kernel/git/next/linux-next.git/net/sched/cls_api.c, branch stable

net/sched: account classifier filter allocations to memcg

2026-08-22T20:02:32+00:00

Allocations in the tc classifier *_change() paths (filter objects, per-CPU counters, and per-filter aux data) use plain GFP_KERNEL without __GFP_ACCOUNT, allowing unprivileged users to pin kernel memory outside memcg charging. The shared tcf_exts_init_ex() action array allocation in cls_api.c was also uncharged; this patch closes it along with the per-classifier filter-object/percpu/aux allocations that remain unaccounted. Add GFP_KERNEL_ACCOUNT to: - the shared tcf_exts_init_ex() action array (cls_api.c), common to every filter of every classifier (32 pointers, 256 bytes); - the filter-object, per-CPU-counter, and per-filter aux allocations in cls_basic, cls_bpf, cls_cgroup, cls_flow, cls_flower, cls_fw, cls_matchall, cls_route and cls_u32; - the u32_init_knode() replace-path knode allocation (cls_u32.c), which allocates the same struct tc_u_knode + sel.keys on every replace of an existing knode and was missed by the create-path-only conversion. Also fix the cls_basic error path: basic_change() inserts fnew into the IDR before allocating the per-CPU counter. If alloc_percpu() fails the errout path kfree'd fnew without idr_remove, leaving a dangling pointer in the IDR. With GFP_KERNEL_ACCOUNT the percpu alloc becomes failable on demand (memcg at memory.max), making the dead path attacker-reachable and burning the handle permanently. Add the idr_remove on the percpu failure path, matching the basic_set_parms failure-path pattern. Note: vega@nebusec.ai provided a poc for basic_cls, but it was easy to extend to the other classifiers. Conditions to recreate the bug: - CONFIG_NET_SCHED, CONFIG_NET_CLS_* (the classifier being used), CONFIG_NET_CLS_ACT, CONFIG_MEMCG, CONFIG_USER_NS, CONFIG_NET_NS. - Unprivileged user in a fresh user+network namespace (unshare -Urn), or root with CAP_NET_ADMIN. - Create a large number of tc filters (e.g. tc filter add dev lo ingress ... ...) while watching a memcg-limited cgroup: system slab grows far faster than memory.current, pinning kernel memory outside memcg charging. Fixes: 0da974f4f303 ("[NET]: Conversions from kmalloc+memset to k(z|c)alloc.") Reported-by: vega@nebusec.ai Signed-off-by: Jamal Hadi Salim Reviewed-by: Breno Leitao Link: https://patch.msgid.link/20260819143733.57538-1-jhs@mojatatu.com Signed-off-by: Jakub Kicinski

net/sched: cls_api: fix teardown of an adopted proto on insert-race loss

2026-08-08T00:22:12+00:00

In tc_new_tfilter() the create branch sets tp_created = 1 before calling tcf_chain_tp_insert_unique(). When the caller loses the race (another request inserted a proto at the same chain/prio first), insert_unique() destroys the caller's own tp_new and returns the winner's proto with an extra reference. tp_created was never cleared, so the loser's errout path treated the winner's live proto as its own and called tcf_chain_tp_delete_empty() on it, silently unlinking an active classifier that the winning request already advertised via RTM_NEWTFILTER. Track the outcome of the insert step in a single tri-state variable so each errout path reacts correctly: - TP_NOT_CREATED: no proto created; pursue the old path. - TP_CREATED: proto inserted successfully; same code path as before. - TP_NOT_OWNED: New - lost the insert race; tp is another request's proto (chain ref already released by tp_new's destroy) Both errout reactions are single expressions derived from the state. This fix is motivated by the Sashiko's automated review of Patch (net/sched: cls_api: Always acquire rtnl_lock when destroying locked classifiers) [1][2]. The review identified the silent-unlink behaviour of an adopted proto's teardown when a request loses the tcf_chain_tp_insert_unique() race. [1] https://sashiko.dev/#/patchset/20260801125632.360365-1-jhs%40mojatatu.com [2] https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260801125632.360365-1-jhs%40mojatatu.com Fixes: 8b64678e0af8 ("net: sched: refactor tp insert/delete for concurrent execution") Reported-by: Sashiko Closes: https://sashiko.dev/#/patchset/20260801125632.360365-1-jhs%40mojatatu.com Closes: https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260801125632.360365-1-jhs%40mojatatu.com Acked-by: Jamal Hadi Salim Signed-off-by: Victor Nogueira Reported-by: TencentOS Corvus AI Tested-by: Aohan Mei Link: https://patch.msgid.link/20260805134049.927864-1-victor@mojatatu.com Signed-off-by: Jakub Kicinski

