In testing the "Fix non-access data TLB cache flush faults" change,
I noticed a significant improvement in glibc build and check times.
This led me to investigate the parisc_cache_flush_threshold setting.
It determines when we switch from line flushing to whole cache flushing.
It turned out that the parisc_cache_flush_threshold setting on
mako and mako2 machines (PA8800 and PA8900 processors) was way too
small. Adjusting this setting provided almost a factor two improvement
in the glibc build and check time.
Signed-off-by: John David Anglin <dave.anglin@xxxxxxxx>
---
diff --git a/arch/parisc/kernel/cache.c b/arch/parisc/kernel/cache.c
index 94150b91c96f..8368fefdd217 100644
--- a/arch/parisc/kernel/cache.c
+++ b/arch/parisc/kernel/cache.c
@@ -388,7 +388,7 @@ void __init parisc_setup_cache_timing(void)
{
unsigned long rangetime, alltime;
unsigned long size;
- unsigned long threshold;
+ unsigned long threshold, threshold2;
alltime = mfctl(16);
flush_data_cache();
@@ -403,8 +403,20 @@ void __init parisc_setup_cache_timing(void)
alltime, size, rangetime);
threshold = L1_CACHE_ALIGN(size * alltime / rangetime);
- if (threshold > cache_info.dc_size)
- threshold = cache_info.dc_size;
+
+ /*
+ * The threshold computed above isn't very reliable since the
+ * flush times depend greatly on the percentage of dirty lines
+ * in the flush range. Further, the whole cache time doesn't
+ * include the time to refill lines that aren't in the mm/vma
+ * being flushed. By timing glibc build and checks on mako cpus,
+ * the following formula seems to work reasonably well. The
+ * value from the timing calculation is too small, and increases
+ * build and check times by almost a factor two.
+ */
+ threshold2 = cache_info.dc_size * num_online_cpus();
+ if (threshold2 > threshold)
+ threshold = threshold2;
if (threshold)
parisc_cache_flush_threshold = threshold;
printk(KERN_INFO "Cache flush threshold set to %lu KiB\n",
Attachment:
signature.asc
Description: PGP signature
