riscv: errata: Add ERRATA_THEAD_WRITE_ONCE fixup
Subsystems (2)
PATCH v4 1/1 riscv: errata: Add ERRATA_THEAD_WRITE_ONCE fixup
From: "Guo Ren (Alibaba DAMO Academy)" <[email protected]>
The early version of XuanTie C910 core has a store merge buffer delay problem. The store merge buffer could improve the store queue performance by merging multi-store requests, but when there are not continued store requests, the prior single store request would be waiting in the store queue for a long time. That would cause significant problems for communication between multi-cores. This problem was found on sg2042 & th1520 platforms with the qspinlock lock torture test.
So appending a fence w.o could immediately flush the store merge buffer and let other cores see the write result.
This will apply the WRITE_ONCE errata to handle the non-standard behavior via appending a fence w.o instruction for WRITE_ONCE().
This problem is only observed on the sg2042 hardware platform by running the lock_torture test program for half an hour. The problem was not found in the user space application, because interrupt can break the livelock.
---
Changelog
v4:
- Add Reviewed-by: Alexandre Ghiti <[email protected]>
(https://lore.kernel.org/linux-riscv/[email protected]/)
- Add Tested-by: Yao Zi <[email protected]>
- Rebase on v7.0-rc7
v3:
https://lore.kernel.org/linux-riscv/[email protected]/
- Rebase on 6.19-rc6.
- Remove errata_list_vendors patch which has been merged.
- Add Tested-by tag.
v2:
https://lore.kernel.org/linux-riscv/[email protected]/
- Add new header file for errata_list_vendors.
- Rebase newest kernel version.
v1:
https://lore.kernel.org/all/[email protected]/
--- arch/riscv/Kconfig.errata | 17 ++++++++++ arch/riscv/errata/thead/errata.c | 20 ++++++++++++ arch/riscv/include/asm/errata_list_vendors.h | 3 +- arch/riscv/include/asm/rwonce.h | 34 ++++++++++++++++++++ include/asm-generic/rwonce.h | 2 ++ 5 files changed, 75 insertions(+), 1 deletion(-) create mode 100644 arch/riscv/include/asm/rwonce.h
arch/riscv/Kconfig.errata +17 -0
@@ -154,4 +154,21 @@ config ERRATA_THEAD_GHOSTWRITE If you don't know what to do here, say "Y". +config ERRATA_THEAD_WRITE_ONCE + bool "Apply T-Head WRITE_ONCE errata" + depends on ERRATA_THEAD + default y + help + The early version of T-Head C9xx cores of sg2042 & th1520 have a store + merge buffer delay problem. The store merge buffer could improve the + store queue performance by merging multi-store requests, but when there + are no continued store requests, the prior single store request would be + waiting in the store queue for a long time. That would cause signifi- + cant problems for communication between multi-cores. Appending a + fence w.o could immediately flush the store merge buffer and let other + cores see the write result. + + This will apply the WRITE_ONCE errata to handle the non-standard beh- + avior via appending a fence w.o instruction for WRITE_ONCE(). + endmenu # "CPU errata selection"
arch/riscv/errata/thead/errata.c +20 -0
