Currently, SMC uses smc->sk.sk_{rcv|snd}buf to create buffers for
send buffer and RMB. And the values of buffer size are from tcp_{w|r}mem
in clcsock.
The buffer size from TCP socket doesn't fit SMC well. Generally, buffers
are usually larger than TCP for SMC-R/-D to get higher performance, for
they are different underlay devices and paths.
So this patch unbinds buffer size from TCP, and introduces two sysctl
knobs to tune them independently. Also, these knobs are per net
namespace and work for containers.
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
SMC-R tests the viability of link by sending out TEST_LINK LLC
messages over RoCE fabric when connections on link have been
idle for a time longer than keepalive interval (testlink time).
But using tcp_keepalive_time as testlink time maybe not quite
suitable because it is default no less than two hours[1], which
is too long for single link to find peer dead. The active host
will still use peer-dead link (QP) sending messages, and can't
find out until get IB_WC_RETRY_EXC_ERR error CQEs, which takes
more time than TEST_LINK timeout (SMC_LLC_WAIT_TIME) normally.
So this patch introduces a independent sysctl for SMC-R to set
link keepalive time, in order to detect link down in time. The
default value is 30 seconds.
[1] https://www.rfc-editor.org/rfc/rfc1122#page-101
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The more sockets we have in the hash table, the longer we spend looking
up the socket. While running a number of small workloads on the same
host, they penalise each other and cause performance degradation.
The root cause might be a single workload that consumes much more
resources than the others. It often happens on a cloud service where
different workloads share the same computing resource.
On EC2 c5.24xlarge instance (196 GiB memory and 524288 (1Mi / 2) ehash
entries), after running iperf3 in different netns, creating 24Mi sockets
without data transfer in the root netns causes about 10% performance
regression for the iperf3's connection.
thash_entries sockets length Gbps
524288 1 1 50.7
24Mi 48 45.1
It is basically related to the length of the list of each hash bucket.
For testing purposes to see how performance drops along the length,
I set 131072 (1Mi / 8) to thash_entries, and here's the result.
thash_entries sockets length Gbps
131072 1 1 50.7
1Mi 8 49.9
2Mi 16 48.9
4Mi 32 47.3
8Mi 64 44.6
16Mi 128 40.6
24Mi 192 36.3
32Mi 256 32.5
40Mi 320 27.0
48Mi 384 25.0
To resolve the socket lookup degradation, we introduce an optional
per-netns hash table for TCP, but it's just ehash, and we still share
the global bhash, bhash2 and lhash2.
With a smaller ehash, we can look up non-listener sockets faster and
isolate such noisy neighbours. In addition, we can reduce lock contention.
We can control the ehash size by a new sysctl knob. However, depending
on workloads, it will require very sensitive tuning, so we disable the
feature by default (net.ipv4.tcp_child_ehash_entries == 0). Moreover,
we can fall back to using the global ehash in case we fail to allocate
enough memory for a new ehash. The maximum size is 16Mi, which is large
enough that even if we have 48Mi sockets, the average list length is 3,
and regression would be less than 1%.
We can check the current ehash size by another read-only sysctl knob,
net.ipv4.tcp_ehash_entries. A negative value means the netns shares
the global ehash (per-netns ehash is disabled or failed to allocate
memory).
# dmesg | cut -d ' ' -f 5- | grep "established hash"
TCP established hash table entries: 524288 (order: 10, 4194304 bytes, vmalloc hugepage)
# sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = 524288 # can be changed by thash_entries
# sysctl net.ipv4.tcp_child_ehash_entries
net.ipv4.tcp_child_ehash_entries = 0 # disabled by default
# ip netns add test1
# ip netns exec test1 sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = -524288 # share the global ehash
# sysctl -w net.ipv4.tcp_child_ehash_entries=100
net.ipv4.tcp_child_ehash_entries = 100
# ip netns add test2
# ip netns exec test2 sysctl net.ipv4.tcp_ehash_entries
net.ipv4.tcp_ehash_entries = 128 # own a per-netns ehash with 2^n buckets
When more than two processes in the same netns create per-netns ehash
concurrently with different sizes, we need to guarantee the size in
one of the following ways:
1) Share the global ehash and create per-netns ehash
First, unshare() with tcp_child_ehash_entries==0. It creates dedicated
netns sysctl knobs where we can safely change tcp_child_ehash_entries
and clone()/unshare() to create a per-netns ehash.
