[PATCH v3 0/5] riscv: vdso: Implement __vdso_futex_robust_try_unlock()

[PATCH v3 0/5] riscv: vdso: Implement __vdso_futex_robust_try_unlock()

[Nam Cao]

This series adds new vdso futex entries, allowing user to address the
robust futex race condition regarding clearing the
robust_list_head::list_op_pending pointer.

Thomas wrote a summary of the problem:
https://lore.kernel.org/lkml/20260602084648.462672743@kernel.org/

The core infrastructure has been implemented along with x86 support. This
series adds RISC-V support.

The first 4 patches prepare compat vdso Makefile. The final patch is the
main show.

v3:
  - add missing include guard
  - switch to if/else for compare-and-exchange vs load-reserved/store-conditional

v2:
  - address Sashiko's comments

Nam Cao (5):
  riscv: compat_vdso: switch to standard kbuild rule
  riscv: compat_vdso: Allow *.c source files
  riscv: compat_vdso: Introduce COMPAT_MARCH variable
  riscv: compat_vdso: Build with zacas if available
  riscv: vdso: Implement __vdso_futex_robust_try_unlock()

 arch/riscv/Kconfig                         |  1 +
 arch/riscv/include/asm/cpufeature-macros.h |  2 +
 arch/riscv/include/asm/futex_robust.h      | 16 +++++
 arch/riscv/include/asm/vdso/futex.h        | 14 ++++
 arch/riscv/kernel/compat_vdso/Makefile     | 27 ++++---
 arch/riscv/kernel/compat_vdso/futex.c      |  3 +
 arch/riscv/kernel/vdso.c                   | 61 ++++++++++++++++
 arch/riscv/kernel/vdso/Makefile            | 14 +++-
 arch/riscv/kernel/vdso/futex.c             | 84 ++++++++++++++++++++++
 arch/riscv/kernel/vdso/vdso.lds.S          |  8 +++
 10 files changed, 219 insertions(+), 11 deletions(-)
 create mode 100644 arch/riscv/include/asm/futex_robust.h
 create mode 100644 arch/riscv/include/asm/vdso/futex.h
 create mode 100644 arch/riscv/kernel/compat_vdso/futex.c
 create mode 100644 arch/riscv/kernel/vdso/futex.c

-- 
2.47.3

[PATCH v3 1/5] riscv: compat_vdso: switch to standard kbuild rule

[Nam Cao]

The compat_vdso Makefile has a custom build rule for *.S files. This
is unnecessary, and makes it hard to extend this Makefile to support
building *.c files.

Switch to standard kbuild rule instead.

Signed-off-by: Nam Cao <namcao@linutronix.de>
---
 arch/riscv/kernel/compat_vdso/Makefile | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)

diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
index 24e37d1ef7ec..8a395b5dd037 100644
--- a/arch/riscv/kernel/compat_vdso/Makefile
+++ b/arch/riscv/kernel/compat_vdso/Makefile
@@ -8,7 +8,6 @@ compat_vdso-syms  = rt_sigreturn
 compat_vdso-syms += getcpu
 compat_vdso-syms += flush_icache
 
-COMPAT_CC := $(CC)
 COMPAT_LD := $(LD)
 
 # binutils 2.35 does not support the zifencei extension, but in the ISA
@@ -43,8 +42,7 @@ $(obj)/compat_vdso.so.dbg: $(obj)/compat_vdso.lds $(obj-compat_vdso) FORCE
 LDFLAGS_compat_vdso.so.dbg = -shared -S -soname=linux-compat_vdso.so.1 \
 	--build-id=sha1 --hash-style=both --eh-frame-hdr
 
-$(obj-compat_vdso): %.o: %.S FORCE
-	$(call if_changed_dep,compat_vdsoas)
+$(obj-compat_vdso) : KBUILD_AFLAGS += $(COMPAT_CC_FLAGS)
 
 # strip rule for the .so file
 $(obj)/%.so: OBJCOPYFLAGS := -S
@@ -66,7 +64,3 @@ quiet_cmd_compat_vdsold = VDSOLD  $@
       cmd_compat_vdsold = $(COMPAT_LD) $(ld_flags) $(COMPAT_LD_FLAGS) -T $(filter-out FORCE,$^) -o $@.tmp && \
                    $(OBJCOPY) $(patsubst %, -G __compat_vdso_%, $(compat_vdso-syms)) $@.tmp $@ && \
                    rm $@.tmp
-
-# actual build commands
-quiet_cmd_compat_vdsoas = VDSOAS  $@
-      cmd_compat_vdsoas = $(COMPAT_CC) $(a_flags) $(COMPAT_CC_FLAGS) -c -o $@ $<
-- 
2.47.3

Re: [PATCH v3 1/5] riscv: compat_vdso: switch to standard kbuild rule

[Thomas Weißschuh]

On Fri, Jul 17, 2026 at 10:06:34AM +0200, Nam Cao wrote:
> The compat_vdso Makefile has a custom build rule for *.S files. This
> is unnecessary, and makes it hard to extend this Makefile to support
> building *.c files.
> 
> Switch to standard kbuild rule instead.
> 
> Signed-off-by: Nam Cao <namcao@linutronix.de>

