux-kbuild@vger.kernel.org>,
	<linux-kernel@vger.kernel.org>, <linux-mm@kvack.org>,
	<linux-riscv@lists.infradead.org>,
	"linux-riscv" <linux-riscv-bounces@lists.infradead.org>,
	<wangziang.ok@bytedance.com>
Subject: [PATCH] RISC-V: store precomputed percpu_offset in the task struct
Date: Thu, 10 Jul 2025 18:40:02 +0200	[thread overview]
Message-ID: <DB8IV9GMLI6X.25TJF0WMVHZEM@ventanamicro.com> (raw)
In-Reply-To: <CAEEQ3wmxJ50PZHVpdexeyy1ELqKw+5mrb+8gRCA4KNj9zsrykA@mail.gmail.com>

2025-07-10T19:47:27+08:00, yunhui cui <cuiyunhui@bytedance.com>:
> On Thu, Jul 10, 2025 at 2:35 PM Radim Krčmář <rkrcmar@ventanamicro.com> wrote:
>> It would be to store the percpu offset in CSR_SCRATCH permanently, do
>> the early exception register shuffling with a percpu area storage, and
>> load the thread pointer from there as well.
>> That method would also eliminate writing CSR_SCRATCH on every exception
>> entry+exit, so maybe it makes sense to try it even if CSRs are slow...
>
> Based on the patch, optimizations for percpu offset have been added,
> with the following data:
> 6.989 7.046 6.976 6.986 7.001 7.017 7.007 7.064 7.008 7.039
> Geometric mean: 7.013248303
> Compared to reusing the scratch register, the performance has improved
> by approximately 0.7%.

Nice, thanks.  The CSR_SCRATCH accesses seem much slower than GPRs, and
possibly even slower than L1 hit -- we might gain more by storing the
precomputed offset in the task struct.

Can you check this patch as well?

(It should be compared against a variant of CSR_SCRATCH that uses the
 TASK_TI_PERCPU_OFFSET optimizations, but we can try to interpolate. :])

---8<---
RISC-V: store precomputed percpu_offset in the task struct

Exploring the memoization trade-off... hoping that __set_task_cpu covers
everything. :)

I didn't put any though into where the percpu_offset should live, and
the naive approach is to put it next to cpu.
This needs more work to not break build on other arches, because I
directly added RISC-V specific code to __set_task_cpu, to save time
figuring out where else it could be.
---
 arch/riscv/include/asm/asm.h         | 6 +-----
 arch/riscv/include/asm/percpu.h      | 8 ++++++++
 arch/riscv/include/asm/thread_info.h | 3 ++-
 arch/riscv/kernel/asm-offsets.c      | 1 +
 arch/riscv/kernel/smpboot.c          | 6 ++++++
 kernel/sched/sched.h                 | 1 +
 6 files changed, 19 insertions(+), 6 deletions(-)
 create mode 100644 arch/riscv/include/asm/percpu.h

diff --git a/arch/riscv/include/asm/asm.h b/arch/riscv/include/asm/asm.h
index a8a2af6dfe9d..2a6b831d9cdf 100644
--- a/arch/riscv/include/asm/asm.h
+++ b/arch/riscv/include/asm/asm.h
@@ -91,11 +91,7 @@
 #endif
 
 .macro asm_per_cpu dst sym tmp
-	REG_L \tmp, TASK_TI_CPU_NUM(tp)
-	slli  \tmp, \tmp, PER_CPU_OFFSET_SHIFT
-	la    \dst, __per_cpu_offset
-	add   \dst, \dst, \tmp
-	REG_L \tmp, 0(\dst)
+	REG_L \tmp, TASK_TI_PERCPU_OFFSET(tp)
 	la    \dst, \sym
 	add   \dst, \dst, \tmp
 .endm
diff --git a/arch/riscv/include/asm/percpu.h b/arch/riscv/include/asm/percpu.h
new file mode 100644
index 000000000000..c37a0fce6ebc
--- /dev/null
+++ b/arch/riscv/include/asm/percpu.h
@@ -0,0 +1,8 @@
+#ifndef __ASM_PERCPU_H
+#define __ASM_PERCPU_H
+
+#define __my_cpu_offset (current_thread_info()->percpu_offset)
+
+#include <asm-generic/percpu.h>
+
+#endif
diff --git a/arch/riscv/include/asm/thread_info.h b/arch/riscv/include/asm/thread_info.h
index f5916a70879a..da776b7a1d02 100644
--- a/arch/riscv/include/asm/thread_info.h
+++ b/arch/riscv/include/asm/thread_info.h
@@ -60,8 +60,9 @@ struct thread_info {
 	 */
 	long			kernel_sp;	/* Kernel stack pointer */
 	long			user_sp;	/* User stack pointer */
-	int			cpu;
+	int			cpu;		// TODO: could be packed better
 	unsigned long		syscall_work;	/* SYSCALL_WORK_ flags */
+	unsigned long		percpu_offset;	// XXX: randomly placed here
 #ifdef CONFIG_SHADOW_CALL_STACK
 	void			*scs_base;
 	void			*scs_sp;
diff --git a/arch/riscv/kernel/asm-offsets.c b/arch/riscv/kernel/asm-offsets.c
index 6e8c0d6feae9..9c7bb4d7e3b3 100644
--- a/arch/riscv/kernel/asm-offsets.c
+++ b/arch/riscv/kernel/asm-offsets.c
@@ -50,6 +50,7 @@ void asm_offsets(void)
 #endif
 
