kernel.git/net/core, branch linux-5.10.y

net: consume xmit errors of GSO frames

2026-03-04T12:20:12Z

[ Upstream commit 7aa767d0d3d04e50ae94e770db7db8197f666970 ] udpgro_frglist.sh and udpgro_bench.sh are the flakiest tests currently in NIPA. They fail in the same exact way, TCP GRO test stalls occasionally and the test gets killed after 10min. These tests use veth to simulate GRO. They attach a trivial ("return XDP_PASS;") XDP program to the veth to force TSO off and NAPI on. Digging into the failure mode we can see that the connection is completely stuck after a burst of drops. The sender's snd_nxt is at sequence number N [1], but the receiver claims to have received (rcv_nxt) up to N + 3 * MSS [2]. Last piece of the puzzle is that senders rtx queue is not empty (let's say the block in the rtx queue is at sequence number N - 4 * MSS [3]). In this state, sender sends a retransmission from the rtx queue with a single segment, and sequence numbers N-4*MSS:N-3*MSS [3]. Receiver sees it and responds with an ACK all the way up to N + 3 * MSS [2]. But sender will reject this ack as TCP_ACK_UNSENT_DATA because it has no recollection of ever sending data that far out [1]. And we are stuck. The root cause is the mess of the xmit return codes. veth returns an error when it can't xmit a frame. We end up with a loss event like this: ------------------------------------------------- | GSO super frame 1 | GSO super frame 2 | |-----------------------------------------------| | seg | seg | seg | seg | seg | seg | seg | seg | | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | ------------------------------------------------- x ok ok | ok ok ok \\ snd_nxt "x" means packet lost by veth, and "ok" means it went thru. Since veth has TSO disabled in this test it sees individual segments. Segment 1 is on the retransmit queue and will be resent. So why did the sender not advance snd_nxt even tho it clearly did send up to seg 8? tcp_write_xmit() interprets the return code from the core to mean that data has not been sent at all. Since TCP deals with GSO super frames, not individual segment the crux of the problem is that loss of a single segment can be interpreted as loss of all. TCP only sees the last return code for the last segment of the GSO frame (in <> brackets in the diagram above). Of course for the problem to occur we need a setup or a device without a Qdisc. Otherwise Qdisc layer disconnects the protocol layer from the device errors completely. We have multiple ways to fix this. 1) make veth not return an error when it lost a packet. While this is what I think we did in the past, the issue keeps reappearing and it's annoying to debug. The game of whack a mole is not great. 2) fix the damn return codes We only talk about NETDEV_TX_OK and NETDEV_TX_BUSY in the documentation, so maybe we should make the return code from ndo_start_xmit() a boolean. I like that the most, but perhaps some ancient, not-really-networking protocol would suffer. 3) make TCP ignore the errors It is not entirely clear to me what benefit TCP gets from interpreting the result of ip_queue_xmit()? Specifically once the connection is established and we're pushing data - packet loss is just packet loss? 4) this fix Ignore the rc in the Qdisc-less+GSO case, since it's unreliable. We already always return OK in the TCQ_F_CAN_BYPASS case. In the Qdisc-less case let's be a bit more conservative and only mask the GSO errors. This path is taken by non-IP-"networks" like CAN, MCTP etc, so we could regress some ancient thing. This is the simplest, but also maybe the hackiest fix? Similar fix has been proposed by Eric in the past but never committed because original reporter was working with an OOT driver and wasn't providing feedback (see Link). Link: https://lore.kernel.org/CANn89iJcLepEin7EtBETrZ36bjoD9LrR=k4cfwWh046GB+4f9A@mail.gmail.com Fixes: 1f59533f9ca5 ("qdisc: validate frames going through the direct_xmit path") Signed-off-by: Jakub Kicinski Reviewed-by: Eric Dumazet Link: https://patch.msgid.link/20260223235100.108939-1-kuba@kernel.org Signed-off-by: Paolo Abeni Signed-off-by: Sasha Levin