Reviewed-by: Thomas Weißschuh <thomas.weissschuh@linutronix.de>

> ---
>  arch/riscv/kernel/compat_vdso/Makefile | 8 +-------
>  1 file changed, 1 insertion(+), 7 deletions(-)
> 
> diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
> index 24e37d1ef7ec..8a395b5dd037 100644
> --- a/arch/riscv/kernel/compat_vdso/Makefile
> +++ b/arch/riscv/kernel/compat_vdso/Makefile
> @@ -8,7 +8,6 @@ compat_vdso-syms  = rt_sigreturn
>  compat_vdso-syms += getcpu
>  compat_vdso-syms += flush_icache
>  
> -COMPAT_CC := $(CC)
>  COMPAT_LD := $(LD)
>  
>  # binutils 2.35 does not support the zifencei extension, but in the ISA
> @@ -43,8 +42,7 @@ $(obj)/compat_vdso.so.dbg: $(obj)/compat_vdso.lds $(obj-compat_vdso) FORCE
>  LDFLAGS_compat_vdso.so.dbg = -shared -S -soname=linux-compat_vdso.so.1 \
>  	--build-id=sha1 --hash-style=both --eh-frame-hdr
>  
> -$(obj-compat_vdso): %.o: %.S FORCE
> -	$(call if_changed_dep,compat_vdsoas)
> +$(obj-compat_vdso) : KBUILD_AFLAGS += $(COMPAT_CC_FLAGS)

Spurious space before ":".

>  # strip rule for the .so file
>  $(obj)/%.so: OBJCOPYFLAGS := -S
> @@ -66,7 +64,3 @@ quiet_cmd_compat_vdsold = VDSOLD  $@
>        cmd_compat_vdsold = $(COMPAT_LD) $(ld_flags) $(COMPAT_LD_FLAGS) -T $(filter-out FORCE,$^) -o $@.tmp && \
>                     $(OBJCOPY) $(patsubst %, -G __compat_vdso_%, $(compat_vdso-syms)) $@.tmp $@ && \
>                     rm $@.tmp
> -
> -# actual build commands
> -quiet_cmd_compat_vdsoas = VDSOAS  $@
> -      cmd_compat_vdsoas = $(COMPAT_CC) $(a_flags) $(COMPAT_CC_FLAGS) -c -o $@ $<

[PATCH v3 2/5] riscv: compat_vdso: Allow *.c source files

[Nam Cao]

Extend to support building *.c files. This allows adding futex.c in a
follow-up commit.

Signed-off-by: Nam Cao <namcao@linutronix.de>
---
 arch/riscv/kernel/compat_vdso/Makefile | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
index 8a395b5dd037..452c623d1a44 100644
--- a/arch/riscv/kernel/compat_vdso/Makefile
+++ b/arch/riscv/kernel/compat_vdso/Makefile
@@ -43,6 +43,7 @@ LDFLAGS_compat_vdso.so.dbg = -shared -S -soname=linux-compat_vdso.so.1 \
 	--build-id=sha1 --hash-style=both --eh-frame-hdr
 
 $(obj-compat_vdso) : KBUILD_AFLAGS += $(COMPAT_CC_FLAGS)
+$(obj-compat_vdso) : KBUILD_CFLAGS += $(COMPAT_CC_FLAGS)
 
 # strip rule for the .so file
 $(obj)/%.so: OBJCOPYFLAGS := -S
-- 
2.47.3

Re: [PATCH v3 2/5] riscv: compat_vdso: Allow *.c source files

[Thomas Weißschuh]

On Fri, Jul 17, 2026 at 10:06:35AM +0200, Nam Cao wrote:
> Extend to support building *.c files. This allows adding futex.c in a
> follow-up commit.
> 
> Signed-off-by: Nam Cao <namcao@linutronix.de>

Reviewed-by: Thomas Weißschuh <thomas.weissschuh@linutronix.de>

> ---
>  arch/riscv/kernel/compat_vdso/Makefile | 1 +
>  1 file changed, 1 insertion(+)

[PATCH v3 3/5] riscv: compat_vdso: Introduce COMPAT_MARCH variable

[Nam Cao]

The -march option is inside COMPAT_CC_FLAGS, making it difficult to extend.

Introduce COMPAT_MARCH variable to allow extension. A follow-up commit
will take advantage of this to add _zacas to -march.