 	OFFSET(TASK_TI_CPU_NUM, task_struct, thread_info.cpu);
+	OFFSET(TASK_TI_PERCPU_OFFSET, task_struct, thread_info.percpu_offset);
 	OFFSET(TASK_THREAD_F0,  task_struct, thread.fstate.f[0]);
 	OFFSET(TASK_THREAD_F1,  task_struct, thread.fstate.f[1]);
 	OFFSET(TASK_THREAD_F2,  task_struct, thread.fstate.f[2]);
diff --git a/arch/riscv/kernel/smpboot.c b/arch/riscv/kernel/smpboot.c
index 601a321e0f17..3c09c8f3e30c 100644
--- a/arch/riscv/kernel/smpboot.c
+++ b/arch/riscv/kernel/smpboot.c
@@ -41,6 +41,11 @@
 
 static DECLARE_COMPLETION(cpu_running);
 
+void __init smp_prepare_boot_cpu(void)
+{
+	current_thread_info()->percpu_offset = per_cpu_offset(smp_processor_id());
+}
+
 void __init smp_prepare_cpus(unsigned int max_cpus)
 {
 	int cpuid;
@@ -183,6 +188,7 @@ int __cpu_up(unsigned int cpu, struct task_struct *tidle)
 {
 	int ret = 0;
 	tidle->thread_info.cpu = cpu;
+	tidle->thread_info.percpu_offset = per_cpu_offset(cpu);
 
 	ret = start_secondary_cpu(cpu, tidle);
 	if (!ret) {
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 475bb5998295..2180a85b1403 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -2199,6 +2199,7 @@ static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
 	 */
 	smp_wmb();
 	WRITE_ONCE(task_thread_info(p)->cpu, cpu);
+	WRITE_ONCE(task_thread_info(p)->percpu_offset, per_cpu_offset(cpu));
 	p->wake_cpu = cpu;
 #endif
 }

_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv

  reply	other threads:[~2025-07-10 16:40 UTC|newest]

Thread overview: 22+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2025-07-04  8:45 [PATCH RFC] RISC-V: Fix a register to store the percpu offset Yunhui Cui
2025-07-04  8:45 ` Yunhui Cui
2025-07-07  7:55 ` Clément Léger
2025-07-07  7:55   ` Clément Léger
2025-07-07 12:50 ` [PATCH] RISC-V: store percpu offset in CSR_SCRATCH Radim Krčmář
2025-07-07 12:50   ` Radim Krčmář
2025-07-08 10:07   ` [External] " yunhui cui
2025-07-08 10:07     ` yunhui cui
2025-07-08 11:10     ` Radim Krčmář
2025-07-08 11:10       ` Radim Krčmář
2025-07-09 11:42       ` yunhui cui
2025-07-09 11:42         ` yunhui cui
2025-07-09 14:20         ` Radim Krčmář
2025-07-09 14:20           ` Radim Krčmář
2025-07-10  3:45           ` yunhui cui
2025-07-10  3:45             ` yunhui cui
2025-07-10  6:35             ` Radim Krčmář
2025-07-10  6:35               ` Radim Krčmář
2025-07-10 11:47               ` yunhui cui
2025-07-10 11:47                 ` yunhui cui
2025-07-10 16:40                 ` Radim Krčmář [this message]
2025-07-10 16:40                   ` [PATCH] RISC-V: store precomputed percpu_offset in the task struct Radim Krčmář

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=DB8IV9GMLI6X.25TJF0WMVHZEM@ventanamicro.com \
    --to=rkrcmar@ventanamicro.com \
    --cc=ajones@ventanamicro.com \
    --cc=alex@ghiti.fr \
    --cc=andybnac@gmail.com \
    --cc=antonb@tenstorrent.com \
    --cc=aou@eecs.berkeley.edu \
    --cc=ben.dooks@codethink.co.uk \
    --cc=bjorn@rivosinc.com \
    --cc=charlie@rivosinc.com \
    --cc=cl@gentwo.org \
    --cc=cleger@rivosinc.com \
    --cc=cuiyunhui@bytedance.com \
    --cc=cyrilbur@tenstorrent.com \
    --cc=debug@rivosinc.com \
    --cc=dennis@kernel.org \
    --cc=haibo1.xu@intel.com \
    --cc=jszhang@kernel.org \
    --cc=linux-kbuild@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=linux-riscv-bounces@lists.infradead.org \
    --cc=linux-riscv@lists.infradead.org \
    --cc=masahiroy@kernel.org \
    --cc=nathan@kernel.org \
    --cc=nicolas.schier@linux.dev \
    --cc=palmer@dabbelt.com \
    --cc=paul.walmsley@sifive.com \
    --cc=puranjay@kernel.org \
    --cc=rostedt@goodmis.org \
    --cc=ruanjinjie@huawei.com \
    --cc=samuel.holland@sifive.com \
    --cc=tj@kernel.org \
    --cc=wangziang.ok@bytedance.com \
    --cc=zhangchunyan@iscas.ac.cn \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.