@@ -168,6 +168,23 @@ static bool errata_probe_ghostwrite(unsigned int stage,return true; } +static bool errata_probe_write_once (unsigned int stage, +unsigned long arch_id,unsigned long impid) +{ +if (!IS_ENABLED (CONFIG_ERRATA_THEAD_WRITE_ONCE )) +return false; + +/* target-c9xx cores report arch_id and impid as 0 */ +if (arch_id !=0 || impid !=0 ) +return false; + +if (stage ==RISCV_ALTERNATIVES_BOOT || + stage ==RISCV_ALTERNATIVES_MODULE ) +return true; + +return false; +} +static u32 thead_errata_probe (unsigned int stage,unsigned long archid,unsigned long impid) {
@@ -183,6 +200,9 @@ static u32 thead_errata_probe(unsigned int stage,errata_probe_ghostwrite (stage, archid, impid); +if (errata_probe_write_once (stage, archid, impid)) + cpu_req_errata |=BIT (ERRATA_THEAD_WRITE_ONCE ); +return cpu_req_errata; }
arch/riscv/include/asm/errata_list_vendors.h +2 -1
@@ -18,7 +18,8 @@#define ERRATA_THEAD_MAE 0#define ERRATA_THEAD_PMU 1#define ERRATA_THEAD_GHOSTWRITE 2 -#define ERRATA_THEAD_NUMBER 3 +#define ERRATA_THEAD_WRITE_ONCE 3 +#define ERRATA_THEAD_NUMBER 4#endif #ifdef CONFIG_ERRATA_MIPS
arch/riscv/include/asm/rwonce.h new file +34 -0
@@ -0,0 +1,34 @@ +/* SPDX-License-Identifier: GPL-2.0 */ + +#ifndef __ASM_RWONCE_H +#define __ASM_RWONCE_H + +#include <linux/compiler_types.h> +#include <asm/alternative-macros.h> +#include <asm/vendorid_list.h> +#include <asm/errata_list_vendors.h> + +#if defined(CONFIG_ERRATA_THEAD_WRITE_ONCE ) && !defined(NO_ALTERNATIVE ) + +#define write_once_fence () \ +do { \ + asm volatile(ALTERNATIVE( \ + "nop", \ + "fence w, o", \ + THEAD_VENDOR_ID, \ + ERRATA_THEAD_WRITE_ONCE, \ + CONFIG_ERRATA_THEAD_WRITE_ONCE) \ + : : : "memory"); \ +} while (0) + +#define __WRITE_ONCE (x, val) \ +do { \ + *(volatile typeof(x) *)&(x) = (val); \ + write_once_fence(); \ +} while (0) + +#endif /* defined(CONFIG_ERRATA_THEAD_WRITE_ONCE) && !defined(NO_ALTERNATIVE) */ + +#include <asm-generic/rwonce.h> + +#endif /* __ASM_RWONCE_H */
include/asm-generic/rwonce.h +2 -0
@@ -50,10 +50,12 @@__READ_ONCE (x); \ }) +#ifndef __WRITE_ONCE#define __WRITE_ONCE (x, val) \ do { \ *(volatile typeof(x) *)&(x) = (val); \ } while (0) +#endif #define WRITE_ONCE (x, val) \ do { \
Signature · 1 line
2.40.1
Quoted 5 lines · On Sun, Apr 12, 2026, at 16:31, [email protected] wrote:
arch/riscv/Kconfig.errata | 17 ++++++++++ arch/riscv/errata/thead/errata.c | 20 ++++++++++++ arch/riscv/include/asm/errata_list_vendors.h | 3 +- arch/riscv/include/asm/rwonce.h | 34 ++++++++++++++++++++ include/asm-generic/rwonce.h | 2 ++
For asm-generic:
arch/riscv/errata/thead/errata.c static bool errata_probe_ghostwrite(unsigned int stage, · line 183
Quoted 13 lines
arch/riscv/errata/thead/errata.c +13
-0+static bool errata_probe_write_once (unsigned int stage, +unsigned long arch_id,unsigned long impid) +{ +if (!IS_ENABLED (CONFIG_ERRATA_THEAD_WRITE_ONCE )) +return false; + +/* target-c9xx cores report arch_id and impid as 0 */ +if (arch_id !=0 || impid !=0 ) +return false; + +if (stage ==RISCV_ALTERNATIVES_BOOT || + stage ==RISCV_ALTERNATIVES_MODULE ) +return true;
Question: Are the affected cores the only ones that report arch_id==impid==0? If there are other known implementations that get the extra barrier but don't need it here, that should probably be mentioned in the Kconfig text and the code comment.
