This new set type allows for intervals in concatenated fields,
which are expressed in the usual way, that is, simple byte
concatenation with padding to 32 bits for single fields, and
given as ranges by specifying start and end elements containing,
each, the full concatenation of start and end values for the
single fields.
Ranges are expanded to composing netmasks, for each field: these
are inserted as rules in per-field lookup tables. Bits to be
classified are divided in 4-bit groups, and for each group, the
lookup table contains 4^2 buckets, representing all the possible
values of a bit group. This approach was inspired by the Grouper
algorithm:
http://www.cse.usf.edu/~ligatti/projects/grouper/
Matching is performed by a sequence of AND operations between
bucket values, with buckets selected according to the value of
packet bits, for each group. The result of this sequence tells
us which rules matched for a given field.
In order to concatenate several ranged fields, per-field rules
are mapped using mapping arrays, one per field, that specify
which rules should be considered while matching the next field.
The mapping array for the last field contains a reference to
the element originally inserted.
The notes in nft_set_pipapo.c cover the algorithm in deeper
detail.
A pure hash-based approach is of no use here, as ranges need
to be classified. An implementation based on "proxying" the
existing red-black tree set type, creating a tree for each
field, was considered, but deemed impractical due to the fact
that elements would need to be shared between trees, at least
as long as we want to keep UAPI changes to a minimum.
A stand-alone implementation of this algorithm is available at:
https://pipapo.lameexcu.se
together with notes about possible future optimisations
(in pipapo.c).
This algorithm was designed with data locality in mind, and can
be highly optimised for SIMD instruction sets, as the bulk of
the matching work is done with repetitive, simple bitwise
operations.
At this point, without further optimisations, nft_concat_range.sh
reports, for one AMD Epyc 7351 thread (2.9GHz, 512 KiB L1D$, 8 MiB
L2$):
TEST: performance
net,port [ OK ]
baseline (drop from netdev hook): 10190076pps
baseline hash (non-ranged entries): 6179564pps
baseline rbtree (match on first field only): 2950341pps
set with 1000 full, ranged entries: 2304165pps
port,net [ OK ]
baseline (drop from netdev hook): 10143615pps
baseline hash (non-ranged entries): 6135776pps
baseline rbtree (match on first field only): 4311934pps
set with 100 full, ranged entries: 4131471pps
net6,port [ OK ]
baseline (drop from netdev hook): 9730404pps
baseline hash (non-ranged entries): 4809557pps
baseline rbtree (match on first field only): 1501699pps
set with 1000 full, ranged entries: 1092557pps
port,proto [ OK ]
baseline (drop from netdev hook): 10812426pps
baseline hash (non-ranged entries): 6929353pps
baseline rbtree (match on first field only): 3027105pps
set with 30000 full, ranged entries: 284147pps
net6,port,mac [ OK ]
baseline (drop from netdev hook): 9660114pps
baseline hash (non-ranged entries): 3778877pps
baseline rbtree (match on first field only): 3179379pps
set with 10 full, ranged entries: 2082880pps
net6,port,mac,proto [ OK ]
baseline (drop from netdev hook): 9718324pps
baseline hash (non-ranged entries): 3799021pps
baseline rbtree (match on first field only): 1506689pps
set with 1000 full, ranged entries: 783810pps
net,mac [ OK ]
baseline (drop from netdev hook): 10190029pps
baseline hash (non-ranged entries): 5172218pps
baseline rbtree (match on first field only): 2946863pps
set with 1000 full, ranged entries: 1279122pps
v4:
- fix build for 32-bit architectures: 64-bit division needs
div_u64() (kbuild test robot <lkp@intel.com>)
v3:
- rework interface for field length specification,
NFT_SET_SUBKEY disappears and information is stored in
description
- remove scratch area to store closing element of ranges,
as elements now come with an actual attribute to specify
the upper range limit (Pablo Neira Ayuso)
- also remove pointer to 'start' element from mapping table,
closing key is now accessible via extension data
- use bytes right away instead of bits for field lengths,
this way we can also double the inner loop of the lookup
function to take care of upper and lower bits in a single
iteration (minor performance improvement)
- make it clearer that set operations are actually atomic
API-wise, but we can't e.g. implement flush() as one-shot
action
- fix type for 'dup' in nft_pipapo_insert(), check for
duplicates only in the next generation, and in general take
care of differentiating generation mask cases depending on
the operation (Pablo Neira Ayuso)
- report C implementation matching rate in commit message, so
that AVX2 implementation can be compared (Pablo Neira Ayuso)
v2:
- protect access to scratch maps in nft_pipapo_lookup() with
local_bh_disable/enable() (Florian Westphal)
- drop rcu_read_lock/unlock() from nft_pipapo_lookup(), it's
already implied (Florian Westphal)
- explain why partial allocation failures don't need handling
in pipapo_realloc_scratch(), rename 'm' to clone and update
related kerneldoc to make it clear we're not operating on
the live copy (Florian Westphal)
- add expicit check for priv->start_elem in
nft_pipapo_insert() to avoid ending up in nft_pipapo_walk()
with a NULL start element, and also zero it out in every
operation that might make it invalid, so that insertion
doesn't proceed with an invalid element (Florian Westphal)
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
3c4287f620