[RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Gustavo A. R. Silva]

Hi all,

While working on tranforming one-element array `peer_chan_list` in
`struct wmi_tdls_peer_capabilities` into a flex-array member

7187 struct wmi_tdls_peer_capabilities {
...
7199         struct wmi_channel peer_chan_list[1];
7200 } __packed;

the following line caught my attention:

./drivers/net/wireless/ath/ath10k/wmi.c:
8920         memset(skb->data, 0, sizeof(*cmd));

Notice that before the flex-array transformation, we are zeroing 128
bytes in `skb->data` because `sizeof(*cmd) == 128`, see below:

$ pahole -C wmi_10_4_tdls_peer_update_cmd drivers/net/wireless/ath/ath10k/wmi.o
struct wmi_10_4_tdls_peer_update_cmd {
	__le32                     vdev_id;              /*     0     4 */
	struct wmi_mac_addr        peer_macaddr;         /*     4     8 */
	__le32                     peer_state;           /*    12     4 */
	__le32                     reserved[4];          /*    16    16 */
	struct wmi_tdls_peer_capabilities peer_capab;    /*    32    96 */

	/* size: 128, cachelines: 2, members: 5 */
};

So, after the flex-array transformation (and the necessary adjustments
to a few other lines of code) we would be zeroing 104 bytes in
`skb->data` because `sizeof(*cmd) == 104`, see below:

$ pahole -C wmi_10_4_tdls_peer_update_cmd drivers/net/wireless/ath/ath10k/wmi.o
struct wmi_10_4_tdls_peer_update_cmd {
	__le32                     vdev_id;              /*     0     4 */
	struct wmi_mac_addr        peer_macaddr;         /*     4     8 */
	__le32                     peer_state;           /*    12     4 */
	__le32                     reserved[4];          /*    16    16 */
	struct wmi_tdls_peer_capabilities peer_capab;    /*    32    72 */

	/* size: 104, cachelines: 2, members: 5 */
	/* last cacheline: 40 bytes */
};

This difference arises because the size of the element type for the
`peer_chan_list` array, which is `sizeof(struct wmi_channel) == 24 `

$ pahole -C wmi_channel drivers/net/wireless/ath/ath10k/wmi.o
struct wmi_channel {
	__le32                     mhz;                  /*     0     4 */
	__le32                     band_center_freq1;    /*     4     4 */
	__le32                     band_center_freq2;    /*     8     4 */

[..]
                                                /*    20     4 */

	/* size: 24, cachelines: 1, members: 6 */
	/* last cacheline: 24 bytes */
};

is included in `sizeof(*cmd)` before the transformation.

So, my question is: do we really need to zero out those extra 24 bytes in
`skb->data`? or is it rather a bug in the original code?

Thanks!
--
Gustavo

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Johannes Berg]

On Tue, 2023-10-24 at 13:50 -0600, Gustavo A. R. Silva wrote:
> Hi all,
> 
> While working on tranforming one-element array `peer_chan_list` in
> `struct wmi_tdls_peer_capabilities` into a flex-array member
> 
> 7187 struct wmi_tdls_peer_capabilities {
> ...
> 7199         struct wmi_channel peer_chan_list[1];
> 7200 } __packed;
> 
> the following line caught my attention:
> 
> ./drivers/net/wireless/ath/ath10k/wmi.c:
> 8920         memset(skb->data, 0, sizeof(*cmd));
> 
> Notice that before the flex-array transformation, we are zeroing 128
> bytes in `skb->data` because `sizeof(*cmd) == 128`, see below:


> So, my question is: do we really need to zero out those extra 24 bytes in
> `skb->data`? or is it rather a bug in the original code?
> 

If we look a step further, I _think_ even that memset is unnecessary?


struct sk_buff *ath10k_wmi_alloc_skb(struct ath10k *ar, u32 len)
{
        struct sk_buff *skb;
        u32 round_len = roundup(len, 4);

        skb = ath10k_htc_alloc_skb(ar, WMI_SKB_HEADROOM + round_len);
        if (!skb)
                return NULL;

        skb_reserve(skb, WMI_SKB_HEADROOM);
        if (!IS_ALIGNED((unsigned long)skb->data, 4))
                ath10k_warn(ar, "Unaligned WMI skb\n");

        skb_put(skb, round_len);
        memset(skb->data, 0, round_len);

        return skb;
}


So shouldn't the outgoing skb be exactly the same?