bpf: Reject narrower access to pointer ctx fields

2026-02-06T15:40:09Z

commit e09299225d5ba3916c91ef70565f7d2187e4cca0 upstream. The following BPF program, simplified from a syzkaller repro, causes a kernel warning: r0 = *(u8 *)(r1 + 169); exit; With pointer field sk being at offset 168 in __sk_buff. This access is detected as a narrower read in bpf_skb_is_valid_access because it doesn't match offsetof(struct __sk_buff, sk). It is therefore allowed and later proceeds to bpf_convert_ctx_access. Note that for the "is_narrower_load" case in the convert_ctx_accesses(), the insn->off is aligned, so the cnt may not be 0 because it matches the offsetof(struct __sk_buff, sk) in the bpf_convert_ctx_access. However, the target_size stays 0 and the verifier errors with a kernel warning: verifier bug: error during ctx access conversion(1) This patch fixes that to return a proper "invalid bpf_context access off=X size=Y" error on the load instruction. The same issue affects multiple other fields in context structures that allow narrow access. Some other non-affected fields (for sk_msg, sk_lookup, and sockopt) were also changed to use bpf_ctx_range_ptr for consistency. Note this syzkaller crash was reported in the "Closes" link below, which used to be about a different bug, fixed in commit fce7bd8e385a ("bpf/verifier: Handle BPF_LOAD_ACQ instructions in insn_def_regno()"). Because syzbot somehow confused the two bugs, the new crash and repro didn't get reported to the mailing list. Fixes: f96da09473b52 ("bpf: simplify narrower ctx access") Fixes: 0df1a55afa832 ("bpf: Warn on internal verifier errors") Reported-by: syzbot+0ef84a7bdf5301d4cbec@syzkaller.appspotmail.com Closes: https://syzkaller.appspot.com/bug?extid=0ef84a7bdf5301d4cbec Signed-off-by: Paul Chaignon Signed-off-by: Martin KaFai Lau Acked-by: Eduard Zingerman Link: https://patch.msgid.link/3b8dcee67ff4296903351a974ddd9c4dca768b64.1753194596.git.paul.chaignon@gmail.com [shung-hsi.yu: offset(struct bpf_sock_ops, skb_hwtstamp) case was dropped becasuse it was only added in v6.2 with commit 9bb053490f1a ("bpf: Add hwtstamp field for the sockops prog")] Signed-off-by: Shung-Hsi Yu Signed-off-by: Greg Kroah-Hartman

bpf: Do not let BPF test infra emit invalid GSO types to stack

2026-02-06T15:40:09Z

commit 04a899573fb87273a656f178b5f920c505f68875 upstream. Yinhao et al. reported that their fuzzer tool was able to trigger a skb_warn_bad_offload() from netif_skb_features() -> gso_features_check(). When a BPF program - triggered via BPF test infra - pushes the packet to the loopback device via bpf_clone_redirect() then mentioned offload warning can be seen. GSO-related features are then rightfully disabled. We get into this situation due to convert___skb_to_skb() setting gso_segs and gso_size but not gso_type. Technically, it makes sense that this warning triggers since the GSO properties are malformed due to the gso_type. Potentially, the gso_type could be marked non-trustworthy through setting it at least to SKB_GSO_DODGY without any other specific assumptions, but that also feels wrong given we should not go further into the GSO engine in the first place. The checks were added in 121d57af308d ("gso: validate gso_type in GSO handlers") because there were malicious (syzbot) senders that combine a protocol with a non-matching gso_type. If we would want to drop such packets, gso_features_check() currently only returns feature flags via netif_skb_features(), so one location for potentially dropping such skbs could be validate_xmit_unreadable_skb(), but then otoh it would be an additional check in the fast-path for a very corner case. Given bpf_clone_redirect() is the only place where BPF test infra could emit such packets, lets reject them right there. Fixes: 850a88cc4096 ("bpf: Expose __sk_buff wire_len/gso_segs to BPF_PROG_TEST_RUN") Fixes: cf62089b0edd ("bpf: Add gso_size to __sk_buff") Reported-by: Yinhao Hu Reported-by: Kaiyan Mei Reported-by: Dongliang Mu Signed-off-by: Daniel Borkmann Signed-off-by: Martin KaFai Lau Link: https://patch.msgid.link/20251020075441.127980-1-daniel@iogearbox.net Signed-off-by: Shung-Hsi Yu Signed-off-by: Greg Kroah-Hartman

