Now that the kernel specifies binutils 2.23 as the minimum version, we can remove ifdefs for AVX2 and ADX throughout. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> Acked-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Doing this probing inside of the Makefiles means we have a maze of ifdefs inside the source code and child Makefiles that need to make proper decisions on this too. Instead, we do it at Kconfig time, like many other compiler and assembler options, which allows us to set up the dependencies normally for full compilation units. In the process, the ADX test changes to use %eax instead of %r10 so that it's valid in both 32-bit and 64-bit mode. Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> Acked-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Nick Desaulniers <ndesaulniers@google.com> Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
If the AVX2 set is available, we can exploit the repetitive
characteristic of this algorithm to provide a fast, vectorised
version by using 256-bit wide AVX2 operations for bucket loads and
bitwise intersections.
In most cases, this implementation consistently outperforms rbtree
set instances despite the fact they are configured to use a given,
single, ranged data type out of the ones used for performance
measurements by the nft_concat_range.sh kselftest.
That script, injecting packets directly on the ingoing device path
with pktgen, reports, averaged over five runs on a single AMD Epyc
7402 thread (3.35GHz, 768 KiB L1D$, 12 MiB L2$), the figures below.
CONFIG_RETPOLINE was not set here.
Note that this is not a fair comparison over hash and rbtree set
types: non-ranged entries (used to have a reference for hash types)
would be matched faster than this, and matching on a single field
only (which is the case for rbtree) is also significantly faster.
However, it's not possible at the moment to choose this set type
for non-ranged entries, and the current implementation also needs
a few minor adjustments in order to match on less than two fields.
---------------.-----------------------------------.------------.
AMD Epyc 7402 | baselines, Mpps | this patch |
1 thread |___________________________________|____________|
3.35GHz | | | | | |
768KiB L1D$ | netdev | hash | rbtree | | |
---------------| hook | no | single | | pipapo |
type entries | drop | ranges | field | pipapo | AVX2 |
---------------|--------|--------|--------|--------|------------|
net,port | | | | | |
1000 | 19.0 | 10.4 | 3.8 | 4.0 | 7.5 +87% |
---------------|--------|--------|--------|--------|------------|
port,net | | | | | |
100 | 18.8 | 10.3 | 5.8 | 6.3 | 8.1 +29% |
---------------|--------|--------|--------|--------|------------|
net6,port | | | | | |
1000 | 16.4 | 7.6 | 1.8 | 2.1 | 4.8 +128% |
---------------|--------|--------|--------|--------|------------|
port,proto | | | | | |
30000 | 19.6 | 11.6 | 3.9 | 0.5 | 2.6 +420% |
---------------|--------|--------|--------|--------|------------|
net6,port,mac | | | | | |
10 | 16.5 | 5.4 | 4.3 | 3.4 | 4.7 +38% |
---------------|--------|--------|--------|--------|------------|
net6,port,mac, | | | | | |
proto 1000 | 16.5 | 5.7 | 1.9 | 1.4 | 3.6 +26% |
---------------|--------|--------|--------|--------|------------|
net,mac | | | | | |
1000 | 19.0 | 8.4 | 3.9 | 2.5 | 6.4 +156% |
---------------'--------'--------'--------'--------'------------'
A similar strategy could be easily reused to implement specialised
versions for other SIMD sets, and I plan to post at least a NEON
version at a later time.
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Placing nftables set support in an extra module is pointless:
1. nf_tables needs dynamic registeration interface for sake of one module
2. nft heavily relies on sets, e.g. even simple rule like
"nft ... tcp dport { 80, 443 }" will not work with _SETS=n.
IOW, either nftables isn't used or both nf_tables and nf_tables_set
modules are needed anyway.
With extra module:
307K net/netfilter/nf_tables.ko
79K net/netfilter/nf_tables_set.ko
text data bss dec filename
146416 3072 545 150033 nf_tables.ko
35496 1817 0 37313 nf_tables_set.ko
This patch:
373K net/netfilter/nf_tables.ko
178563 4049 545 183157 nf_tables.ko
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
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>
This patch adds the dataplane hardware offload to the flowtable infrastructure. Three new flags represent the hardware state of this flow: * FLOW_OFFLOAD_HW: This flow entry resides in the hardware. * FLOW_OFFLOAD_HW_DYING: This flow entry has been scheduled to be remove from hardware. This might be triggered by either packet path (via TCP RST/FIN packet) or via aging. * FLOW_OFFLOAD_HW_DEAD: This flow entry has been already removed from the hardware, the software garbage collector can remove it from the software flowtable. This patch supports for: * IPv4 only. * Aging via FLOW_CLS_STATS, no packet and byte counter synchronization at this stage. This patch also adds the action callback that specifies how to convert the flow entry into the flow_rule object that is passed to the driver. Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Signed-off-by: David S. Miller <davem@davemloft.net>
Several header-files, Kconfig files and Makefiles have trailing white-space. Remove it. In netfilter/Kconfig, indent the type of CONFIG_NETFILTER_NETLINK_ACCT correctly. There are semicolons at the end of two function definitions in include/net/netfilter/nf_conntrack_acct.h and include/net/netfilter/nf_conntrack_ecache.h. Remove them. Fix indentation in nf_conntrack_l4proto.h. Signed-off-by: Jeremy Sowden <jeremy@azazel.net> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
This patch adds hardware offload support for nftables through the existing netdev_ops->ndo_setup_tc() interface, the TC_SETUP_CLSFLOWER classifier and the flow rule API. This hardware offload support is available for the NFPROTO_NETDEV family and the ingress hook. Each nftables expression has a new ->offload interface, that is used to populate the flow rule object that is attached to the transaction object. There is a new per-table NFT_TABLE_F_HW flag, that is set on to offload an entire table, including all of its chains. This patch supports for basic metadata (layer 3 and 4 protocol numbers), 5-tuple payload matching and the accept/drop actions; this also includes basechain hardware offload only. Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org> Signed-off-by: David S. Miller <davem@davemloft.net>
Add synproxy support for nf_tables. This behaves like the iptables synproxy target but it is structured in a way that allows us to propose improvements in the future. Signed-off-by: Fernando Fernandez Mancera <ffmancera@riseup.net> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
No need to have separate modules for this. before: text data bss dec filename 2038 1168 0 3206 net/ipv4/netfilter/ipt_MASQUERADE.ko 1526 1024 0 2550 net/ipv6/netfilter/ip6t_MASQUERADE.ko after: text data bss dec filename 2521 1296 0 3817 net/netfilter/xt_MASQUERADE.ko Signed-off-by: Florian Westphal <fw@strlen.de> Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>