2) Control write on sysctl by BPF
We can use BPF_PROG_TYPE_CGROUP_SYSCTL to allow/deny read/write on
sysctl knobs.
Note that the global ehash allocated at the boot time is spread over
available NUMA nodes, but inet_pernet_hashinfo_alloc() will allocate
pages for each per-netns ehash depending on the current process's NUMA
policy. By default, the allocation is done in the local node only, so
the per-netns hash table could fully reside on a random node. Thus,
depending on the NUMA policy the netns is created with and the CPU the
current thread is running on, we could see some performance differences
for highly optimised networking applications.
Note also that the default values of two sysctl knobs depend on the ehash
size and should be tuned carefully:
tcp_max_tw_buckets : tcp_child_ehash_entries / 2
tcp_max_syn_backlog : max(128, tcp_child_ehash_entries / 128)
As a bonus, we can dismantle netns faster. Currently, while destroying
netns, we call inet_twsk_purge(), which walks through the global ehash.
It can be potentially big because it can have many sockets other than
TIME_WAIT in all netns. Splitting ehash changes that situation, where
it's only necessary for inet_twsk_purge() to clean up TIME_WAIT sockets
in each netns.
With regard to this, we do not free the per-netns ehash in inet_twsk_kill()
to avoid UAF while iterating the per-netns ehash in inet_twsk_purge().
Instead, we do it in tcp_sk_exit_batch() after calling tcp_twsk_purge() to
keep it protocol-family-independent.
In the future, we could optimise ehash lookup/iteration further by removing
netns comparison for the per-netns ehash.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will soon introduce an optional per-netns ehash and access hash
tables via net->ipv4.tcp_death_row->hashinfo instead of &tcp_hashinfo
in most places.
It could harm the fast path because dereferences of two fields in net
and tcp_death_row might incur two extra cache line misses. To save one
dereference, let's place tcp_death_row back in netns_ipv4 and fetch
hashinfo via net->ipv4.tcp_death_row"."hashinfo.
Note tcp_death_row was initially placed in netns_ipv4, and commit
fbb8295248 ("tcp: allocate tcp_death_row outside of struct netns_ipv4")
changed it to a pointer so that we can fire TIME_WAIT timers after freeing
net. However, we don't do so after commit 04c494e68a ("Revert "tcp/dccp:
get rid of inet_twsk_purge()""), so we need not define tcp_death_row as a
pointer.
Also, we move refcount_dec_and_test(&tw_refcount) from tcp_sk_exit() to
tcp_sk_exit_batch() as a debug check.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drivers/net/ethernet/freescale/fec.h
7d650df99d ("net: fec: add pm_qos support on imx6q platform")
40c79ce13b ("net: fec: add stop mode support for imx8 platform")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Florian Westphal says:
====================
netfilter: bug fixes for net
1. Fix IP address check in irc DCC conntrack helper, this should check
the opposite direction rather than the destination address of the
packets' direction, from David Leadbeater.
2. bridge netfilter needs to drop dst references, from Harsh Modi.
This was fine back in the day the code was originally written,
but nowadays various tunnels can pre-set metadata dsts on packets.
3. Remove nf_conntrack_helper sysctl and the modparam toggle, users
need to explicitily assign the helpers to use via nftables or
iptables. Conntrack helpers, by design, may be used to add dynamic
port redirections to internal machines, so its necessary to restrict
which hosts/peers are allowed to use them.
It was discovered that improper checking in the irc DCC helper makes
it possible to trigger the 'please do dynamic port forward'
from outside by embedding a 'DCC' in a PING request; if the client
echos that back a expectation/port forward gets added.
The auto-assign-for-everything mechanism has been in "please don't do this"
territory since 2012. From Pablo.
4. Fix a memory leak in the netdev hook error unwind path, also from Pablo.
* git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf:
netfilter: nf_conntrack_irc: Fix forged IP logic
netfilter: nf_tables: clean up hook list when offload flags check fails
netfilter: br_netfilter: Drop dst references before setting.
netfilter: remove nf_conntrack_helper sysctl and modparam toggles
====================
Link: https://lore.kernel.org/r/20220901071238.3044-1-fw@strlen.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Because per host rate limiting has been proven problematic (side channel
attacks can be based on it), per host rate limiting of challenge acks ideally
should be per netns and turned off by default.