net: sock: fix hardened usercopy panic in sock_recv_errqueue

2026-01-19T12:12:10Z

[ Upstream commit 2a71a1a8d0ed718b1c7a9ac61f07e5755c47ae20 ] skbuff_fclone_cache was created without defining a usercopy region, [1] unlike skbuff_head_cache which properly whitelists the cb[] field. [2] This causes a usercopy BUG() when CONFIG_HARDENED_USERCOPY is enabled and the kernel attempts to copy sk_buff.cb data to userspace via sock_recv_errqueue() -> put_cmsg(). The crash occurs when: 1. TCP allocates an skb using alloc_skb_fclone() (from skbuff_fclone_cache) [1] 2. The skb is cloned via skb_clone() using the pre-allocated fclone [3] 3. The cloned skb is queued to sk_error_queue for timestamp reporting 4. Userspace reads the error queue via recvmsg(MSG_ERRQUEUE) 5. sock_recv_errqueue() calls put_cmsg() to copy serr->ee from skb->cb [4] 6. __check_heap_object() fails because skbuff_fclone_cache has no usercopy whitelist [5] When cloned skbs allocated from skbuff_fclone_cache are used in the socket error queue, accessing the sock_exterr_skb structure in skb->cb via put_cmsg() triggers a usercopy hardening violation: [ 5.379589] usercopy: Kernel memory exposure attempt detected from SLUB object 'skbuff_fclone_cache' (offset 296, size 16)! [ 5.382796] kernel BUG at mm/usercopy.c:102! [ 5.383923] Oops: invalid opcode: 0000 [#1] SMP KASAN NOPTI [ 5.384903] CPU: 1 UID: 0 PID: 138 Comm: poc_put_cmsg Not tainted 6.12.57 #7 [ 5.384903] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.3-0-ga6ed6b701f0a-prebuilt.qemu.org 04/01/2014 [ 5.384903] RIP: 0010:usercopy_abort+0x6c/0x80 [ 5.384903] Code: 1a 86 51 48 c7 c2 40 15 1a 86 41 52 48 c7 c7 c0 15 1a 86 48 0f 45 d6 48 c7 c6 80 15 1a 86 48 89 c1 49 0f 45 f3 e8 84 27 88 ff <0f> 0b 490 [ 5.384903] RSP: 0018:ffffc900006f77a8 EFLAGS: 00010246 [ 5.384903] RAX: 000000000000006f RBX: ffff88800f0ad2a8 RCX: 1ffffffff0f72e74 [ 5.384903] RDX: 0000000000000000 RSI: 0000000000000004 RDI: ffffffff87b973a0 [ 5.384903] RBP: 0000000000000010 R08: 0000000000000000 R09: fffffbfff0f72e74 [ 5.384903] R10: 0000000000000003 R11: 79706f6372657375 R12: 0000000000000001 [ 5.384903] R13: ffff88800f0ad2b8 R14: ffffea00003c2b40 R15: ffffea00003c2b00 [ 5.384903] FS: 0000000011bc4380(0000) GS:ffff8880bf100000(0000) knlGS:0000000000000000 [ 5.384903] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 5.384903] CR2: 000056aa3b8e5fe4 CR3: 000000000ea26004 CR4: 0000000000770ef0 [ 5.384903] PKRU: 55555554 [ 5.384903] Call Trace: [ 5.384903] [ 5.384903] __check_heap_object+0x9a/0xd0 [ 5.384903] __check_object_size+0x46c/0x690 [ 5.384903] put_cmsg+0x129/0x5e0 [ 5.384903] sock_recv_errqueue+0x22f/0x380 [ 5.384903] tls_sw_recvmsg+0x7ed/0x1960 [ 5.384903] ? srso_alias_return_thunk+0x5/0xfbef5 [ 5.384903] ? schedule+0x6d/0x270 [ 5.384903] ? srso_alias_return_thunk+0x5/0xfbef5 [ 5.384903] ? mutex_unlock+0x81/0xd0 [ 5.384903] ? __pfx_mutex_unlock+0x10/0x10 [ 5.384903] ? __pfx_tls_sw_recvmsg+0x10/0x10 [ 5.384903] ? _raw_spin_lock_irqsave+0x8f/0xf0 [ 5.384903] ? _raw_read_unlock_irqrestore+0x20/0x40 [ 5.384903] ? srso_alias_return_thunk+0x5/0xfbef5 The crash offset 296 corresponds to skb2->cb within skbuff_fclones: - sizeof(struct sk_buff) = 232 - offsetof(struct sk_buff, cb) = 40 - offset of skb2.cb in fclones = 232 + 40 = 272 - crash offset 296 = 272 + 24 (inside sock_exterr_skb.ee) This patch uses a local stack variable as a bounce buffer to avoid the hardened usercopy check failure. [1] https://elixir.bootlin.com/linux/v6.12.62/source/net/ipv4/tcp.c#L885 [2] https://elixir.bootlin.com/linux/v6.12.62/source/net/core/skbuff.c#L5104 [3] https://elixir.bootlin.com/linux/v6.12.62/source/net/core/skbuff.c#L5566 [4] https://elixir.bootlin.com/linux/v6.12.62/source/net/core/skbuff.c#L5491 [5] https://elixir.bootlin.com/linux/v6.12.62/source/mm/slub.c#L5719 Fixes: 6d07d1cd300f ("usercopy: Restrict non-usercopy caches to size 0") Reported-by: Xiang Mei Signed-off-by: Weiming Shi Reviewed-by: Eric Dumazet Link: https://patch.msgid.link/20251223203534.1392218-2-bestswngs@gmail.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

