qdisc_get_rtab() and qdisc_put_rtab() mutate the process-global singly
linked list qdisc_rtab_list and a plain non-atomic 'int refcnt' with no
lock. This was only safe because every caller historically held the RTNL
mutex, which serialized all rate-table lookups, inserts and frees.
That invariant no longer holds. cls_flower sets
TCF_PROTO_OPS_DOIT_UNLOCKED, so tc_new_tfilter() keeps rtnl_held == false
for it and sets TCA_ACT_FLAGS_NO_RTNL. That flag propagates through
tcf_exts_validate_ex() -> tcf_action_init() -> tcf_action_init_1() ->
tcf_police_init(), which calls qdisc_get_rtab()/qdisc_put_rtab() with the
RTNL mutex NOT held. Two RTM_NEWTFILTER requests on different CPUs, each
adding a flower filter with a police action carrying the same rate, then
race on qdisc_rtab_list and on the non-atomic refcnt, leading to a
use-after-free / double-free of the kmalloc-2k struct qdisc_rate_table.
qdisc_rtab_list is a single global (not per-netns), so the corrupted
object is shared system-wide.
BUG: KASAN: slab-use-after-free in qdisc_put_rtab+0x12f/0x160
qdisc_put_rtab+0x12f/0x160
tcf_police_init+0xda9/0x1590
tcf_action_init_1+0x460/0x6b0
tcf_action_init+0x439/0xa40
tcf_exts_validate_ex+0x42d/0x550
fl_change+0xddd/0x7da0
tc_new_tfilter+0xaa7/0x2420
rtnetlink_rcv_msg+0x95e/0xe90
which belongs to the cache kmalloc-2k of size 2048
Protect qdisc_rtab_list and the refcount with a dedicated spinlock. The
(sleeping, GFP_KERNEL) allocation in qdisc_get_rtab() is performed before
taking the lock; if a concurrent inserter added an identical table in the
meantime the freshly allocated one is freed under the lock, so no
duplicate is leaked. qdisc_put_rtab() now decrements the refcount and
unlinks under the same lock.
Fixes: 470502de5b ("net: sched: unlock rules update API")
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Aldo Ariel Panzardo <qwe.aldo@gmail.com>
Cc: stable@vger.kernel.org
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260715114114.446841-1-qwe.aldo@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When a TC filter attached to a qdisc filter chain returns
TC_ACT_REDIRECT (ex: via an eBPF program calling bpf_redirect() or an
act_bpf action), the redirect was silently lost i.e no qdisc classify
function handled TC_ACT_REDIRECT, so the packet fell through the
switch and was enqueued normally instead of being redirected.
This has been broken since bpf_redirect() was introduced for TC in
commit 27b29f6305 ("bpf: add bpf_redirect() helper"). We got lucky
for a long time because bpf_net_context was a per-CPU variable that
was always available.
commit 401cb7dae8 ("net: Reference bpf_redirect_info via task_struct
on PREEMPT_RT.") turned bpf_net_context into a task_struct member that
is only set up by explicit callers. Without a caller setting it up,
bpf_redirect() itself crashes with a NULL pointer dereference in
bpf_net_ctx_get_ri(). However, even with bpf_net_context available,
TC_ACT_REDIRECT from qdisc filter chains cannot be honored without
adding skb_do_redirect() calls to every qdisc classify function, which
would require changes across net/sched/. Isolate it to ebpf core where
it belongs.
Instead, add a tcf_classify_qdisc() inline helper in pkt_cls.h, as a
wrapper around tcf_classify() for use by qdisc classify functions and
tcf_qevent_handle(). When the classify verdict is TC_ACT_REDIRECT,
the wrapper converts it to TC_ACT_SHOT, dropping the packet rather
than letting it continue silently. Dropping is preferred over
letting the packet through because the user immediately sees packet
loss. Silently passing the packet through would hide the problem and
leave the user wondering why their redirect is not working.
The clsact fast path, tc_run() continues to call tcf_classify() directly
and is unaffected: TC_ACT_REDIRECT is returned as-is and handled by
sch_handle_egress/ingress() calling skb_do_redirect() as before.
Fixes: 27b29f6305 ("bpf: add bpf_redirect() helper")
Fixes: 401cb7dae8 ("net: Reference bpf_redirect_info via task_struct on PREEMPT_RT.")
Tested-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://patch.msgid.link/20260706185609.330006-3-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The bpf_redirect*() helpers and skb_do_redirect() obtain the per-task
bpf_redirect_info via bpf_net_ctx_get_ri(), which dereferences the
current->bpf_net_context unconditionally. That context is established
on the paths that run tc BPF such as sch_handle_{ingress,egress}(),
*except* for the case where {cls,act}_bpf was attached to a proper
qdisc. A program running from there reaches the NULL deref in two ways:
* It calls bpf_redirect() directly, which dereferences the context at
the top of the helper:
tc qdisc add dev eth0 root handle 1: red limit 1MB min 10KB max 20KB \
avpkt 1000 burst 100 qevent early_drop block 10
tc filter add block 10 pref 1 bpf obj redirect.o
* It simply returns TC_ACT_REDIRECT without helper call: tcf_qevent_handle()
then dispatches to skb_do_redirect(), which dereferences the context
Rather than extending bpf_net_context management into the qdisc path,
make the redirect helpers refuse to operate when no context exists, and
have tcf_qevent_handle() drop a TC_ACT_REDIRECT verdict instead of
calling skb_do_redirect(). Previous behaviour was a crash, so nothing
regresses by not supporting it.
Fixes: 401cb7dae8 ("net: Reference bpf_redirect_info via task_struct on PREEMPT_RT.")
Fixes: 3625750f05 ("net: sched: Introduce helpers for qevent blocks")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Link: https://patch.msgid.link/20260706185609.330006-2-daniel@iogearbox.net
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
CAKE accepts signed overhead values and stores them in an s16, but the
adjusted packet length calculation uses unsigned arithmetic. A negative
effective length can therefore wrap to a large value.
