All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH nf v2 0/2] netfilter: nf_tables: only disable tables on a dormant transition
@ 2026-11-17 16:14 Ma Xinmeng
  2026-11-18 10:08 ` Xuanqiang Luo
  0 siblings, 1 reply; 2+ messages in thread
From: Ma Xinmeng @ 2026-11-17 16:14 UTC (permalink / raw)
  To: Pablo Neira Ayuso, Florian Westphal; +Cc: netfilter-devel, netdev, Phil Sutter

Hi,

This is v2 of the fix for the nftables WARN below. v1 was sent as
[PATCH net] and tracked a per-basechain "registered" flag. Xuanqiang
suggested fixing the commit condition instead: a dormant table never has
a registered basechain hook, and nf_tables_commit() should not call
nf_tables_table_disable() for a table that is already dormant.

v2 drops the new field and changes only the NEWTABLE commit path:

- if (table->flags & NFT_TABLE_F_DORMANT)
+ if (table->flags & __NFT_TABLE_F_WAS_AWAKEN)
      nf_tables_table_disable(net, table);

__NFT_TABLE_F_WAS_AWAKEN records a real awake-to-dormant transition.
Since table adoption added __NFT_TABLE_F_WAS_ORPHAN, __NFT_TABLE_F_UPDATE
can be set for an adoption that leaves the table dormant, so the old
check disabled the table and unregistered a hook that was never
registered.

The WARN was found by FuzzAnything's AI-assisted kernel fuzzer while
fuzzing nftables netlink messages in a user and network namespace, and
I reproduced it at runtime on 7.3.0-rc4, 7.3.1, 7.4-rc1, 7.4-rc2, and
7.4-rc3. The baseline 7.4-rc3 dmesg is:

  hook not found, pf 2 num 3
  WARNING: net/netfilter/core.c:522 at __nf_unregister_net_hook+0x71/0x1b0
  Call Trace:
    nft_table_disable+0x78/0xa0
    nf_tables_commit+0x1c03/0x1ec0
    nfnetlink_rcv_batch+0x939/0xaf0

v2 was tested on 7.4.0-rc3-c027fixv2: the PoC returns 0 and dmesg has no
WARN. I also checked a real awake-to-dormant transition: after the table
became dormant, a ping left the basechain counter at 0, so the hook was
still unregistered correctly.

Changes in v2:
- remove the per-basechain registered field
- check __NFT_TABLE_F_WAS_AWAKEN instead of NFT_TABLE_F_DORMANT
- add Fixes: 31bf508be656 ("netfilter: nf_tables: Implement table adoption support")
- use [PATCH nf] as requested
- add a selftest as 2/2

The selftest uses the nft owner/persist table flags to create the
dormant orphan, and checks the kernel taint state rather than parsing
dmesg. It fails on unpatched 7.4-rc3 and passes on v2.

Thanks,
Ma Xinmeng

Ma Xinmeng (2):
  netfilter: nf_tables: only disable tables on a dormant transition
  selftests: netfilter: test re-owning an already dormant table

 net/netfilter/nf_tables_api.c                 |  2 +-
 .../testing/selftests/net/netfilter/Makefile  |  1 +
 .../net/netfilter/nft_table_dormant_adopt.sh  | 57 +++++++++++++++++++
 3 files changed, 59 insertions(+), 1 deletion(-)
 create mode 100755 tools/testing/selftests/net/netfilter/nft_table_dormant_adopt.sh

-- 
2.56.0.windows.4.niugit


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH nf v2 0/2] netfilter: nf_tables: only disable tables on a dormant transition
  2026-11-17 16:14 [PATCH nf v2 0/2] netfilter: nf_tables: only disable tables on a dormant transition Ma Xinmeng
@ 2026-11-18 10:08 ` Xuanqiang Luo
  0 siblings, 0 replies; 2+ messages in thread
From: Xuanqiang Luo @ 2026-11-18 10:08 UTC (permalink / raw)
  To: Ma Xinmeng
  Cc: netfilter-devel, netdev, Phil Sutter, Pablo Neira Ayuso,
	Florian Westphal

On 2026/11/18 at 1:14 Ma Xinmeng wrote:
> Hi,
> 
> This is v2 of the fix for the nftables WARN below. v1 was sent as
> [PATCH net] and tracked a per-basechain "registered" flag. Xuanqiang
> suggested fixing the commit condition instead: a dormant table never has
> a registered basechain hook, and nf_tables_commit() should not call
> nf_tables_table_disable() for a table that is already dormant.
> 
> v2 drops the new field and changes only the NEWTABLE commit path:
> 
> - if (table->flags & NFT_TABLE_F_DORMANT)
> + if (table->flags & __NFT_TABLE_F_WAS_AWAKEN)
>        nf_tables_table_disable(net, table);
> 
> __NFT_TABLE_F_WAS_AWAKEN records a real awake-to-dormant transition.
> Since table adoption added __NFT_TABLE_F_WAS_ORPHAN, __NFT_TABLE_F_UPDATE
> can be set for an adoption that leaves the table dormant, so the old
> check disabled the table and unregistered a hook that was never
> registered.
> 
> The WARN was found by FuzzAnything's AI-assisted kernel fuzzer while
> fuzzing nftables netlink messages in a user and network namespace, and
> I reproduced it at runtime on 7.3.0-rc4, 7.3.1, 7.4-rc1, 7.4-rc2, and
> 7.4-rc3. The baseline 7.4-rc3 dmesg is:
> 
>    hook not found, pf 2 num 3
>    WARNING: net/netfilter/core.c:522 at __nf_unregister_net_hook+0x71/0x1b0
>    Call Trace:
>      nft_table_disable+0x78/0xa0
>      nf_tables_commit+0x1c03/0x1ec0
>      nfnetlink_rcv_batch+0x939/0xaf0
> 
> v2 was tested on 7.4.0-rc3-c027fixv2: the PoC returns 0 and dmesg has no
> WARN. I also checked a real awake-to-dormant transition: after the table
> became dormant, a ping left the basechain counter at 0, so the hook was
> still unregistered correctly.
> 
> Changes in v2:
> - remove the per-basechain registered field
> - check __NFT_TABLE_F_WAS_AWAKEN instead of NFT_TABLE_F_DORMANT
> - add Fixes: 31bf508be656 ("netfilter: nf_tables: Implement table adoption support")
> - use [PATCH nf] as requested
> - add a selftest as 2/2
> 

It would be nice to add a link to v1 here.

> The selftest uses the nft owner/persist table flags to create the
> dormant orphan, and checks the kernel taint state rather than parsing
> dmesg. It fails on unpatched 7.4-rc3 and passes on v2.
> 
> Thanks,
> Ma Xinmeng
> 
> Ma Xinmeng (2):
>    netfilter: nf_tables: only disable tables on a dormant transition
>    selftests: netfilter: test re-owning an already dormant table
> 
>   net/netfilter/nf_tables_api.c                 |  2 +-
>   .../testing/selftests/net/netfilter/Makefile  |  1 +
>   .../net/netfilter/nft_table_dormant_adopt.sh  | 57 +++++++++++++++++++
>   3 files changed, 59 insertions(+), 1 deletion(-)
>   create mode 100755 tools/testing/selftests/net/netfilter/nft_table_dormant_adopt.sh
> 


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-18 10:08 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-17 16:14 [PATCH nf v2 0/2] netfilter: nf_tables: only disable tables on a dormant transition Ma Xinmeng
2026-11-18 10:08 ` Xuanqiang Luo

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.