summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorAndi Kleen <andi@firstfloor.org>2008-03-12 03:53:32 +0100
committerIngo Molnar <mingo@elte.hu>2008-04-17 17:41:30 +0200
commit8346ea17aa20e9864b0f7dc03d55f3cd5620b8c1 (patch)
tree23a0e515b5aa3c5611f2b821ecc3a2b08c463858
parent1de87bd40e119d26533b5135677901990390bfa9 (diff)
downloadop-kernel-dev-8346ea17aa20e9864b0f7dc03d55f3cd5620b8c1.zip
op-kernel-dev-8346ea17aa20e9864b0f7dc03d55f3cd5620b8c1.tar.gz
x86: split large page mapping for AMD TSEG
On AMD SMM protected memory is part of the address map, but handled internally like an MTRR. That leads to large pages getting split internally which has some performance implications. Check for the AMD TSEG MSR and split the large page mapping on that area explicitely if it is part of the direct mapping. There is also SMM ASEG, but it is in the first 1MB and already covered by the earlier split first page patch. Idea for this came from an earlier patch by Andreas Herrmann On a RevF dual Socket Opteron system kernbench shows a clear improvement from this: (together with the earlier patches in this series, especially the split first 2MB patch) [lower is better] no split stddev split stddev delta Elapsed Time 87.146 (0.727516) 84.296 (1.09098) -3.2% User Time 274.537 (4.05226) 273.692 (3.34344) -0.3% System Time 34.907 (0.42492) 34.508 (0.26832) -1.1% Percent CPU 322.5 (38.3007) 326.5 (44.5128) +1.2% => About 3.2% improvement in elapsed time for kernbench. With GB pages on AMD Fam1h the impact of splitting is much higher of course, since it would split two full GB pages (together with the first 1MB split patch) instead of two 2MB pages. I could not benchmark a clear difference in kernbench on gbpages, so I kept it disabled for that case That was only limited benchmarking of course, so if someone was interested in running more tests for the gbpages case that could be revisited (contributions welcome) I didn't bother implementing this for 32bit because it is very unlikely the 32bit lowmem mapping overlaps into the TSEG near 4GB and the 2MB low split is already handled for both. [ mingo@elte.hu: do it on gbpages kernels too, there's no clear reason why it shouldnt help there. ] Signed-off-by: Andi Kleen <ak@suse.de> Acked-by: andreas.herrmann3@amd.com Cc: mingo@elte.hu Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Ingo Molnar <mingo@elte.hu>
-rw-r--r--arch/x86/kernel/setup_64.c13
-rw-r--r--include/asm-x86/msr-index.h1
2 files changed, 14 insertions, 0 deletions
diff --git a/arch/x86/kernel/setup_64.c b/arch/x86/kernel/setup_64.c
index 3d76dbd..b542597 100644
--- a/arch/x86/kernel/setup_64.c
+++ b/arch/x86/kernel/setup_64.c
@@ -729,6 +729,19 @@ static void __cpuinit init_amd(struct cpuinfo_x86 *c)
if (amd_apic_timer_broken())
disable_apic_timer = 1;
+
+ if (c == &boot_cpu_data && c->x86 >= 0xf && c->x86 <= 0x11) {
+ unsigned long long tseg;
+
+ /*
+ * Split up direct mapping around the TSEG SMM area.
+ * Don't do it for gbpages because there seems very little
+ * benefit in doing so.
+ */
+ if (!rdmsrl_safe(MSR_K8_TSEG_ADDR, &tseg) &&
+ (tseg >> PMD_SHIFT) < (max_pfn_mapped >> (PMD_SHIFT-PAGE_SHIFT)))
+ set_memory_4k((unsigned long)__va(tseg), 1);
+ }
}
void __cpuinit detect_ht(struct cpuinfo_x86 *c)
diff --git a/include/asm-x86/msr-index.h b/include/asm-x86/msr-index.h
index af4e07f..09413ad 100644
--- a/include/asm-x86/msr-index.h
+++ b/include/asm-x86/msr-index.h
@@ -112,6 +112,7 @@
#define MSR_K8_SYSCFG 0xc0010010
#define MSR_K8_HWCR 0xc0010015
#define MSR_K8_ENABLE_C1E 0xc0010055
+#define MSR_K8_TSEG_ADDR 0xc0010112
#define K8_MTRRFIXRANGE_DRAM_ENABLE 0x00040000 /* MtrrFixDramEn bit */
#define K8_MTRRFIXRANGE_DRAM_MODIFY 0x00080000 /* MtrrFixDramModEn bit */
#define K8_MTRR_RDMEM_WRMEM_MASK 0x18181818 /* Mask: RdMem|WrMem */
OpenPOWER on IntegriCloud