Such configurations make rate accounting depend on integer wraparound
rather than on the packet size userspace intended to model. A static
netlink lower bound is not enough because packets reaching CAKE can be
smaller than any reasonable manual-overhead allowance.
Fold the signed overhead adjustment into the existing datapath MPU clamp
so negative adjusted lengths are clamped before link-layer framing
adjustments.
Fixes: a729b7f0bd ("sch_cake: Add overhead compensation support to the rate shaper")
Assisted-by: Codex:gpt-5.5-cyber-preview
Signed-off-by: Samuel Moelius <sam.moelius@trailofbits.com>
Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>
Link: https://patch.msgid.link/20260702000758.297407.e5c888d9d99d.cake-overhead-underflow@trailofbits.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
There is a TOCTOU race condition in flower lockless approach between sizing
a flow_rule buffer and filling it.
zdi-disclosures@trendmicro.com reports:
The cls_flower classifier operates with TCF_PROTO_OPS_DOIT_UNLOCKED
(fl_change runs without RTNL), while RTM_NEWACTION holds RTNL, so the
independent locking domains make the race reachable in practice. KASAN
confirms:
BUG: KASAN: slab-out-of-bounds in tcf_pedit_offload_act_setup+0x81b/0x930
Write of size 4 at addr ffff888001f27520 by task poc-toctou/312
The buggy address is located 0 bytes to the right of
allocated 288-byte region [ffff888001f27400, ffff888001f27520)
(cache kmalloc-512)
Note: The result is a heap OOB write attacker-controlled content into the
adjacent slab object (requires CAP_NET_ADMIN).
The fix introduces reading tcfp_nkeys under act->tcfa_lock in all places
using a new tcf_pedit_nkeys_locked() which replaces the old tcf_pedit_nkeys().
Additionally we close the remaining TOCTOU window between the sizing read and
the fill reads by more careful accounting.
Rather than silently truncating the key count, which leads to incorrect
action semantics offloaded to hardware and secondary OOB writes if
the remaining capacity is zero or consumed by prior actions, we enforce
remaining capacity checks and return -ENOSPC if the required space exceeds
the remaining capacity.
Fixes: 71d0ed7079 ("net/act_pedit: Support using offset relative to the conventional network headers")
Reported-by: zdi-disclosures@trendmicro.com
Tested-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260701161912.125355-1-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This is a followup based on sashiko comments [1] on commit e5b811fe79
("net/sched: sch_teql: Introduce slaves_lock to avoid race condition and UAF")
Use plain rcu_read_lock()/spin_lock() in teql_master_xmit() instead of the
_bh variants, since ndo_start_xmit is already invoked with BH disabled
by the core stack and the _bh primitives can warn in_hardirq() when xmit
is reached through netpoll or a softirq xmit path with hard IRQs disabled.
Moves rcu_read_lock() after restart: label + adds rcu_read_unlock() before
goto restart (fixes the unbounded RCU hold across retries)
[1] https://sashiko.dev/#/patchset/20260628111229.669751-1-jhs%40mojatatu.com
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Fixes: e5b811fe79 ("net/sched: sch_teql: Introduce slaves_lock to avoid race condition and UAF")
Link: https://patch.msgid.link/20260630150922.238714-1-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
HHF reset does not clear the classifier state used to identify heavy
hitters. Packets after reset can therefore be scheduled using flow
history from before the reset.
The reset operation should return the qdisc to an empty state.
Clear the heavy-hitter classifier tables when HHF is reset.
Fixes: 10239edf86 ("net-qdisc-hhf: Heavy-Hitter Filter (HHF) qdisc")
Assisted-by: Codex:gpt-5.5-cyber-preview
Signed-off-by: Samuel Moelius <sam.moelius@trailofbits.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
DualPI2 leaves previous classification state attached to an skb when
filter classification returns no match. The enqueue path can then act
on stale state from an earlier classification attempt.
A filter miss should fall back to the default class without reusing old
per-packet classification data.
Initialize the classification result to CLASSIC before running the
classifier. Explicit L4S, priority, and successful filter
classification can still override that default.
Fixes: 8f9516daed ("sched: Add enqueue/dequeue of dualpi2 qdisc")
Assisted-by: Codex:gpt-5.5-cyber-preview
Signed-off-by: Samuel Moelius <sam.moelius@trailofbits.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tcf_bpf_act() can run from the tc egress path, which holds only
rcu_read_lock_bh(), but reads prog->filter with rcu_dereference() and
trips lockdep:
WARNING: suspicious RCU usage
net/sched/act_bpf.c:47 suspicious rcu_dereference_check() usage!
1 lock held by syz.2.1588/12756:
#0: (rcu_read_lock_bh){....}-{1:3}, at: __dev_queue_xmit net/core/dev.c:4792
tcf_bpf_act+0x6ae/0x940 net/sched/act_bpf.c:47
tcf_classify+0x6e4/0x1080 net/sched/cls_api.c:1860
sch_handle_egress net/core/dev.c:4545 [inline]
__dev_queue_xmit+0x2185/0x2c00 net/core/dev.c:4808
packet_sendmsg+0x3dfa/0x5120 net/packet/af_packet.c:3114
The other tc actions and cls_bpf already use rcu_dereference_bh() here.
Do the same.
Fixes: 1f211a1b92 ("net, sched: add clsact qdisc")
Signed-off-by: Sechang Lim <rhkrqnwk98@gmail.com>
Reviewed-by: Amery Hung <ameryhung@gmail.com>
Link: https://patch.msgid.link/20260629154112.1164986-1-rhkrqnwk98@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The teql master->slaves singly linked list is not protected against
multiple writes. It can be mod'ed concurently from teql_master_xmit(),
teql_dequeue(), teql_init() and teql_destroy() without holding any list
lock or RCU protection.
zdi-disclosures@trendmicro.com has demonstrated that the qdisc is freed
after an RCU grace period, but teql_master_xmit() running on another
CPU can still hold a stale pointer into the list, resulting in a
slab-use-after-free:
BUG: KASAN: slab-use-after-free in teql_master_xmit+0xf0f/0x16b0
Read of size 8 at addr ffff888013fb0440 by task poc/332
Freed 512-byte region [ffff888013fb0400, ffff888013fb0600) (kmalloc-512)
The fix?
