This converts some of the visually simpler cases that have been split
over multiple lines. I only did the ones that are easy to verify the
resulting diff by having just that final GFP_KERNEL argument on the next
line.
Somebody should probably do a proper coccinelle script for this, but for
me the trivial script actually resulted in an assertion failure in the
middle of the script. I probably had made it a bit _too_ trivial.
So after fighting that far a while I decided to just do some of the
syntactically simpler cases with variations of the previous 'sed'
scripts.
The more syntactically complex multi-line cases would mostly really want
whitespace cleanup anyway.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is the exact same thing as the 'alloc_obj()' version, only much
smaller because there are a lot fewer users of the *alloc_flex()
interface.
As with alloc_obj() version, this was done entirely with mindless brute
force, using the same script, except using 'flex' in the pattern rather
than 'objs*'.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This was done entirely with mindless brute force, using
git grep -l '\<k[vmz]*alloc_objs*(.*, GFP_KERNEL)' |
xargs sed -i 's/\(alloc_objs*(.*\), GFP_KERNEL)/\1)/'
to convert the new alloc_obj() users that had a simple GFP_KERNEL
argument to just drop that argument.
Note that due to the extreme simplicity of the scripting, any slightly
more complex cases spread over multiple lines would not be triggered:
they definitely exist, but this covers the vast bulk of the cases, and
the resulting diff is also then easier to check automatically.
For the same reason the 'flex' versions will be done as a separate
conversion.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This is the result of running the Coccinelle script from
scripts/coccinelle/api/kmalloc_objs.cocci. The script is designed to
avoid scalar types (which need careful case-by-case checking), and
instead replace kmalloc-family calls that allocate struct or union
object instances:
Single allocations: kmalloc(sizeof(TYPE), ...)
are replaced with: kmalloc_obj(TYPE, ...)
Array allocations: kmalloc_array(COUNT, sizeof(TYPE), ...)
are replaced with: kmalloc_objs(TYPE, COUNT, ...)
Flex array allocations: kmalloc(struct_size(PTR, FAM, COUNT), ...)
are replaced with: kmalloc_flex(*PTR, FAM, COUNT, ...)
(where TYPE may also be *VAR)
The resulting allocations no longer return "void *", instead returning
"TYPE *".
Signed-off-by: Kees Cook <kees@kernel.org>
Although unlikely, recent support for IPIP tunnels increases chances of
reaching this WARN_ON_ONCE if userspace manages to build a sufficiently
long forward path.
Remove it.
Fixes: ddb94eafab ("net: resolve forwarding path from virtual netdevice and HW destination address")
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
Signed-off-by: Florian Westphal <fw@strlen.de>
We should follow the prepare/commit approach for queue configuration.
The qcfg struct should be added to dev->cfg rather than directly to
queue objects so that we can clone and discard the pending config
easily.
Remove the qcfg in struct netdev_rx_queue, and switch remaining callers
to netdev_queue_config(). netdev_queue_config() will construct the qcfg
on the fly based on device defaults and state of the queue.
ndo_default_qcfg becomes optional because having the callback itself
does not have any meaningful semantics to us.
Reviewed-by: Dragos Tatulea <dtatulea@nvidia.com>
Link: https://patch.msgid.link/20260122005113.2476634-5-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some/most devices implementing gso_partial need to disable the GSO partial
features when the IP ID can't be mangled; to that extend each of them
implements something alike the following[1]:
if (skb->encapsulation && !(features & NETIF_F_TSO_MANGLEID))
features &= ~NETIF_F_TSO;
in the ndo_features_check() op, which leads to a bit of duplicate code.
Later patch in the series will implement GSO partial support for virtual
devices, and the current status quo will require more duplicate code and
a new indirect call in the TX path for them.
Introduce the mangleid_features mask, allowing the core to disable NIC
features based on/requiring MANGLEID, without any further intervention
from the driver.
The same functionality could be alternatively implemented adding a single
boolean flag to the struct net_device, but would require an additional
checks in ndo_features_check().
Also note that [1] is incorrect if the NIC additionally implements
NETIF_F_GSO_UDP_L4, mangleid_features transparently handle even such a
case.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/5a7cdaeea40b0a29b88e525b6c942d73ed3b8ce7.1769011015.git.pabeni@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Pavel Begunkov says:
====================
Add support for providers with large rx buffer
Many modern NICs support configurable receive buffer lengths, and zcrx and
memory providers can use buffers larger than 4K to improve performance.
When paired with hw-gro larger rx buffer sizes can drastically reduce
the number of buffers traversing the stack and save a lot of processing
time. It also allows to give to users larger contiguous chunks of data.
Single stream benchmarks showed up to ~30% CPU util improvement.
