Re: [RFC] wifi: mwifiex: Asking for some light on this, please :)

From: Gustavo A. R. Silva
Date: Tue Aug 22 2023 - 15:23:18 EST


Hi Dan,

Thanks a lot for the feedback!

Please, see my comments below.

On 8/22/23 11:00, Dan Williams wrote:
On Tue, 2023-08-15 at 18:52 -0600, Gustavo A. R. Silva wrote:
Hi all,

While working on flex-array transformations I ran into the following
implementation:

drivers/net/wireless/marvell/mwifiex/fw.h:775:
struct mwifiex_ie_types_rxba_sync {
        struct mwifiex_ie_types_header header;
        u8 mac[ETH_ALEN];
        u8 tid;
        u8 reserved;
        __le16 seq_num;
        __le16 bitmap_len;
        u8 bitmap[1];
} __packed;

`bitmap` is currently being used as a fake-flex array and we should
transform it into a proper flexible-array member.

However, while doing that, I noticed something in the following function
that's not clear to me and I wanted to ask you for feedback:

drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c:907:
void mwifiex_11n_rxba_sync_event(struct mwifiex_private *priv,
                                 u8 *event_buf, u16 len)
{
        struct mwifiex_ie_types_rxba_sync *tlv_rxba = (void *)event_buf;
        u16 tlv_type, tlv_len;
        struct mwifiex_rx_reorder_tbl *rx_reor_tbl_ptr;
        u8 i, j;
        u16 seq_num, tlv_seq_num, tlv_bitmap_len;
        int tlv_buf_left = len;
        int ret;
        u8 *tmp;

        mwifiex_dbg_dump(priv->adapter, EVT_D, "RXBA_SYNC event:",
                         event_buf, len);
        while (tlv_buf_left >= sizeof(*tlv_rxba)) {

                tlv_type = le16_to_cpu(tlv_rxba->header.type);
                tlv_len  = le16_to_cpu(tlv_rxba->header.len);

                if (tlv_type != TLV_TYPE_RXBA_SYNC) {
                        mwifiex_dbg(priv->adapter, ERROR,
                                    "Wrong TLV id=0x%x\n", tlv_type);
                        return;
                }

                tlv_seq_num = le16_to_cpu(tlv_rxba->seq_num);
                tlv_bitmap_len = le16_to_cpu(tlv_rxba->bitmap_len);

This seems superfluous since couldn't the bitmap_len be calculated from
the tlv_len and sizeof(*tlv_rxba)? But whatever, sure.

Seems like there should be some input validation here to ensure that
tlv_bitmap_len and tlv_len don't overrun event_buf's memory though, if
the firmware is hosed or malicious.

But that's not your problem since you're not touching this code.

                mwifiex_dbg(priv->adapter, INFO,
                            "%pM tid=%d seq_num=%d bitmap_len=%d\n",
                            tlv_rxba->mac, tlv_rxba->tid, tlv_seq_num,
                            tlv_bitmap_len);

                rx_reor_tbl_ptr =
                        mwifiex_11n_get_rx_reorder_tbl(priv, tlv_rxba->tid,
                                                       tlv_rxba->mac);
                if (!rx_reor_tbl_ptr) {
                        mwifiex_dbg(priv->adapter, ERROR,
                                    "Can not find rx_reorder_tbl!");
                        return;
                }

                for (i = 0; i < tlv_bitmap_len; i++) {
                        for (j = 0 ; j < 8; j++) {
                                if (tlv_rxba->bitmap[i] & (1 << j)) {
                                        seq_num = (MAX_TID_VALUE - 1) &
                                                (tlv_seq_num + i * 8 + j);

                                        mwifiex_dbg(priv->adapter, ERROR,
                                                    "drop packet,seq=%d\n",
                                                    seq_num);

                                        ret = mwifiex_11n_rx_reorder_pkt
                                        (priv, seq_num, tlv_rxba->tid,
                                         tlv_rxba->mac, 0, NULL);

                                        if (ret)
                                                mwifiex_dbg(priv->adapter,
                                                            ERROR,
                                                            "Fail to drop packet");
                                }
                        }
                }

                tlv_buf_left -= (sizeof(*tlv_rxba) + tlv_len);

Now we have to subtract the size of the whole TLV (including the header
and flexarray) from the remaining bytes in event_buf.

But this looks pretty sketchy. Marvell TLVs have a header (the TL of
the TLV) and header->len says how long the V is. Most Marvell kernel
driver code (mwifiex, libertas, etc) does something like this:

pos += ssid_tlv->header + ssid_tlv->header.len;

but tlv_rxba is much more than just the header; I think this code is
going to *over* count how many bytes were just consumed.

I'm not the only one thinking it's sketchy:

https://www.spinics.net/lists/linux-wireless/msg174231.html

                tmp = (u8 *)tlv_rxba + tlv_len + sizeof(*tlv_rxba);
What's the relation between tlv_len, sizeof(*tlv_rxba) and tlv_bitmap_len?

Isn't `sizeof(*tlv_rxba) + tlv_len` and `tlv_len + sizeof(*tlv_rxba)`
double-counting some fields in `struct mwifiex_ie_types_rxba_sync`?