Add a per-master slaves_lock spinlock that serializes all mutations of
master->slaves and the NEXT_SLAVE() links in teql_destroy() and
teql_qdisc_init(). teql_master_xmit() also takes the same slaves_lock
around those updates.
Annotate master->slaves and the per-slave ->next pointer with __rcu and
use the appropriate RCU accessors everywhere they are touched:
rcu_assign_pointer() on the writer side (under slaves_lock),
rcu_dereference_protected() for the writer-side loads (also under
slaves_lock), rcu_dereference_bh() for the loads in teql_master_xmit() and
rtnl_dereference() for the loads in teql_master_open()/teql_master_mtu(),
which run under RTNL.
Pair this with rcu_read_lock_bh()/rcu_read_unlock_bh() around the list
traversal in teql_master_xmit(), so that readers either observe a fully
linked list or are deferred until the in-flight mutation completes. The two
early-return paths in teql_master_xmit() are updated to release the RCU-bh
read-side critical section before returning, since leaving it held would
disable BH on that CPU for good.
Fixes: 1da177e4c3 ("Linux-2.6.12-rc2")
Reported-by: zdi-disclosures@trendmicro.com
Tested-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260628111229.669751-1-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
multiq_dequeue() takes a packet from a band's child with a direct
->dequeue() call after multiq_peek() peeked it. When the child is
non-work-conserving the peek stashes the skb in the child's gso_skb, so
the direct dequeue returns a different skb and orphans the stash,
desyncing the child's qlen/backlog. With a qfq child reached through a
peeking parent (e.g. tbf) this re-enters the child on an emptied list and
dereferences NULL, panicking the kernel from softirq on ordinary egress.
Take the packet through qdisc_dequeue_peeked(), as sch_prio already does
and as sch_red and sch_sfb were just fixed to do. The helper is a no-op
when the child has no stash, so a work-conserving child is unaffected.
Fixes: 77be155cba ("pkt_sched: Add peek emulation for non-work-conserving qdiscs.")
Cc: stable@vger.kernel.org
Signed-off-by: Bryam Vargas <hexlabsecurity@proton.me>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260625-b4-disp-31bcb279-v1-2-85c40b83c529@proton.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When taprio's software path peeks a non-work-conserving child qdisc, the
child stashes the peeked skb in its gso_skb; taprio_dequeue_from_txq()
then takes the packet with a direct child ->dequeue() call, which ignores
that stash, orphans the peeked skb and desyncs the child's qlen/backlog.
With a qfq child this re-enters the child on an emptied list and
dereferences NULL, panicking the kernel from softirq on ordinary egress.
Take the packet through qdisc_dequeue_peeked(), as sch_red and sch_sfb
now do. The helper returns the child's stashed skb first and is a no-op
when there is none, so a work-conserving child is unaffected and the
gated path now consumes the skb whose length was charged to the budget.
Fixes: 5a781ccbd1 ("tc: Add support for configuring the taprio scheduler")
Cc: stable@vger.kernel.org
Cc: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Bryam Vargas <hexlabsecurity@proton.me>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260625-b4-disp-31bcb279-v1-1-85c40b83c529@proton.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tcf_classify() can return TC_ACT_CONSUMED while the skb is held by the
defragmentation engine (e.g. act_ct on out-of-order fragments). When
that happens the skb is no longer owned by the caller and must not be
touched again.
tcf_qevent_handle() did not handle TC_ACT_CONSUMED: it fell through the
switch and returned the skb to the caller as if classification had
passed. The only qdisc that wires up qevents today is RED, via three call sites
(qe_mark on RED_PROB_MARK/HARD_MARK, qe_early_drop on congestion_drop)
red_enqueue() was continuing to operate on an skb it no longer owns in this
case -- enqueueing it, dropping it, or updating statistics. Resulting in a UAF.
tc qdisc add dev eth0 root handle 1: red ... qevent early_drop block 10
tc filter add block 10 ... action ct
(with ct defrag enabled and traffic that produces out-of-order
fragments, e.g. a fragmented UDP stream)
Handle TC_ACT_CONSUMED in tcf_qevent_handle() the same way the ingress
and egress fast paths do: treat it as stolen and return NULL without
touching the skb. Unlike the TC_ACT_STOLEN case, the skb must not be
dropped/freed here, as it is no longer owned by us.
Fixes: 3f14b377d0 ("net/sched: act_ct: fix skb leak and crash on ooo frags")
Reported-by: Zero Day Initiative <zdi-disclosures@trendmicro.com>
Tested-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260620130749.226642-1-jhs@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Marek Szyprowski reported a deadlock during system resume when virtio_net
driver is used.
The deadlock occurs because netif_device_attach() is called while holding
dev->tx_global_lock (via netif_tx_lock_bh() in virtnet_restore_up()).
netif_device_attach() calls __netdev_watchdog_up(), which now also tries
to acquire dev->tx_global_lock to synchronize with dev_watchdog().
This recursive lock acquisition results in a deadlock.
Fix this by removing the tx_global_lock acquisition from netdev_watchdog_up().
The critical state (watchdog_timer and watchdog_ref_held) is already
protected by dev->watchdog_lock, which was introduced in the blamed commit.
Fixes: 8eed5519e4 ("net: watchdog: fix refcount tracking races")
Reported-by: Marek Szyprowski <m.szyprowski@samsung.com>
Closes: https://lore.kernel.org/netdev/a443376e-5187-4268-93b3-58047ef113a8@samsung.com/
Signed-off-by: Eric Dumazet <edumazet@google.com>
Tested-by: Marek Szyprowski <m.szyprowski@samsung.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20260622110108.69541-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When DualPI2 splits a GSO skb into N segments, it propagates N
additional packets to its parent before returning NET_XMIT_SUCCESS.