Anyway, just looking at the code out of curiosity, I don't actually know
anything about this driver :)

johannes

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Gustavo A. R. Silva]

On 10/24/23 14:11, Johannes Berg wrote:
> On Tue, 2023-10-24 at 13:50 -0600, Gustavo A. R. Silva wrote:
>> Hi all,
>>
>> While working on tranforming one-element array `peer_chan_list` in
>> `struct wmi_tdls_peer_capabilities` into a flex-array member
>>
>> 7187 struct wmi_tdls_peer_capabilities {
>> ...
>> 7199         struct wmi_channel peer_chan_list[1];
>> 7200 } __packed;
>>
>> the following line caught my attention:
>>
>> ./drivers/net/wireless/ath/ath10k/wmi.c:
>> 8920         memset(skb->data, 0, sizeof(*cmd));
>>
>> Notice that before the flex-array transformation, we are zeroing 128
>> bytes in `skb->data` because `sizeof(*cmd) == 128`, see below:
> 
> 
>> So, my question is: do we really need to zero out those extra 24 bytes in
>> `skb->data`? or is it rather a bug in the original code?
>>
> 
> If we look a step further, I _think_ even that memset is unnecessary?

It seems we run into the same issue in the function below, even in the
case this `memset()` is unnecessary (which it seems it's not):

	8920         memset(skb->data, 0, sizeof(*cmd));

Notice that if `cap->peer_chan_len == 0` or `cap->peer_chan_len == 1`,
in the original code, we have `len == sizeof(*cmd) == 128`:

drivers/net/wireless/ath/ath10k/wmi.c:
8911         /* tdls peer update cmd has place holder for one channel*/
8912         chan_len = cap->peer_chan_len ? (cap->peer_chan_len - 1) : 0;
8913
8914         len = sizeof(*cmd) + chan_len * sizeof(*chan);
8915
8916         skb = ath10k_wmi_alloc_skb(ar, len);

> 
> 
> struct sk_buff *ath10k_wmi_alloc_skb(struct ath10k *ar, u32 len)
> {
>          struct sk_buff *skb;
>          u32 round_len = roundup(len, 4);
> 
>          skb = ath10k_htc_alloc_skb(ar, WMI_SKB_HEADROOM + round_len);
>          if (!skb)
>                  return NULL;
> 
>          skb_reserve(skb, WMI_SKB_HEADROOM);
>          if (!IS_ALIGNED((unsigned long)skb->data, 4))
>                  ath10k_warn(ar, "Unaligned WMI skb\n");
> 
>          skb_put(skb, round_len);

so `round_len == roundup(len, 4) == 128` at the moment of this
`memset()` call:

>          memset(skb->data, 0, round_len);

which take us back to the same problem, this time in the `memset()` above,
because after the flex-array transformation we would have:

--- a/drivers/net/wireless/ath/ath10k/wmi.c
+++ b/drivers/net/wireless/ath/ath10k/wmi.c
@@ -8905,13 +8905,10 @@ ath10k_wmi_10_4_gen_tdls_peer_update(struct ath10k *ar,
         struct wmi_channel *chan;
         struct sk_buff *skb;
         u32 peer_qos;
-       int len, chan_len;
+       size_t len;
         int i;

-       /* tdls peer update cmd has place holder for one channel*/
-       chan_len = cap->peer_chan_len ? (cap->peer_chan_len - 1) : 0;
-
-       len = sizeof(*cmd) + chan_len * sizeof(*chan);
+       len = struct_size(cmd, peer_capab.peer_chan_list, cap->peer_chan_len);

         skb = ath10k_wmi_alloc_skb(ar, len);
         if (!skb)

which makes `round_len == roundup(len, 4) == struct_size(cmd,...,...) == 104`
when `cap->peer_chan_len == 0`

> So shouldn't the outgoing skb be exactly the same?

