On 10/09/18 17:46, Eric Dumazet wrote:

On 10/09/2018 03:48 PM, Florian Fainelli wrote:

quoted

This is based on Stephen's v4.14 patches, with the necessary merge
conflicts, and the lack of timer_setup() on the 4.9 baseline.

Perf results on a gigabit capable system, before and after are below.

Series can also be found here:

https://github.com/ffainelli/linux/commits/fragment-stack-v4.9


   PerfTop:     457 irqs/sec  kernel:74.4%  exact:  0.0% [4000Hz cycles],  (all, 4 CPUs)
-------------------------------------------------------------------------------

    29.62%  [kernel]       [k] ip_defrag                  
     6.57%  [kernel]       [k] arch_cpu_idle              
     1.72%  [kernel]       [k] v7_dma_inv_range           
     1.68%  [kernel]       [k] __netif_receive_skb_core   
     1.43%  [kernel]       [k] fib_table_lookup           
     1.30%  [kernel]       [k] finish_task_switch         
     1.08%  [kernel]       [k] ip_rcv                     
     1.01%  [kernel]       [k] skb_release_data           
     0.99%  [kernel]       [k] __slab_free                
     0.96%  [kernel]       [k] bcm_sysport_poll           
     0.88%  [kernel]       [k] __netdev_alloc_skb         
     0.87%  [kernel]       [k] tick_nohz_idle_enter       
     0.86%  [kernel]       [k] dev_gro_receive            
     0.85%  [kernel]       [k] _raw_spin_unlock_irqrestore
     0.84%  [kernel]       [k] __memzero                  
     0.74%  [kernel]       [k] tick_nohz_idle_exit        
     0.73%  ld-2.24.so     [.] do_lookup_x                
     0.66%  [kernel]       [k] kmem_cache_free            
     0.66%  [kernel]       [k] bcm_sysport_rx_refill      
     0.65%  [kernel]       [k] eth_type_trans             


After patching:

  PerfTop:     170 irqs/sec  kernel:86.5%  exact:  0.0% [4000Hz cycles],  (all, 4 CPUs)
-------------------------------------------------------------------------------

     7.79%  [kernel]       [k] arch_cpu_idle              
     5.14%  [kernel]       [k] v7_dma_inv_range           
     4.20%  [kernel]       [k] ip_defrag                  
     3.89%  [kernel]       [k] __netif_receive_skb_core   
     3.65%  [kernel]       [k] fib_table_lookup           
     2.16%  [kernel]       [k] finish_task_switch         
     1.93%  [kernel]       [k] _raw_spin_unlock_irqrestore
     1.90%  [kernel]       [k] ip_rcv                     
     1.84%  [kernel]       [k] bcm_sysport_poll           
     1.83%  [kernel]       [k] __memzero                  
     1.65%  [kernel]       [k] __netdev_alloc_skb         
     1.60%  [kernel]       [k] __slab_free                
     1.49%  [kernel]       [k] __do_softirq               
     1.49%  [kernel]       [k] bcm_sysport_rx_refill      
     1.31%  [kernel]       [k] dma_cache_maint_page       
     1.25%  [kernel]       [k] tick_nohz_idle_enter       
     1.24%  [kernel]       [k] ip_route_input_noref       
     1.17%  [kernel]       [k] eth_type_trans             
     1.06%  [kernel]       [k] fib_validate_source        
     1.03%  [kernel]       [k] inet_frag_find    

Dan Carpenter (1):
  ipv4: frags: precedence bug in ip_expire()

Eric Dumazet (22):
  inet: frags: change inet_frags_init_net() return value
  inet: frags: add a pointer to struct netns_frags
  inet: frags: refactor ipfrag_init()
  inet: frags: refactor ipv6_frag_init()
  inet: frags: refactor lowpan_net_frag_init()
  ipv6: export ip6 fragments sysctl to unprivileged users
  rhashtable: add schedule points
  inet: frags: use rhashtables for reassembly units
  inet: frags: remove some helpers
  inet: frags: get rif of inet_frag_evicting()
  inet: frags: remove inet_frag_maybe_warn_overflow()
  inet: frags: break the 2GB limit for frags storage
  inet: frags: do not clone skb in ip_expire()
  ipv6: frags: rewrite ip6_expire_frag_queue()
  rhashtable: reorganize struct rhashtable layout
  inet: frags: reorganize struct netns_frags
  inet: frags: get rid of ipfrag_skb_cb/FRAG_CB
  inet: frags: fix ip6frag_low_thresh boundary
  net: speed up skb_rbtree_purge()
  net: pskb_trim_rcsum() and CHECKSUM_COMPLETE are friends
  net: add rb_to_skb() and other rb tree helpers
  net: sk_buff rbnode reorg

Florian Westphal (1):
  ipv6: defrag: drop non-last frags smaller than min mtu

Peter Oskolkov (4):
  ip: discard IPv4 datagrams with overlapping segments.
  net: modify skb_rbtree_purge to return the truesize of all purged
    skbs.
  ip: add helpers to process in-order fragments faster.
  ip: process in-order fragments efficiently

Taehee Yoo (1):
  ip: frags: fix crash in ip_do_fragment()

Strange, I do not see "ip: use rb trees for IP frag queue." in this list ?

And it was not in Stephen's backport to 4.14 either, wait, looks like it
was somehow squashed into "net: sk_buff rbnode reorg". Stephen, was
there a reason for that?

Let me go back and add bffa72cf7f9df842f0016ba03586039296b4caaf as well
as eeea10b83a139451130df1594f26710c8fa390c8 to the rebase todo and see
how things go from there.

Thanks for taking a look.
-- 
Florian