The parent then accounts for the original skb once more, leaving its
qlen one larger than the number of packets actually queued.
With QFQ as the parent, after all real packets are dequeued, QFQ still
has a non-zero qlen while its in-service aggregate has no active
classes. qfq_choose_next_agg() returns NULL and qfq_dequeue() passes
the result to qfq_peek_skb(), causing a NULL pointer dereference.
Follow the same pattern used by tbf_segment() and taprio: count only
successfully queued segments, propagate the difference between the
original skb and those segments, and return NET_XMIT_SUCCESS whenever
at least one segment was queued.
Fixes: 8f9516daed ("sched: Add enqueue/dequeue of dualpi2 qdisc")
Cc: stable@vger.kernel.org
Signed-off-by: Xingquan Liu <b1n@b1n.io>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Link: https://patch.msgid.link/20260619151447.223640-1-b1n@b1n.io
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tcf_ct_fill_params() calls nf_connlabels_get() (setting put_labels) when
TCA_CT_LABELS is present, but two later error sites use a bare return
instead of "goto err", skipping the err: nf_connlabels_put() cleanup.
They also precede the "p->put_labels = put_labels" assignment, so the
tcf_ct_params_free() fallback does not release the count either. Each
failed RTM_NEWACTION on these paths leaks one nf_connlabels reference:
net->ct.labels_used is incremented and never released. The action is
reachable with CAP_NET_ADMIN over the netns, i.e. from an unprivileged
user namespace on default-userns kernels.
Impact: an unprivileged user with CAP_NET_ADMIN over a network namespace
(e.g. via user namespaces) leaks one nf_connlabels reference per failed
RTM_NEWACTION on the two error paths; net->ct.labels_used is never
released.
The err: label is safe to reach from both sites: p->tmpl is still NULL
there (kzalloc'd, not yet assigned) and nf_ct_put(NULL) is a no-op, so
no inline release is needed.
Fixes: 70f06c115b ("sched: act_ct: switch to per-action label counting")
Signed-off-by: Michael Bommarito <michael.bommarito@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260617215708.1115818-1-michael.bommarito@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tcf_ct_handle_fragments() calls nf_ct_handle_fragments() without saving
and restoring skb->cb. The defrag helper clears IPCB/IP6CB, which aliases
the tc_skb_cb/qdisc_skb_cb control buffer. Fragmented traffic through
act_ct therefore loses qdisc metadata such as pkt_segs and can trigger
WARN_ON_ONCE() in qdisc_pkt_segs() when panic_on_warn is enabled.
Save and restore the full tc_skb_cb around nf_ct_handle_fragments(),
matching the pattern used by ovs_ct_handle_fragments().
Fixes: ec624fe740 ("net/sched: Extend qdisc control block with tc control block")
Cc: stable@vger.kernel.org
Reported-by: Yuan Tan <yuantan098@gmail.com>
Reported-by: Yifan Wu <yifanwucs@gmail.com>
Reported-by: Juefei Pu <tomapufckgml@gmail.com>
Reported-by: Xin Liu <bird@lzu.edu.cn>
Signed-off-by: Zihan Xi <xizh2024@lzu.edu.cn>
Signed-off-by: Ren Wei <n05ec@lzu.edu.cn>
Link: https://patch.msgid.link/510c51217fd7aaf29c6dc298bab8d643fe229b1c.1781358692.git.xizh2024@lzu.edu.cn
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Merge in late fixes in preparation for the net-next PR.
Conflicts:
net/tls/tls_sw.c
406e8a651a ("net: skmsg: preserve sg.copy across SG transforms")
79511603a6 ("tls: remove dead sockmap (psock) handling from the SW path")
drivers/net/ethernet/microsoft/mana/mana_en.c
f8fd56977e ("net: mana: guard TX wq object destroy with INVALID_MANA_HANDLE check")
d07efe5a6e ("net: mana: Use per-queue allocation for tx_qp to reduce allocation size")
https://lore.kernel.org/ajAPXu-C_PuTgV-a@sirena.org.uk
No adjacent changes.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Similar to commit add641e7de ("sched: act_csum: don't mangle TCP and
UDP GSO packets"), UDP tunnel GSO packets going through act_csum
shouldn't have their checksum calculated at this point, because it will
be done after segmentation. Setting the checksum in act_csum modifies
skb->ip_summed and prevents inner IP csum offload from kicking in,
resulting in a packet with a bad checksum.
Add UDP tunnel GSO packets to the exceptions, and also add UDP GSO
(SKB_GSO_UDP_L4), as the same logic as in the commit mentioned above
applies to UDP GSO too.
Signed-off-by: Alice Mikityanska <alice@isovalent.com>
Reviewed-by: Davide Caratti <dcaratti@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260611192955.604661-2-alice.kernel@fastmail.im
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When a qdisc is added by name, the kernel tries to autoload its module
via request_qdisc_module(), which calls:
request_module(NET_SCH_ALIAS_PREFIX "%s", name);
i.e. it asks modprobe to resolve the "net-sch-<kind>" alias (e.g.
"net-sch-dualpi2") rather than the module's file name. Since dualpi2
was shipped without this alias, the autoload fails:
tc qdisc add dev lo root handle 1: dualpi2
Error: Specified qdisc kind is unknown.
Fix this by adding the missing alias so the qdisc is autoloaded on demand
like the others.
Fixes: 320d031ad6 ("sched: Struct definition and parsing of dualpi2 qdisc")
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: Pedro Tammela <pctammela@mojatatu.com>
Link: https://patch.msgid.link/20260611205849.3287640-1-victor@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Blamed commit converted the untracked dev_hold()/dev_put() calls
in the watchdog code to use the tracked dev_hold_track()/dev_put_track()
(which were later renamed/interfaced to netdev_hold() and netdev_put()).
By introducing dev->watchdog_dev_tracker to store the
reference tracking information without adding synchronization
between netdev_watchdog_up() and dev_watchdog(), it enabled the
race condition where this pointer could be overwritten or freed
concurrently, leading to the list corruption crash syzbot reported:
list_del corruption, ffff888114a18c00->next is NULL
kernel BUG at lib/list_debug.c:52 !