Signed-off-by: Nam Cao <namcao@linutronix.de>
---
 arch/riscv/kernel/compat_vdso/Makefile | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
index 452c623d1a44..63580b33e3ce 100644
--- a/arch/riscv/kernel/compat_vdso/Makefile
+++ b/arch/riscv/kernel/compat_vdso/Makefile
@@ -13,10 +13,13 @@ COMPAT_LD := $(LD)
 # binutils 2.35 does not support the zifencei extension, but in the ISA
 # spec 20191213, G stands for IMAFD_ZICSR_ZIFENCEI.
 ifdef CONFIG_TOOLCHAIN_NEEDS_EXPLICIT_ZICSR_ZIFENCEI
-	COMPAT_CC_FLAGS := -march=rv32g -mabi=ilp32
+	COMPAT_MARCH := rv32g
 else
-	COMPAT_CC_FLAGS := -march=rv32imafd -mabi=ilp32
+	COMPAT_MARCH := rv32imafd
 endif
+
+COMPAT_CC_FLAGS := -march=$(COMPAT_MARCH) -mabi=ilp32
+
 COMPAT_LD_FLAGS := -melf32lriscv
 
 # Disable attributes, as they're useless and break the build.
-- 
2.47.3

Re: [PATCH v3 3/5] riscv: compat_vdso: Introduce COMPAT_MARCH variable

[Thomas Weißschuh]

On Fri, Jul 17, 2026 at 10:06:36AM +0200, Nam Cao wrote:
> The -march option is inside COMPAT_CC_FLAGS, making it difficult to extend.
> 
> Introduce COMPAT_MARCH variable to allow extension. A follow-up commit
> will take advantage of this to add _zacas to -march.
> 
> Signed-off-by: Nam Cao <namcao@linutronix.de>

Reviewed-by: Thomas Weißschuh <thomas.weissschuh@linutronix.de>

> ---
>  arch/riscv/kernel/compat_vdso/Makefile | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)

[PATCH v3 4/5] riscv: compat_vdso: Build with zacas if available

[Nam Cao]

The compat vdso Makefile does not enable zacas, disallowing source
files from using zacas instructions. This is an obstacle to adding
robust __vdso_futex_robust_list32_try_unlock() in a follow-up commit.

Build with zacas, if compiler supports it.

Signed-off-by: Nam Cao <namcao@linutronix.de>
---
 arch/riscv/kernel/compat_vdso/Makefile | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
index 63580b33e3ce..ffbf84aeb081 100644
--- a/arch/riscv/kernel/compat_vdso/Makefile
+++ b/arch/riscv/kernel/compat_vdso/Makefile
@@ -18,6 +18,10 @@ else
 	COMPAT_MARCH := rv32imafd
 endif
 
+ifneq ($(call cc-option,-mabi=ilp32 -march=rv32ima_zacas),)
+COMPAT_MARCH := $(COMPAT_MARCH)_zacas
+endif
+
 COMPAT_CC_FLAGS := -march=$(COMPAT_MARCH) -mabi=ilp32
 
 COMPAT_LD_FLAGS := -melf32lriscv
-- 
2.47.3

Re: [PATCH v3 4/5] riscv: compat_vdso: Build with zacas if available

[Thomas Weißschuh]

On Fri, Jul 17, 2026 at 10:06:37AM +0200, Nam Cao wrote:
> The compat vdso Makefile does not enable zacas, disallowing source
> files from using zacas instructions. This is an obstacle to adding
> robust __vdso_futex_robust_list32_try_unlock() in a follow-up commit.

"obstacle" is a bit hard. It prevents using the optimized variant.

> Build with zacas, if compiler supports it.
> 
> Signed-off-by: Nam Cao <namcao@linutronix.de>

Reviewed-by: Thomas Weißschuh <thomas.weissschuh@linutronix.de>

> ---
>  arch/riscv/kernel/compat_vdso/Makefile | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
> index 63580b33e3ce..ffbf84aeb081 100644
> --- a/arch/riscv/kernel/compat_vdso/Makefile
> +++ b/arch/riscv/kernel/compat_vdso/Makefile
> @@ -18,6 +18,10 @@ else
>  	COMPAT_MARCH := rv32imafd
>  endif
>  
> +ifneq ($(call cc-option,-mabi=ilp32 -march=rv32ima_zacas),)
> +COMPAT_MARCH := $(COMPAT_MARCH)_zacas
> +endif

Could be :

ifeq ($(call cc-option-yn,-mabi=ilp32 -march=rv32ima_zacas),y)

> +
>  COMPAT_CC_FLAGS := -march=$(COMPAT_MARCH) -mabi=ilp32
>  
>  COMPAT_LD_FLAGS := -melf32lriscv
> -- 
> 2.47.3
> 

Re: [PATCH v3 4/5] riscv: compat_vdso: Build with zacas if available

[Vivian Wang]

Hi Nam,

On 7/17/26 16:06, Nam Cao wrote:

> The compat vdso Makefile does not enable zacas, disallowing source
> files from using zacas instructions. This is an obstacle to adding
> robust __vdso_futex_robust_list32_try_unlock() in a follow-up commit.
>
> Build with zacas, if compiler supports it.
>
> Signed-off-by: Nam Cao <namcao@linutronix.de>
> ---
>  arch/riscv/kernel/compat_vdso/Makefile | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
> index 63580b33e3ce..ffbf84aeb081 100644
> --- a/arch/riscv/kernel/compat_vdso/Makefile
> +++ b/arch/riscv/kernel/compat_vdso/Makefile
> @@ -18,6 +18,10 @@ else
>  	COMPAT_MARCH := rv32imafd
>  endif
>  
> +ifneq ($(call cc-option,-mabi=ilp32 -march=rv32ima_zacas),)
> +COMPAT_MARCH := $(COMPAT_MARCH)_zacas
> +endif
> +

I'm wondering if you would be interested in using .option arch, +zacas
for this, instead of enabling it for everything. 

