zone->lock and zone->lru_lock are two of the hottest locks in the
kernel. Their usage patterns are quite independent. And they have
just been put into the same structure. It is essential that they not
fall into the same cacheline.
That could be fixed by padding with L1_CACHE_BYTES. But the problem
with this is that a kernel which was configured for (say) a PIII will
perform poorly on SMP PIV. This will cause problems for kernel
vendors. For example, RH currently ship PII and Athlon binaries. To
get best SMP performance they will end up needing to ship a lot of
differently configured kernels.
To solve this we need to know, at compile time, the maximum L1 size
which this kernel will ever run on.
This patch adds L1_CACHE_SHIFT_MAX to every architecture's cache.h.
Of course it'll break when newer chips come out with increased
cacheline sizes. Better suggestions are welcome.
asm-alpha/cache.h | 1 +
asm-arm/cache.h | 2 ++
asm-cris/cache.h | 2 ++
asm-i386/cache.h | 2 ++
asm-ia64/cache.h | 2 ++
asm-m68k/cache.h | 2 ++
asm-mips/cache.h | 1 +
asm-mips64/cache.h | 1 +
asm-parisc/cache.h | 1 +
asm-ppc/cache.h | 1 +
asm-ppc64/cache.h | 1 +
asm-s390/cache.h | 1 +
asm-s390x/cache.h | 1 +
asm-sh/cache.h | 2 ++
asm-sparc/cache.h | 1 +
asm-sparc64/cache.h | 1 +
asm-x86_64/cache.h | 1 +
linux/cache.h | 9 +++++++++
18 files changed, 32 insertions(+)
--- 2.5.31/include/linux/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/linux/cache.h Sun Aug 11 00:20:35 2002
@@ -44,4 +44,13 @@
#endif /* CONFIG_SMP */
#endif
+#if !defined(____cacheline_maxaligned_in_smp)
+#if defined(CONFIG_SMP)
+#define ____cacheline_maxaligned_in_smp \
+ __attribute__((__aligned__(1 << (L1_CACHE_SHIFT_MAX))))
+#else
+#define ____cacheline_maxaligned_in_smp
+#endif
+#endif
+
#endif /* __LINUX_CACHE_H */
--- 2.5.31/include/asm-alpha/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-alpha/cache.h Sun Aug 11 00:20:35 2002
@@ -20,5 +20,6 @@
#define L1_CACHE_ALIGN(x) (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
#define SMP_CACHE_BYTES L1_CACHE_BYTES
+#define L1_CACHE_SHIFT_MAX 6 /* largest L1 which this arch supports */
#endif
--- 2.5.31/include/asm-arm/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-arm/cache.h Sun Aug 11 00:20:35 2002
@@ -16,4 +16,6 @@
__section__(".data.cacheline_aligned")))
#endif
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
+
#endif
--- 2.5.31/include/asm-cris/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-cris/cache.h Sun Aug 11 00:20:35 2002
@@ -7,4 +7,6 @@
#define L1_CACHE_BYTES 32
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
+
#endif /* _ASM_CACHE_H */
--- 2.5.31/include/asm-i386/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-i386/cache.h Sun Aug 11 00:20:35 2002
@@ -10,4 +10,6 @@
#define L1_CACHE_SHIFT (CONFIG_X86_L1_CACHE_SHIFT)
#define L1_CACHE_BYTES (1 << L1_CACHE_SHIFT)
+#define L1_CACHE_SHIFT_MAX 7 /* largest L1 which this arch supports */
+
#endif
--- 2.5.31/include/asm-ia64/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-ia64/cache.h Sun Aug 11 00:20:35 2002
@@ -12,6 +12,8 @@
#define L1_CACHE_SHIFT CONFIG_IA64_L1_CACHE_SHIFT
#define L1_CACHE_BYTES (1 << L1_CACHE_SHIFT)
+#define L1_CACHE_SHIFT_MAX 7 /* largest L1 which this arch supports */
+
#ifdef CONFIG_SMP
# define SMP_CACHE_SHIFT L1_CACHE_SHIFT
# define SMP_CACHE_BYTES L1_CACHE_BYTES
--- 2.5.31/include/asm-m68k/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-m68k/cache.h Sun Aug 11 00:20:35 2002
@@ -8,4 +8,6 @@
#define L1_CACHE_SHIFT 4
#define L1_CACHE_BYTES (1<< L1_CACHE_SHIFT)
+#define L1_CACHE_SHIFT_MAX 4 /* largest L1 which this arch supports */
+
#endif