Oops: invalid opcode: 0000 [#1] SMP KASAN PTI
CPU: 1 UID: 0 PID: 91 Comm: kworker/u8:5 Not tainted syzkaller #0 PREEMPT(lazy)
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 05/09/2026
Workqueue: events_unbound linkwatch_event
RIP: 0010:__list_del_entry_valid_or_report.cold+0x22/0x2a lib/list_debug.c:52
Call Trace:
<TASK>
__list_del_entry_valid include/linux/list.h:132 [inline]
__list_del_entry include/linux/list.h:246 [inline]
list_move_tail include/linux/list.h:341 [inline]
ref_tracker_free+0x1a7/0x6c0 lib/ref_tracker.c:329
netdev_tracker_free include/linux/netdevice.h:4491 [inline]
netdev_put include/linux/netdevice.h:4508 [inline]
netdev_put include/linux/netdevice.h:4504 [inline]
netdev_watchdog_down net/sched/sch_generic.c:600 [inline]
dev_deactivate_many+0x28c/0xfe0 net/sched/sch_generic.c:1363
dev_deactivate+0x109/0x1d0 net/sched/sch_generic.c:1397
linkwatch_do_dev net/core/link_watch.c:184 [inline]
linkwatch_do_dev+0xd3/0x120 net/core/link_watch.c:166
__linkwatch_run_queue+0x3a5/0x810 net/core/link_watch.c:240
linkwatch_event+0x8f/0xc0 net/core/link_watch.c:314
process_one_work+0xa0e/0x1980 kernel/workqueue.c:3314
process_scheduled_works kernel/workqueue.c:3397 [inline]
worker_thread+0x5ef/0xe50 kernel/workqueue.c:3478
kthread+0x370/0x450 kernel/kthread.c:436
ret_from_fork+0x69a/0xc80 arch/x86/kernel/process.c:158
ret_from_fork_asm+0x1a/0x30 arch/x86/entry/entry_64.S:245
This patch has three coordinated parts:
1) Add dev->watchdog_lock and dev->watchdog_ref_held to serialize watchdog operations.
2) Remove netdev_watchdog_up() call from netif_carrier_on():
This ensures netdev_watchdog_up() is only called from process/BH context
(via linkwatch workqueue dev_activate()), allowing us to use
spin_lock_bh() for synchronization.
3) Synchronize watchdog up and watchdog timer:
Protect netdev_watchdog_up() with tx_global_lock and watchdog_lock.
Only allocate a new tracker in netdev_watchdog_up() if one is
not already present.
In dev_watchdog(), ensure we don't release the tracker if the
timer was rescheduled either by dev_watchdog() itself or concurrently
by netdev_watchdog_up().
Fixes: f12bf6f3f9 ("net: watchdog: add net device refcount tracker")
Reported-by: syzbot+381d82bbf0253710b35d@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/netdev/6a26b751.c25708ab.1b19ef.0013.GAE@google.com/T/#u
Tested-by: syzbot+3479efbc2821cb2a79f2@syzkaller.appspotmail.com
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260611152737.2580480-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The extended IPv4 L4 header mode in act_pedit can select TCP or UDP
header fields without confirming that the IPv4 protocol field matches
the selected transport header.
That lets a rule written for TCP or UDP modify unrelated payload bytes
in a packet carrying a different protocol.
Verify that the IPv4 header is long enough, that the protocol matches
the selected TCP or UDP header, and that the packet is not a non-initial
fragment before applying TCP or UDP extended header edits.
Cc: stable+noautosel@kernel.org # in real rule sets the match confirms this before calling the action
Signed-off-by: Samuel Moelius <sam.moelius@trailofbits.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
update_vf() is called from two places for the same class during a single
dequeue when the class's child qdisc (e.g. codel/fq_codel) drops its last
packets while dequeuing:
1. The child calls qdisc_tree_reduce_backlog(), which, now that the child
is empty, invokes hfsc_qlen_notify() -> update_vf(cl, 0, 0) and turns
the class passive (cl_nactive is decremented up the hierarchy).
2. hfsc_dequeue() then calls update_vf(cl, qdisc_pkt_len(skb), cur_time)
to charge the dequeued bytes.
On the second call the class is already passive, but its child qdisc is
still empty, so update_vf() arms go_passive again:
if (cl->qdisc->q.qlen == 0 && cl->cl_flags & HFSC_FSC)
go_passive = 1;
The leaf is then skipped by the cl_nactive == 0 check inside the loop,
which does not clear go_passive, so the stale go_passive propagates to the
parent and decrements its cl_nactive a second time. A parent that still
has other active children is driven to cl_nactive == 0 and removed from
the vttree, even though those siblings are still backlogged. They are
never dequeued again and the qdisc stalls.
Fix this by only arming go_passive when the class is actually active, so an
already-passive class no longer triggers a second passive transition. The
byte accounting (cl->cl_total += len) still runs for every ancestor, so
dequeued bytes continue to be counted exactly once.
Fixes: 51eb3b6554 ("sch_hfsc: make hfsc_qlen_notify() idempotent")
Reported-by: Anirudh Gupta <anirudhrudr@gmail.com>
Closes: https://lore.kernel.org/netdev/CAN2cbVe79oj0O9==m4+4x3v+O+qzRagA=2=wkrp9i9=CqYvyZA@mail.gmail.com/
Tested-by: Anirudh Gupta <anirudhrudr@gmail.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Link: https://patch.msgid.link/20260610132824.3027549-1-victor@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The flow classifier falls back to addr_fold() for fields that are missing
from packet headers. In map mode, userspace controls mask, xor, rshift,
addend and divisor, and can observe the resulting classid through class
statistics. This allows a tc classifier in a user/network namespace to
recover the 32-bit folded value of skb->sk, skb_dst() or skb_nfct().