It seems it's not.

> 
> Anyway, just looking at the code out of curiosity, I don't actually know
> anything about this driver :)
> 
> johannes

--
Gustavo

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Johannes Berg]

On Tue, 2023-10-24 at 14:41 -0600, Gustavo A. R. Silva wrote:
> 
> It seems we run into the same issue in the function below, even in the
> case this `memset()` is unnecessary (which it seems it's not):
> 
> 	8920         memset(skb->data, 0, sizeof(*cmd));
> 
> Notice that if `cap->peer_chan_len == 0` or `cap->peer_chan_len == 1`,
> in the original code, we have `len == sizeof(*cmd) == 128`:

Right.

> -       /* tdls peer update cmd has place holder for one channel*/
> -       chan_len = cap->peer_chan_len ? (cap->peer_chan_len - 1) : 0;
> -
> -       len = sizeof(*cmd) + chan_len * sizeof(*chan);
> +       len = struct_size(cmd, peer_capab.peer_chan_list, cap->peer_chan_len);
> 
>          skb = ath10k_wmi_alloc_skb(ar, len);
>          if (!skb)
> 
> which makes `round_len == roundup(len, 4) == struct_size(cmd,...,...) == 104`
> when `cap->peer_chan_len == 0`

And yeah, that's really the issue, it only matters for ==0. For a moment
there I thought that doesn't even make sense, but it looks like it never
even becomes non-zero.

No idea then, sorry. You'd hope firmware doesn't care about the actual
message size if the inner data says "0 entries", but who knows? And how
many firmware versions are there? :)

So I guess you'd want to stay compatible, even if it means having a

	chan_len = min(cap->peer_chan_len, 1);

for the struct_size()?

johannes

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Gustavo A. R. Silva]

[+CC Manikanta Pubbisetty ]

As Johannes[1] pointed out, this `memset()` is probably unnecessary:

./drivers/net/wireless/ath/ath10k/wmi.c:
8920         memset(skb->data, 0, sizeof(*cmd));

However, the same exact issue[2] is present at the line below inside
function `ath10k_wmi_alloc_skb()`:

drivers/net/wireless/ath/ath10k/wmi.c:
1799         memset(skb->data, 0, round_len);


Thanks
--
Gustavo

[1] https://lore.kernel.org/linux-hardening/26b15f4702cef17fe70b496a62f03735874bd16a.camel@sipsolutions.net/
[2] https://lore.kernel.org/linux-hardening/07e9bb04-f9fc-46d5-bfb9-a00a63a707c0@embeddedor.com/