net/sched: cls_api: Always acquire rtnl_lock when destroying locked classifiers

2026-08-04T14:35:45+00:00

Another challenge with unlocked filters. There is a short window in tc_new_tfilter where a tcf_proto can be found and briefly referenced by a totally unrelated, unlocked classifier's request and cause a race. Feng created a poc which created this race with two threads, one creating a u32 filter and other a flower filter in the same chain/prio: 1. Both threads enter tc_new_tfilter, both find the chain empty, both drop filter_chain_lock 2. u32 finishes tcf_proto_create("u32") first, calls tcf_chain_tp_insert_unique() -> inserts u32_tp into the chain 3. flower finishes tcf_proto_create("flower") later, calls tcf_chain_tp_insert_unique() -> tcf_chain_tp_find() now sees u32_tp already there, takes a reference on it, destroys flower's own tp_new and returns u32_tp to the caller. Flower then hits the kind mismatch check (because it requested for kind "flower" but tp->ops->kind is "u32") and goes through the errout path which calls tcf_proto_put() on u32_tp. If the u32 thread has already gone through its own errout (its change() call failed on the PoC's empty options) and dropped its create and insert refs, flower's put is the last one and drops u32_tp's refcnt to zero. At this point tp->ops->destroy() runs in a context that never took rtnl_lock. When that happens, it might cause a UAF like the following (illustrated by the PoC): [ +0.000710] BUG: KASAN: slab-use-after-free in u32_init (net/sched/cls_u32.c:393) [ +0.000281] Read of size 8 at addr ffff888120022f00 by task poc_feng_xue/524 Call Trace: u32_init (net/sched/cls_u32.c:393) tc_new_tfilter (net/sched/cls_api.c:2378) Allocated by task 526: u32_init (net/sched/cls_u32.c:378) tc_new_tfilter (net/sched/cls_api.c:2378) Freed by task 522: kfree u32_destroy (net/sched/cls_u32.c:662) tcf_proto_destroy (net/sched/cls_api.c:446) tcf_proto_put (net/sched/cls_api.c:459) tc_new_tfilter (net/sched/cls_api.c:2459) Fix this by having tcf_proto_destroy() take rtnl_lock around tp->ops->destroy() for locked classifiers whenever rtnl is not held. To explain why I used a temp variable "not_lockless" I'd like to point to a semi-related note on rtnl_held vs TCF_PROTO_OPS_DOIT_UNLOCKED (adding here for future cleanup if deemed necessary): The rtnl_held parameter and the TCF_PROTO_OPS_DOIT_UNLOCKED flag are redundant sources of truth for whether rtnl_lock is held. Among the nine classifier destroy(..rtnl_held..) callbacks, only flower consults the rtnl_held parameter which it propagates to tc_setup_cb_destroy() and tc_setup_cb_call(). The other eight (u32, flow, bpf, cgroup, route, basic, fw, mall) ignore it entirely;-> those that call tc_setup_cb_destroy() (u32, bpf, mall) hardcode true always instead of forwarding the parameter. A future cleanup should remove the rtnl_held parameter from the destroy callback signature entirely and have callers rely solely on their knowledge whether they are running in an unlocked context. Fixes: 12db03b65c2b ("net: sched: extend proto ops to support unlocked classifiers") Reported-by: Feng Xue Tested-by: Victor Nogueira Signed-off-by: Jamal Hadi Salim Link: https://patch.msgid.link/20260801125632.360365-1-jhs@mojatatu.com Signed-off-by: Paolo Abeni