Align with standard kernel practices for pointer hashing and replace the
XOR folding with a keyed siphash (which is cryptographically secure)
Fixes: e5dfb81518 ("[NET_SCHED]: Add flow classifier")
Reported-by: Kyle Zeng <kylebot@openai.com>
Tested-by: Kyle Zeng <kylebot@openai.com>
Tested-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260610101839.14135-1-jhs@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
sch_cbs and sch_taprio print ecmd.base.speed, even if
netif_get_link_ksettings() failed. When netif_get_link_ksettings()
fails the ecmd may not be initialized.
Use the always-initialized speed variable instead.
The semantics change slightly because UNKNOWN will
never be printed, but that doesn't seem important
enough to complicate the code for.
This is a _dbg() print, anyway.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Link: https://patch.msgid.link/20260609183353.1109641-1-kuba@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
FQ currently assumes skb->tstamp holds monotonic time, as used by TCP.
Users with ns_capable CAP_NET_ADMIN can transmit skbs using SO_TXTIME
with CLOCK_MONOTONIC, CLOCK_REALTIME or CLOCK_TAI clockids as of
commit 80b14dee2b ("net: Add a new socket option for a future
transmit time.")
More recently, skbs also gained tstamp_type to explicitly communicate
the clockid of skb->tstamp, with commit 4d25ca2d68 ("net: Rename
mono_delivery_time to tstamp_type for scalabilty"), commit
1693c5db6a ("net: Add additional bit to support clockid_t timestamp
type") and a few others.
Detect other clocks and convert to monotonic for use in FQ. That is,
convert fq_skb_cb(skb)->time_to_send. Do not convert skb->tstamp
itself. Network device clocks are more commonly synchronized to TAI.
Conversion may be imprecise due to clock adjustment (e.g., adjfreq)
between when SCM_TSTAMP is set and when it is converted in fq_enqueue.
The common codepath is short, so skew will be well below common pacing
operation. Even in edge cases, bursts (too soon) or beyond horizon
(too late) are indistinguishable from network conditions. To which
senders must be robust, as long as infrequent.
Avoid overflow due to negative offsets becoming huge when converting
from signed ktime_t to u64 time_to_send. Bound lower to mono 1 and
upper to now + q->horizon. This protects against bad input, e.g.,
from BPF programs.
Detect legacy BPF programs that program skb->tstamp without setting
skb->tstamp_type. Here tstamp_type is zero (SKB_CLOCK_REALTIME), but
the value will be unrealistic for realtime in the 21st century. Follow
existing TIME_UPTIME_SEC_MAX as bound between mono and realtime.
Signed-off-by: Willem de Bruijn <willemb@google.com>
----
Changes
v1 -> v2
- replace Fixes tag with references inside the commit message
Link: https://patch.msgid.link/20260604194221.3319080-3-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cbs_set_port_rate() and taprio_set_picos_per_byte() are reached from
two paths and both already hold the device's ops lock:
*_change(), via tc_modify_qdisc() which calls netdev_lock_ops(dev)
before dispatching to the qdisc ops.
*_dev_notifier() on NETDEV_UP / NETDEV_CHANGE, where caller
holds the ops lock across the notifier chain.
Switch to netif_get_link_ksettings() to avoid deadlock once
__ethtool_get_link_ksettings() starts taking the netdev lock.
Link: https://patch.msgid.link/20260603012840.2254293-9-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
tcf_pedit_act() computes the COW range for skb_ensure_writable()
once before the key loop using tcfp_off_max_hint, but the hint does
not account for the runtime header offset added by typed keys. This
can leave part of the write region un-COW'd.
Fix by moving skb_ensure_writable() inside the per-key loop where
the actual write offset is known, and add overflow checking on the
offset arithmetic. For negative offsets (e.g. Ethernet header edits
at ingress), use skb_cow() to COW the headroom instead. Guard
offset_valid() against INT_MIN, where negation is undefined.
Fixes: 8b796475fd ("net/sched: act_pedit: really ensure the skb is writable")
Reported-by: Yiming Qian <yimingqian591@gmail.com>
Reported-by: Keenan Dong <keenanat2000@gmail.com>
Reported-by: Han Guidong <2045gemini@gmail.com>
Reported-by: Zhang Cen <rollkingzzc@gmail.com>
Reviewed-by: Han Guidong <2045gemini@gmail.com>
Tested-by: Han Guidong <2045gemini@gmail.com>
Reviewed-by: Davide Caratti <dcaratti@redhat.com>
Tested-by: Davide Caratti <dcaratti@redhat.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Tested-by: Toke Høiland-Jørgensen <toke@redhat.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Tested-by: Victor Nogueira <victor@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Signed-off-by: Rajat Gupta <rajat.gupta@oss.qualcomm.com>
Link: https://patch.msgid.link/20260531123221.48732-1-jhs@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When NEWTFILTER and DELFILTER are run concurrently it is possible to create a
race with an associated action.
Let's illustrate with CPU0 running NEWTFILTER and CPU1 running DELFILTER:
0: mutex_lock() <-- holds the idr lock
0: rcu_read_lock()
0: p = idr_find(idr, index) <-- action p is valid (RCU protects IDR)
0: mutex_unlock() <-- releases the idr lock
1: refcount_dec_and_mutex_lock() <-- refcnt 1->0, mutex held
1: idr_remove(idr, index) <-- Action removed from IDR
1: mutex_unlock() <-- mutex released allowing us to delete the action
1: tcf_action_cleanup(p); kfree(p) <-- Kfrees p immediately, no deferral
0: refcount_inc_not_zero(&p->tcfa_refcnt) <-- ouch, UAF p points to freed memory
This patch fixes the race condition between NEWTFILTER and DELFILTER by
adding struct rcu_head to tc_action used in the deferral and introducing a
call_rcu() in the delete path to defer the final kfree().
Note: this is a revert of commit d7fb60b9ca ("net_sched: get rid of tcfa_rcu")
but also modernization/simplification to directly use kfree_rcu().