On 10/24/23 13:50, Gustavo A. R. Silva wrote:
> Hi all,
> 
> While working on tranforming one-element array `peer_chan_list` in
> `struct wmi_tdls_peer_capabilities` into a flex-array member
> 
> 7187 struct wmi_tdls_peer_capabilities {
> ...
> 7199         struct wmi_channel peer_chan_list[1];
> 7200 } __packed;
> 
> the following line caught my attention:
> 
> ./drivers/net/wireless/ath/ath10k/wmi.c:
> 8920         memset(skb->data, 0, sizeof(*cmd));
> 
> Notice that before the flex-array transformation, we are zeroing 128
> bytes in `skb->data` because `sizeof(*cmd) == 128`, see below:
> 
> $ pahole -C wmi_10_4_tdls_peer_update_cmd drivers/net/wireless/ath/ath10k/wmi.o
> struct wmi_10_4_tdls_peer_update_cmd {
>      __le32                     vdev_id;              /*     0     4 */
>      struct wmi_mac_addr        peer_macaddr;         /*     4     8 */
>      __le32                     peer_state;           /*    12     4 */
>      __le32                     reserved[4];          /*    16    16 */
>      struct wmi_tdls_peer_capabilities peer_capab;    /*    32    96 */
> 
>      /* size: 128, cachelines: 2, members: 5 */
> };
> 
> So, after the flex-array transformation (and the necessary adjustments
> to a few other lines of code) we would be zeroing 104 bytes in
> `skb->data` because `sizeof(*cmd) == 104`, see below:
> 
> $ pahole -C wmi_10_4_tdls_peer_update_cmd drivers/net/wireless/ath/ath10k/wmi.o
> struct wmi_10_4_tdls_peer_update_cmd {
>      __le32                     vdev_id;              /*     0     4 */
>      struct wmi_mac_addr        peer_macaddr;         /*     4     8 */
>      __le32                     peer_state;           /*    12     4 */
>      __le32                     reserved[4];          /*    16    16 */
>      struct wmi_tdls_peer_capabilities peer_capab;    /*    32    72 */
> 
>      /* size: 104, cachelines: 2, members: 5 */
>      /* last cacheline: 40 bytes */
> };
> 
> This difference arises because the size of the element type for the
> `peer_chan_list` array, which is `sizeof(struct wmi_channel) == 24 `
> 
> $ pahole -C wmi_channel drivers/net/wireless/ath/ath10k/wmi.o
> struct wmi_channel {
>      __le32                     mhz;                  /*     0     4 */
>      __le32                     band_center_freq1;    /*     4     4 */
>      __le32                     band_center_freq2;    /*     8     4 */
> 
> [..]
>                                                 /*    20     4 */
> 
>      /* size: 24, cachelines: 1, members: 6 */
>      /* last cacheline: 24 bytes */
> };
> 
> is included in `sizeof(*cmd)` before the transformation.
> 
> So, my question is: do we really need to zero out those extra 24 bytes in
> `skb->data`? or is it rather a bug in the original code?
> 
> Thanks!
> -- 
> Gustavo
> 
> 

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Gustavo A. R. Silva]

On 10/24/23 14:49, Johannes Berg wrote:
> On Tue, 2023-10-24 at 14:41 -0600, Gustavo A. R. Silva wrote:
>>
>> It seems we run into the same issue in the function below, even in the
>> case this `memset()` is unnecessary (which it seems it's not):
>>
>> 	8920         memset(skb->data, 0, sizeof(*cmd));
>>
>> Notice that if `cap->peer_chan_len == 0` or `cap->peer_chan_len == 1`,
>> in the original code, we have `len == sizeof(*cmd) == 128`:
> 
> Right.
> 
>> -       /* tdls peer update cmd has place holder for one channel*/
>> -       chan_len = cap->peer_chan_len ? (cap->peer_chan_len - 1) : 0;
>> -
>> -       len = sizeof(*cmd) + chan_len * sizeof(*chan);
>> +       len = struct_size(cmd, peer_capab.peer_chan_list, cap->peer_chan_len);
>>
>>           skb = ath10k_wmi_alloc_skb(ar, len);
>>           if (!skb)
>>
>> which makes `round_len == roundup(len, 4) == struct_size(cmd,...,...) == 104`
>> when `cap->peer_chan_len == 0`
> 
> And yeah, that's really the issue, it only matters for ==0. For a moment
> there I thought that doesn't even make sense, but it looks like it never
> even becomes non-zero.
> 
> No idea then, sorry. You'd hope firmware doesn't care about the actual
> message size if the inner data says "0 entries", but who knows? And how
> many firmware versions are there? :)
> 
> So I guess you'd want to stay compatible, even if it means having a
> 
> 	chan_len = min(cap->peer_chan_len, 1);
> 
> for the struct_size()?

Yeah, that's an alternative.

I'll wait for the maintainers to chime in and see if they have a different
opinion.

Thanks
--
Gustavo

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Jeff Johnson]