E.g. comparison for 4K vs 32K buffers using a 200Gbit NIC:
packets=23987040 (MB=2745098), rps=199559 (MB/s=22837)
CPU %usr %nice %sys %iowait %irq %soft %idle
0 1.53 0.00 27.78 2.72 1.31 66.45 0.22
packets=24078368 (MB=2755550), rps=200319 (MB/s=22924)
CPU %usr %nice %sys %iowait %irq %soft %idle
0 0.69 0.00 8.26 31.65 1.83 57.00 0.57
This series adds net infrastructure for memory providers configuring
the size and implements it for bnxt. It's an opt-in feature for drivers,
they should advertise support for the parameter in the qops and must check
if the hardware supports the given size. It's limited to memory providers
as it drastically simplifies implementation. It doesn't affect the fast
path zcrx uAPI, and the user exposed parameter is defined in zcrx terms,
which allows it to be flexible and adjusted in the future.
A liburing example can be found at [2]
full branch:
[1] https://github.com/isilence/linux.git zcrx/large-buffers-v8
Liburing example:
[2] https://github.com/isilence/liburing.git zcrx/rx-buf-len
* tag 'net-queue-rx-buf-len-v9' of https://github.com/isilence/linux:
io_uring/zcrx: document area chunking parameter
selftests: iou-zcrx: test large chunk sizes
eth: bnxt: support qcfg provided rx page size
eth: bnxt: adjust the fill level of agg queues with larger buffers
eth: bnxt: store rx buffer size per queue
net: pass queue rx page size from memory provider
net: add bare bone queue configs
net: reduce indent of struct netdev_queue_mgmt_ops members
net: memzero mp params when closing a queue
====================
Link: https://patch.msgid.link/
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This reverts commit 77b9c4a438, reversing
changes made to 4515ec4ad5:
931420a2fc ("selftests/net: Add netkit container tests")
ab771c938d ("selftests/net: Make NetDrvContEnv support queue leasing")
6be87fbb27 ("selftests/net: Add env for container based tests")
61d99ce3df ("selftests/net: Add bpf skb forwarding program")
920da36341 ("netkit: Add xsk support for af_xdp applications")
eef51113f8 ("netkit: Add netkit notifier to check for unregistering devices")
b5ef109d22 ("netkit: Implement rtnl_link_ops->alloc and ndo_queue_create")
b5c3fa4a0b ("netkit: Add single device mode for netkit")
0073d2fd67 ("xsk: Proxy pool management for leased queues")
1ecea95dd3 ("xsk: Extend xsk_rcv_check validation")
804bf334d0 ("net: Proxy netdev_queue_get_dma_dev for leased queues")
0caa9a8dde ("net: Proxy net_mp_{open,close}_rxq for leased queues")
ff8889ff91 ("net, ethtool: Disallow leased real rxqs to be resized")
9e2103f361 ("net: Add lease info to queue-get response")
31127dedde ("net: Implement netdev_nl_queue_create_doit")
a5546e18f7 ("net: Add queue-create operation")
The series will conflict with io_uring work, and the code needs more
polish.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Implement netdev_nl_queue_create_doit which creates a new rx queue in a
virtual netdev and then leases it to a rx queue in a physical netdev.
Example with ynl client:
# ./pyynl/cli.py \
--spec ~/netlink/specs/netdev.yaml \
--do queue-create \
--json '{"ifindex": 8, "type": "rx", "lease": {"ifindex": 4, "queue": {"type": "rx", "id": 15}}}'
{'id': 1}
Note that the netdevice locking order is always from the virtual to
the physical device.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Co-developed-by: David Wei <dw@davidwei.uk>
Signed-off-by: David Wei <dw@davidwei.uk>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
Link: https://patch.msgid.link/20260115082603.219152-3-daniel@iogearbox.net
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
We'll need to pass extra parameters when allocating a queue for memory
providers. Define a new structure for queue configurations, and pass it
to qapi callbacks. It's empty for now, actual parameters will be added
in following patches.
Configurations should persist across resets, and for that they're
default-initialised on device registration and stored in struct
netdev_rx_queue. We also add a new qapi callback for defaulting a given
config. It must be implemented if a driver wants to use queue configs
and is optional otherwise.
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
In blamed commit, I added a check against the temporary queue
built in __dev_xmit_skb(). Idea was to drop packets early,
before any spinlock was acquired.
if (unlikely(defer_count > READ_ONCE(q->limit))) {
kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_DROP);
return NET_XMIT_DROP;
}
It turned out that HTB Qdisc has a zero q->limit.
HTB limits packets on a per-class basis.
Some of our tests became flaky.
