Re: [PATCH net v2] net: zero the tail of ha->addr when allocating address list entries
From: Eric Dumazet
Date: Thu Nov 19 2026 - 03:29:38 EST
Le jeu. 19 nov. 2026 à 05:02, Henry Martin <bsdhenrymartin@xxxxxxxxx> a écrit :
>
> __hw_addr_create() allocates entries with kmalloc() and initializes only
> addr_len bytes of ha->addr. The dump paths emit each entry using the
> current dev->addr_len, so an entry that was created with a smaller
> addr_len is dumped with uninitialized tail bytes.
>
> dev->addr_len can grow after the entry was created: tun's TUNSETLINK
> changes a device from ARPHRD_NONE (addr_len 0) to ARPHRD_IP6GRE
> (addr_len 16). On v7.4-rc3, with 64 entries armed on a down tun device
> and the link type then set to ARPHRD_IP6GRE, RTM_GETMULTICAST returned
> 16 bytes per entry of which 37 were non-zero and 21 started with a
> kernel heap pointer (e.g. 0xffff8188_04857ee0); /proc/net/dev_mcast
> exposed the same bytes. This discloses stale heap contents to readers
> of those interfaces.
How did you get 64 entries on one device while addr_len was 0 ?
__hw_addr_add_ex() does memcmp(addr, ha->addr, 0), so every add of the
same type finds the existing node and only bumps its refcount.
Please describe what was really run (64 devices ? another starting
link type ?)
>
> Allocate with kzalloc(), so the tail is always zeroed and a dump can
> only emit bytes that were written or zeroed at creation. This also
> makes the explicit ha->sync_cnt = 0 initialization redundant.
>
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: a748ee242681 ("net: move address list functions to a separate file")
This commit only moved code around.
For tun, the issue became reachable with
cca8ea3b05c9 ("net: tun: set tun->dev->addr_len during TUNSETLINK processing")
> Link: https://lists.openwall.net/linux-kernel/2026/09/10/320
Please point to your v1 instead:
https://lore.kernel.org/netdev/20261116024056.1274402-1-bsdhenrymartin@xxxxxxxxx/
Also please mention the companion patch, which flushes dev->uc and
dev->mc in TUNSETLINK when addr_len changes:
https://lore.kernel.org/netdev/20261118125443.1137177-1-edumazet@xxxxxxxxxx/
Your patch is still useful with it: tun is not the only driver
changing dev->addr_len without flushing these lists (qmi_wwan raw_ip,
hdlc protocol attach/detach), so the changelog should not present tun
as the only case.
Also, the bot asked how the issue was discovered, please include this
in the changelog.
The code change itself looks good to me, please send a v3 with
the changelog fixed (after the usual 24 hours).
Thanks.
pw-bot: cr
> Signed-off-by: Henry Martin <bsdhenrymartin@xxxxxxxxx>
> Reviewed-by: Daniel Machon <daniel.machon@xxxxxxxxxxxxx>
> ---
> v1 -> v2:
> - allocate with kzalloc() and drop the manual ha->sync_cnt = 0, as
> suggested by Eric Dumazet and Daniel Machon
> - add the runtime observation to the changelog
>
> net/core/dev_addr_lists.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/net/core/dev_addr_lists.c b/net/core/dev_addr_lists.c
> index d827cc21867b..9613c8a9ed7f 100644
> --- a/net/core/dev_addr_lists.c
> +++ b/net/core/dev_addr_lists.c
> @@ -96,7 +96,7 @@ __hw_addr_create(const unsigned char *addr, int addr_len,
> alloc_size = sizeof(*ha);
> if (alloc_size < L1_CACHE_BYTES)
> alloc_size = L1_CACHE_BYTES;
> - ha = kmalloc(alloc_size, GFP_ATOMIC);
> + ha = kzalloc(alloc_size, GFP_ATOMIC);
> if (!ha)
> return NULL;
> memcpy(ha->addr, addr, addr_len);
> @@ -104,7 +104,6 @@ __hw_addr_create(const unsigned char *addr, int addr_len,
> ha->refcount = 1;
> ha->global_use = global;
> ha->synced = sync ? 1 : 0;
> - ha->sync_cnt = 0;
>
> return ha;
> }
> --
> 2.43.7
>