Re: [PATCH net v2] netfilter: flowtable: flush delete work after final GC

[Pablo Neira Ayuso]

On Mon, Sep 28, 2026 at 12:32:48AM +0800, Chengfeng Ye wrote:
> nf_flow_table_free() can return while delete work still holds a pointer to
> the flowtable. Its caller can then free the flowtable before the worker
> accesses it, causing a use-after-free.
> 
> nf_flow_offload_del() sets NF_FLOW_HW_DYING only after allocating the work
> item. If this GFP_ATOMIC allocation fails during the first teardown GC
> pass, the flow remains eligible for deletion in the final GC pass inside
> nf_flow_table_offload_flush_cleanup(). That pass runs after the delete
> workqueue has been flushed, so a successful retry queues work which is
> not waited for:
> 
>   teardown worker                         delete worker
>   first GC: work allocation fails
>     NF_FLOW_HW_DYING remains clear
>   flush delete workqueue
>   final GC: allocation succeeds
>     queue FLOW_CLS_DESTROY work
>   destroy rhashtable
>   free flowtable
>                                           access offload->flowtable
> 
> Flush the delete workqueue again after the final GC to complete this work
> before the flowtable can be freed. All flows have already been marked for
> teardown, so this GC pass only queues delete work, and the delete worker
> does not queue further offload work.
> 
> KASAN reported:
> 
>   BUG: KASAN: slab-use-after-free in flow_offload_work_handler+0xbe8/0xe30
>   Read of size 8 at addr ffff888109c9fd98 by task kworker/u16:3/397
>   Workqueue: nf_ft_offload_del flow_offload_work_handler
>   Call Trace:
>    flow_offload_work_handler+0xbe8/0xe30
>    process_one_work+0x63a/0x1070
>    worker_thread+0x45b/0xd10
>   Allocated by task 87:
>    nf_tables_newflowtable+0x5d0/0x22f0
>    nfnetlink_rcv_batch+0x1396/0x1d00
>   Freed by task 11:
>    kfree+0x131/0x3c0
>    nf_tables_trans_destroy_work+0xb26/0xeb0
>    process_one_work+0x63a/0x1070
>   Last potentially related work creation:
>    __queue_work+0x68e/0x1030
>    flow_offload_del+0x74c/0xad0
>    nf_flow_offload_gc_step+0x264/0x8e0
>    nf_flow_table_gc_run+0xcd/0x150
>    nf_flow_table_offload_flush_cleanup+0x5c/0x70
>    nf_flow_table_free+0x280/0x350
>    nf_tables_flowtable_destroy+0x71/0x270

For the record, proposed alternative to this patch:

https://patch.msgid.link/netfilter-devel/20261029230657.302417-1-pablo@xxxxxxxxxxxxx/

> Fixes: c921ffe85333 ("netfilter: flowtable: Fix flushing of offloaded flows on free")
> Cc: stable@xxxxxxxxxxxxxxx
> Assisted-by: GPT-6-Astra
> Signed-off-by: Chengfeng Ye <nicoyip.dev@xxxxxxxxx>
> ---
> Changes in v2:
> - Rebase onto current mainline; the independently revalidated one-line fix
>   is unchanged from v1.
> 
> v1: https://lore.kernel.org/r/20260824115829.205118-1-nicoyip.dev@xxxxxxxxx/
> 
>  net/netfilter/nf_flow_table_offload.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/net/netfilter/nf_flow_table_offload.c b/net/netfilter/nf_flow_table_offload.c
> index 6757fd89c1f1..728220ba3882 100644
> --- a/net/netfilter/nf_flow_table_offload.c
> +++ b/net/netfilter/nf_flow_table_offload.c
> @@ -1172,6 +1172,7 @@ void nf_flow_table_offload_flush_cleanup(struct nf_flowtable *flowtable)
>  	if (nf_flowtable_hw_offload(flowtable)) {
>  		flush_workqueue(nf_flow_offload_del_wq);
>  		nf_flow_table_gc_run(flowtable);
> +		flush_workqueue(nf_flow_offload_del_wq);
>  	}
>  }
>  
> -- 
> 2.43.0
>