bpf, sockmap: Don't let sock_map_{close,destroy,unhash} call itself

2026-01-19T12:12:09Z

commit 5b4a79ba65a1ab479903fff2e604865d229b70a9 upstream. sock_map proto callbacks should never call themselves by design. Protect against bugs like [1] and break out of the recursive loop to avoid a stack overflow in favor of a resource leak. [1] https://lore.kernel.org/all/00000000000073b14905ef2e7401@google.com/ Suggested-by: Eric Dumazet Signed-off-by: Jakub Sitnicki Acked-by: John Fastabend Link: https://lore.kernel.org/r/20230113-sockmap-fix-v2-1-1e0ee7ac2f90@cloudflare.com Signed-off-by: Alexei Starovoitov [Harinadh: Modified to apply on v5.10.y ] Signed-off-by: Harinadh Dommaraju Signed-off-by: Greg Kroah-Hartman

net: netpoll: fix incorrect refcount handling causing incorrect cleanup

2025-12-06T21:08:21Z

[ Upstream commit 49c8d2c1f94cc2f4d1a108530d7ba52614b874c2 ] commit efa95b01da18 ("netpoll: fix use after free") incorrectly ignored the refcount and prematurely set dev->npinfo to NULL during netpoll cleanup, leading to improper behavior and memory leaks. Scenario causing lack of proper cleanup: 1) A netpoll is associated with a NIC (e.g., eth0) and netdev->npinfo is allocated, and refcnt = 1 - Keep in mind that npinfo is shared among all netpoll instances. In this case, there is just one. 2) Another netpoll is also associated with the same NIC and npinfo->refcnt += 1. - Now dev->npinfo->refcnt = 2; - There is just one npinfo associated to the netdev. 3) When the first netpolls goes to clean up: - The first cleanup succeeds and clears np->dev->npinfo, ignoring refcnt. - It basically calls `RCU_INIT_POINTER(np->dev->npinfo, NULL);` - Set dev->npinfo = NULL, without proper cleanup - No ->ndo_netpoll_cleanup() is either called 4) Now the second target tries to clean up - The second cleanup fails because np->dev->npinfo is already NULL. * In this case, ops->ndo_netpoll_cleanup() was never called, and the skb pool is not cleaned as well (for the second netpoll instance) - This leaks npinfo and skbpool skbs, which is clearly reported by kmemleak. Revert commit efa95b01da18 ("netpoll: fix use after free") and adds clarifying comments emphasizing that npinfo cleanup should only happen once the refcount reaches zero, ensuring stable and correct netpoll behavior. Cc: # 3.17.x Cc: Jay Vosburgh Fixes: efa95b01da18 ("netpoll: fix use after free") Signed-off-by: Breno Leitao Reviewed-by: Simon Horman Link: https://patch.msgid.link/20251107-netconsole_torture-v10-1-749227b55f63@debian.org Signed-off-by: Jakub Kicinski [ Adjust context ] Signed-off-by: Sasha Levin Signed-off-by: Greg Kroah-Hartman