This is a long due followup of following commits:
083ae30828 ("tcp: enable per-socket rate limiting of all 'challenge acks'")
f2b2c582e8 ("tcp: mitigate ACK loops for connections as tcp_sock")
75ff39ccc1 ("tcp: make challenge acks less predictable")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jason Baron <jbaron@akamai.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
__nf_ct_try_assign_helper() remains in place but it now requires a
template to configure the helper.
A toggle to disable automatic helper assignment was added by:
a900689264 ("netfilter: nf_ct_helper: allow to disable automatic helper assignment")
in 2012 to address the issues described in "Secure use of iptables and
connection tracking helpers". Automatic conntrack helper assignment was
disabled by:
3bb398d925 ("netfilter: nf_ct_helper: disable automatic helper assignment")
back in 2016.
This patch removes the sysctl and modparam toggles, users now have to
rely on explicit conntrack helper configuration via ruleset.
Update tools/testing/selftests/netfilter/nft_conntrack_helper.sh to
check that auto-assignment does not happen anymore.
Acked-by: Aaron Conole <aconole@redhat.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
DECnet is an obsolete network protocol that receives more attention
from kernel janitors than users. It belongs in computer protocol
history museum not in Linux kernel.
It has been "Orphaned" in kernel since 2010. The iproute2 support
for DECnet was dropped in 5.0 release. The documentation link on
Sourceforge says it is abandoned there as well.
Leave the UAPI alone to keep userspace programs compiling.
This means that there is still an empty neighbour table
for AF_DECNET.
The table of /proc/sys/net entries was updated to match
current directories and reformatted to be alphabetical.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: David Ahern <dsahern@kernel.org>
Acked-by: Nikolay Aleksandrov <razor@blackwall.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
To avoid allocation of the conntrack extension area when possible,
the default behaviour was changed to only allocate the event extension
if a userspace program is subscribed to a notification group.
Problem is that while 'conntrack -E' does enable the event allocation
behind the scenes, 'conntrack -E expect' does not: no expectation events
are delivered unless user sets
"net.netfilter.nf_conntrack_events" back to 1 (always on).
Fix the autodetection to also consider EXP type group.
We need to track the 6 event groups (3+3, new/update/destroy for events and
for expectations each) independently, else we'd disable events again
if an expectation group becomes empty while there is still an active
event group.
Fixes: 2794cdb0b9 ("netfilter: nfnetlink: allow to detect if ctnetlink listeners exist")
Reported-by: Yi Chen <yiche@redhat.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
This patch adds missing includes to headers under include/net.
All these problems are currently masked by the existing users
including the missing dependency before the broken header.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pablo Neira Ayuso says:
====================
Netfilter/IPVS updates for net-next
The following patchset contains Netfilter/IPVS updates for net-next:
1) Simplify nf_ct_get_tuple(), from Jackie Liu.
2) Add format to request_module() call, from Bill Wendling.
3) Add /proc/net/stats/nf_flowtable to monitor in-flight pending
hardware offload objects to be processed, from Vlad Buslov.
4) Missing rcu annotation and accessors in the netfilter tree,
from Florian Westphal.
5) Merge h323 conntrack helper nat hooks into single object,
also from Florian.
6) A batch of update to fix sparse warnings treewide,
from Florian Westphal.
7) Move nft_cmp_fast_mask() where it used, from Florian.
8) Missing const in nf_nat_initialized(), from James Yonan.
9) Use bitmap API for Maglev IPVS scheduler, from Christophe Jaillet.
10) Use refcount_inc instead of _inc_not_zero in flowtable,
from Florian Westphal.
11) Remove pr_debug in xt_TPROXY, from Nathan Cancellor.
