> 
> 
> On 2/4/22 11:08 AM, Lorenzo Bianconi wrote:
> > > > 
> > > 
> > > [...]
> > > 
> > > > > > 
> > > > > > In kernel, the nr_frags checking is against MAX_SKB_FRAGS,
> > > > > > but if /proc/sys/net/core/max_skb_flags is 2 or more less
> > > > > > than MAX_SKB_FRAGS, the test won't fail, right?
> > > > > 
> > > > > yes, you are right. Should we use the same definition used in
> > > > > include/linux/skbuff.h instead? Something like:
> > > > > 
> > > > > if (65536 / page_size + 1 < 16)
> > > > > 	max_skb_flags = 16;
> > > > > else
> > > > > 	max_skb_flags = 65536/page_size + 1;
> > > > 
> > > > The maximum packet size limit 64KB won't change anytime soon.
> > > > So the above should work. Some comments to explain why using
> > > > the above formula will be good.
> > > 
> > > ack, I will do in v2.
> > 
> > I can see there is a on-going discussion here [0] about increasing
> > MAX_SKB_FRAGS. I guess we can put on-hold this patch and see how
> > MAX_SKB_FRAGS will be changed.
> 
> Thanks for the link. The new patch is going to increase
> MAX_SKB_FRAGS and it is possible that will be changed again
> (maybe under some config options).
> 
> The default value for
> /proc/sys/net/core/max_skb_flags is MAX_SKB_FRAGS and I suspect
> anybody is bothering to change it. So your patch is okay to me.
> Maybe change a little bit -ENOMEM error message. current,
>   ASSERT_EQ(err, -ENOMEM, "unsupported buffer size");
> to
>   ASSERT_EQ(err, -ENOMEM, "unsupported buffer size, possible non-default
> /proc/sys/net/core/max_skb_flags?");

ack, I am fine with it.
@Alexei, Andrii: any hints about it?

Regards,
Lorenzo

> 
> > 
> > Regards,
> > Lorenzo
> > 
> > [0] https://lore.kernel.org/all/202202031315.B425Ipe8-lkp@intel.com/t/#ma1b2c7e71fe9bc69e24642a62dadf32fda7d5f03
> > 
> > > 
> > > Regards,
> > > Lorenzo
> > > 
> > > > 
> > > > > 
> > > > > Regards,
> > > > > Lorenzo
> > > > > 
> > > > > > 
> > > > > > > +
> > > > > > > +	num = fscanf(f, "%d", &max_skb_frags);
> > > > > > > +	fclose(f);
> > > > > > > +
> > > > > > > +	if (!ASSERT_EQ(num, 1, "max_skb_frags read failed"))
> > > > > > > +		goto out;
> > > > > > > +
> > > > > > > +	/* xdp_buff linear area size is always set to 4096 in the
> > > > > > > +	 * bpf_prog_test_run_xdp routine.
> > > > > > > +	 */
> > > > > > > +	buf_size = 4096 + (max_skb_frags + 1) * sysconf(_SC_PAGE_SIZE);
> > > > > > > +	buf = malloc(buf_size);
> > > > > > > +	if (!ASSERT_OK_PTR(buf, "alloc buf"))
> > > > > > > +		goto out;
> > > > > > > +
> > > > > > > +	memset(buf, 0, buf_size);
> > > > > > > +	offset = (__u32 *)buf;
> > > > > > > +	*offset = 16;
> > > > > > > +	buf[*offset] = 0xaa;
> > > > > > > +	buf[*offset + 15] = 0xaa;
> > > > > > > +
> > > > > > > +	topts.data_in = buf;
> > > > > > > +	topts.data_out = buf;
> > > > > > > +	topts.data_size_in = buf_size;
> > > > > > > +	topts.data_size_out = buf_size;
> > > > > > > +
> > > > > > > +	err = bpf_prog_test_run_opts(prog_fd, &topts);
> > > > > > > +	ASSERT_EQ(err, -ENOMEM, "unsupported buffer size");
> > > > > > > +	free(buf);
> > > > > > >     out:
> > > > > > >     	bpf_object__close(obj);
> > > > > > >     }
> > > > > > 
> > > > 
> > 
> >