On 6/1/23 12:14, Peter Zijlstra wrote:
On Wed, May 31, 2023 at 04:21:22PM +0200, Arnd Bergmann wrote:
It would be nice to have the hack more localized to parisc
and guarded with a CONFIG_GCC_VERSION check so we can kill
it off in the future, once we drop either gcc-10 or parisc
support.
I vote for dropping parisc -- it's the only 64bit arch that doesn't have
sane atomics.
Of course I'm against dropping parisc.
Anyway, the below seems to work -- build tested with GCC-10.1
I don't think we need to care about gcc-10 on parisc.
Debian and Gentoo are the only supported distributions, while Debian
requires gcc-12 to build > 6.x kernels, and I assume Gentoo uses at least
gcc-12 as well.
So raising the gcc limit for parisc only (at least temporarily for now)
should be fine and your workaround below wouldn't be necessary, right?
Helge
---
Subject: parisc/percpu: Work around the lack of __SIZEOF_INT128__
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Tue May 30 22:27:40 CEST 2023
HPPA64 is unique in not providing __SIZEOF_INT128__ across all
supported compilers, specifically it only started doing this with
GCC-11.
Since the per-cpu ops are universally availably, and
this_cpu_{,try_}cmpxchg128() is expected to be available on all 64bit
architectures a wee bodge is in order.
Sadly, while C reverts to memcpy() for assignment of POD types, it does
not revert to memcmp() for for equality. Therefore frob that manually.
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
arch/parisc/include/asm/percpu.h | 77 +++++++++++++++++++++++++++++++++++++++
1 file changed, 77 insertions(+)
--- /dev/null
+++ b/arch/parisc/include/asm/percpu.h
@@ -0,0 +1,77 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_PARISC_PERCPU_H
+#define _ASM_PARISC_PERCPU_H
+
+#include <linux/types.h>
+
+#if defined(CONFIG_64BIT) && CONFIG_GCC_VERSION < 1100000
+
+/*
+ * GCC prior to 11 does not provide __SIZEOF_INT128__ on HPPA64
+ * as such we need to provide an alternative implementation of
+ * {raw,this}_cpu_{,try_}cmpxchg128().
+ *
+ * This obviously doesn't function as u128 should, but for the purpose
+ * of per-cpu cmpxchg128 it might just do.
+ */
+typedef struct {
+ u64 a, b;
+} u128 __attribute__((aligned(16)));
+
+#define raw_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval) \
+({ \
+ typeof(pcp) *__p = raw_cpu_ptr(&(pcp)); \
+ typeof(pcp) __val = *__p, __old = *(ovalp); \
+ bool __ret; \
+ if (!__builtin_memcmp(&__val, &__old, sizeof(pcp))) { \
+ *__p = nval; \
+ __ret = true; \
+ } else { \
+ *(ovalp) = __val; \
+ __ret = false; \
+ } \
+ __ret; \
+})
+
+#define raw_cpu_generic_cmpxchg_memcmp(pcp, oval, nval) \
+({ \
+ typeof(pcp) __old = (oval); \
+ raw_cpu_generic_try_cmpxchg_memcpy(pcp, &__old, nval); \
+ __old; \
+})
+
+#define raw_cpu_cmpxchg128(pcp, oval, nval) \
+ raw_cpu_generic_cmpxchg_memcmp(pcp, oval, nval)
+#define raw_cpu_try_cmpxchg128(pcp, ovalp, nval) \
+ raw_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval)
+
+#define this_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval) \
+({ \
+ bool __ret; \
+ unsigned long __flags; \
+ raw_local_irq_save(__flags); \
+ __ret = raw_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval); \
+ raw_local_irq_restore(__flags); \
+ __ret; \
+})
+
+#define this_cpu_generic_cmpxchg_memcmp(pcp, oval, nval) \
+({ \
+ typeof(pcp) __ret; \
+ unsigned long __flags; \
+ raw_local_irq_save(__flags); \
+ __ret = raw_cpu_generic_cmpxchg_memcmp(pcp, oval, nval); \
+ raw_local_irq_restore(__flags); \
+ __ret; \
+})
+
+#define this_cpu_cmpxchg128(pcp, oval, nval) \
+ this_cpu_generic_cmpxchg_memcmp(pcp, oval, nval)
+#define this_cpu_try_cmpxchg128(pcp, ovalp, nval) \
+ this_cpu_generic_try_cmpxchg_memcmp(pcp, ovalp, nval)
+
+#endif /* !__SIZEOF_INT128__ */
+
+#include <asm-generic/percpu.h>
+
+#endif /* _ASM_PARISC_PERCPU_H */