* git://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf-next:
netfilter: xt_TPROXY: remove pr_debug invocations
netfilter: flowtable: prefer refcount_inc
netfilter: ipvs: Use the bitmap API to allocate bitmaps
netfilter: nf_nat: in nf_nat_initialized(), use const struct nf_conn *
netfilter: nf_tables: move nft_cmp_fast_mask to where its used
netfilter: nf_tables: use correct integer types
netfilter: nf_tables: add and use BE register load-store helpers
netfilter: nf_tables: use the correct get/put helpers
netfilter: x_tables: use correct integer types
netfilter: nfnetlink: add missing __be16 cast
netfilter: nft_set_bitmap: Fix spelling mistake
netfilter: h323: merge nat hook pointers into one
netfilter: nf_conntrack: use rcu accessors where needed
netfilter: nf_conntrack: add missing __rcu annotations
netfilter: nf_flow_table: count pending offload workqueue tasks
net/sched: act_ct: set 'net' pointer when creating new nf_flow_table
netfilter: conntrack: use correct format characters
netfilter: conntrack: use fallthrough to cleanup
====================
Link: https://lore.kernel.org/r/20220720230754.209053-1-pablo@netfilter.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch introduces the sysctl smcr_buf_type for setting
the type of SMC-R sndbufs and RMBs.
Valid values includes:
- SMCR_PHYS_CONT_BUFS, which means use physically contiguous
buffers for better performance and is the default value.
- SMCR_VIRT_CONT_BUFS, which means use virtually contiguous
buffers in case of physically contiguous memory is scarce.
- SMCR_MIXED_BUFS, which means first try to use physically
contiguous buffers. If not available, then use virtually
contiguous buffers.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To improve hardware offload debuggability count pending 'add', 'del' and
'stats' flow_table offload workqueue tasks. Counters are incremented before
scheduling new task and decremented when workqueue handler finishes
executing. These counters allow user to diagnose congestion on hardware
offload workqueues that can happen when either CPU is starved and workqueue
jobs are executed at lower rate than new ones are added or when
hardware/driver can't keep up with the rate.
Implement the described counters as percpu counters inside new struct
netns_ft which is stored inside struct net. Expose them via new procfs file
'/proc/net/stats/nf_flowtable' that is similar to existing 'nf_conntrack'
file.
Signed-off-by: Vlad Buslov <vladbu@nvidia.com>
Signed-off-by: Oz Shlomo <ozsh@nvidia.com>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
This commit adds a per netns hash table for AF_UNIX, which size is fixed
as UNIX_HASH_SIZE for now.
The first implementation defines a per-netns hash table as a single array
of lock and list:
struct unix_hashbucket {
spinlock_t lock;
struct hlist_head head;
};
struct netns_unix {
struct unix_hashbucket *hash;
...
};
But, Eric pointed out memory cost that the structure has holes because of
sizeof(spinlock_t), which is 4 (or more if LOCKDEP is enabled). [0] It
could be expensive on a host with thousands of netns and few AF_UNIX
sockets. For this reason, a per-netns hash table uses two dense arrays.
struct unix_table {
spinlock_t *locks;
struct hlist_head *buckets;
};
struct netns_unix {
struct unix_table table;
...
};
Note the length of the list has a significant impact rather than lock
contention, so having shared locks can be an option. But, per-netns
locks and lists still perform better than the global locks and per-netns
lists. [1]
Also, this patch adds a change so that struct netns_unix disappears from
struct net if CONFIG_UNIX is disabled.
[0]: https://lore.kernel.org/netdev/CANn89iLVxO5aqx16azNU7p7Z-nz5NrnM5QTqOzueVxEnkVTxyg@mail.gmail.com/
[1]: https://lore.kernel.org/netdev/20220617175215.1769-1-kuniyu@amazon.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Pablo Neira Ayuso says:
====================
Netfilter updates for net-next
This is v2 including deadlock fix in conntrack ecache rework
reported by Jakub Kicinski.
The following patchset contains Netfilter updates for net-next,
mostly updates to conntrack from Florian Westphal.
1) Add a dedicated list for conntrack event redelivery.
2) Include event redelivery list in conntrack dumps of dying type.
3) Remove per-cpu dying list for event redelivery, not used anymore.
4) Add netns .pre_exit to cttimeout to zap timeout objects before
synchronize_rcu() call.
5) Remove nf_ct_unconfirmed_destroy.
6) Add generation id for conntrack extensions for conntrack
timeout and helpers.
7) Detach timeout policy from conntrack on cttimeout module removal.
8) Remove __nf_ct_unconfirmed_destroy.
9) Remove unconfirmed list.