net/sched: Handle TC_ACT_REDIRECT from qdisc filter chains

2026-07-21T01:16:43+00:00

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 27b29f63058d ("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 401cb7dae813 ("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: 27b29f63058d ("bpf: add bpf_redirect() helper") Fixes: 401cb7dae813 ("net: Reference bpf_redirect_info via task_struct on PREEMPT_RT.") Tested-by: Victor Nogueira Signed-off-by: Jamal Hadi Salim Signed-off-by: Daniel Borkmann Reviewed-by: Sebastian Andrzej Siewior Link: https://patch.msgid.link/20260706185609.330006-3-daniel@iogearbox.net Signed-off-by: Jakub Kicinski

bpf: Reject redirect helpers without a bpf_net_context

2026-07-21T01:16:42+00:00

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: 401cb7dae813 ("net: Reference bpf_redirect_info via task_struct on PREEMPT_RT.") Fixes: 3625750f05ec ("net: sched: Introduce helpers for qevent blocks") Signed-off-by: Daniel Borkmann Reviewed-by: Sebastian Andrzej Siewior Link: https://patch.msgid.link/20260706185609.330006-2-daniel@iogearbox.net Signed-off-by: Jakub Kicinski

net/sched: act_pedit: fix TOCTOU heap OOB write in tc offload

2026-07-06T10:52:01+00:00

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: 71d0ed7079df ("net/act_pedit: Support using offset relative to the conventional network headers") Reported-by: zdi-disclosures@trendmicro.com Tested-by: Victor Nogueira Signed-off-by: Jamal Hadi Salim Link: https://patch.msgid.link/20260701161912.125355-1-jhs@mojatatu.com Signed-off-by: Paolo Abeni

net/sched: cls_api: Handle TC_ACT_CONSUMED in tcf_qevent_handle

2026-06-24T02:08:36+00:00

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: 3f14b377d01d ("net/sched: act_ct: fix skb leak and crash on ooo frags") Reported-by: Zero Day Initiative Tested-by: Victor Nogueira Signed-off-by: Jamal Hadi Salim Link: https://patch.msgid.link/20260620130749.226642-1-jhs@mojatatu.com Signed-off-by: Jakub Kicinski

net: sched: cls_api: fix tc_chain_fill_node to initialize tcm_info to zero to prevent an info-leak

2026-03-31T00:56:40+00:00

When building netlink messages, tc_chain_fill_node() never initializes the tcm_info field of struct tcmsg. Since the allocation is not zeroed, kernel heap memory is leaked to userspace through this 4-byte field. The fix simply zeroes tcm_info alongside the other fields that are already initialized. Fixes: 32a4f5ecd738 ("net: sched: introduce chain object to uapi") Signed-off-by: Yochai Eisenrich Acked-by: Jamal Hadi Salim Link: https://patch.msgid.link/20260328211436.1010152-1-echelonh@gmail.com Signed-off-by: Jakub Kicinski

net/sched: Only allow act_ct to bind to clsact/ingress qdiscs and shared blocks

2026-02-28T03:06:21+00:00

As Paolo said earlier [1]: "Since the blamed commit below, classify can return TC_ACT_CONSUMED while the current skb being held by the defragmentation engine. As reported by GangMin Kim, if such packet is that may cause a UaF when the defrag engine later on tries to tuch again such packet." act_ct was never meant to be used in the egress path, however some users are attaching it to egress today [2]. Attempting to reach a middle ground, we noticed that, while most qdiscs are not handling TC_ACT_CONSUMED, clsact/ingress qdiscs are. With that in mind, we address the issue by only allowing act_ct to bind to clsact/ingress qdiscs and shared blocks. That way it's still possible to attach act_ct to egress (albeit only with clsact). [1] https://lore.kernel.org/netdev/674b8cbfc385c6f37fb29a1de08d8fe5c2b0fbee.1771321118.git.pabeni@redhat.com/ [2] https://lore.kernel.org/netdev/cc6bfb4a-4a2b-42d8-b9ce-7ef6644fb22b@ovn.org/ Reported-by: GangMin Kim Fixes: 3f14b377d01d ("net/sched: act_ct: fix skb leak and crash on ooo frags") CC: stable@vger.kernel.org Signed-off-by: Victor Nogueira Acked-by: Jamal Hadi Salim Link: https://patch.msgid.link/20260225134349.1287037-1-victor@mojatatu.com Signed-off-by: Jakub Kicinski

Convert more 'alloc_obj' cases to default GFP_KERNEL arguments

2026-02-22T04:03:00+00:00

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