OK. So, based on your feedback, it seems that my assumptions were correct.

So, first I'll send the following fix:

diff --git a/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c b/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
index 391793a16adc..9eade3aa2918 100644
--- a/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
+++ b/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
@@ -965,8 +965,8 @@ void mwifiex_11n_rxba_sync_event(struct mwifiex_private *priv,
}
}

- tlv_buf_left -= (sizeof(*tlv_rxba) + tlv_len);
- tmp = (u8 *)tlv_rxba + tlv_len + sizeof(*tlv_rxba);
+ tlv_buf_left -= (sizeof(tlv_rxba->header) + tlv_len);
+ tmp = (u8 *)tlv_rxba + tlv_len + sizeof(tlv_rxba->header);
tlv_rxba = (struct mwifiex_ie_types_rxba_sync *)tmp;
}
}

Then, I'll do the flex-array transformation on top of the fix above:

diff --git a/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c b/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
index 9eade3aa2918..cb5a399cd56a 100644
--- a/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
+++ b/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
@@ -918,7 +918,7 @@ void mwifiex_11n_rxba_sync_event(struct mwifiex_private *priv,

mwifiex_dbg_dump(priv->adapter, EVT_D, "RXBA_SYNC event:",
event_buf, len);
- while (tlv_buf_left >= sizeof(*tlv_rxba)) {
+ while (tlv_buf_left > sizeof(*tlv_rxba)) {
tlv_type = le16_to_cpu(tlv_rxba->header.type);
tlv_len = le16_to_cpu(tlv_rxba->header.len);
if (tlv_type != TLV_TYPE_RXBA_SYNC) {
diff --git a/drivers/net/wireless/marvell/mwifiex/fw.h b/drivers/net/wireless/marvell/mwifiex/fw.h
index f2168fac95ed..8e6db904e5b2 100644
--- a/drivers/net/wireless/marvell/mwifiex/fw.h
+++ b/drivers/net/wireless/marvell/mwifiex/fw.h
@@ -779,7 +779,7 @@ struct mwifiex_ie_types_rxba_sync {
u8 reserved;
__le16 seq_num;
__le16 bitmap_len;
- u8 bitmap[1];
+ u8 bitmap[];
} __packed;

struct chan_band_param_set {

This happilly results in no binary output differences before/after changes. :)

Finally, to top it off, I can send this sanity check:

diff --git a/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c b/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
index cb5a399cd56a..237d0ee3573f 100644
--- a/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
+++ b/drivers/net/wireless/marvell/mwifiex/11n_rxreorder.c
@@ -929,6 +929,13 @@ void mwifiex_11n_rxba_sync_event(struct mwifiex_private *priv,

tlv_seq_num = le16_to_cpu(tlv_rxba->seq_num);
tlv_bitmap_len = le16_to_cpu(tlv_rxba->bitmap_len);
+ if (sizeof(*tlv_rxba) + tlv_bitmap_len > tlv_buf_left) {
+ mwifiex_dbg(priv->adapter, ERROR,
+ "TLV size (%ld) overflows event_buf (%d)\n",
+ sizeof(*tlv_rxba) + tlv_bitmap_len,
+ tlv_buf_left);
+ return;
+ }
mwifiex_dbg(priv->adapter, INFO,
"%pM tid=%d seq_num=%d bitmap_len=%d\n",
tlv_rxba->mac, tlv_rxba->tid, tlv_seq_num,

I wanted to used `sizeof(*tlv_rxba) + tlv_bitmap_len` here instead of
`sizeof(tlv_rxba->header) + tlv_len` to avoid any issues in case there
is any (buggy) discrepancy between `tlv_len` and `tlv_bitmap_len`.
This is when for some (weird) reason
`tlv_len - (sizeof(*tlv_rxba) - sizeof(tlv_rxba->header)) != tlv_bitmap_len`

What do you think?

Thanks!
--
Gustavo



Shouldn't this be something like this, instead (before the flex-array
transformation, of course):

-               tlv_buf_left -= (sizeof(*tlv_rxba) + tlv_len);
-               tmp = (u8 *)tlv_rxba + tlv_len + sizeof(*tlv_rxba);
+               tlv_buf_left -= (sizeof(*tlv_rxba) + tlv_bitmap_len - 1);
+               tmp = (u8 *)tlv_rxba + tlv_bitmap_len + sizeof(*tlv_rxba - 1);

If my assertion about tlv->header.len is correct then we can do:

tlv_buf_left -= sizeof(tlv_rxba->header) + tlv_len;
tmp = (u8 *)tlv_rxba + sizeof(tlv_rxba->header) + tlv_len;



                tlv_rxba = (struct mwifiex_ie_types_rxba_sync *)tmp;

This is silly; instead of tmp we could do:

u16 bytes_used;

...

bytes_used = sizeof(tlv_rxba->header) + tlv_len;
tlv_buf_left -= bytes_used;
tlv_rxba += bytes_used;

(with appropriate casting).

Dan

        }
}

Thanks in advance for any feedback!

--
Gustavo