10) Remove unconditional local_bh_disable in init_conntrack().
11) Consolidate conntrack iterator nf_ct_iterate_cleanup().
12) Detect if ctnetlink listeners exist to short-circuit event
path early.
13) Un-inline nf_ct_ecache_ext_add().
14) Add nf_conntrack_events autodetect ctnetlink listener mode
and make it default.
15) Add nf_ct_ecache_exist() to check for event cache extension.
16) Extend flowtable reverse route lookup to include source, iif,
tos and mark, from Sven Auhagen.
17) Do not verify zero checksum UDP packets in nf_reject,
from Kevin Mitchell.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
At this time, every new conntrack gets the 'event cache extension'
enabled for it.
This is because the 'net.netfilter.nf_conntrack_events' sysctl defaults
to 1.
Changing the default to 0 means that commands that rely on the event
notification extension, e.g. 'conntrack -E' or conntrackd, stop working.
We COULD detect if there is a listener by means of
'nfnetlink_has_listeners()' and only add the extension if this is true.
The downside is a dependency from conntrack module to nfnetlink module.
This adds a different way: inc/dec a counter whenever a ctnetlink group
is being (un)subscribed and toggle a flag in struct net.
Next patches will take advantage of this and will only add the event
extension if the flag is set.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Its no longer needed. Entries that need event redelivery are placed
on the new pernet dying list.
The advantage is that there is no need to take additional spinlock on
conntrack removal unless event redelivery failed or the conntrack entry
was never added to the table in the first place (confirmed bit not set).
The IPS_CONFIRMED bit now needs to be set as soon as the entry has been
unlinked from the unconfirmed list, else the destroy function may
attempt to unlink it a second time.
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Reads and Writes to ip6_rt_gc_expire always have been racy,
as syzbot reported lately [1]
There is a possible risk of under-flow, leading
to unexpected high value passed to fib6_run_gc(),
although I have not observed this in the field.
Hosts hitting ip6_dst_gc() very hard are under pretty bad
state anyway.
[1]
BUG: KCSAN: data-race in ip6_dst_gc / ip6_dst_gc
read-write to 0xffff888102110744 of 4 bytes by task 13165 on cpu 1:
ip6_dst_gc+0x1f3/0x220 net/ipv6/route.c:3311
dst_alloc+0x9b/0x160 net/core/dst.c:86
ip6_dst_alloc net/ipv6/route.c:344 [inline]
icmp6_dst_alloc+0xb2/0x360 net/ipv6/route.c:3261
mld_sendpack+0x2b9/0x580 net/ipv6/mcast.c:1807
mld_send_cr net/ipv6/mcast.c:2119 [inline]
mld_ifc_work+0x576/0x800 net/ipv6/mcast.c:2651
process_one_work+0x3d3/0x720 kernel/workqueue.c:2289
worker_thread+0x618/0xa70 kernel/workqueue.c:2436
kthread+0x1a9/0x1e0 kernel/kthread.c:376
ret_from_fork+0x1f/0x30
read-write to 0xffff888102110744 of 4 bytes by task 11607 on cpu 0:
ip6_dst_gc+0x1f3/0x220 net/ipv6/route.c:3311
dst_alloc+0x9b/0x160 net/core/dst.c:86
ip6_dst_alloc net/ipv6/route.c:344 [inline]
icmp6_dst_alloc+0xb2/0x360 net/ipv6/route.c:3261
mld_sendpack+0x2b9/0x580 net/ipv6/mcast.c:1807
mld_send_cr net/ipv6/mcast.c:2119 [inline]
mld_ifc_work+0x576/0x800 net/ipv6/mcast.c:2651
process_one_work+0x3d3/0x720 kernel/workqueue.c:2289
worker_thread+0x618/0xa70 kernel/workqueue.c:2436
kthread+0x1a9/0x1e0 kernel/kthread.c:376
ret_from_fork+0x1f/0x30
value changed: 0x00000bb3 -> 0x00000ba9
Reported by Kernel Concurrency Sanitizer on:
CPU: 0 PID: 11607 Comm: kworker/0:21 Not tainted 5.18.0-rc1-syzkaller-00037-g42e7a03d3bad-dirty #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Workqueue: mld mld_ifc_work
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reported-by: syzbot <syzkaller@googlegroups.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20220413181333.649424-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ipsec-next
Steffen Klassert says:
====================
pull request (net-next): ipsec-next 2022-03-19
1) Delete duplicated functions that calls same xfrm_api_check.