Add a new sysctl : net.core.qdisc_max_burst to control
how many packets can be stored in the temporary lockless queue.
Also add a new QDISC_BURST_DROP drop reason to better diagnose
future issues.
Thanks Neal !
Fixes: 100dfa74ca ("net: dev_queue_xmit() llist adoption")
Reported-and-bisected-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Neal Cardwell <ncardwell@google.com>
Link: https://patch.msgid.link/20260107104159.3669285-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add missing entries in netdev_lock_type[] and netdev_lock_name[] :
CAN, MCTP, RAWIP, CAIF, IP6GRE, 6LOWPAN, NETLINK, VSOCKMON,
IEEE802154_MONITOR.
Also add a WARN_ONCE() in netdev_lock_pos() to help future bug hunting
next time a protocol is added without updating these arrays.
Fixes: 1a33e10e4a ("net: partially revert dynamic lockdep key changes")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260108093244.830280-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some arches (like x86) do not inline spin_unlock_irqrestore().
backlog_unlock_irq_restore() is in RPS/RFS critical path,
we prefer using spin_unlock() + local_irq_restore() for
optimal performance.
Also change backlog_unlock_irq_restore() second argument
to avoid a pointless dereference.
No difference in net/core/dev.o code size.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260105163054.13698-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some qdisc like cake, codel, fq_codel might drop packets
in their dequeue() method.
This is currently problematic because dequeue() runs with
the qdisc spinlock held. Freeing skbs can be extremely expensive.
Add qdisc_dequeue_drop() method and a new TCQ_F_DEQUEUE_DROPS
so that these qdiscs can opt-in to defer the skb frees
after the socket spinlock is released.
TCQ_F_DEQUEUE_DROPS is an attempt to not penalize other qdiscs
with an extra cache line miss.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251121083256.674562-14-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Most qdiscs need to read skb->priority at enqueue time().
In commit 100dfa74ca ("net: dev_queue_xmit() llist adoption")
I added a prefetch(next), lets add another one for the second
half of skb.
Note that skb->priority and skb->hash share a common cache line,
so this patch helps qdiscs needing both fields.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251121083256.674562-11-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
skb_defer_free_flush() is becoming more important these days.
Add a prefetch operation to reduce latency a bit on some
platforms like AMD EPYC 7B12.
On more recent cpus, a stall happens when reading skb_shinfo().
Avoiding it will require a more elaborate strategy.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://patch.msgid.link/20251106085500.2438951-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Revert struct sockaddr from flexible array to fixed 14-byte "sa_data",
to solve over 36,000 -Wflex-array-member-not-at-end warnings, since
struct sockaddr is embedded within many network structs.
With socket/proto sockaddr-based internal APIs switched to use struct
sockaddr_unsized, there should be no more uses of struct sockaddr that
depend on reading beyond the end of struct sockaddr::sa_data that might
trigger bounds checking.
Comparing an x86_64 "allyesconfig" vmlinux build before and after this
patch showed no new "ud1" instructions from CONFIG_UBSAN_BOUNDS nor any
new "field-spanning" memcpy CONFIG_FORTIFY_SOURCE instrumentations.
Cc: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Kees Cook <kees@kernel.org>
Link: https://patch.msgid.link/20251104002617.2752303-8-kees@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
deliver_skb() should not be inlined as is it not called
in the fast path.
Add unlikely() clauses giving hints to the compiler about this fact.
Before this patch:
size net/core/dev.o
text data bss dec hex filename
121794 13330 176 135300 21084 net/core/dev.o
__netif_receive_skb_core() size on x86_64 : 4080 bytes.
After:
size net/core/dev.o
text data bss dec hex filenamee
120330 13338 176 133844 20ad4 net/core/dev.o
__netif_receive_skb_core() size on x86_64 : 2781 bytes.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251103165256.1712169-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add a new state NAPI_STATE_THREADED_BUSY_POLL to the NAPI state enum to
enable and disable threaded busy polling.
When threaded busy polling is enabled for a NAPI, enable
NAPI_STATE_THREADED also.
When the threaded NAPI is scheduled, set NAPI_STATE_IN_BUSY_POLL to
signal napi_complete_done not to rearm interrupts.
Whenever NAPI_STATE_THREADED_BUSY_POLL is unset, the
NAPI_STATE_IN_BUSY_POLL will be unset, napi_complete_done unsets the
NAPI_STATE_SCHED_THREADED bit also, which in turn will make the kthread
go to sleep.
Signed-off-by: Samiullah Khawaja <skhawaja@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Acked-by: Martin Karsten <mkarsten@uwaterloo.ca>
Tested-by: Martin Karsten <mkarsten@uwaterloo.ca>
Link: https://patch.msgid.link/20251028203007.575686-2-skhawaja@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add the ability to append the incoming IP interface information to
ICMPv6 error messages in accordance with RFC 5837 and RFC 4884. This is
required for more meaningful traceroute results in unnumbered networks.
