From: David Miller <davem@xxxxxxxxxxxxx>
Date: Fri, 14 Jul 2017 11:17:36 -0700 (PDT)
> I'm still pretty sure that unmaps are taking an unreasonable amount of
> time to execute. Our current range flush implementation is incredibly
> stupid, and could be improved by orders of magnitude. It allocates an
> entire kernel stack frame, just so that it can call __flush_tlb_pending().
>
> In fact, we can end up doing this full trap entry/exit just for
> purging 2 or 3 pages.
So I wanted to dig deeper into this, so I wrote a simple tracker for
flush_tlb_pending() which keeps of how often we do flushes of various
sizes.
The patch is below, the output from /proc/cpuinfo looks like:
TLB PENDING : Total 2441449 Chart [1365735 435807 105896 134569 134121 89929 55970 123867 0 0 0 0 0 0 0]
First is the total number of flushes, then there is a histogram of counts by power
of 2. So the first slot is number of 1 entry flushes, the second slot is the
number of 2 or 3 entry flushes, the third is the number of 4, 5, 6, or 7
entry flushes. And so on.
I did three kernel builds after a fresh boot and ended up with this (annotated):
Total 2437888
1: 1363857
2: 435389
4: 105582
8: 134329
16: 133978
32: 89821
64: 55833
128: 123544
256: 0
512: 0
1024: 0
2048: 0
4096: 0
8192: 0
16384: 0
Can you run with this patch on the test case that triggered these mondo
timeouts and send the /proc/cpuinfo "TLB PENDING" line to this list?
Thanks!
diff --git a/arch/sparc/include/asm/tlb_64.h b/arch/sparc/include/asm/tlb_64.h
index 4cb392f..7a9473c 100644
--- a/arch/sparc/include/asm/tlb_64.h
+++ b/arch/sparc/include/asm/tlb_64.h
@@ -7,6 +7,13 @@
#include <asm/tlbflush.h>
#include <asm/mmu_context.h>
+#define TLB_PENDING_STATS_SLOTS 15
+struct tlb_pending_stats {
+ unsigned long total;
+ unsigned long num_flushes[TLB_PENDING_STATS_SLOTS];
+};
+extern struct tlb_pending_stats tpstats;
+
#ifdef CONFIG_SMP
void smp_flush_tlb_pending(struct mm_struct *,
unsigned long, unsigned long *);
diff --git a/arch/sparc/kernel/cpu.c b/arch/sparc/kernel/cpu.c
index 493e023..706ffbf 100644
--- a/arch/sparc/kernel/cpu.c
+++ b/arch/sparc/kernel/cpu.c
@@ -20,6 +20,7 @@
#include <asm/psr.h>
#include <asm/mbus.h>
#include <asm/cpudata.h>
+#include <asm/tlb.h>
#include "kernel.h"
#include "entry.h"
@@ -378,6 +379,7 @@ static int show_cpuinfo(struct seq_file *m, void *__unused)
"ncpus active\t: %d\n"
"D$ parity tl1\t: %u\n"
"I$ parity tl1\t: %u\n"
+ "TLB PENDING\t: Total %lu Chart [%lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu]\n"
#ifndef CONFIG_SMP
"Cpu0ClkTck\t: %016lx\n"
#endif
@@ -392,7 +394,12 @@ static int show_cpuinfo(struct seq_file *m, void *__unused)
ncpus_probed,
num_online_cpus(),
dcache_parity_tl1_occurred,
- icache_parity_tl1_occurred
+ icache_parity_tl1_occurred,
+ tpstats.total,
+ tpstats.num_flushes[0], tpstats.num_flushes[1], tpstats.num_flushes[2], tpstats.num_flushes[3],
+ tpstats.num_flushes[4], tpstats.num_flushes[5], tpstats.num_flushes[6], tpstats.num_flushes[7],
+ tpstats.num_flushes[8], tpstats.num_flushes[9], tpstats.num_flushes[10], tpstats.num_flushes[11],
+ tpstats.num_flushes[12], tpstats.num_flushes[13], tpstats.num_flushes[14]
#ifndef CONFIG_SMP
, cpu_data(0).clock_tick
#endif
diff --git a/arch/sparc/mm/tlb.c b/arch/sparc/mm/tlb.c
index ee8066c..8a8a2e2 100644
--- a/arch/sparc/mm/tlb.c
+++ b/arch/sparc/mm/tlb.c
@@ -20,6 +20,8 @@
static DEFINE_PER_CPU(struct tlb_batch, tlb_batch);
+struct tlb_pending_stats tpstats;
+
void flush_tlb_pending(void)
{
struct tlb_batch *tb = &get_cpu_var(tlb_batch);
@@ -31,6 +33,14 @@ void flush_tlb_pending(void)
flush_tsb_user(tb);
if (CTX_VALID(mm->context)) {
+ u32 slot = tb->tlb_nr;
+
+ tpstats.total++;
+ slot = ilog2(slot);
+ if (slot > 14)
+ slot = 14;
+ tpstats.num_flushes[slot]++;
+
if (tb->tlb_nr == 1) {
global_flush_tlb_page(mm, tb->vaddrs[0]);
} else {
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
