Re: netfilter regression on aarch64 16k page size

[Pablo Neira Ayuso <pablo@netfilter.org>]

Hi,

Thanks for your report, it is an unalign atomic that results from
this.

I will post a patch asap to address this.

On Fri, Dec 20, 2024 at 12:15:36PM +0000, Amby @ Hyperbeam wrote:
> Greetings,
> We are seeing a regression in 6.6.66 which I have narrowed down to this commit:
> stable: 86c27603514cb8ead29857365cdd145404ee9706
> upstream: 7ffc7481153bbabf3332c6a19b289730c7e1edf5
> 
> Kernel version: 6.6.66 on aarch64 with 16k page size
> Last known good version: 6.6.65
> 
> Steps to repro:
> - run a 16k page size kernel (check with getconf PAGESIZE)
> - try to load an nftables config file on the problem
> 
> Expected:
> - no errors
> 
> Actual:
> - system enters a broken state, with the following trace in dmesg:
> [   40.939230] Unable to handle kernel paging request at virtual
> address ffff00015ad7e4cc
> [   40.939841] Mem abort info:
> [   40.940079]   ESR = 0x0000000096000021
> [   40.940389]   EC = 0x25: DABT (current EL), IL = 32 bits
> [   40.940820]   SET = 0, FnV = 0
> [   40.941042]   EA = 0, S1PTW = 0
> [   40.941289]   FSC = 0x21: alignment fault
> [   40.941570] Data abort info:
> [   40.941805]   ISV = 0, ISS = 0x00000021, ISS2 = 0x00000000
> [   40.942229]   CM = 0, WnR = 0, TnD = 0, TagAccess = 0
> [   40.942857]   GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0
> [   40.943313] swapper pgtable: 16k pages, 48-bit VAs, pgdp=00000000474f0000
> [   40.943865] [ffff00015ad7e4cc] pgd=180000043f7e8003,
> p4d=180000043f7e8003, pud=180000043f7e4003, pmd=180000043f52c003,
> pte=006800019ad7cf07
> [   40.945664] Internal error: Oops: 0000000096000021 [#1] SMP
> [   40.946055] Modules linked in: zstd zram zsmalloc nf_log_syslog
> nft_log nft_reject_inet nf_reject_ipv4 nf_reject_ipv6 nft_reject
> nft_ct nf_conntrack nf_defrag_ipv6 nf_defrag_ipv4 nf_tables
> crct10dif_ce polyval_ce polyval_generic ghash_ce tcp_bbr sch_fq fuse
> nfnetlink vsock_loopback vmw_vsock_virtio_transport_common
> vmw_vsock_vmci_transport vmw_vmci vsock bpf_preload qemu_fw_cfg
> ip_tables squashfs virtio_net net_failover virtio_blk gpio_keys
> failover virtio_mmio virtio_scsi virtio_console virtio_balloon
> virtio_gpu virtio_dma_buf megaraid_sas
> [   40.950262] CPU: 7 PID: 116 Comm: kworker/7:1 Not tainted 6.6.67-1-lts #1
> [   40.951542] Hardware name: netcup KVM Server, BIOS VPS 2000 ARM G11
> 08/07/2024
> [   40.952287] Workqueue: events_power_efficient nft_rhash_gc [nf_tables]
> [   40.952798] pstate: 20401005 (nzCv daif +PAN -UAO -TCO -DIT +SSBS BTYPE=--)
> [   40.953401] pc : nft_rhash_gc+0x208/0x2c0 [nf_tables]
> [   40.954054] lr : nft_rhash_gc+0x134/0x2c0 [nf_tables]
> [   40.954806] sp : ffff800081fb3cf0
> [   40.955138] x29: ffff800081fb3d50 x28: 0000000000000000 x27: 0000000000000000
> [   40.955974] x26: ffff0000cc760ef0 x25: ffff0000ccba9c80 x24: ffff0000cc758f78
> [   40.956750] x23: 0000000000000010 x22: ffff0000ca789000 x21: ffff0000cc760f78
> [   40.957455] x20: ffffd62d3ac3be40 x19: ffff00015ad7e4c0 x18: 0000000000000000
> [   40.959005] x17: 0000000000000000 x16: ffffd62d37926880 x15: 000000400002fef8
> [   40.959614] x14: ffffffffffffffff x13: 0000000000000030 x12: ffff0000cc760ef0
> [   40.960197] x11: 0000000000000000 x10: ffff800081fb3d08 x9 : ffff0000cba89fe0
> [   40.960739] x8 : 0000000000000003 x7 : 0000000000000000 x6 : 0000000000000000
> [   40.961287] x5 : 0000000000000040 x4 : ffff0000cba89ff0 x3 : ffff00015ad7e4c0
> [   40.961814] x2 : ffff00015ad7e4cc x1 : ffff00015ad7e4cc x0 : 0000000000000004
> [   40.962339] Call trace:
> [   40.962531]  nft_rhash_gc+0x208/0x2c0 [nf_tables]
> [   40.962911]  process_one_work+0x178/0x3e0
> [   40.963710]  worker_thread+0x2ac/0x3e0
> [   40.964530]  kthread+0xf0/0x108
> [   40.966259]  ret_from_fork+0x10/0x20
> [   40.967287] Code: 54fff9a4 d503201f d2800080 91003261 (f820303f)
> [   40.968245] ---[ end trace 0000000000000000 ]---
> 
> faddr2line gave me the following:
> nft_rhash_gc+0x208/0x2c0:
> __lse_atomic64_or at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./arch/arm64/include/asm/atomic_lse.h:132
> (inlined by) arch_atomic64_or at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./arch/arm64/include/asm/atomic.h:65
> (inlined by) raw_atomic64_or at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./include/linux/atomic/atomic-arch-fallback.h:3771
> (inlined by) raw_atomic_long_or at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./include/linux/atomic/atomic-long.h:1069
> (inlined by) arch_set_bit at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./include/asm-generic/bitops/atomic.h:18
> (inlined by) set_bit at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./include/asm-generic/bitops/instrumented-atomic.h:29
> (inlined by) nft_set_elem_dead at
> /root/gg/setup/linux-lts/src/linux-6.6.67/./include/net/netfilter/nf_tables.h:1576
> (inlined by) nft_rhash_gc at
> /root/gg/setup/linux-lts/src/linux-6.6.67/net/netfilter/nft_set_hash.c:375
> 
> By looking at the diff between 6.6.65 and 6.6.66 I was able to narrow
> it down to the above commit and I can confirm that reverting it fixes
> the issue.
> 
> Best
> -- 
> Amby Balaji
> Co-founder & CTO
> Hyperbeam, Inc.