The feature is disabled by default and controlled via a new sysctl
("net.ipv6.icmp.errors_extension_mask") which accepts a bitmask of ICMP
extensions to append to ICMP error messages. Currently, only a single
value is supported, but the interface and the implementation should be
able to support more extensions, if needed.
Clone the skb and copy the relevant data portions before modifying the
skb as the caller of icmp6_send() still owns the skb after the function
returns. This should be fine since by default ICMP error messages are
rate limited to 1000 per second and no more than 1 per second per
specific host.
Trim or pad the packet to 128 bytes before appending the ICMP extension
structure in order to be compatible with legacy applications that assume
that the ICMP extension structure always starts at this offset (the
minimum length specified by RFC 4884).
Since commit 20e1954fe2 ("ipv6: RFC 4884 partial support for SIT/GRE
tunnels") it is possible for icmp6_send() to be called with an skb that
already contains ICMP extensions. This can happen when we receive an
ICMPv4 message with extensions from a tunnel and translate it to an
ICMPv6 message towards an IPv6 host in the overlay network. I could not
find an RFC that supports this behavior, but it makes sense to not
overwrite the original extensions that were appended to the packet.
Therefore, avoid appending extensions if the length field in the
provided ICMPv6 header is already filled.
Export netdev_copy_name() using EXPORT_IPV6_MOD_GPL() to make it
available to IPv6 when it is built as a module.
Reviewed-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20251027082232.232571-3-idosch@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add likely() and unlikely() clauses for the common cases:
Device is running.
Queue is not full.
Queue is less than half capacity.
Add max_backlog parameter to skb_flow_limit() to avoid
a second READ_ONCE(net_hotdata.max_backlog).
skb_flow_limit() does not need the backlog_lock protection,
and can be called before we acquire the lock, for even better
resistance to attacks.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20251024090517.3289181-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some high level software drivers need to compute features from lower
devices. But each has their own implementations and may lost some
feature compute. Let's use one common function to compute features
for kinds of these devices.
The new helper uses the current bond implementation as the reference
one, as the latter already handles all the relevant aspects: netdev
features, TSO limits and dst retention.
Suggested-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Hangbin Liu <liuhangbin@gmail.com>
Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Link: https://patch.msgid.link/20251017034155.61990-2-liuhangbin@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cpus serving NIC interrupts and specifically TX completions are often
trapped in also restarting a busy qdisc (because qdisc was stopped by BQL
or the driver's own flow control).
When they call netdev_tx_completed_queue() or netif_tx_wake_queue(),
they call __netif_schedule() so that the queue can be run
later from net_tx_action() (involving NET_TX_SOFTIRQ)
Quite often, by the time the cpu reaches net_tx_action(), another cpu
grabbed the qdisc spinlock from __dev_xmit_skb(), and we spend too much
time spinning on this lock.
We can detect in __netif_schedule() if a cpu is already at a specific
point in __dev_xmit_skb() where we have the guarantee the queue will
be run.
This patch gives a 13 % increase of throughput on an IDPF NIC (200Gbit),
32 TX qeues, sending UDP packets of 120 bytes.
This also helps __qdisc_run() to not force a NET_TX_SOFTIRQ
if another thread is waiting in __dev_xmit_skb()
Before:
sar -n DEV 5 5|grep eth1|grep Average
Average: eth1 1496.44 52191462.56 210.00 13369396.90 0.00 0.00 0.00 54.76
After:
sar -n DEV 5 5|grep eth1|grep Average
Average: eth1 1457.88 59363099.96 205.08 15206384.35 0.00 0.00 0.00 62.29
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20251017145334.3016097-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Remove busylock spinlock and use a lockless list (llist)
to reduce spinlock contention to the minimum.
Idea is that only one cpu might spin on the qdisc spinlock,
while others simply add their skb in the llist.
After this patch, we get a 300 % improvement on heavy TX workloads.
- Sending twice the number of packets per second.
- While consuming 50 % less cycles.
Note that this also allows in the future to submit batches
to various qdisc->enqueue() methods.
Tested:
- Dual Intel(R) Xeon(R) 6985P-C (480 hyper threads).
- 100Gbit NIC, 30 TX queues with FQ packet scheduler.
