TLB flush counters gone in 2.5.35-bk?

TLB flush counters gone in 2.5.35-bk?

[Roberto Nibali]

Hello,

I was just browsing over the latest bk tree when I saw the following change:

--- a/include/asm-generic/tlb.h Thu Aug 29 13:27:24 2002
+++ b/include/asm-generic/tlb.h Mon Sep  9 14:58:18 2002
@@ -21,7 +21,7 @@
   * and page free order so much..
   */
  #ifdef CONFIG_SMP
-  #define FREE_PTE_NR  507
+  #define FREE_PTE_NR  506
    #define tlb_fast_mode(tlb) ((tlb)->nr == ~0U)
  #else
    #define FREE_PTE_NR  1
@@ -40,8 +40,6 @@
         unsigned int            fullmm; /* non-zero means full mm flush */
         unsigned long           freed;
         struct page *           pages[FREE_PTE_NR];
-       unsigned long           flushes;/* stats: count avoided flushes */
-       unsigned long           avoided_flushes;
  } mmu_gather_t;

  /* Users of the generic TLB shootdown code must declare this storage 
space. */
@@ -67,17 +65,10 @@

  static inline void tlb_flush_mmu(mmu_gather_t *tlb, unsigned long 
start, unsigned long en
d)
  {
-       unsigned long nr;
-
-       if (!tlb->need_flush) {
-               tlb->avoided_flushes++;
+       if (!tlb->need_flush)
                 return;
-       }
         tlb->need_flush = 0;
-       tlb->flushes++;
-
         tlb_flush(tlb);
-       nr = tlb->nr;
         if (!tlb_fast_mode(tlb)) {
                 free_pages_and_swap_cache(tlb->pages, tlb->nr);
                 tlb->nr = 0;

Why was that done? I'm actually about to conduct some tests where I 
think that I need this information to check the L1 <-> L2 caching size 
influence on kernel data structures. What is the problem with the 
existing counters, did I miss some discussion on LKML?

Best regards,
Roberto Nibali, ratz
-- 
echo '[q]sa[ln0=aln256%Pln256/snlbx]sb3135071790101768542287578439snlbxq'|dc

Re: TLB flush counters gone in 2.5.35-bk?

[Andrew Morton]

Roberto Nibali wrote:
> 
> Hello,
> 
> I was just browsing over the latest bk tree when I saw the following change:
> 
> ...
> -       unsigned long           flushes;/* stats: count avoided flushes */
> -       unsigned long           avoided_flushes;

That was some statistical/debug code to evaluate how useful
that particular optimisation was being.  Answer: it saves
30-35% of the global TLB invalidations coming out of there.

But it had served its purpose.

Re: TLB flush counters gone in 2.5.35-bk?

[Andi Kleen]

Roberto Nibali <ratz@drugphish.ch> writes:

> Why was that done? I'm actually about to conduct some tests where I 
> think that I need this information to check the L1 <-> L2 caching size 
> influence on kernel data structures. What is the problem with the 
> existing counters, did I miss some discussion on LKML?

You can easily get the same information from the CPU performance counters
(e.g. via oprofile) 

-Andi

Re: TLB flush counters gone in 2.5.35-bk?

[Roberto Nibali]

> You can easily get the same information from the CPU performance counters
> (e.g. via oprofile) 

Thanks for the pointer Andi, I should have thought of oprofile before.

You wouldn't happen to know the equivalent counter event to a 
tlb_flush_mmu() for a PIII by any chance, would you? :). I've checked 
op_help and only found the ITLB_MISS. I look at the L2_* related cpu 
counters but can't find a TLB flush counter.

I'm reading through Appendix A of the IA-32 Architecture Vol 3 manual 
(it's actually very interesting), but I haven't found it either so far. 
Do I have to check for the INVLPG instructions?

Best regards,
Roberto Nibali, ratz
-- 
echo '[q]sa[ln0=aln256%Pln256/snlbx]sb3135071790101768542287578439snlbxq'|dc