--- 2.5.31/include/asm-mips/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-mips/cache.h Sun Aug 11 00:20:35 2002
@@ -35,5 +35,6 @@ struct cache_desc {
#endif
#define SMP_CACHE_BYTES L1_CACHE_BYTES
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
#endif /* _ASM_CACHE_H */
--- 2.5.31/include/asm-mips64/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-mips64/cache.h Sun Aug 11 00:20:35 2002
@@ -11,5 +11,6 @@
/* bytes per L1 cache line */
#define L1_CACHE_BYTES (1 << CONFIG_L1_CACHE_SHIFT)
+#define L1_CACHE_SHIFT_MAX 7 /* largest L1 which this arch supports */
#endif /* _ASM_CACHE_H */
--- 2.5.31/include/asm-parisc/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-parisc/cache.h Sun Aug 11 00:20:35 2002
@@ -34,6 +34,7 @@
#define L1_CACHE_ALIGN(x) (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
#define SMP_CACHE_BYTES L1_CACHE_BYTES
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
#define __cacheline_aligned __attribute__((__aligned__(L1_CACHE_BYTES)))
--- 2.5.31/include/asm-ppc/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-ppc/cache.h Sun Aug 11 00:20:35 2002
@@ -28,6 +28,7 @@
#define L1_CACHE_BYTES L1_CACHE_LINE_SIZE
#define SMP_CACHE_BYTES L1_CACHE_BYTES
+#define L1_CACHE_SHIFT_MAX 7 /* largest L1 which this arch supports */
#define L1_CACHE_ALIGN(x) (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
#define L1_CACHE_PAGES 8
--- 2.5.31/include/asm-ppc64/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-ppc64/cache.h Sun Aug 11 00:20:35 2002
@@ -12,5 +12,6 @@
#define L1_CACHE_BYTES (1 << L1_CACHE_SHIFT)
#define SMP_CACHE_BYTES L1_CACHE_BYTES
+#define L1_CACHE_SHIFT_MAX 7 /* largest L1 which this arch supports */
#endif
--- 2.5.31/include/asm-s390/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-s390/cache.h Sun Aug 11 00:20:35 2002
@@ -13,5 +13,6 @@
#define L1_CACHE_BYTES 256
#define L1_CACHE_SHIFT 8
+#define L1_CACHE_SHIFT_MAX 8 /* largest L1 which this arch supports */
#endif
--- 2.5.31/include/asm-s390x/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-s390x/cache.h Sun Aug 11 00:20:35 2002
@@ -13,5 +13,6 @@
#define L1_CACHE_BYTES 256
#define L1_CACHE_SHIFT 8
+#define L1_CACHE_SHIFT_MAX 8 /* largest L1 which this arch supports */
#endif
--- 2.5.31/include/asm-sh/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-sh/cache.h Sun Aug 11 00:20:35 2002
@@ -14,4 +14,6 @@
#define L1_CACHE_BYTES 32
#endif
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
+
#endif /* __ASM_SH_CACHE_H */
--- 2.5.31/include/asm-sparc/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-sparc/cache.h Sun Aug 11 00:20:35 2002
@@ -13,6 +13,7 @@
#define L1_CACHE_SHIFT 5
#define L1_CACHE_BYTES 32
#define L1_CACHE_ALIGN(x) ((((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1)))
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
#define SMP_CACHE_BYTES 32
--- 2.5.31/include/asm-sparc64/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-sparc64/cache.h Sun Aug 11 00:20:35 2002
@@ -9,6 +9,7 @@
#define L1_CACHE_BYTES 32 /* Two 16-byte sub-blocks per line. */
#define L1_CACHE_ALIGN(x) (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
+#define L1_CACHE_SHIFT_MAX 5 /* largest L1 which this arch supports */
#define SMP_CACHE_BYTES_SHIFT 6
#define SMP_CACHE_BYTES (1 << SMP_CACHE_BYTES_SHIFT) /* L2 cache line size. */
--- 2.5.31/include/asm-x86_64/cache.h~l1-max-size Sun Aug 11 00:20:35 2002
+++ 2.5.31-akpm/include/asm-x86_64/cache.h Sun Aug 11 00:20:35 2002
@@ -9,5 +9,6 @@
/* L1 cache line size */
#define L1_CACHE_SHIFT (CONFIG_X86_L1_CACHE_SHIFT)
#define L1_CACHE_BYTES (1 << L1_CACHE_SHIFT)
+#define L1_CACHE_SHIFT_MAX 6 /* largest L1 which this arch supports */
#endif
.
