[PATCH] arm64: futex: Consolidate 'old == new' check in __lsui_cmpxchg32()
Will Deacon
will at kernel.org
Tue Jun 2 07:19:35 PDT 2026
On Tue, May 19, 2026 at 04:09:02PM +0100, Catalin Marinas wrote:
> On Tue, May 19, 2026 at 10:08:22AM +0100, Will Deacon wrote:
> > The LSUI futex implementation relies on a cmpxchg() loop to implement
> > FUTEX_OP_XOR, as the architecture doesn't provide unprivileged *EOR
> > atomics. Since the unprivileged 'CAST' instructions used to implement
> > the cmpxchg() can only operate on 64-bit memory locations, the
> > __lsui_cmpxchg32() helper function performs a song and dance to marshall
> > the 32-bit futex value into the correct part of a 64-bit register and
> > fill the remaining bytes with the neighbouring data.
>
> IIRC, the reason for the current __lsui_cmpxchg32() was not EOR but the
> expected futex_atomic_cmpxchg_inatomic() semantics. Looking at it again,
> we have wake_futex_pi() that does something else if the ret is 0 but the
> value differs. Looking at it again, the caller of wake_futex_pi()
> retries on -EAGAIN anyway, so I don't see a correctness issue, it will
> eventually hit the condition.
Hmm, but I think that means my patch does change the behaviour of
wake_futex_pi() in an undesirable way. For example, futex_unlock_pi()
will go round the retry loop for any change in the futex value, whereas
before we would go back to userspace only if the TID changed.
So I think we should swallow the -EAGAIN for the CAS-based cmpxchg() if
the futex word has changed, along the lines of the diff below.
Will
--->8
diff --git a/arch/arm64/include/asm/futex.h b/arch/arm64/include/asm/futex.h
index db84a7b2de74..79c6d86c38a9 100644
--- a/arch/arm64/include/asm/futex.h
+++ b/arch/arm64/include/asm/futex.h
@@ -215,14 +215,14 @@ __lsui_futex_atomic_eor(int oparg, u32 __user *uaddr, int *oval)
static __always_inline int
__lsui_futex_cmpxchg(u32 __user *uaddr, u32 oldval, u32 newval, u32 *oval)
{
+ u32 curval = oldval;
int ret;
- /*
- * Callers of futex_atomic_cmpxchg_inatomic() already retry on
- * -EAGAIN, no need for another loop of max retries.
- */
- ret = __lsui_cmpxchg32(uaddr, &oldval, newval);
- *oval = oldval;
+ ret = __lsui_cmpxchg32(uaddr, &curval, newval);
+ if (ret == -EAGAIN && curval != oldval)
+ ret = 0;
+
+ *oval = curval;
return ret;
}
#endif /* CONFIG_ARM64_LSUI */
More information about the linux-arm-kernel
mailing list