Let's illustrate the new restored code path:
0: rcu_read_lock()
1: refcount_dec_and_mutex_lock() <-- refcnt 1->0, mutex held
1: idr_remove(idr, index)
1: mutex_unlock()
1: call_rcu(&p->tcfa_rcu, tcf_action_rcu_free) <-- defer kfree after grace period
0: p = idr_find(idr, index)
0: refcount_inc_not_zero(&p->tcfa_refcnt) <-- fails, refcnt already 0
1: rcu_read_unlock() <-- release so freeing can run after grace period
After CPU1 calls idr_remove(), the object is no longer reachable through the IDR.
CPU0's subsequent idr_find() will return NULL, and even if it still held a
stale pointer, the immediate kfree() is now deferred until after the RCU grace
period, so no UAF can occur.
Fixes: d7fb60b9ca ("net_sched: get rid of tcfa_rcu")
Suggested-by: Jakub Kicinski <kuba@kernel.org>
Reported-by: Kyle Zeng <kylebot@openai.com>
Tested-by: Victor Nogueira <victor@mojatatu.com>
Tested-by: syzbot@syzkaller.appspotmail.com
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Tested-by: Kyle Zeng <kylebot@openai.com>
Reviewed-by: Pedro Tammela <pctammela@mojatatu.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Link: https://patch.msgid.link/20260531160812.68020-1-jhs@mojatatu.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since retval is set as TC_ACT_STOLEN in the mirred redirect case, returning
retval in cases where redirect failed will make the callers not register
the skb as being dropped.
Fix this by returning TC_ACT_SHOT instead in such scenarios.
Fixes: 16085e48cb ("net/sched: act_mirred: Create function tcf_mirred_to_dev and improve readability")
Reported-by: Sashiko <sashiko-bot@kernel.org>
Closes: https://sashiko.dev/#/patchset/20260413082027.2244884-1-hxzene%40gmail.com
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Link: https://patch.msgid.link/20260525122556.973584-8-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
tcf_mirred_act() checks sched_mirred_nest against MIRRED_NEST_LIMIT (4)
to prevent deep recursion. However, when the action uses blockcast
(tcfm_blockid != 0), the function returns at the tcf_blockcast() call
BEFORE reaching the counter increment. As a result, the recursion
counter never advances and the limit check is entirely bypassed.
When two devices share a TC egress block with a mirred blockcast rule,
a packet egressing on device A is mirrored to device B via blockcast;
device B's egress TC re-enters tcf_mirred_act() via blockcast and
mirrors back to A, creating an unbounded recursion loop:
tcf_mirred_act -> tcf_blockcast -> tcf_mirred_to_dev -> dev_queue_xmit
-> sch_handle_egress -> tcf_classify -> tcf_mirred_act -> (repeat)
This recursion continues until the kernel stack overflows.
The bug is reachable from an unprivileged user via
unshare(CLONE_NEWUSER | CLONE_NEWNET): user namespaces grant
CAP_NET_ADMIN in the new network namespace, which is sufficient to
create dummy devices, attach clsact qdiscs with shared blocks, and
install mirred blockcast filters.
BUG: TASK stack guard page was hit at ffffc90000b7fff8
Oops: stack guard page: 0000 [#1] SMP KASAN NOPTI
CPU: 2 UID: 1000 PID: 169 Comm: poc Not tainted 7.0.0-rc7-next-20260410
RIP: 0010:xas_find+0x17/0x480
Call Trace:
xa_find+0x17b/0x1d0
tcf_mirred_act+0x640/0x1060
tcf_action_exec+0x400/0x530
basic_classify+0x128/0x1d0
tcf_classify+0xd83/0x1150
tc_run+0x328/0x620
__dev_queue_xmit+0x797/0x3100
tcf_mirred_to_dev+0x7b1/0xf70
tcf_mirred_act+0x68a/0x1060
[repeating ~30+ times until stack overflow]
Kernel panic - not syncing: Fatal exception in interrupt
Fix this by incrementing sched_mirred_nest before calling
tcf_blockcast() and decrementing it on return, mirroring the
non-blockcast path. This ensures subsequent recursive entries see the
updated counter and are correctly limited by MIRRED_NEST_LIMIT.
Fixes: fe946a751d ("net/sched: act_mirred: add loop detection")
Signed-off-by: Kito Xu (veritas501) <hxzene@gmail.com>
Link: https://patch.msgid.link/20260525122556.973584-7-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When mirred redirects to ingress (from either ingress or egress) the loop
state from sched_mirred_dev array dev is lost because of 1) the packet
deferral into the backlog and 2) the fact the sched_mirred_dev array is
cleared. In such cases, if there was a loop we won't discover it.
Here's a simple test to reproduce:
ip a add dev port0 10.10.10.11/24
tc qdisc add dev port0 clsact
tc filter add dev port0 egress protocol ip \
prio 10 matchall action mirred ingress redirect dev port1
tc qdisc add dev port1 clsact
tc filter add dev port1 ingress protocol ip \
prio 10 matchall action mirred egress redirect dev port0
ping -c 1 -W0.01 10.10.10.10
Fixes: fe946a751d ("net/sched: act_mirred: add loop detection")
Tested-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260525122556.973584-6-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When netem duplicates a packet it re-enqueues the copy at the root qdisc.
If another netem sits in the tree the copy can be duplicated
again, recursing until the stack or memory is exhausted.
The original duplication guard temporarily zeroed q->duplicate around
the re-enqueue, but that does not cover all cases because it is
per-qdisc state shared across all concurrent enqueue paths
and is not safe without additional locking.
Use the skb tc_depth field introduced in an earlier patch:
- increment it on the duplicate before re-enqueue
- skip duplication for any skb whose tc_depth is already non-zero.
This marks the packet itself rather than mutating qdisc state,
therefore it is safe regardless of tree topology or concurrency.