I recently found out that we add and *only* add Zacas and Zabha to the
global -march for kernel code, just for use in assembly code in
arch/riscv/include/asm/cmpxchg.h.

So this is *probably* not a correctness problem, since compilers
probably don't do atomics on their own accord. However, I personally
think it would be a good idea to make everything consistently use
.option arch. This makes it so there's no question on what to do to get
extension instructions in assembly code - you always do. I have a patch
WIP that would make cmpxchg.h use .option arch as well that I will send
shortly.

You would need to write some more boilerplate:

.option push
.option arch, +zacas
amocas.w.rl ...
.option pop

Compilation of this code needs to be guarded behind something like if
(IS_ENABLED(CONFIG_RISCV_ISA_ZABHA)), which AFAICT your code does.

But in exchange for this boilerplate, it would be consistent with stuff
like Zbb, and there would be no more need to be plumbing -march around
in Makefiles.

What do you think?

Vivian "dramforever" Wang

Re: [PATCH v3 4/5] riscv: compat_vdso: Build with zacas if available

[Vivian Wang]

On 7/17/26 20:38, Vivian Wang wrote:

> Hi Nam,
> [...]
>
> I recently found out that we add and *only* add Zacas and Zabha to the

So ... okay, this is not strictly true, since we add F, D, V sometimes
as well, but those have their own state and need special handling so
it's not the same thing. And also they're only added for *some*
compilation units.

> global -march for kernel code, just for use in assembly code in
> arch/riscv/include/asm/cmpxchg.h.

[PATCH v3 5/5] riscv: vdso: Implement __vdso_futex_robust_try_unlock()

[Nam Cao]

This is the RISC-V port for __vdso_futex_robust_try_unlock(). It is based
on the x86's implementation in commit 61cfc8e372d1 ("x86/vdso: Prepare for
robust futex unlock support") and commit a2274cc0091e ("x86/vdso: Implement
__vdso_futex_robust_try_unlock()").

When the FUTEX_ROBUST_UNLOCK mechanism is used for unlocking (PI-)futexes,
the unlock sequence in userspace looks like this:

  1)    robust_list_set_op_pending(mutex);
  2)    robust_list_remove(mutex);

        lval = gettid();
  3)    if (atomic_try_cmpxchg(&mutex->lock, lval, 0))
  4)            robust_list_clear_op_pending();
        else
  5)            sys_futex(OP,...FUTEX_ROBUST_UNLOCK);

That still leaves a minimal race window between #3 and #4 where the mutex
could be acquired by some other task which observes that it is the last
user and:

  1) unmaps the mutex memory
  2) maps a different file, which ends up covering the same address

When then the original task exits before reaching #4 then the kernel robust
list handling observes the pending op entry and tries to fix up user space.

In case that the newly mapped data contains the TID of the exiting thread
at the address of the mutex/futex the kernel will set the owner died bit in
that memory and therefore corrupt unrelated data.

Provide a VDSO function which exposes the critical section window in the
VDSO symbol table. The resulting addresses are updated in the task's mm
when the VDSO is (re)map()'ed.

The core code detects when a task was interrupted within the critical
section and is about to deliver a signal. It then invokes an architecture
specific function which determines whether the pending op pointer has to be
cleared or not. The unlock assembly sequence on 64-bit is:

        mv      a5, a0           ; move LOCK to a5
    .Lretry
        lr.w    a0, (a5)         ; load *LOCK
        bne     a0, a1, 24 .Lend ; if (*LOCK != TID) then exit
        sc.w.rl t0, zero, (a5)   ; try changing *LOCK from TID to zero
    .Lstart
        bnez    t0, .Lretry      ; retry if store-conditional failed
        sd      zero, 0(a2)      ; clear POP
    .Lend
        sext.w	a0,a0
        ret

So if the kernel sees that the user task is between .Lstart and .Lend,
t0 can be checked to determine if the pending op pointer (POP) should
be cleared.

When the kernel is built with ZACAS, another unlock assembly sequence
is provided using ZACAS instructions. If the CPU supports ZACAS, that
alternative sequence is used and the critical section is also changed
to that.

There are two entry points to handle the different robust list pending op
pointer sizes:

        __vdso_futex_robust_list64_try_unlock()
        __vdso_futex_robust_list32_try_unlock()

The 32-bit VDSO and compat VDSO provides
__vdso_futex_robust_list32_try_unlock().

Unlike x86, the RISC-V 64-bit VDSO provides only provides
__vdso_futex_robust_list64_try_unlock(). Because RISC-V threads can
only have one list_op_pending pointer type.