Arnd
Quoted 23 lines · Arnd Bergmann
Quoted 5 lines · On Sun, Apr 12, 2026, at 16:31, [email protected] wrote:
arch/riscv/Kconfig.errata | 17 ++++++++++ arch/riscv/errata/thead/errata.c | 20 ++++++++++++ arch/riscv/include/asm/errata_list_vendors.h | 3 +- arch/riscv/include/asm/rwonce.h | 34 ++++++++++++++++++++ include/asm-generic/rwonce.h | 2 ++For asm-generic:
Acked-by Arnd Bergmann <[email protected]>Quoted 11 lines
arch/riscv/errata/thead/errata.c +11
-0+static bool errata_probe_write_once (unsigned int stage, +unsigned long arch_id,unsigned long impid) +{ +if (!IS_ENABLED (CONFIG_ERRATA_THEAD_WRITE_ONCE )) +return false; + +/* target-c9xx cores report arch_id and impid as 0 */ +if (arch_id !=0 || impid !=0 ) +return false; + + if (stage == RISCV_ALTERNATIVES_BOOT ||
Quoted 8 lines
arch/riscv/errata/thead/errata.c static bool errata_probe_ghostwrite(unsigned int stage, · line 183
Quoted 2 lines
arch/riscv/errata/thead/errata.c +2
-0+ stage == RISCV_ALTERNATIVES_MODULE) +return true;Question: Are the affected cores the only ones that report arch_id==impid==0? If there are other known implementations that get the extra barrier but don't need it here, that should probably be mentioned in the
Kconfigtext and the code comment.
Thank you for your question.
So far, we have only observed this issue on the SG2042 chip. However, in theory, it affects all older c9xx cores that report arch_id == 0 and impid == 0.
Since these early versions are relatively few in number and share the same signature (arch_id = 0 && impid = 0), we decided to apply the workaround to all cores that match this condition to ensure complete coverage and avoid missing any affected silicon.
Signature · 2 lines
Best Regards GUO Ren
Quoted 31 lines · On Sun, Apr 12, 2026 at 10:31:16AM -0400, [email protected] wrote:
From: "Guo Ren (Alibaba DAMO Academy)" <[email protected]>
The early version of XuanTie C910 core has a store merge buffer delay problem. The store merge buffer could improve the store queue performance by merging multi-store requests, but when there are not continued store requests, the prior single store request would be waiting in the store queue for a long time. That would cause significant problems for communication between multi-cores. This problem was found on sg2042 & th1520 platforms with the qspinlock lock torture test.
So appending a fence w.o could immediately flush the store merge buffer and let other cores see the write result.
This will apply the
WRITE_ONCEerrata to handle the non-standard behavior via appending a fence w.o instruction forWRITE_ONCE().This problem is only observed on the sg2042 hardware platform by running the
lock_torturetest program for half an hour. The problem was not found in the user space application, because interrupt can break the livelock.Reviewed-by Leonardo Bras <[email protected]>Reviewed-by Alexandre Ghiti <[email protected]>Tested-by Han Gao <[email protected]>Tested-by Yao Zi <[email protected]>Cc Chen Wang <[email protected]>Cc Inochi Amaoto <[email protected]>Cc Xiaoguang Xing <[email protected]>Cc Paul Walmsley <[email protected]>Signed-off-by Guo Ren (Alibaba DAMO Academy) <[email protected]>
Quoted 1 line
---
Quoted 168 lines
Changelog
v4:
- Add Reviewed-by: Alexandre Ghiti <[email protected]>
(https://lore.kernel.org/linux-riscv/[email protected]/)
- Add Tested-by: Yao Zi <[email protected]>
- Rebase on v7.0-rc7v3:
https://lore.kernel.org/linux-riscv/[email protected]/
- Rebase on 6.19-rc6.
- Removeerrata_list_vendorspatch which has been merged.
- Add Tested-by tag.v2:
https://lore.kernel.org/linux-riscv/[email protected]/
- Add new header file forerrata_list_vendors.
