With CONFIG_MITIGATION_RETPOLINE=y dst_mtu() is a bit fat,
because it is generic.
Indeed, clang does not always inline it.
Add dst4_mtu() and dst6_mtu() helpers for callers that
expect either ipv4_mtu() or ip6_mtu() to be called.
These helpers are always inlined.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260130210303.3888261-6-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In ip_output() skb->dev is updated from the skb_dst(skb)->dev
this can become invalid when the interface is unregistered and freed,
Introduced new skb_dst_dev_rcu() function to be used instead of
skb_dst_dev() within rcu_locks in ip_output.This will ensure that
all the skb's associated with the dev being deregistered will
be transnmitted out first, before freeing the dev.
Given that ip_output() is called within an rcu_read_lock()
critical section or from a bottom-half context, it is safe to introduce
an RCU read-side critical section within it.
Multiple panic call stacks were observed when UL traffic was run
in concurrency with device deregistration from different functions,
pasting one sample for reference.
[496733.627565][T13385] Call trace:
[496733.627570][T13385] bpf_prog_ce7c9180c3b128ea_cgroupskb_egres+0x24c/0x7f0
[496733.627581][T13385] __cgroup_bpf_run_filter_skb+0x128/0x498
[496733.627595][T13385] ip_finish_output+0xa4/0xf4
[496733.627605][T13385] ip_output+0x100/0x1a0
[496733.627613][T13385] ip_send_skb+0x68/0x100
[496733.627618][T13385] udp_send_skb+0x1c4/0x384
[496733.627625][T13385] udp_sendmsg+0x7b0/0x898
[496733.627631][T13385] inet_sendmsg+0x5c/0x7c
[496733.627639][T13385] __sys_sendto+0x174/0x1e4
[496733.627647][T13385] __arm64_sys_sendto+0x28/0x3c
[496733.627653][T13385] invoke_syscall+0x58/0x11c
[496733.627662][T13385] el0_svc_common+0x88/0xf4
[496733.627669][T13385] do_el0_svc+0x2c/0xb0
[496733.627676][T13385] el0_svc+0x2c/0xa4
[496733.627683][T13385] el0t_64_sync_handler+0x68/0xb4
[496733.627689][T13385] el0t_64_sync+0x1a4/0x1a8
Changes in v3:
- Replaced WARN_ON() with WARN_ON_ONCE(), as suggested by Willem de Bruijn.
- Dropped legacy lines mistakenly pulled in from an outdated branch.
Changes in v2:
- Addressed review comments from Eric Dumazet
- Used READ_ONCE() to prevent potential load/store tearing
- Added skb_dst_dev_rcu() and used along with rcu_read_lock() in ip_output
Signed-off-by: Sharath Chandra Vurukala <quic_sharathv@quicinc.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250730105118.GA26100@hu-sharathv-hyd.qualcomm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
dst->dev is read locklessly in many contexts,
and written in dst_dev_put().
Fixing all the races is going to need many changes.
We probably will have to add full RCU protection.
Add three helpers to ease this painful process.
static inline struct net_device *dst_dev(const struct dst_entry *dst)
{
return READ_ONCE(dst->dev);
}
static inline struct net_device *skb_dst_dev(const struct sk_buff *skb)
{
return dst_dev(skb_dst(skb));
}
static inline struct net *skb_dst_dev_net(const struct sk_buff *skb)
{
return dev_net(skb_dst_dev(skb));
}
static inline struct net *skb_dst_dev_net_rcu(const struct sk_buff *skb)
{
return dev_net_rcu(skb_dst_dev(skb));
}
Fixes: 4a6ce2b6f2 ("net: introduce a new function dst_dev_put()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250630121934.3399505-7-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
dst_dev_put() can overwrite dst->output while other
cpus might read this field (for instance from dst_output())
Add READ_ONCE()/WRITE_ONCE() annotations to suppress
potential issues.
We will likely need RCU protection in the future.
Fixes: 4a6ce2b6f2 ("net: introduce a new function dst_dev_put()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250630121934.3399505-6-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add static inline dst_dev_overhead() function to include/net/dst.h. This
helper function is used by ioam6_iptunnel, rpl_iptunnel and
seg6_iptunnel to get the dev's overhead based on a cache entry
(dst_entry). If the cache is empty, the default and generic value
skb->mac_len is returned. Otherwise, LL_RESERVED_SPACE() over dst's dev
is returned.
Signed-off-by: Justin Iurman <justin.iurman@uliege.be>
Cc: Alexander Lobakin <aleksander.lobakin@intel.com>
Cc: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Since commit 1202cdd66531("Remove DECnet support from kernel") has been
merged, all callers pass in the initial_ref value of 1 when they call
dst_alloc(). Therefore, remove initial_ref when the dst_alloc() is
declared and replace initial_ref with 1 in dst_alloc().
Also when all callers call dst_init(), the value of initial_ref is 1.