Signed-off-by: Nam Cao <namcao@linutronix.de>
---
 arch/riscv/Kconfig                         |  1 +
 arch/riscv/include/asm/cpufeature-macros.h |  2 +
 arch/riscv/include/asm/futex_robust.h      | 16 +++++
 arch/riscv/include/asm/vdso/futex.h        | 14 ++++
 arch/riscv/kernel/compat_vdso/Makefile     |  7 +-
 arch/riscv/kernel/compat_vdso/futex.c      |  3 +
 arch/riscv/kernel/vdso.c                   | 61 ++++++++++++++++
 arch/riscv/kernel/vdso/Makefile            | 14 +++-
 arch/riscv/kernel/vdso/futex.c             | 84 ++++++++++++++++++++++
 arch/riscv/kernel/vdso/vdso.lds.S          |  8 +++
 10 files changed, 208 insertions(+), 2 deletions(-)
 create mode 100644 arch/riscv/include/asm/futex_robust.h
 create mode 100644 arch/riscv/include/asm/vdso/futex.h
 create mode 100644 arch/riscv/kernel/compat_vdso/futex.c
 create mode 100644 arch/riscv/kernel/vdso/futex.c

diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
index c2c2f5f3ed5e..588963bef804 100644
--- a/arch/riscv/Kconfig
+++ b/arch/riscv/Kconfig
@@ -166,6 +166,7 @@ config RISCV
 	select HAVE_FUNCTION_GRAPH_TRACER if HAVE_DYNAMIC_FTRACE_WITH_ARGS
 	select HAVE_FUNCTION_GRAPH_FREGS
 	select HAVE_FUNCTION_TRACER if HAVE_DYNAMIC_FTRACE
+	select HAVE_FUTEX_ROBUST_UNLOCK
 	select HAVE_EBPF_JIT if MMU
 	select HAVE_GENERIC_TIF_BITS
 	select HAVE_GUP_FAST if MMU
diff --git a/arch/riscv/include/asm/cpufeature-macros.h b/arch/riscv/include/asm/cpufeature-macros.h
index a8103edbf51f..f522c6c3cd97 100644
--- a/arch/riscv/include/asm/cpufeature-macros.h
+++ b/arch/riscv/include/asm/cpufeature-macros.h
@@ -6,6 +6,8 @@
 #ifndef _ASM_CPUFEATURE_MACROS_H
 #define _ASM_CPUFEATURE_MACROS_H
 
+#include <linux/types.h>
+
 #include <asm/hwcap.h>
 #include <asm/alternative-macros.h>
 
diff --git a/arch/riscv/include/asm/futex_robust.h b/arch/riscv/include/asm/futex_robust.h
new file mode 100644
index 000000000000..16ec58e925c8
--- /dev/null
+++ b/arch/riscv/include/asm/futex_robust.h
@@ -0,0 +1,16 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_RISCV_FUTEX_ROBUST_H
+#define _ASM_RISCV_FUTEX_ROBUST_H
+
+#include <asm/ptrace.h>
+#include <asm/vdso/futex.h>
+
+static inline void __user *arch_futex_robust_unlock_get_pop(struct pt_regs *regs)
+{
+	if (cpu_supports_zacas())
+		return (regs->a0 == regs->a1) ? (void __user *)regs->a2 : NULL;
+	else
+		return (regs->t0 == 0) ? (void __user *)regs->a2 : NULL;
+}
+
+#endif /* _ASM_RISCV_FUTEX_ROBUST_H */
diff --git a/arch/riscv/include/asm/vdso/futex.h b/arch/riscv/include/asm/vdso/futex.h
new file mode 100644
index 000000000000..9e173f45f3e2
--- /dev/null
+++ b/arch/riscv/include/asm/vdso/futex.h
@@ -0,0 +1,14 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+
+#ifndef __ASM_VDSO_FUTEX_H
+#define __ASM_VDSO_FUTEX_H
+
+#include <asm/cpufeature-macros.h>
+
+static inline bool cpu_supports_zacas(void)
+{
+	return IS_ENABLED(CONFIG_RISCV_ISA_ZACAS) && IS_ENABLED(CONFIG_TOOLCHAIN_HAS_ZACAS) &&
+		riscv_has_extension_unlikely(RISCV_ISA_EXT_ZACAS);
+}
+
+#endif /* __ASM_VDSO_FUTEX_H */
diff --git a/arch/riscv/kernel/compat_vdso/Makefile b/arch/riscv/kernel/compat_vdso/Makefile
index ffbf84aeb081..5736297e6281 100644
--- a/arch/riscv/kernel/compat_vdso/Makefile
+++ b/arch/riscv/kernel/compat_vdso/Makefile
@@ -31,7 +31,12 @@ COMPAT_CC_FLAGS += $(call cc-option,-mno-riscv-attribute)
 COMPAT_CC_FLAGS += $(call as-option,-Wa$(comma)-mno-arch-attr)
 
 # Files to link into the compat_vdso
-obj-compat_vdso = $(patsubst %, %.o, $(compat_vdso-syms)) note.o
+obj-compat_vdso := $(patsubst %, %.o, $(compat_vdso-syms)) note.o
+
+ifdef CONFIG_FUTEX_ROBUST_UNLOCK
+compat_vdso-syms += __vdso_futex_robust_list32_try_unlock
+obj-compat_vdso += futex.o
+endif
 
