Skip to content

L
linux
Switch branch/tag
  1. 06 Jan, 2015 1 commit
  3. 03 Jan, 2015 10 commits
  5. 02 Jan, 2015 14 commits
  7. 01 Jan, 2015 7 commits
  9. 31 Dec, 2014 8 commits
    • David S. Miller's avatar
      igb_ptp: Include clocksource.h to get CLOCKSOURCE_MASK. · 9aacfb20
      David S. Miller authored 
      Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      9aacfb20
    • David S. Miller's avatar
      e1000e: Include clocksource.h to get CLOCKSOURCE_MASK. · 54da5083
      David S. Miller authored 
      Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      54da5083
    • David S. Miller's avatar
      Merge branch 'fib_trie-next' · e495f78d
      David S. Miller authored 
      Alexander Duyck says:

====================
fib_trie: Reduce time spent in fib_table_lookup by 35 to 75%

These patches are meant to address several performance issues I have seen
in the fib_trie implementation, and fib_table_lookup specifically.  With
these changes in place I have seen a reduction of up to 35 to 75% for the
total time spent in fib_table_lookup depending on the type of search being
performed.

On a VM running in my Corei7-4930K system with a trie of maximum depth of 7
this resulted in a reduction of over 370ns per packet in the total time to
process packets received from an ixgbe interface and route them to a dummy
interface.  This represents a failed lookup in the local trie followed by
a successful search in the main trie.

				Baseline	Refactor
  ixgbe->dummy routing		1.20Mpps	2.21Mpps
  ------------------------------------------------------------
  processing time per packet		835ns		453ns
  fib_table_lookup		50.1%	418ns	25.0%	113ns
  check_leaf.isra.9		 7.9%	 66ns	   --	 --
  ixgbe_clean_rx_irq		 5.3%	 44ns	 9.8%	 44ns
  ip_route_input_noref		 2.9%	 25ns	 4.6%	 21ns
  pvclock_clocksource_read	 2.6%	 21ns	 4.6%	 21ns
  ip_rcv			 2.6%	 22ns	 4.0%	 18ns

In the simple case of receiving a frame and dropping it before it can reach
the socket layer I saw a reduction of 40ns per packet.  This represents a
trip through the local trie with the correct leaf found with no need for
any backtracing.

				Baseline	Refactor
  ixgbe->local receive		2.65Mpps	2.96Mpps
  ------------------------------------------------------------
  processing time per packet		377ns		337ns
  fib_table_lookup		25.1%	 95ns	25.8%	 87ns
  ixgbe_clean_rx_irq		 8.7%	 33ns	 9.0%	 30ns
  check_leaf.isra.9		 7.2%	 27ns	   --	 --
  ip_rcv			 5.7%	 21ns	 6.5%	 22ns

These changes have resulted in several functions being inlined such as
check_leaf and fib_find_node, but due to the code simplification the
overall size of the code has been reduced.

   text	   data	    bss	    dec	    hex	filename
  16932	    376	     16	  17324	   43ac	net/ipv4/fib_trie.o - before
  15259	    376	      8	  15643	   3d1b	net/ipv4/fib_trie.o - after

Changes since RFC:
  Replaced this_cpu_ptr with correct call to this_cpu_inc in patch 1
  Changed test for leaf_info mismatch to (key ^ n->key) & li->mask_plen in patch 10
====================
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      e495f78d
    • Alexander Duyck's avatar
      fib_trie: Add tracking value for suffix length · 5405afd1
      Alexander Duyck authored 
      This change adds a tracking value for the maximum suffix length of all
prefixes stored in any given tnode.  With this value we can determine if we
need to backtrace or not based on if the suffix is greater than the pos
value.

By doing this we can reduce the CPU overhead for lookups in the local table
as many of the prefixes there are 32b long and have a suffix length of 0
meaning we can immediately backtrace to the root node without needing to
test any of the nodes between it and where we ended up.
Signed-off-by: default avatarAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      5405afd1
    • Alexander Duyck's avatar
      fib_trie: Remove checks for index >= tnode_child_length from tnode_get_child · 21d1f11d
      Alexander Duyck authored 
      For some reason the compiler doesn't seem to understand that when we are in
a loop that runs from tnode_child_length - 1 to 0 we don't expect the value
of tn->bits to change.  As such every call to tnode_get_child was rerunning
tnode_chile_length which ended up consuming quite a bit of space in the
resultant assembly code.

I have gone though and verified that in all cases where tnode_get_child
is used we are either winding though a fixed loop from tnode_child_length -
1 to 0, or are in a fastpath case where we are verifying the value by
either checking for any remaining bits after shifting index by bits and
testing for leaf, or by using tnode_child_length.

size net/ipv4/fib_trie.o
Before:
   text	   data	    bss	    dec	    hex	filename
  15506	    376	      8	  15890	   3e12	net/ipv4/fib_trie.o

After:
   text	   data	    bss	    dec	    hex	filename
  14827	    376	      8	  15211	   3b6b	net/ipv4/fib_trie.o
Signed-off-by: default avatarAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      21d1f11d
    • Alexander Duyck's avatar
      fib_trie: inflate/halve nodes in a more RCU friendly way · 12c081a5
      Alexander Duyck authored 
      This change pulls the node_set_parent functionality out of put_child_reorg
and instead leaves that to the function to take care of as well.  By doing
this we can fully construct the new cluster of tnodes and all of the
pointers out of it before we start routing pointers into it.

I am suspecting this will likely fix some concurency issues though I don't
have a good test to show as such.
Signed-off-by: default avatarAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      12c081a5
    • Alexander Duyck's avatar
      fib_trie: Push tnode flushing down to inflate/halve · fc86a93b
      Alexander Duyck authored 
      This change pushes the tnode freeing down into the inflate and halve
functions.  It makes more sense here as we have a better grasp of what is
going on and when a given cluster of nodes is ready to be freed.

I believe this may address a bug in the freeing logic as well.  For some
reason if the freelist got to a certain size we would call
synchronize_rcu().  I'm assuming that what they meant to do is call
synchronize_rcu() after they had handed off that much memory via
call_rcu().  As such that is what I have updated the behavior to be.
Signed-off-by: default avatarAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      fc86a93b
    • Alexander Duyck's avatar
      fib_trie: Push assignment of child to parent down into inflate/halve · ff181ed8
      Alexander Duyck authored 
      This change makes it so that the assignment of the tnode to the parent is
handled directly within whatever function is currently handling the node be
it inflate, halve, or resize.  By doing this we can avoid some of the need
to set NULL pointers in the tree while we are resizing the subnodes.
Signed-off-by: default avatarAlexander Duyck <alexander.h.duyck@redhat.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      ff181ed8
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7