Fixes: 0afb51e728 ("[PKT_SCHED]: netem: reinsert for duplication")
Reported-by: William Liu <will@willsroot.io>
Reported-by: Savino Dicanosa <savy@syst3mfailure.io>
Closes: https://lore.kernel.org/netdev/8DuRWwfqjoRDLDmBMlIfbrsZg9Gx50DHJc1ilxsEBNe2D6NMoigR_eIRIG0LOjMc3r10nUUZtArXx4oZBIdUfZQrwjcQhdinnMis_0G7VEk=@willsroot.io/
Co-developed-by: Victor Nogueira <victor@mojatatu.com>
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: William Liu <will@willsroot.io>
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260525122556.973584-5-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This reverts commit ec8e0e3d7a.
The original patch rejects any tree containing two netems when
either has duplication set, even when they sit on unrelated classes
of the same classful parent. That broke configurations that have
worked since netem was introduced.
The re-entrancy problem the original commit was trying to solve is
handled by later patch using tc_depth flag.
Doing this revert will (re)expose the original bug with multiple
netem duplication. When this patch is backported make sure
and get the full series.
Fixes: ec8e0e3d7a ("net/sched: Restrict conditions for adding duplicating netems to qdisc tree")
Reported-by: Ji-Soo Chung <jschung2@proton.me>
Reported-by: Gerlinde <lrGerlinde@mailfence.com>
Closes: https://bugzilla.kernel.org/show_bug.cgi?id=220774
Reported-by: zyc zyc <zyc199902@zohomail.cn>
Closes: https://lore.kernel.org/all/19adda5a1e2.12410b78222774.9191120410578703463@zohomail.cn/
Reported-by: Manas Ghandat <ghandatmanas@gmail.com>
Closes: https://lore.kernel.org/netdev/f69b2c8f-8325-4c2e-a011-6dbc089f30e4@gmail.com/
Reviewed-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://patch.msgid.link/20260525122556.973584-3-jhs@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When inserting an EDT packet with time before flow->time_next_packet,
update the flow and possibly queue next delivery time.
Reinsert the flow into the q->delayed rb-tree to position correctly
and to have fq_check_throttled set wake-up at the right next time.
Factor RB tree insertion out fq_flow_set_throttled to avoid open
coding twice.
EDT packets do not take precedence over queue rate limit. Skip this
new step if a queue limit is set. EDT packets do take precedence over
per-socket rate limits, as can be seen from fq_dequeue reading
sk_pacing_rate if !skb->tstamp.
With this change the so_txtime selftest sends packets in the expected
order.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260526134109.2624493-1-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Quan Sun reported [1] a stack overflow in cls_bpf_offload_cmd().
Reproducer on netdevsim: add a skip_sw cls_bpf filter, set the
bpf_tc_accept debugfs knob to 0, then `tc filter replace`. The replace
calls tc_setup_cb_replace() which fails. cls_bpf_offload_cmd() then
swaps prog/oldprog and recursively calls itself to roll back. But
bpf_tc_accept=0 makes the rollback fail too, which triggers yet another
rollback frame with the same arguments, and so on until the stack is
exhausted.
bpf_tc_accept is just a convenient knob for the reproducer. Any driver
whose tc_setup_cb_replace() fails twice in a row can hit the same loop,
so this is not a netdevsim-only issue.
Two ways to fix it:
1) Have the rollback call tc_setup_cb_add() on oldprog instead of
re-entering cls_bpf_offload_cmd().
2) Mark the rollback frame with a flag and skip a second-level
rollback from inside it.
Go with (2). It is the smaller change and keeps the original behaviour:
the rollback still goes through tc_setup_cb_replace(), so the driver
gets one real chance to restore its state. If that attempt also fails,
we just return the original error instead of recursing.
[1]: https://lore.kernel.org/bpf/ce5a6005-3c5e-4696-9e05-eba9461dc860@std.uestc.edu.cn/T/#u
Fixes: 102740bd94 ("cls_bpf: fix offload assumptions after callback conversion")
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Jiayuan Chen <jiayuan.chen@linux.dev>
Link: https://patch.msgid.link/20260526025529.24382-1-jiayuan.chen@linux.dev
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
cl->quantum does not need to be protected by RTNL or qdisc spinlock.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Link: https://patch.msgid.link/20260522110356.1403343-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Drivers might have error messages to propagate to user space. Propagate
the netlink extack so that they can inform user space in a verbal way of
their limitations.
Signed-off-by: David Yang <mmyangfl@gmail.com>
Link: https://patch.msgid.link/20260521010320.208138-3-mmyangfl@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use NL_SET_ERR_MSG_WEAK so any existing error message (probably from the
underlying driver) is used instead of the generic fallback.
Signed-off-by: David Yang <mmyangfl@gmail.com>
Link: https://patch.msgid.link/20260521010320.208138-2-mmyangfl@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drr_dump_class_stats() runs without qdisc spinlock held.
Add missing READ_ONCE()/WRITE_ONCE() annotations around cl->deficit.
Fixes: edb09eb17e ("net: sched: do not acquire qdisc spinlock in qdisc/class stats dump")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260519094618.2632073-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
htb_dump_class_stats() and htb_offload_aggregate_stats()
call gnet_stats_basic_sync_init(&cl->bstats) which
is wrong on 32bit arches when syncp is cleared.
Make sure to acquire qdisc spinlock and use
_bstats_set() to ease future lockless dumps.
Fixes: 8327158624 ("sch_htb: Stats for offloaded HTB")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Maxim Mikityanskiy <maximmi@mellanox.com>
Cc: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260518090518.629245-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
htb_dump_class_stats() will soon run locklessly.
(no RTNL, not qdisc spinlock).
Add READ_ONCE()/WRITE_ONCE() annotations on these fields:
- cl->overlimits
- cl->drops
- cl->tokens
- cl->ctokens
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260514095935.3926276-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
htb_dump_class_stats() will soon run locklessly.
(no RTNL, not qdisc spinlock).
Remove cl->xstats and replace it with two fields:
- xstats_lends
- xstats_borrows
Then use READ_ONCE()/WRITE_ONCE() annotations on them, and change
htb_dump_class_stats to use a private struct tc_htb_xstats.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20260514095935.3926276-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