 # Build rules
 targets := $(obj-compat_vdso) compat_vdso.so compat_vdso.so.dbg compat_vdso.lds
diff --git a/arch/riscv/kernel/compat_vdso/futex.c b/arch/riscv/kernel/compat_vdso/futex.c
new file mode 100644
index 000000000000..ad05ded43955
--- /dev/null
+++ b/arch/riscv/kernel/compat_vdso/futex.c
@@ -0,0 +1,3 @@
+// SPDX-License-Identifier: GPL-2.0-only
+
+#include "../vdso/futex.c"
diff --git a/arch/riscv/kernel/vdso.c b/arch/riscv/kernel/vdso.c
index 9c2f5e442338..2610363f9a02 100644
--- a/arch/riscv/kernel/vdso.c
+++ b/arch/riscv/kernel/vdso.c
@@ -11,6 +11,7 @@
 #include <linux/slab.h>
 #include <linux/binfmts.h>
 #include <linux/err.h>
+#include <linux/futex.h>
 #include <asm/page.h>
 #include <asm/vdso.h>
 #include <linux/vdso_datastore.h>
@@ -33,11 +34,69 @@ static struct __vdso_info vdso_info;
 static struct __vdso_info compat_vdso_info;
 #endif
 
+#define _CONCAT3(a, b, c)	a ## b ## c
+#define CONCAT3(a, b, c)	_CONCAT3(a, b, c)
+
+#if defined(CONFIG_RISCV_ISA_ZACAS) && defined(CONFIG_TOOLCHAIN_HAS_ZACAS)
+#define FUTEX_CAS_SET_VDSO_CS_RANGE(vdso, fd, idx, xlen, symbol)				\
+({												\
+	void *start = symbol(vdso, CONCAT3(futex_list, xlen, _try_unlock_cs_cas_start));	\
+	void *end   = symbol(vdso, CONCAT3(futex_list, xlen, _try_unlock_cs_cas_end));		\
+												\
+	futex_set_vdso_cs_range(fd, idx, (uintptr_t)start, (uintptr_t)end, xlen == 32);		\
+})
+#else
+#define FUTEX_CAS_SET_VDSO_CS_RANGE(...) BUILD_BUG()
+#endif
+
+#define FUTEX_LRSC_SET_VDSO_CS_RANGE(vdso, fd, idx, xlen, symbol)				\
+({												\
+	void *start = symbol(vdso, CONCAT3(futex_list, xlen, _try_unlock_cs_lrsc_start));	\
+	void *end   = symbol(vdso, CONCAT3(futex_list, xlen, _try_unlock_cs_lrsc_end));		\
+												\
+	futex_set_vdso_cs_range(fd, idx, (uintptr_t)start, (uintptr_t)end, xlen == 32);		\
+})
+
+#define FUTEX_SET_VDSO_CS_RANGE(vdso, fd, idx, xlen, symbol)					\
+({												\
+	if (cpu_supports_zacas())								\
+		FUTEX_CAS_SET_VDSO_CS_RANGE(vdso, fd, idx, xlen, symbol);			\
+	else											\
+		FUTEX_LRSC_SET_VDSO_CS_RANGE(vdso, fd, idx, xlen, symbol);			\
+})
+
+#ifdef CONFIG_FUTEX_ROBUST_UNLOCK
+
+/* Allow parameters to expand first */
+#define __VDSO_SYMBOL(vdso, name)		VDSO_SYMBOL(vdso, name)
+#define __COMPAT_VDSO_SYMBOL(vdso, name)	COMPAT_VDSO_SYMBOL(vdso, name)
+
+static void vdso_futex_robust_unlock_update_ips(void)
+{
+	unsigned long vdso = (unsigned long) current->mm->context.vdso;
+	struct futex_mm_data *fd = &current->mm->futex;
+
+	futex_reset_cs_ranges(fd);
+
+	if (!is_compat_task())
+		FUTEX_SET_VDSO_CS_RANGE(vdso, fd, 0, __riscv_xlen, __VDSO_SYMBOL);
+
+#ifdef CONFIG_COMPAT
+	if (is_compat_task())
+		FUTEX_SET_VDSO_CS_RANGE(vdso, fd, 0, 32, __COMPAT_VDSO_SYMBOL);
+#endif
+}
+#else
+static inline void vdso_futex_robust_unlock_update_ips(void) {}
+#endif
+
 static int vdso_mremap(const struct vm_special_mapping *sm,
 		       struct vm_area_struct *new_vma)
 {
 	current->mm->context.vdso = (void *)new_vma->vm_start;
 
+	vdso_futex_robust_unlock_update_ips();
+
 	return 0;
 }
 
@@ -147,6 +206,8 @@ static int __setup_additional_pages(struct mm_struct *mm,
 	if (IS_ERR(ret))
 		goto up_fail;
 
+	vdso_futex_robust_unlock_update_ips();
+
 	return 0;
 
 up_fail:
diff --git a/arch/riscv/kernel/vdso/Makefile b/arch/riscv/kernel/vdso/Makefile
index 8dbf2532a573..d4238db8c6d3 100644
--- a/arch/riscv/kernel/vdso/Makefile
+++ b/arch/riscv/kernel/vdso/Makefile
@@ -27,12 +27,24 @@ asflags-y += -DVDSO_CFI=1
 endif
 
