Re: [PATCH] ipv4: guard inetdev_event() against a NULL from inetdev_init()

[Yunseong Kim]

Hi Eric,

Thanks for the review.

On 21/09/2026 6:09 pm, Eric Dumazet wrote:
> On Mon, Sep 21, 2026 at 5:58 PM Yunseong Kim <yunseong.kim@xxxxxxxx> wrote:
>>
>> On NETDEV_REGISTER, inetdev_event() only checks the inetdev_init()
>> return value with IS_ERR():
>>
>>     in_dev = inetdev_init(dev);
>>     if (IS_ERR(in_dev))
>>         return notifier_from_errno(PTR_ERR(in_dev));
>>     if (dev->flags & IFF_LOOPBACK) {
>>         IN_DEV_CONF_SET(in_dev, NOXFRM, 1);
>>     ...
>>
>> Because IS_ERR(NULL) is false, a NULL return flows straight into
>> IN_DEV_CONF_SET() -> ipv4_devconf_set(), whose
>> set_bit(index, in_dev->cnf.state) dereferences NULL.
>>
>> inetdev_init() is only kept from returning NULL by its trailing
>>
>>     out:
>>         return in_dev ?: ERR_PTR(err);
>>
>> which relies on every failure path that leaves in_dev NULL having also
>> set a non-zero err. That invariant is fragile and lives in the producer,
>> while the consumer's IS_ERR()-only check silently depends on it: any
>> future inetdev_init() failure path that returns NULL (directly, or by
>> leaving err == 0) becomes a NULL dereference at this call site rather
>> than a clean error return.
>>
>> Before commit 20e61da7ffcf ("ipv4: fail early when creating netdev
>> named all or default") this call site used "if (!in_dev)", which caught
>> a NULL return; that commit converted inetdev_init() to the ERR_PTR()
>> convention and switched the check to IS_ERR(), dropping the NULL
>> handling here.
>>
>> Decouple the caller from that invariant by using IS_ERR_OR_NULL() and
>> translating a NULL return to -ENOMEM, so a NULL can no longer be
>> dereferenced regardless of how inetdev_init() signals failure.
>>
>> Cc: stable@xxxxxxxxxxxxxxx
>> Cc: syzkaller-bugs@xxxxxxxxxxxxxxxx
>> Signed-off-by: Yunseong Kim <yunseong.kim@xxxxxxxx>
>> ---
>>
>> Syzkaller reproducer:
>> # {Threaded:true Repeat:true RepeatTimes:0 Procs:8 Slowdown:1 Sandbox:none SandboxArg:0 Leak:false NetInjection:false NetDevices:true NetReset:true Cgroups:true BinfmtMisc:true CloseFDs:true KCSAN:false DevlinkP
>> CI:false NicVF:false USB:false VhciInjection:false Wifi:false IEEE802154:false Sysctl:true Swap:true UseTmpDir:true HandleSegv:true Trace:false CallComments:true LegacyOptions:{Collide:false Fault:false FaultCal
>> l:0 FaultNth:0}}
>> r0 = creat(&(0x7f00000000c0)='./file0\x00', 0x26)
>> ioctl$RNDADDTOENTCNT(r0, 0x40045201, 0x0)
>> r1 = socket$inet6_udplite(0xa, 0x2, 0x88)
>> setsockopt$sock_int(r1, 0x1, 0x1d, &(0x7f0000000380), 0x4) (async, rerun: 64)
>> unshare(0x42020000) (async, rerun: 64)
>> truncate(&(0x7f0000000040)='./file0\x00', 0x9) (rerun: 64)
>> ioctl$sock_SIOCBRADDBR(0xffffffffffffffff, 0x89a0, &(0x7f0000000000)='syzkaller0\x00')
>> socket$inet_udplite(0x2, 0x2, 0x88)
>> rename(&(0x7f0000000640)='./file0aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
>> aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa\x00', &(0x7f0000000780)='./file1\x00') (async)
>> ioctl$BTRFS_IOC_BALANCE_PROGRESS(0xffffffffffffffff, 0x84009422, &(0x7f00000004c0)={0x0, 0x0, {0x0, @struct, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, @struct}, {}, {0x0, @struct}})
> 
> 
> inetdev_init() can not return NULL.

Okay, Thanks for pointing out.

> err is initialized to -ENOMEM, and the only place it is assigned is
> inside "if (err)" after devinet_sysctl_register(). Therefore every path
> reaching "return in_dev ?: ERR_PTR(err);" with in_dev == NULL has a
> non-zero err, and IS_ERR() is sufficient.

Ack, thanks again for the review. This issue is always reproducible, and I'm
currently trying to figure it out.

> Please do not send patches for bugs that can not happen, especially
> with a Cc: stable and no Fixes: tag.

Ack, I will take more time to why this bug happened.

> Also, the syzkaller reproducer you pasted has nothing to do with this
> code path, and there is no syzbot report or Reported-by: here. Please
> do not decorate patches with unrelated material.
> 
> If you used an LLM to generate this, please verify the claim before sending.

Oh, This isn't an LLM slob. My misunderstood the code problem. I didn't mean
to cause any extra burden by misusing the LLM. Sorry about that.

> See Documentation/process/researcher-guidelines.rst

Okay, I'll read the documentation more carefully.

> pw-bot: rejected

Thanks!

Best regards,
Yunseong