From Leon Romanovsky.
2) Align userland API of the default policy structure to the
internal structures. From Nicolas Dichtel.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>
Back when tcp_tso_autosize() and TCP pacing were introduced,
our focus was really to reduce burst sizes for long distance
flows.
The simple heuristic of using sk_pacing_rate/1024 has worked
well, but can lead to too small packets for hosts in the same
rack/cluster, when thousands of flows compete for the bottleneck.
Neal Cardwell had the idea of making the TSO burst size
a function of both sk_pacing_rate and tcp_min_rtt()
Indeed, for local flows, sending bigger bursts is better
to reduce cpu costs, as occasional losses can be repaired
quite fast.
This patch is based on Neal Cardwell implementation
done more than two years ago.
bbr is adjusting max_pacing_rate based on measured bandwidth,
while cubic would over estimate max_pacing_rate.
/proc/sys/net/ipv4/tcp_tso_rtt_log can be used to tune or disable
this new feature, in logarithmic steps.
Tested:
100Gbit NIC, two hosts in the same rack, 4K MTU.
600 flows rate-limited to 20000000 bytes per second.
Before patch: (TSO sizes would be limited to 20000000/1024/4096 -> 4 segments per TSO)
~# echo 0 >/proc/sys/net/ipv4/tcp_tso_rtt_log
~# nstat -n;perf stat ./super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000;nstat|egrep "TcpInSegs|TcpOutSegs|TcpRetransSegs|Delivered"
96005
Performance counter stats for './super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000':
65,945.29 msec task-clock # 2.845 CPUs utilized
1,314,632 context-switches # 19935.279 M/sec
5,292 cpu-migrations # 80.249 M/sec
940,641 page-faults # 14264.023 M/sec
201,117,030,926 cycles # 3049769.216 GHz (83.45%)
17,699,435,405 stalled-cycles-frontend # 8.80% frontend cycles idle (83.48%)
136,584,015,071 stalled-cycles-backend # 67.91% backend cycles idle (83.44%)
53,809,530,436 instructions # 0.27 insn per cycle
# 2.54 stalled cycles per insn (83.36%)
9,062,315,523 branches # 137422329.563 M/sec (83.22%)
153,008,621 branch-misses # 1.69% of all branches (83.32%)
23.182970846 seconds time elapsed
TcpInSegs 15648792 0.0
TcpOutSegs 58659110 0.0 # Average of 3.7 4K segments per TSO packet
TcpExtTCPDelivered 58654791 0.0
TcpExtTCPDeliveredCE 19 0.0
After patch:
~# echo 9 >/proc/sys/net/ipv4/tcp_tso_rtt_log
~# nstat -n;perf stat ./super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000;nstat|egrep "TcpInSegs|TcpOutSegs|TcpRetransSegs|Delivered"
96046
Performance counter stats for './super_netperf 600 -H otrv6 -l 20 -- -K dctcp -q 20000000':
48,982.58 msec task-clock # 2.104 CPUs utilized
186,014 context-switches # 3797.599 M/sec
3,109 cpu-migrations # 63.472 M/sec
941,180 page-faults # 19214.814 M/sec
153,459,763,868 cycles # 3132982.807 GHz (83.56%)
12,069,861,356 stalled-cycles-frontend # 7.87% frontend cycles idle (83.32%)
120,485,917,953 stalled-cycles-backend # 78.51% backend cycles idle (83.24%)
36,803,672,106 instructions # 0.24 insn per cycle
# 3.27 stalled cycles per insn (83.18%)
5,947,266,275 branches # 121417383.427 M/sec (83.64%)
87,984,616 branch-misses # 1.48% of all branches (83.43%)
23.281200256 seconds time elapsed
TcpInSegs 1434706 0.0
TcpOutSegs 58883378 0.0 # Average of 41 4K segments per TSO packet
TcpExtTCPDelivered 58878971 0.0
TcpExtTCPDeliveredCE 9664 0.0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://lore.kernel.org/r/20220309015757.2532973-1-eric.dumazet@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