- echo 64 >/sys/kernel/slab/skbuff_small_head/cpu_partial (avoid contention in mm)
- 240 concurrent "netperf -t UDP_STREAM -- -m 120 -n"
Before:
16 Mpps (41 Mpps if each thread is pinned to a different cpu)
vmstat 2 5
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
243 0 0 2368988672 51036 1100852 0 0 146 1 242 60 0 9 91 0 0
244 0 0 2368988672 51036 1100852 0 0 536 10 487745 14718 0 52 48 0 0
244 0 0 2368988672 51036 1100852 0 0 512 0 503067 46033 0 52 48 0 0
244 0 0 2368988672 51036 1100852 0 0 512 0 494807 12107 0 52 48 0 0
244 0 0 2368988672 51036 1100852 0 0 702 26 492845 10110 0 52 48 0 0
Lock contention (1 second sample taken on 8 cores)
perf lock record -C0-7 sleep 1; perf lock contention
contended total wait max wait avg wait type caller
442111 6.79 s 162.47 ms 15.35 us spinlock dev_hard_start_xmit+0xcd
5961 9.57 ms 8.12 us 1.60 us spinlock __dev_queue_xmit+0x3a0
244 560.63 us 7.63 us 2.30 us spinlock do_softirq+0x5b
13 25.09 us 3.21 us 1.93 us spinlock net_tx_action+0xf8
If netperf threads are pinned, spinlock stress is very high.
perf lock record -C0-7 sleep 1; perf lock contention
contended total wait max wait avg wait type caller
964508 7.10 s 147.25 ms 7.36 us spinlock dev_hard_start_xmit+0xcd
201 268.05 us 4.65 us 1.33 us spinlock __dev_queue_xmit+0x3a0
12 26.05 us 3.84 us 2.17 us spinlock do_softirq+0x5b
@__dev_queue_xmit_ns:
[256, 512) 21 | |
[512, 1K) 631 | |
[1K, 2K) 27328 |@ |
[2K, 4K) 265392 |@@@@@@@@@@@@@@@@ |
[4K, 8K) 417543 |@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8K, 16K) 826292 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[16K, 32K) 733822 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32K, 64K) 19055 |@ |
[64K, 128K) 17240 |@ |
[128K, 256K) 25633 |@ |
[256K, 512K) 4 | |
After:
29 Mpps (57 Mpps if each thread is pinned to a different cpu)
vmstat 2 5
procs -----------memory---------- ---swap-- -----io---- -system-- ------cpu-----
r b swpd free buff cache si so bi bo in cs us sy id wa st
78 0 0 2369573632 32896 1350988 0 0 22 0 331 254 0 8 92 0 0
75 0 0 2369573632 32896 1350988 0 0 22 50 425713 280199 0 23 76 0 0
104 0 0 2369573632 32896 1350988 0 0 290 0 430238 298247 0 23 76 0 0
86 0 0 2369573632 32896 1350988 0 0 132 0 428019 291865 0 24 76 0 0
90 0 0 2369573632 32896 1350988 0 0 502 0 422498 278672 0 23 76 0 0
perf lock record -C0-7 sleep 1; perf lock contention
contended total wait max wait avg wait type caller
2524 116.15 ms 486.61 us 46.02 us spinlock __dev_queue_xmit+0x55b
5821 107.18 ms 371.67 us 18.41 us spinlock dev_hard_start_xmit+0xcd
2377 9.73 ms 35.86 us 4.09 us spinlock ___slab_alloc+0x4e0
923 5.74 ms 20.91 us 6.22 us spinlock ___slab_alloc+0x5c9
121 3.42 ms 193.05 us 28.24 us spinlock net_tx_action+0xf8
6 564.33 us 167.60 us 94.05 us spinlock do_softirq+0x5b
If netperf threads are pinned (~54 Mpps)
perf lock record -C0-7 sleep 1; perf lock contention
32907 316.98 ms 195.98 us 9.63 us spinlock dev_hard_start_xmit+0xcd
4507 61.83 ms 212.73 us 13.72 us spinlock __dev_queue_xmit+0x554
2781 23.53 ms 40.03 us 8.46 us spinlock ___slab_alloc+0x5c9
3554 18.94 ms 34.69 us 5.33 us spinlock ___slab_alloc+0x4e0
233 9.09 ms 215.70 us 38.99 us spinlock do_softirq+0x5b
153 930.66 us 48.67 us 6.08 us spinlock net_tx_action+0xfd
84 331.10 us 14.22 us 3.94 us spinlock ___slab_alloc+0x5c9
140 323.71 us 9.94 us 2.31 us spinlock ___slab_alloc+0x4e0
@__dev_queue_xmit_ns:
[128, 256) 1539830 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[256, 512) 2299558 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[512, 1K) 483936 |@@@@@@@@@@ |
[1K, 2K) 265345 |@@@@@@ |
[2K, 4K) 145463 |@@@ |
[4K, 8K) 54571 |@ |
[8K, 16K) 10270 | |
[16K, 32K) 9385 | |
[32K, 64K) 7749 | |
[64K, 128K) 26799 | |
[128K, 256K) 2665 | |
[256K, 512K) 665 | |
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Tested-by: Jamal Hadi Salim <jhs@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20251014171907.3554413-7-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This is a followup of commit 726e9e8b94 ("tcp: refine
skb->ooo_okay setting") and of prior commit in this series
("net: control skb->ooo_okay from skb_set_owner_w()")
skb->ooo_okay might never be set for bulk flows that always
have at least one skb in a qdisc queue of NIC queue,
especially if TX completion is delayed because of a stressed cpu.
