Re: [PATCH bpf-next 2/6] bpf, net: rework cookie generator as per-cpu one

[Daniel Borkmann <daniel@iogearbox.net>]

On 9/25/20 9:49 AM, Eric Dumazet wrote:
> On 9/25/20 12:03 AM, Daniel Borkmann wrote:
>> On 9/24/20 8:58 PM, Eric Dumazet wrote:
>>> On 9/24/20 8:21 PM, Daniel Borkmann wrote:
>> [...]
>>>> diff --git a/include/linux/cookie.h b/include/linux/cookie.h
>>>> new file mode 100644
>>>> index 000000000000..2488203dc004
>>>> --- /dev/null
>>>> +++ b/include/linux/cookie.h
>>>> @@ -0,0 +1,41 @@
>>>> +/* SPDX-License-Identifier: GPL-2.0 */
>>>> +#ifndef __LINUX_COOKIE_H
>>>> +#define __LINUX_COOKIE_H
>>>> +
>>>> +#include <linux/atomic.h>
>>>> +#include <linux/percpu.h>
>>>> +
>>>> +struct gen_cookie {
>>>> +    u64 __percpu    *local_last;
>>>> +    atomic64_t     shared_last ____cacheline_aligned_in_smp;
>>>> +};
>>>> +
>>>> +#define COOKIE_LOCAL_BATCH    4096
>>>> +
>>>> +#define DEFINE_COOKIE(name)                    \
>>>> +    static DEFINE_PER_CPU(u64, __##name);            \
>>>> +    static struct gen_cookie name = {            \
>>>> +        .local_last    = &__##name,            \
>>>> +        .shared_last    = ATOMIC64_INIT(0),        \
>>>> +    }
>>>> +
>>>> +static inline u64 gen_cookie_next(struct gen_cookie *gc)
>>>> +{
>>>> +    u64 *local_last = &get_cpu_var(*gc->local_last);
>>>> +    u64 val = *local_last;
>>>> +
>>>> +    if (__is_defined(CONFIG_SMP) &&
>>>> +        unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {
>>>> +        s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
>>>> +                           &gc->shared_last);
>>>> +        val = next - COOKIE_LOCAL_BATCH;
>>>> +    }
>>>> +    val++;
>>>> +    if (unlikely(!val))
>>>> +        val++;
>>>> +    *local_last = val;
>>>> +    put_cpu_var(local_last);
>>>> +    return val;
>>>
>>> This is not interrupt safe.
>>>
>>> I think sock_gen_cookie() can be called from interrupt context.
>>>
>>> get_next_ino() is only called from process context, that is what I used get_cpu_var()
>>> and put_cpu_var()
>>
>> Hmm, agree, good point. Need to experiment a bit more .. initial thinking
>> potentially something like the below could do where we fall back to atomic
>> counter iff we encounter nesting (which should be an extremely rare case
>> normally).
>>
>> BPF progs where this can be called from are non-preemptible, so we could
>> actually move the temp preempt_disable/enable() from get/put_cpu_var() into
>> a wrapper func for slow path non-BPF users as well.
>>
>> static inline u64 gen_cookie_next(struct gen_cookie *gc)
>> {
>>          u64 val;
> 
> I presume you would use a single structure to hold level_nesting and local_last
> in the same cache line.
> 
> struct pcpu_gen_cookie {
>      int level_nesting;
>      u64 local_last;
> } __aligned(16);

Yes.

>>          if (likely(this_cpu_inc_return(*gc->level_nesting) == 1)) {
>>                  u64 *local_last = this_cpu_ptr(gc->local_last);
>>
>>                  val = *local_last;
>>                  if (__is_defined(CONFIG_SMP) &&
>>                      unlikely((val & (COOKIE_LOCAL_BATCH - 1)) == 0)) {
>>                          s64 next = atomic64_add_return(COOKIE_LOCAL_BATCH,
>>                                                         &gc->shared_last);
>>                          val = next - COOKIE_LOCAL_BATCH;
>>                  }
>>                  val++;
> 
>>                  if (unlikely(!val))
>>                          val++;
> 
> Note that we really expect this wrapping will never happen, with 64bit value.
> (We had to take care of the wrapping in get_next_ino() as it was dealing with 32bit values)

Agree, all local counters will start off at 0, but we inc right after the batch and
thus never run into it anyway and neither via overflow. Will remove.

>>                  *local_last = val;
>>          } else {
>>                  val = atomic64_add_return(COOKIE_LOCAL_BATCH,
>>                                            &gc->shared_last);
> 
> Or val = atomic64_dec_return(&reverse_counter)
> 
> With reverse_counter initial value set to ATOMIC64_INIT(0) ?
> 
> This will start sending 'big cookies like 0xFFFFFFFFxxxxxxxx' to make sure applications
> are not breaking with them, after few months of uptime.
> 
> This would also not consume COOKIE_LOCAL_BATCH units per value,
> but this seems minor based on the available space.

Excellent idea, I like it given it doesn't waste COOKIE_LOCAL_BATCH space. Thanks for
the feedback!

>>          }
>>          this_cpu_dec(*gc->level_nesting);
>>          return val;
>> }
>>
>> Thanks,
>> Daniel