betmap-libpcap crashes with jumbo packets set in "/sys/modules/netmap_lin/parameters/buff_size" about netmap HOT 7 CLOSED

Hi morphyno,
we fixed the problem with jumbo frames and ixgbe, but without libpcap.
This fixes are into "next" branch of netmap. Maybe this problem is related.

Can you try with "next" branch?

Thanks,
Stefano

Hi Stefano:

Jumbo packets does seem to work with the "next" branch of netmap. However, 
here's my observation.

I set "buf_size" to 9000, in dmesg, I see bufsz is set to 9088

[ 2834.998629] 050.285168 [ 460] ixgbe_netmap_configure_srrctl bufsz: 9088 
srrctl: 8

However, I see actual MTU is 8196 Bytes.

Because as soon as I send packets bigger than 8196 bytes, I see fragmentation.
892 bytes is a pretty big difference.

Is buf_size a software buffer size that includes other objects other than skb?

Hi morphyno,
bufsz is set to 9088 for alignment reason.

For the MTU, I don't understand if you are set it up to 9000 in both side (tx 
and rx).
Can you explain the test environment? (tools for traffic generation, NIC set 
up, tx and rx NIC interconnection)

Cheers,
Stefano

Hi Stefano:

I actually avoided using jumbo packets for some time. I'm using tcpreplay over 
netmap to play pcaps that contain jumbo packets (largest is 9200 bytes). I'm 
able to do so when I set /sys/module/netmap/parameters/buf_size to 9300. I also 
use tcpdump over netmap-libpcap to capture the packets in a loopback (from my 
device under test).

tcpeplay -> Netmap:ethX -> DUT > Netmap:eth7 -> tcpdump

I'm noticing something very interesting. When I attempt to invoke but tcpreplay 
and tcpdump both in netmap mode(they both use difference interfaces). I see the 
following errors in dmesg. 

[2549737.221434] 057.187385 [ 798] generic_netmap_dtor       Restored native NA 
          (null)
[2549738.286908] 058.252026 [1786] netmap_interp_ringid      eth0: tx [0,12) rx 
[0,12) id 0
[2549738.350512] 058.315582 [ 356] nm_mem_assign_group       iommu_group 0
[2549738.416472] 058.381491 [ 507] netmap_obj_malloc         no more netmap_buf 
objects
[2549738.482315] 058.447283 [ 672] netmap_new_bufs           no more buffers 
after 444 of 512
[2549738.548963] 058.513879 [1360] netmap_mem_rings_create   Cannot allocate 
buffers for tx_ring
[2549738.617586] 058.582448 [1516] netmap_mem_global_deref   refcount = 1

This only happens if I have a buffer size set to a very large number. Is there 
any other parameter I need to adjust accordingly with buff_size?

Attached is the dmesg, tcpdump over netmap-libpcap is using eth13, tcpreplay 
over netmap is using eth12