The so-called "strange attractors" has caused many performance
issues (see for instance 9b462d02d6 ("tcp: TCP Small Queues
and strange attractors")), we need to do better.
We have tried very hard to avoid reorders because TCP was
not dealing with them nicely a decade ago.
Use the new net.core.txq_reselection_ms sysctl to let
flows follow XPS and select a more efficient queue.
After this patch, we no longer have to make sure threads
are pinned to cpus, they now can be migrated without
adding too much spinlock/qdisc/TX completion pressure anymore.
TX completion part was problematic, because it added false sharing
on various socket fields, but also added false sharing and spinlock
contention in mm layers. Calling skb_orphan() from ndo_start_xmit()
is not an option unfortunately.
Note for later:
1) move sk->sk_tx_queue_mapping closer
to sk_tx_queue_mapping_jiffies for better cache locality.
2) Study if 9b462d02d6 ("tcp: TCP Small Queues
and strange attractors") could be revised.
Tested:
Used a host with 32 TX queues, shared by groups of 8 cores.
XPS setup :
echo ff >/sys/class/net/eth1/queue/tx-0/xps_cpus
echo ff00 >/sys/class/net/eth1/queue/tx-1/xps_cpus
echo ff0000 >/sys/class/net/eth1/queue/tx-2/xps_cpus
echo ff000000 >/sys/class/net/eth1/queue/tx-3/xps_cpus
echo ff,00000000 >/sys/class/net/eth1/queue/tx-4/xps_cpus
echo ff00,00000000 >/sys/class/net/eth1/queue/tx-5/xps_cpus
echo ff0000,00000000 >/sys/class/net/eth1/queue/tx-6/xps_cpus
echo ff000000,00000000 >/sys/class/net/eth1/queue/tx-7/xps_cpus
...
Launched a tcp_stream with 15 threads and 1000 flows, initially affined to core 0-15
taskset -c 0-15 tcp_stream -T15 -F1000 -l1000 -c -H target_host
Checked that only queues 0 and 1 are used as instructed by XPS :
tc -s qdisc show dev eth1|grep backlog|grep -v "backlog 0b 0p"
backlog 123489410b 1890p
backlog 69809026b 1064p
backlog 52401054b 805p
Then force each thread to run on cpu 1,9,17,25,33,41,49,57,65,73,81,89,97,105,113,121
C=1;PID=`pidof tcp_stream`;for P in `ls /proc/$PID/task`; do taskset -pc $C $P; C=$(($C + 8));done
Set txq_reselection_ms to 1000
echo 1000 > /proc/sys/net/core/txq_reselection_ms
Check that the flows have migrated nicely:
tc -s qdisc show dev eth1|grep backlog|grep -v "backlog 0b 0p"
backlog 130508314b 1916p
backlog 8584380b 126p
backlog 8584380b 126p
backlog 8379990b 123p
backlog 8584380b 126p
backlog 8487484b 125p
backlog 8584380b 126p
backlog 8448120b 124p
backlog 8584380b 126p
backlog 8720640b 128p
backlog 8856900b 130p
backlog 8584380b 126p
backlog 8652510b 127p
backlog 8448120b 124p
backlog 8516250b 125p
backlog 7834950b 115p
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20251013152234.842065-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since blamed commit, unregister_netdevice_many_notify() takes the netdev
mutex if the device needs it.
If the device list is too long, this will lock more device mutexes than
lockdep can handle:
unshare -n \
bash -c 'for i in $(seq 1 100);do ip link add foo$i type dummy;done'
BUG: MAX_LOCK_DEPTH too low!
turning off the locking correctness validator.
depth: 48 max: 48!
48 locks held by kworker/u16:1/69:
#0: ..148 ((wq_completion)netns){+.+.}-{0:0}, at: process_one_work
#1: ..d40 (net_cleanup_work){+.+.}-{0:0}, at: process_one_work
#2: ..bd0 (pernet_ops_rwsem){++++}-{4:4}, at: cleanup_net
#3: ..aa8 (rtnl_mutex){+.+.}-{4:4}, at: default_device_exit_batch
#4: ..cb0 (&dev_instance_lock_key#3){+.+.}-{4:4}, at: unregister_netdevice_many_notify
[..]