- Rebase newest kernel version.v1:
https://lore.kernel.org/all/[email protected]/--- arch/riscv/Kconfig.errata | 17 ++++++++++ arch/riscv/errata/thead/errata.c | 20 ++++++++++++ arch/riscv/include/asm/errata_list_vendors.h | 3 +- arch/riscv/include/asm/rwonce.h | 34 ++++++++++++++++++++ include/asm-generic/rwonce.h | 2 ++ 5 files changed, 75 insertions(+), 1 deletion(-) create mode 100644 arch/riscv/include/asm/rwonce.harch/riscv/Kconfig.errata +17
-0@@ -154,4 +154,21 @@ config ERRATA_THEAD_GHOSTWRITE If you don't know what to do here, say "Y". +config ERRATA_THEAD_WRITE_ONCE + bool "Apply T-Head WRITE_ONCE errata" + depends on ERRATA_THEAD + default y + help + The early version of T-Head C9xx cores of sg2042 & th1520 have a store + merge buffer delay problem. The store merge buffer could improve the + store queue performance by merging multi-store requests, but when there + are no continued store requests, the prior single store request would be + waiting in the store queue for a long time. That would cause signifi- + cant problems for communication between multi-cores. Appending a + fence w.o could immediately flush the store merge buffer and let other + cores see the write result. + + This will apply the WRITE_ONCE errata to handle the non-standard beh- + avior via appending a fence w.o instruction for WRITE_ONCE(). + endmenu # "CPU errata selection"arch/riscv/errata/thead/errata.c +20
-0@@ -168,6 +168,23 @@ static bool errata_probe_ghostwrite(unsigned int stage,return true; } +static bool errata_probe_write_once (unsigned int stage, +unsigned long arch_id,unsigned long impid) +{ +if (!IS_ENABLED (CONFIG_ERRATA_THEAD_WRITE_ONCE )) +return false; + +/* target-c9xx cores report arch_id and impid as 0 */ +if (arch_id !=0 || impid !=0 ) +return false; + +if (stage ==RISCV_ALTERNATIVES_BOOT || + stage ==RISCV_ALTERNATIVES_MODULE ) +return true; + +return false; +} +static u32 thead_errata_probe (unsigned int stage,unsigned long archid,unsigned long impid) {@@ -183,6 +200,9 @@ static u32 thead_errata_probe(unsigned int stage,errata_probe_ghostwrite (stage, archid, impid); +if (errata_probe_write_once (stage, archid, impid)) + cpu_req_errata |=BIT (ERRATA_THEAD_WRITE_ONCE ); +return cpu_req_errata; }arch/riscv/include/asm/errata_list_vendors.h +2
-1@@ -18,7 +18,8 @@#define ERRATA_THEAD_MAE 0#define ERRATA_THEAD_PMU 1#define ERRATA_THEAD_GHOSTWRITE 2 -#define ERRATA_THEAD_NUMBER 3 +#define ERRATA_THEAD_WRITE_ONCE 3 +#define ERRATA_THEAD_NUMBER 4#endif #ifdef CONFIG_ERRATA_MIPS arch/riscv/include/asm/rwonce.h new file +34
-0@@ -0,0 +1,34 @@ +/* SPDX-License-Identifier: GPL-2.0 */ + +#ifndef __ASM_RWONCE_H +#define __ASM_RWONCE_H + +#include <linux/compiler_types.h> +#include <asm/alternative-macros.h> +#include <asm/vendorid_list.h> +#include <asm/errata_list_vendors.h> + +#if defined(CONFIG_ERRATA_THEAD_WRITE_ONCE ) && !defined(NO_ALTERNATIVE ) + +#define write_once_fence () \ +do { \ + asm volatile(ALTERNATIVE( \ + "nop", \ + "fence w, o", \ + THEAD_VENDOR_ID, \ + ERRATA_THEAD_WRITE_ONCE, \ + CONFIG_ERRATA_THEAD_WRITE_ONCE) \ + : : : "memory"); \ +} while (0) + +#define __WRITE_ONCE (x, val) \ +do { \ + *(volatile typeof(x) *)&(x) = (val); \ + write_once_fence(); \ +} while (0) + +#endif /* defined(CONFIG_ERRATA_THEAD_WRITE_ONCE) && !defined(NO_ALTERNATIVE) */ + +#include <asm-generic/rwonce.h> + +#endif /* __ASM_RWONCE_H */ include/asm-generic/rwonce.h +2
-0@@ -50,10 +50,12 @@__READ_ONCE (x); \ }) +#ifndef __WRITE_ONCE#define __WRITE_ONCE (x, val) \ do { \ *(volatile typeof(x) *)&(x) = (val); \ } while (0) +#endif #define WRITE_ONCE (x, val) \ do { \Signature · 1 line
2.40.1