page_pool: Clamp pool size to max 16K pages

2025-12-06T21:08:15Z

[ Upstream commit a1b501a8c6a87c9265fd03bd004035199e2e8128 ] page_pool_init() returns E2BIG when the page_pool size goes above 32K pages. As some drivers are configuring the page_pool size according to the MTU and ring size, there are cases where this limit is exceeded and the queue creation fails. The page_pool size doesn't have to cover a full queue, especially for larger ring size. So clamp the size instead of returning an error. Do this in the core to avoid having each driver do the clamping. The current limit was deemed to high [1] so it was reduced to 16K to avoid page waste. [1] https://lore.kernel.org/all/1758532715-820422-3-git-send-email-tariqt@nvidia.com/ Signed-off-by: Dragos Tatulea Reviewed-by: Tariq Toukan Link: https://patch.msgid.link/20250926131605.2276734-2-dtatulea@nvidia.com Signed-off-by: Paolo Abeni Signed-off-by: Sasha Levin

net: call cond_resched() less often in __release_sock()

2025-12-06T21:08:12Z

[ Upstream commit 16c610162d1f1c332209de1c91ffb09b659bb65d ] While stress testing TCP I had unexpected retransmits and sack packets when a single cpu receives data from multiple high-throughput flows. super_netperf 4 -H srv -T,10 -l 3000 & Tcpdump extract: 00:00:00.000007 IP6 clnt > srv: Flags [.], seq 26062848:26124288, ack 1, win 66, options [nop,nop,TS val 651460834 ecr 3100749131], length 61440 00:00:00.000006 IP6 clnt > srv: Flags [.], seq 26124288:26185728, ack 1, win 66, options [nop,nop,TS val 651460834 ecr 3100749131], length 61440 00:00:00.000005 IP6 clnt > srv: Flags [P.], seq 26185728:26243072, ack 1, win 66, options [nop,nop,TS val 651460834 ecr 3100749131], length 57344 00:00:00.000006 IP6 clnt > srv: Flags [.], seq 26243072:26304512, ack 1, win 66, options [nop,nop,TS val 651460844 ecr 3100749141], length 61440 00:00:00.000005 IP6 clnt > srv: Flags [.], seq 26304512:26365952, ack 1, win 66, options [nop,nop,TS val 651460844 ecr 3100749141], length 61440 00:00:00.000007 IP6 clnt > srv: Flags [P.], seq 26365952:26423296, ack 1, win 66, options [nop,nop,TS val 651460844 ecr 3100749141], length 57344 00:00:00.000006 IP6 clnt > srv: Flags [.], seq 26423296:26484736, ack 1, win 66, options [nop,nop,TS val 651460853 ecr 3100749150], length 61440 00:00:00.000005 IP6 clnt > srv: Flags [.], seq 26484736:26546176, ack 1, win 66, options [nop,nop,TS val 651460853 ecr 3100749150], length 61440 00:00:00.000005 IP6 clnt > srv: Flags [P.], seq 26546176:26603520, ack 1, win 66, options [nop,nop,TS val 651460853 ecr 3100749150], length 57344 00:00:00.003932 IP6 clnt > srv: Flags [P.], seq 26603520:26619904, ack 1, win 66, options [nop,nop,TS val 651464844 ecr 3100753141], length 16384 00:00:00.006602 IP6 clnt > srv: Flags [.], seq 24862720:24866816, ack 1, win 66, options [nop,nop,TS val 651471419 ecr 3100759716], length 4096 00:00:00.013000 IP6 clnt > srv: Flags [.], seq 24862720:24866816, ack 