Add a helper to close and then unlock a list of net_devices.
Devices that are not up have to be skipped - netif_close_many always
removes them from the list without any other actions taken, so they'd
remain in locked state.
Close devices whenever we've used up half of the tracking slots or we
processed entire list without hitting the limit.
Fixes: 7e4d784f58 ("net: hold netdev instance lock during rtnetlink operations")
Signed-off-by: Florian Westphal <fw@strlen.de>
Link: https://patch.msgid.link/20251013185052.14021-1-fw@strlen.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Cross-merge networking fixes after downstream PR (net-6.17-rc7).
No conflicts.
Adjacent changes:
drivers/net/ethernet/mellanox/mlx5/core/en/fs.h
9536fbe10c ("net/mlx5e: Add PSP steering in local NIC RX")
7601a0a462 ("net/mlx5e: Add a miss level for ipsec crypto offload")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Provide a callback to validate skb's originating from tcp timewait
socks before passing to the device layer. Full socks have a
sk_validate_xmit_skb member for checking that a device is capable of
performing offloads required for transmitting an skb. With psp, tcp
timewait socks will inherit the crypto state from their corresponding
full socks. Any ACKs or RSTs that originate from a tcp timewait sock
carrying psp state should be psp encapsulated.
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Daniel Zahka <daniel.zahka@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250917000954.859376-8-daniel.zahka@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Move definition of sk_validate_xmit_skb() from net/core/sock.c to
net/core/dev.c.
This change is in preparation of the next patch, where
sk_validate_xmit_skb() will need to cast sk to a tcp_timewait_sock *,
and access member fields. Including linux/tcp.h from linux/sock.h
creates a circular dependency, and dev.c is the only current call site
of this function.
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Daniel Zahka <daniel.zahka@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250917000954.859376-7-daniel.zahka@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Hosts under DOS attack can suffer from false sharing
in enqueue_to_backlog() : atomic_inc(&sd->dropped).
This is because sd->dropped can be touched from many cpus,
possibly residing on different NUMA nodes.
Generalize the sk_drop_counters infrastucture
added in commit c51613fa27 ("net: add sk->sk_drop_counters")
and use it to replace softnet_data.dropped
with NUMA friendly softnet_data.drop_counters.
This adds 64 bytes per cpu, maybe more in the future
if we increase the number of counters (currently 2)
per 'struct numa_drop_counters'.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250909121942.1202585-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
napi_stop_kthread waits for the NAPI_STATE_SCHED_THREADED to be unset
before stopping the kthread. But it uses test_bit with the
NAPIF_STATE_SCHED_THREADED and that might stop the kthread early before
the flag is unset.
Use the NAPI_* variant of the NAPI state bits in test_bit instead.
Tested:
./tools/testing/selftests/net/nl_netdev.py
TAP version 13
1..7
ok 1 nl_netdev.empty_check
ok 2 nl_netdev.lo_check
ok 3 nl_netdev.page_pool_check
ok 4 nl_netdev.napi_list_check
ok 5 nl_netdev.dev_set_threaded
ok 6 nl_netdev.napi_set_threaded
ok 7 nl_netdev.nsim_rxq_reset_down
# Totals: pass:7 fail:0 xfail:0 xpass:0 skip:0 error:0
./tools/testing/selftests/drivers/net/napi_threaded.py
TAP version 13
1..2
ok 1 napi_threaded.change_num_queues
ok 2 napi_threaded.enable_dev_threaded_disable_napi_threaded
# Totals: pass:2 fail:0 xfail:0 xpass:0 skip:0 error:0
Fixes: 689883de94 ("net: stop napi kthreads when THREADED napi is disabled")
Signed-off-by: Samiullah Khawaja <skhawaja@google.com>
Link: https://patch.msgid.link/20250910203716.1016546-1-skhawaja@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch fixes an issue where two different flows on the same RXq
produce the same hash resulting in continuous flow overwrites.
Flow #1: A packet for Flow #1 comes in, kernel calls the steering
function. The driver gives back a filter id. The kernel saves
this filter id in the selected slot. Later, the driver's
service task checks if any filters have expired and then
installs the rule for Flow #1.
Flow #2: A packet for Flow #2 comes in. It goes through the same steps.
But this time, the chosen slot is being used by Flow #1. The
driver gives a new filter id and the kernel saves it in the
same slot. When the driver's service task runs, it runs through
all the flows, checks if Flow #1 should be expired, the kernel
returns True as the slot has a different filter id, and then
the driver installs the rule for Flow #2.