 # Files to link into the vdso
-obj-vdso = $(patsubst %, %.o, $(vdso-syms)) note.o
+obj-vdso := $(patsubst %, %.o, $(vdso-syms)) note.o
 
 ifdef CONFIG_VDSO_GETRANDOM
 obj-vdso += vgetrandom-chacha.o
 endif
 
+ifdef CONFIG_FUTEX_ROBUST_UNLOCK
+
+obj-vdso += futex.o
+
+ifdef CONFIG_64BIT
+vdso-syms += __vdso_futex_robust_list64_try_unlock
+else
+vdso-syms += __vdso_futex_robust_list32_try_unlock
+endif
+
+endif
+
 ccflags-y := -fno-stack-protector
 ccflags-y += -DDISABLE_BRANCH_PROFILING
 ccflags-y += -fno-builtin
diff --git a/arch/riscv/kernel/vdso/futex.c b/arch/riscv/kernel/vdso/futex.c
new file mode 100644
index 000000000000..7da77e635012
--- /dev/null
+++ b/arch/riscv/kernel/vdso/futex.c
@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: GPL-2.0-only
+#include <linux/stringify.h>
+#include <vdso/futex.h>
+#include <asm/hwcap.h>
+#include <asm/cpufeature-macros.h>
+#include <asm/vdso/futex.h>
+
+#define LABEL(pop_size, which) __stringify(__vdso_futex_list##pop_size##_try_unlock_cs_##which)
+
+#define futex_robust_try_unlock_cas(pop_size, store_pop, lock, tid, pop)	\
+({										\
+	/*									\
+	 * arch_futex_robust_unlock_get_pop() assumes the variables are in	\
+	 * those registers. So make sure.					\
+	 *									\
+	 * tid and pop are in a1 and a2 at function entry according to the	\
+	 * calling convention, so it likely still works if we remove _tid	\
+	 * and _pop. But technically compiler is allowed to move tid and pop	\
+	 * to different registers, and _tid and _pop do not generate any	\
+	 * extra instructions so it does not hurt to keep them.			\
+	 */									\
+	register __u32  ret asm ("a0") = tid;					\
+	register __u32 _tid asm ("a1") = tid;					\
+	register void *_pop asm ("a2") = pop;					\
+										\
+	asm volatile (								\
+		"	amocas.w.rl	%[ret], zero, (%[lock])\n"		\
+		LABEL(pop_size, cas_start)":"					\
+		"	bne %[ret], %[tid], "LABEL(pop_size, cas_end)"\n"	\
+		"	"store_pop" zero, (%[pop])\n"				\
+		LABEL(pop_size, cas_end)":"					\
+		: [ret]  "+&r" (ret)						\
+		: [tid]  "r"   (_tid),						\
+		  [lock] "r"   (lock),						\
+		  [pop]  "r"   (_pop)						\
+		: "memory"							\
+	);									\
+										\
+	ret;									\
+})
+
+#define futex_robust_try_unlock_lrsc(pop_size, store_pop, lock, tid, pop)	\
+({										\
+	register void *_pop asm ("a2") = pop;					\
+	__u32 ret;								\
+										\
+	asm volatile (								\
+		"1:	lr.w %[ret], (%[lock])\n"				\
+		"	bne %[ret], %[tid], "LABEL(pop_size, lrsc_end)"\n"	\
+		"	sc.w.rl	t0, x0, (%[lock])\n"				\
+		LABEL(pop_size, lrsc_start)":"					\
+		"	bnez t0, 1b\n"						\
+		"	"store_pop" zero, (%[pop])\n"				\
+		LABEL(pop_size, lrsc_end)":"					\
+		: [ret]  "=&r" (ret)						\
+		: [tid]  "r"   (tid),						\
+		  [lock] "r"   (lock),						\
+		  [pop]  "r"   (_pop)						\
+		: "t0", "memory"						\
+	);									\
+										\
+	ret;									\
+})
+
+
+#if __riscv_xlen == 64
+__u32 __vdso_futex_robust_list64_try_unlock(__u32 *lock, __u32 tid, __u64 *pop)
+{
+	if (cpu_supports_zacas())
+		return futex_robust_try_unlock_cas(64, "sd", lock, tid, pop);
+	else
+		return futex_robust_try_unlock_lrsc(64, "sd", lock, tid, pop);
+}
+#endif
+
+#if __riscv_xlen == 32
+__u32 __vdso_futex_robust_list32_try_unlock(__u32 *lock, __u32 tid, __u32 *pop)
+{
+	if (cpu_supports_zacas())
+		return futex_robust_try_unlock_cas(32, "sw", lock, tid, pop);
+	else
+		return futex_robust_try_unlock_lrsc(32, "sw", lock, tid, pop);
+}
+#endif
diff --git a/arch/riscv/kernel/vdso/vdso.lds.S b/arch/riscv/kernel/vdso/vdso.lds.S
index c29ef12a63bb..9acfa9ef5bab 100644
--- a/arch/riscv/kernel/vdso/vdso.lds.S
+++ b/arch/riscv/kernel/vdso/vdso.lds.S
@@ -82,6 +82,14 @@ VERSION
 #endif
 #if defined(CONFIG_VDSO_GETRANDOM) && !defined(COMPAT_VDSO)
 		__vdso_getrandom;
+#endif
+
+#ifdef CONFIG_FUTEX_ROBUST_UNLOCK
+#if defined(CONFIG_32BIT) || defined(COMPAT_VDSO)
+		__vdso_futex_robust_list32_try_unlock;
+#else
+		__vdso_futex_robust_list64_try_unlock;
+#endif
 #endif
 	local: *;
 	};
-- 
2.47.3

Re: [PATCH v3 5/5] riscv: vdso: Implement __vdso_futex_robust_try_unlock()

[Thomas Weißschuh]

On Fri, Jul 17, 2026 at 10:06:38AM +0200, Nam Cao wrote:
(...)