On 10/24/2023 7:37 PM, Gustavo A. R. Silva wrote:
> 
> 
> On 10/24/23 14:49, Johannes Berg wrote:
>> On Tue, 2023-10-24 at 14:41 -0600, Gustavo A. R. Silva wrote:
>>>
>>> It seems we run into the same issue in the function below, even in the
>>> case this `memset()` is unnecessary (which it seems it's not):
>>>
>>>     8920         memset(skb->data, 0, sizeof(*cmd));
>>>
>>> Notice that if `cap->peer_chan_len == 0` or `cap->peer_chan_len == 1`,
>>> in the original code, we have `len == sizeof(*cmd) == 128`:
>>
>> Right.
>>
>>> -       /* tdls peer update cmd has place holder for one channel*/
>>> -       chan_len = cap->peer_chan_len ? (cap->peer_chan_len - 1) : 0;
>>> -
>>> -       len = sizeof(*cmd) + chan_len * sizeof(*chan);
>>> +       len = struct_size(cmd, peer_capab.peer_chan_list, 
>>> cap->peer_chan_len);
>>>
>>>           skb = ath10k_wmi_alloc_skb(ar, len);
>>>           if (!skb)
>>>
>>> which makes `round_len == roundup(len, 4) == struct_size(cmd,...,...) 
>>> == 104`
>>> when `cap->peer_chan_len == 0`
>>
>> And yeah, that's really the issue, it only matters for ==0. For a moment
>> there I thought that doesn't even make sense, but it looks like it never
>> even becomes non-zero.
>>
>> No idea then, sorry. You'd hope firmware doesn't care about the actual
>> message size if the inner data says "0 entries", but who knows? And how
>> many firmware versions are there? :)
>>
>> So I guess you'd want to stay compatible, even if it means having a
>>
>>     chan_len = min(cap->peer_chan_len, 1);
>>
>> for the struct_size()?
> 
> Yeah, that's an alternative.
> 
> I'll wait for the maintainers to chime in and see if they have a different
> opinion.

I'm seeing clarification from the development team.

/jeff

Re: [RFC - is this a bug?] wifi: ath10k: Asking for some light on this, please :)

[Jeff Johnson]

On 10/25/2023 8:52 AM, Jeff Johnson wrote:
> On 10/24/2023 7:37 PM, Gustavo A. R. Silva wrote:
>>
>>
>> On 10/24/23 14:49, Johannes Berg wrote:
>>> On Tue, 2023-10-24 at 14:41 -0600, Gustavo A. R. Silva wrote:
>>>>
>>>> It seems we run into the same issue in the function below, even in the
>>>> case this `memset()` is unnecessary (which it seems it's not):
>>>>
>>>>     8920         memset(skb->data, 0, sizeof(*cmd));
>>>>
>>>> Notice that if `cap->peer_chan_len == 0` or `cap->peer_chan_len == 1`,
>>>> in the original code, we have `len == sizeof(*cmd) == 128`:
>>>
>>> Right.
>>>
>>>> -       /* tdls peer update cmd has place holder for one channel*/
>>>> -       chan_len = cap->peer_chan_len ? (cap->peer_chan_len - 1) : 0;
>>>> -
>>>> -       len = sizeof(*cmd) + chan_len * sizeof(*chan);
>>>> +       len = struct_size(cmd, peer_capab.peer_chan_list, 
>>>> cap->peer_chan_len);
>>>>
>>>>           skb = ath10k_wmi_alloc_skb(ar, len);
>>>>           if (!skb)
>>>>
>>>> which makes `round_len == roundup(len, 4) == struct_size(cmd,...,...) 
>>>> == 104`
>>>> when `cap->peer_chan_len == 0`
>>>
>>> And yeah, that's really the issue, it only matters for ==0. For a moment
>>> there I thought that doesn't even make sense, but it looks like it never
>>> even becomes non-zero.
>>>
>>> No idea then, sorry. You'd hope firmware doesn't care about the actual
>>> message size if the inner data says "0 entries", but who knows? And how
>>> many firmware versions are there? :)
>>>
>>> So I guess you'd want to stay compatible, even if it means having a
>>>
>>>     chan_len = min(cap->peer_chan_len, 1);
>>>
>>> for the struct_size()?
>>
>> Yeah, that's an alternative.
>>
>> I'll wait for the maintainers to chime in and see if they have a different
>> opinion.
> 
> I'm seeing clarification from the development team.
> 
> /jeff
> 

I was not able to get a response from the firmware team.

I have gone ahead and created a series of patches to fix the remaining
flexible array issues in ath10k including the one discussed here. I
should be able to post those sometime this week.

/jeff