Flow #1: Another packet for Flow #1 comes in. The same thing repeats.
The slot is overwritten with a new filter id for Flow #1.
This causes a repeated cycle of flow programming for missed packets,
wasting CPU cycles while not improving performance. This problem happens
at higher rates when the RPS table is small, but tests show it still
happens even with 12,000 connections and an RPS size of 16K per queue
(global table size = 144x16K = 64K).
This patch prevents overwriting an rps_dev_flow entry if it is active.
The intention is that it is better to do aRFS for the first flow instead
of hurting all flows on the same hash. Without this, two (or more) flows
on one RX queue with the same hash can keep overwriting each other. This
causes the driver to reprogram the flow repeatedly.
Changes:
1. Add a new 'hash' field to struct rps_dev_flow.
2. Add rps_flow_is_active(): a helper function to check if a flow is
active or not, extracted from rps_may_expire_flow(). It is further
simplified as per reviewer feedback.
3. In set_rps_cpu():
- Avoid overwriting by programming a new filter if:
- The slot is not in use, or
- The slot is in use but the flow is not active, or
- The slot has an active flow with the same hash, but target CPU
differs.
- Save the hash in the rps_dev_flow entry.
4. rps_may_expire_flow(): Use earlier extracted rps_flow_is_active().
Testing & results:
- Driver: ice (E810 NIC), Kernel: net-next
- #CPUs = #RXq = 144 (1:1)
- Number of flows: 12K
- Eight RPS settings from 256 to 32768. Though RPS=256 is not ideal,
it is still sufficient to cover 12K flows (256*144 rx-queues = 64K
global table slots)
- Global Table Size = 144 * RPS (effectively equal to 256 * RPS)
- Each RPS test duration = 8 mins (org code) + 8 mins (new code).
- Metrics captured on client
Legend for following tables:
Steer-C: #times ndo_rx_flow_steer() was Called by set_rps_cpu()
Steer-L: #times ice_arfs_flow_steer() Looped over aRFS entries
Add: #times driver actually programmed aRFS (ice_arfs_build_entry())
Del: #times driver deleted the flow (ice_arfs_del_flow_rules())
Units: K = 1,000 times, M = 1 million times
|-------|---------|------| Org Code |---------|---------|
| RPS | Latency | CPU | Add | Del | Steer-C | Steer-L |
|-------|---------|------|--------|--------|---------|---------|
| 256 | 227.0 | 93.2 | 1.6M | 1.6M | 121.7M | 267.6M |
| 512 | 225.9 | 94.1 | 11.5M | 11.2M | 65.7M | 199.6M |
| 1024 | 223.5 | 95.6 | 16.5M | 16.5M | 27.1M | 187.3M |
| 2048 | 222.2 | 96.3 | 10.5M | 10.5M | 12.5M | 115.2M |
| 4096 | 223.9 | 94.1 | 5.5M | 5.5M | 7.2M | 65.9M |
| 8192 | 224.7 | 92.5 | 2.7M | 2.7M | 3.0M | 29.9M |
| 16384 | 223.5 | 92.5 | 1.3M | 1.3M | 1.4M | 13.9M |
| 32768 | 219.6 | 93.2 | 838.1K | 838.1K | 965.1K | 8.9M |
|-------|---------|------| New Code |---------|---------|
| 256 | 201.5 | 99.1 | 13.4K | 5.0K | 13.7K | 75.2K |
| 512 | 202.5 | 98.2 | 11.2K | 5.9K | 11.2K | 55.5K |
| 1024 | 207.3 | 93.9 | 11.5K | 9.7K | 11.5K | 59.6K |
| 2048 | 207.5 | 96.7 | 11.8K | 11.1K | 15.5K | 79.3K |
| 4096 | 206.9 | 96.6 | 11.8K | 11.7K | 11.8K | 63.2K |
| 8192 | 205.8 | 96.7 | 11.9K | 11.8K | 11.9K | 63.9K |
| 16384 | 200.9 | 98.2 | 11.9K | 11.9K | 11.9K | 64.2K |
| 32768 | 202.5 | 98.0 | 11.9K | 11.9K | 11.9K | 64.2K |
|-------|---------|------|--------|--------|---------|---------|
Some observations:
1. Overall Latency improved: (1790.19-1634.94)/1790.19*100 = 8.67%
2. Overall CPU increased: (777.32-751.49)/751.45*100 = 3.44%
3. Flow Management (add/delete) remained almost constant at ~11K
compared to values in millions.
Signed-off-by: Krishna Kumar <krikku@gmail.com>
Link: https://patch.msgid.link/20250825031005.3674864-2-krikku@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