> Signed-off-by: Nam Cao <namcao@linutronix.de>

One question about the toolchain detection.
With that cleared up:

Reviewed-by: Thomas Weißschuh <thomas.weissschuh@linutronix.de>

(Minus the asm bits...)

> ---
>  arch/riscv/Kconfig                         |  1 +
>  arch/riscv/include/asm/cpufeature-macros.h |  2 +
>  arch/riscv/include/asm/futex_robust.h      | 16 +++++
>  arch/riscv/include/asm/vdso/futex.h        | 14 ++++
>  arch/riscv/kernel/compat_vdso/Makefile     |  7 +-
>  arch/riscv/kernel/compat_vdso/futex.c      |  3 +
>  arch/riscv/kernel/vdso.c                   | 61 ++++++++++++++++
>  arch/riscv/kernel/vdso/Makefile            | 14 +++-
>  arch/riscv/kernel/vdso/futex.c             | 84 ++++++++++++++++++++++
>  arch/riscv/kernel/vdso/vdso.lds.S          |  8 +++
>  10 files changed, 208 insertions(+), 2 deletions(-)
>  create mode 100644 arch/riscv/include/asm/futex_robust.h
>  create mode 100644 arch/riscv/include/asm/vdso/futex.h
>  create mode 100644 arch/riscv/kernel/compat_vdso/futex.c
>  create mode 100644 arch/riscv/kernel/vdso/futex.c

(...)

> diff --git a/arch/riscv/include/asm/futex_robust.h b/arch/riscv/include/asm/futex_robust.h
> new file mode 100644
> index 000000000000..16ec58e925c8
> --- /dev/null
> +++ b/arch/riscv/include/asm/futex_robust.h
> @@ -0,0 +1,16 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef _ASM_RISCV_FUTEX_ROBUST_H
> +#define _ASM_RISCV_FUTEX_ROBUST_H
> +
> +#include <asm/ptrace.h>
> +#include <asm/vdso/futex.h>
> +
> +static inline void __user *arch_futex_robust_unlock_get_pop(struct pt_regs *regs)
> +{
> +	if (cpu_supports_zacas())
> +		return (regs->a0 == regs->a1) ? (void __user *)regs->a2 : NULL;
> +	else
> +		return (regs->t0 == 0) ? (void __user *)regs->a2 : NULL;
> +}

Can it ever happen that the toolchain has differing zacas support between
32-bit and 64-bit? Then cpu_supports_zacas() may not be sufficient.
If it is guaranteed to be in sync I would prefer to reuse
CONFIG_TOOLCHAIN_HAS_ZACAS instead of the cc-option check in the Makefile
to make that clear and error out early if it should ever change.

Also this assumes that compat vDSO alternative patching works.
But this is not the case. My series [0] to implement it has not yet been
applied. This revision is missing the note about this dependency.

[0] https://lore.kernel.org/all/20260630-riscv-vdso32-alternative-v1-0-a32fd89b7b1c@linutronix.de/

> +#endif /* _ASM_RISCV_FUTEX_ROBUST_H */

(...)

> diff --git a/arch/riscv/kernel/vdso/Makefile b/arch/riscv/kernel/vdso/Makefile
> index 8dbf2532a573..d4238db8c6d3 100644
> --- a/arch/riscv/kernel/vdso/Makefile
> +++ b/arch/riscv/kernel/vdso/Makefile
> @@ -27,12 +27,24 @@ asflags-y += -DVDSO_CFI=1
>  endif
>  
>  # Files to link into the vdso
> -obj-vdso = $(patsubst %, %.o, $(vdso-syms)) note.o
> +obj-vdso := $(patsubst %, %.o, $(vdso-syms)) note.o
>  
>  ifdef CONFIG_VDSO_GETRANDOM
>  obj-vdso += vgetrandom-chacha.o
>  endif
>  
> +ifdef CONFIG_FUTEX_ROBUST_UNLOCK
> +
> +obj-vdso += futex.o
> +
> +ifdef CONFIG_64BIT
> +vdso-syms += __vdso_futex_robust_list64_try_unlock
> +else
> +vdso-syms += __vdso_futex_robust_list32_try_unlock
> +endif

Could be :

vdso-syms += __vdso_futex_robust_list$(BITS)_try_unlock

> +
> +endif
> +
>  ccflags-y := -fno-stack-protector
>  ccflags-y += -DDISABLE_BRANCH_PROFILING
>  ccflags-y += -fno-builtin

(...)