1, win 66, options [nop,nop,TS val 651484421 ecr 3100772718], length 4096 00:00:00.000416 IP6 srv > clnt: Flags [.], ack 26619904, win 1393, options [nop,nop,TS val 3100773185 ecr 651484421,nop,nop,sack 1 {24862720:24866816}], length 0 After analysis, it appears this is because of the cond_resched() call from __release_sock(). When current thread is yielding, while still holding the TCP socket lock, it might regain the cpu after a very long time. Other peer TLP/RTO is firing (multiple times) and packets are retransmit, while the initial copy is waiting in the socket backlog or receive queue. In this patch, I call cond_resched() only once every 16 packets. Modern TCP stack now spends less time per packet in the backlog, especially because ACK are no longer sent (commit 133c4c0d3717 "tcp: defer regular ACK while processing socket backlog") Before: clnt:/# nstat -n;sleep 10;nstat|egrep "TcpOutSegs|TcpRetransSegs|TCPFastRetrans|TCPTimeouts|Probes|TCPSpuriousRTOs|DSACK" TcpOutSegs 19046186 0.0 TcpRetransSegs 1471 0.0 TcpExtTCPTimeouts 1397 0.0 TcpExtTCPLossProbes 1356 0.0 TcpExtTCPDSACKRecv 1352 0.0 TcpExtTCPSpuriousRTOs 114 0.0 TcpExtTCPDSACKRecvSegs 1352 0.0 After: clnt:/# nstat -n;sleep 10;nstat|egrep "TcpOutSegs|TcpRetransSegs|TCPFastRetrans|TCPTimeouts|Probes|TCPSpuriousRTOs|DSACK" TcpOutSegs 19218936 0.0 Signed-off-by: Eric Dumazet Reviewed-by: Kuniyuki Iwashima Link: https://patch.msgid.link/20250903174811.1930820-1-edumazet@google.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

net: Call trace_sock_exceed_buf_limit() for memcg failure with SK_MEM_RECV.

2025-12-06T21:08:11Z

[ Upstream commit 9d85c565a7b7c78b732393c02bcaa4d5c275fe58 ] Initially, trace_sock_exceed_buf_limit() was invoked when __sk_mem_raise_allocated() failed due to the memcg limit or the global limit. However, commit d6f19938eb031 ("net: expose sk wmem in sock_exceed_buf_limit tracepoint") somehow suppressed the event only when memcg failed to charge for SK_MEM_RECV, although the memcg failure for SK_MEM_SEND still triggers the event. Let's restore the event for SK_MEM_RECV. Signed-off-by: Kuniyuki Iwashima Reviewed-by: Eric Dumazet Reviewed-by: Shakeel Butt Link: https://patch.msgid.link/20250815201712.1745332-5-kuniyu@google.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

net: rtnetlink: fix module reference count leak issue in rtnetlink_rcv_msg

2025-10-29T13:01:25Z

commit 5b22f62724a0a09e00d301abf5b57b0c12be8a16 upstream. When bulk delete command is received in the rtnetlink_rcv_msg function, if bulk delete is not supported, module_put is not called to release the reference counting. As a result, module reference count is leaked. Fixes: a6cec0bcd342 ("net: rtnetlink: add bulk delete support flag") Signed-off-by: Zhengchao Shao Acked-by: Nikolay Aleksandrov Link: https://lore.kernel.org/r/20220815024629.240367-1-shaozhengchao@huawei.com Signed-off-by: Jakub Kicinski Signed-off-by: Greg Kroah-Hartman