On Sunday 11 August 2002 09:40, Andrew Morton wrote:
> zone->lock and zone->lru_lock are two of the hottest locks in the
> kernel. Their usage patterns are quite independent. And they have
> just been put into the same structure. It is essential that they not
> fall into the same cacheline.
>
> That could be fixed by padding with L1_CACHE_BYTES. But the problem
> with this is that a kernel which was configured for (say) a PIII will
> perform poorly on SMP PIV. This will cause problems for kernel
> vendors. For example, RH currently ship PII and Athlon binaries. To
> get best SMP performance they will end up needing to ship a lot of
> differently configured kernels.
>
> To solve this we need to know, at compile time, the maximum L1 size
> which this kernel will ever run on.
>
> This patch adds L1_CACHE_SHIFT_MAX to every architecture's cache.h.
>
> Of course it'll break when newer chips come out with increased
> cacheline sizes. Better suggestions are welcome.
I think you're being too paranoid. You pushed the performance degradation
from the PIV to the PIII (because it will tend to hit more cachelines than it
should) and you won't be able to build a kernel that is optimal for the PIII
any more. I'd say that is PIII kernel is *supposed* to suck to some degree
when run on a PIV, otherwise why bother having the PIV option?
I expect the performance difference you're talking about is marginal anyway.
Maybe you've measured it?
--
Daniel
Daniel Phillips wrote:
>
> On Sunday 11 August 2002 09:40, Andrew Morton wrote:
> > zone->lock and zone->lru_lock are two of the hottest locks in the
> > kernel. Their usage patterns are quite independent. And they have
> > just been put into the same structure. It is essential that they not
> > fall into the same cacheline.
> >
> > That could be fixed by padding with L1_CACHE_BYTES. But the problem
> > with this is that a kernel which was configured for (say) a PIII will
> > perform poorly on SMP PIV. This will cause problems for kernel
> > vendors. For example, RH currently ship PII and Athlon binaries. To
> > get best SMP performance they will end up needing to ship a lot of
> > differently configured kernels.
> >
> > To solve this we need to know, at compile time, the maximum L1 size
> > which this kernel will ever run on.
> >
> > This patch adds L1_CACHE_SHIFT_MAX to every architecture's cache.h.
> >
> > Of course it'll break when newer chips come out with increased
> > cacheline sizes. Better suggestions are welcome.
>
> I think you're being too paranoid.
Staring at too many horrific profile outputs does that to one.
These are *the* two big locks.
> You pushed the performance degradation
> from the PIV to the PIII (because it will tend to hit more cachelines than it
> should)
The buddy info is all in one cacheline and the LRU info is in another.
So there's no loss to PIII here. But those two things are soooo hot
that paranoia is warranted.
> and you won't be able to build a kernel that is optimal for the PIII
> any more. I'd say that is PIII kernel is *supposed* to suck to some degree
> when run on a PIV, otherwise why bother having the PIV option?
>
> I expect the performance difference you're talking about is marginal anyway.
> Maybe you've measured it?
No, I haven't. NUMA boxes don't need it if the node-local allocation is
working right. But if they go cross-node much, it'll help. On high-performance
UMA SMP, allowing those two particular locks to fall into the same cacheline
is a big goofup.