Hi,
On 27/03/2022 09:27, Xiaomeng Tong wrote:
> The bug is here:
> rcu_assign_pointer(ar->tx_ampdu_iter,
> (struct carl9170_sta_tid *) &ar->tx_ampdu_list);
yeah, so... I know there's currently a big discussion revolving
around LISTs due to incoming the GNU89 to GNU11 switch. I'm not
currently aware that something related to this had updated
INIT_LIST_HEAD + friends. So, please tell me if there is extra
information that has to be considered.
> The 'ar->tx_ampdu_iter' is used as a list iterator variable
> which point to a structure object containing the list HEAD
> (&ar->tx_ampdu_list), not as the HEAD itself.
>
> The only use case of 'ar->tx_ampdu_iter' is as a base pos
> for list_for_each_entry_continue_rcu in carl9170_tx_ampdu().
> If the iterator variable holds the *wrong* HEAD value here
> (has not been modified elsewhere before), this will lead to
> an invalid memory access.
>
> Using list_entry_rcu to get the right list iterator variable
> and reassign it, to fix this bug.
> Note: use 'ar->tx_ampdu_list.next' instead of '&ar->tx_ampdu_list'
> to avoid compiler error.
>
> Cc: [email protected]
> Fixes: fe8ee9ad80b28 ("carl9170: mac80211 glue and command interface")
> Signed-off-by: Xiaomeng Tong <[email protected]>
> ---
> drivers/net/wireless/ath/carl9170/main.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/wireless/ath/carl9170/main.c b/drivers/net/wireless/ath/carl9170/main.c
> index 49f7ee1c912b..a287937bf666 100644
> --- a/drivers/net/wireless/ath/carl9170/main.c
> +++ b/drivers/net/wireless/ath/carl9170/main.c
> @@ -1756,6 +1756,7 @@ static const struct ieee80211_ops carl9170_ops = {
>
> void *carl9170_alloc(size_t priv_size)
> {
> + struct carl9170_sta_tid *tid_info;
> struct ieee80211_hw *hw;
> struct ar9170 *ar;
> struct sk_buff *skb;
> @@ -1815,8 +1816,9 @@ void *carl9170_alloc(size_t priv_size)
> INIT_DELAYED_WORK(&ar->stat_work, carl9170_stat_work);
> INIT_DELAYED_WORK(&ar->tx_janitor, carl9170_tx_janitor);
> INIT_LIST_HEAD(&ar->tx_ampdu_list);
> - rcu_assign_pointer(ar->tx_ampdu_iter,
> - (struct carl9170_sta_tid *) &ar->tx_ampdu_list);
> + tid_info = list_entry_rcu(ar->tx_ampdu_list.next,
> + struct carl9170_sta_tid, list);
> + rcu_assign_pointer(ar->tx_ampdu_iter, tid_info);
I've tested this. I've added the following pr_info that would
print the (raw) pointer of both your new method (your patch)
and the old (current code) one:
pr_info("new:%px\n", list_entry_rcu(ar->tx_ampdu_list.next,struct carl9170_sta_tid, list)); // tid_info
pr_info("old:%px\n", (struct carl9170_sta_tid *) &ar->tx_ampdu_list);
and run it on AR9170 USB Stick
[ 216.547932] usb 2-10: SerialNumber: 12345
[ 216.673629] usb 2-10: reset high-speed USB device number 10 using xhci_hcd
[ 216.853488] new:ffff9394268a38e0
[ 216.853496] old:ffff9394268a38e0
[ 216.858174] usb 2-10: driver API: 1.9.9 2016-02-15 [1-1]
[ 216.858186] usb 2-10: firmware API: 1.9.9 2021-02-05
phew, what a relieve :). Both the new and old pointers are the same.
So, the tx_ampdu_list is empty, as it was just initialized to
(list->next = list->prev = list).
And you are right about the iter being suspeciously bogus. But I think
this is true for both the new and the old way. There is no real
carl9170_sta_tid* tid associated with that empty entry and if some code
would expect a valid carl9170_sta_tid* there, it would certainly cause
crashes&burns.
The carl9170_tx_ampdu() and carl9170_ampdu_gc() code is really
careful though and checks whenever the list is empty or not
before doing any list traversing with the tx_ampdu_iter.
Any thoughts or insights?
Cheers,
Christian