Therefore, remove the input parameter initial_ref of the dst_init() and
replace initial_ref with the value 1 in dst_init.
Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Link: https://lore.kernel.org/r/20230911125045.346390-1-shaozhengchao@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Under high contention dst_entry::__refcnt becomes a significant bottleneck.
atomic_inc_not_zero() is implemented with a cmpxchg() loop, which goes into
high retry rates on contention.
Switch the reference count to rcuref_t which results in a significant
performance gain. Rename the reference count member to __rcuref to reflect
the change.
The gain depends on the micro-architecture and the number of concurrent
operations and has been measured in the range of +25% to +130% with a
localhost memtier/memcached benchmark which amplifies the problem
massively.
Running the memtier/memcached benchmark over a real (1Gb) network
connection the conversion on top of the false sharing fix for struct
dst_entry::__refcnt results in a total gain in the 2%-5% range over the
upstream baseline.
Reported-by: Wangyang Guo <wangyang.guo@intel.com>
Reported-by: Arjan Van De Ven <arjan.van.de.ven@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20230307125538.989175656@linutronix.de
Link: https://lore.kernel.org/r/20230323102800.215027837@linutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
dst_entry::__refcnt is highly contended in scenarios where many connections
happen from and to the same IP. The reference count is an atomic_t, so the
reference count operations have to take the cache-line exclusive.
Aside of the unavoidable reference count contention there is another
significant problem which is caused by that: False sharing.
perf top identified two affected read accesses. dst_entry::lwtstate and
rtable::rt_genid.
dst_entry:__refcnt is located at offset 64 of dst_entry, which puts it into
a seperate cacheline vs. the read mostly members located at the beginning
of the struct.
That prevents false sharing vs. the struct members in the first 64
bytes of the structure, but there is also
dst_entry::lwtstate
which is located after the reference count and in the same cache line. This
member is read after a reference count has been acquired.
struct rtable embeds a struct dst_entry at offset 0. struct dst_entry has a
size of 112 bytes, which means that the struct members of rtable which
follow the dst member share the same cache line as dst_entry::__refcnt.
Especially
rtable::rt_genid
is also read by the contexts which have a reference count acquired
already.
When dst_entry:__refcnt is incremented or decremented via an atomic
operation these read accesses stall. This was found when analysing the
memtier benchmark in 1:100 mode, which amplifies the problem extremly.
Move the rt[6i]_uncached[_list] members out of struct rtable and struct
rt6_info into struct dst_entry to provide padding and move the lwtstate
member after that so it ends up in the same cache line.
The resulting improvement depends on the micro-architecture and the number
of CPUs. It ranges from +20% to +120% with a localhost memtier/memcached
benchmark.
[ tglx: Rearrange struct ]
Signed-off-by: Wangyang Guo <wangyang.guo@intel.com>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20230323102800.042297517@linutronix.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Long standing KCSAN issues are caused by data-race around
some dev->stats changes.
Most performance critical paths already use per-cpu
variables, or per-queue ones.
It is reasonable (and more correct) to use atomic operations
for the slow paths.
This patch adds an union for each field of net_device_stats,
so that we can convert paths that are not yet protected
by a spinlock or a mutex.
netdev_stats_to_stats64() no longer has an #if BITS_PER_LONG==64
Note that the memcpy() we were using on 64bit arches
had no provision to avoid load-tearing,
while atomic_long_read() is providing the needed protection
at no cost.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
All uses of dst_hold_and_use() have
been removed since commit 1202cdd665 ("Remove DECnet support
from kernel"), so remove it.
Signed-off-by: Gaosheng Cui <cuigaosheng1@huawei.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We want to track all dev_hold()/dev_put() to ease leak hunting.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Similar to the previous patch, but covering the dst field:
the slow_gro flag is additionally set when a dst is attached
to the skb
RFC -> v1:
- use the existing flag instead of adding a new one
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move generic blackhole dst ops to the core and use them from both
ipv4_dst_blackhole_ops and ip6_dst_blackhole_ops where possible. No
functional change otherwise. We need these also in other locations
and having to define them over and over again is not great.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch avoids the indirect call for the common case:
ip6_dst_check and ipv4_dst_check
Signed-off-by: Brian Vazquez <brianvv@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch avoids the indirect call for the common case:
ip6_mtu and ipv4_mtu
Signed-off-by: Brian Vazquez <brianvv@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch avoids the indirect call for the common case:
ip6_output and ip_output
Signed-off-by: Brian Vazquez <brianvv@google.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Eric Dumazet
Sharath Chandra Vurukala
Jakub Kicinski
Justin Iurman
Simon Horman
Sebastian Andrzej Siewior
Yan Zhai
Zhengchao Shao
Thomas Gleixner
Wangyang Guo
Gaosheng Cui
Paolo Abeni
Daniel Borkmann
Brian Vazquez