On Fri, 2015-05-29 at 12:05 +0200, Florian Westphal wrote:
> Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> > > How? What address? You mean relative offset to counter start?
> >
> > No, the address itself. This is what I coded years ago.
> > > when rule x matches, I need to increment the corresponding counter
> > > for that rule.
> >
> > Right. You have in the rule itself the storage for xt_counter. Thats 16
> > bytes. You only need 8 bytes (on 64bit arch) to store the percpu
> > address.
>
> > Instead of leaving nothing in it, place the percpu address of the
> > corresponding counter. access to it is in the cache line needed to
> > analyze the entry anyway. Its free.
>
> I'm dense. So, what you're saying is that I should
>
> alloc_percpu(sizeof(struct xt_counters));
>
> for each rule and store the resulting address in ipt_entry?
> Then it would be possible to get the correct location for current cpu
> when counter has to be incremented.
>
> Thanks Eric.
That was my plan.
When looking at perf record/report, ipt_do_table spends a lot of time in
setup phase to fetch all these indirections :
IP_NF_ASSERT(table->valid_hooks & (1 << hook));
local_bh_disable();
addend = xt_write_recseq_begin();
private = table->private;
cpu = smp_processor_id();
/*
* Ensure we load private-> members after we've fetched the base
* pointer.
*/
smp_read_barrier_depends();
table_base = private->entries[cpu];
jumpstack = (struct ipt_entry **)private->jumpstack[cpu];
stackptr = per_cpu_ptr(private->stackptr, cpu);
origptr = *stackptr;
By storing the pointer to the percpu storage inside the ipt_entry,
you do not need an extra dereference on x86_64.
- ADD_COUNTER(e->counters, skb->len, 1);
+ ADD_COUNTER(counters[e->comefrom], skb->len, 1);
Would be :
ADD_COUNTER(e->pcu_addr, skb->len, 1);
If you alias/union pcu_addr and counters
You might need temporary arrays at load/dump times, but not in fast path.
x86_64 would then use
load e->pcpu_addr into register R (assume it is %rax)
load skb->len into %rdx
incq %gs:(%rax) // pcnt++
addq %rdx,%gs:8(%rax) // bcnt += skb->len
--
To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
