drivers/net/wireless/ath/ath12k/mac.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
ath12k_mac_setup_channels_rates() frees band channel arrays on failure
but either leaves the pointers non-NULL or clears the wrong band. Later
ath12k_mac_cleanup_unregister() frees the same pointers again.
Clear the correct sbands[].channels pointers after kfree(), including
a copy-paste bug that nulled 2 GHz after freeing 6 GHz channels.
Fixes: acc152f9be20 ("wifi: ath12k: combine channel list for split-phy devices in single-wiphy")
Signed-off-by: Linkai Gong <gonglinkai@kylinos.cn>
---
drivers/net/wireless/ath/ath12k/mac.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/drivers/net/wireless/ath/ath12k/mac.c b/drivers/net/wireless/ath/ath12k/mac.c
index a0928890671a..5468a8d2d5d5 100644
--- a/drivers/net/wireless/ath/ath12k/mac.c
+++ b/drivers/net/wireless/ath/ath12k/mac.c
@@ -14275,6 +14275,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
sizeof(ath12k_6ghz_channels), GFP_KERNEL);
if (!channels) {
kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
+ ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
return -ENOMEM;
}
@@ -14325,7 +14326,9 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
GFP_KERNEL);
if (!channels) {
kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
+ ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
+ ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
return -ENOMEM;
}
@@ -14365,7 +14368,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
- ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
+ ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
kfree(channels);
band->channels = NULL;
return ret;
--
2.25.1On 7/31/2026 3:08 PM, Linkai Gong wrote:
> ath12k_mac_setup_channels_rates() frees band channel arrays on failure
> but either leaves the pointers non-NULL or clears the wrong band. Later
> ath12k_mac_cleanup_unregister() frees the same pointers again.
>
> Clear the correct sbands[].channels pointers after kfree(), including
> a copy-paste bug that nulled 2 GHz after freeing 6 GHz channels.
>
> Fixes: acc152f9be20 ("wifi: ath12k: combine channel list for split-phy devices in single-wiphy")
> Signed-off-by: Linkai Gong <gonglinkai@kylinos.cn>
> ---
> drivers/net/wireless/ath/ath12k/mac.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/wireless/ath/ath12k/mac.c b/drivers/net/wireless/ath/ath12k/mac.c
> index a0928890671a..5468a8d2d5d5 100644
> --- a/drivers/net/wireless/ath/ath12k/mac.c
> +++ b/drivers/net/wireless/ath/ath12k/mac.c
> @@ -14275,6 +14275,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> sizeof(ath12k_6ghz_channels), GFP_KERNEL);
> if (!channels) {
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> return -ENOMEM;
> }
>
> @@ -14325,7 +14326,9 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> GFP_KERNEL);
> if (!channels) {
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
> return -ENOMEM;
> }
>
Above two hunks seem to fix older allocation-failure paths which
already existed before acc152f9be20. In current history those lines
blame back to d889913205cf ("wifi: ath12k: driver for Qualcomm IEEE
802.11ax devices").
So this probably should either be split into two fixes, or the commit
message/Fixes tag should be adjusted.
But unless there is another reachable path where these stale pointers
are later freed or dereferenced, do we really need a Fixes tag here?
> @@ -14365,7 +14368,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
> - ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> + ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
> kfree(channels);
> band->channels = NULL;
> return ret;
--
Ramesh
On Mon, Aug 10, 2026 at 03:44:54PM +0530, Rameshkumar Sundaram wrote: > So this probably should either be split into two fixes, or the commit > message/Fixes tag should be adjusted. > > But unless there is another reachable path where these stale pointers > are later freed or dereferenced, do we really need a Fixes tag here? Thanks. Agree on both points , the first two hunks predate acc152f9be20 , and with Baochen's clarification there is no reachable second free, so a Fixes tag does not make sense. v2 will drop Fixes and keep this as one small hardening patch. Thanks, Linkai
On 7/31/2026 5:38 PM, Linkai Gong wrote:
> ath12k_mac_setup_channels_rates() frees band channel arrays on failure
> but either leaves the pointers non-NULL or clears the wrong band. Later
> ath12k_mac_cleanup_unregister() frees the same pointers again.
The radio that frees its sbands[].channels inside setup_channels_rates() is precisely the
one excluded from cleanup_unregister() — the loop is for (j = 0; j < i; j++) and the
failing radio is index i. The earlier radios (0..i-1) all fully succeeded, so their
channel arrays are valid and freed exactly once. So this is actually a dangling pointer
issue, not a reachable double-free — nothing frees the failing radio's pointers a second
time. Please reword the message to something like "error paths leave dangling pointers + a
copy-paste typo; harden by consistently nulling after kfree()."
>
> Clear the correct sbands[].channels pointers after kfree(), including
> a copy-paste bug that nulled 2 GHz after freeing 6 GHz channels.
>
> Fixes: acc152f9be20 ("wifi: ath12k: combine channel list for split-phy devices in single-wiphy")
> Signed-off-by: Linkai Gong <gonglinkai@kylinos.cn>
> ---
> drivers/net/wireless/ath/ath12k/mac.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/wireless/ath/ath12k/mac.c b/drivers/net/wireless/ath/ath12k/mac.c
> index a0928890671a..5468a8d2d5d5 100644
> --- a/drivers/net/wireless/ath/ath12k/mac.c
> +++ b/drivers/net/wireless/ath/ath12k/mac.c
> @@ -14275,6 +14275,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> sizeof(ath12k_6ghz_channels), GFP_KERNEL);
> if (!channels) {
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> return -ENOMEM;
> }
>
> @@ -14325,7 +14326,9 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> GFP_KERNEL);
> if (!channels) {
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
> return -ENOMEM;
> }
>
> @@ -14365,7 +14368,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
> - ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> + ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
> kfree(channels);
> band->channels = NULL;
> return ret;
actual code change looks good
On Mon, Aug 10, 2026 at 03:07:25PM +0530, Baochen Qiang wrote: > So this is actually a dangling pointer > issue, not a reachable double-free — nothing frees the failing radio's pointers a second > time. Please reword the message to something like "error paths leave dangling pointers + a > copy-paste typo; harden by consistently nulling after kfree()." Thanks. You are right, I misread the cleanup loop. Will reword the commit message in v2. Code change stays the same. Thanks, Linkai
On 7/31/2026 2:38 AM, Linkai Gong wrote:
> ath12k_mac_setup_channels_rates() frees band channel arrays on failure
> but either leaves the pointers non-NULL or clears the wrong band. Later
> ath12k_mac_cleanup_unregister() frees the same pointers again.
>
> Clear the correct sbands[].channels pointers after kfree(), including
> a copy-paste bug that nulled 2 GHz after freeing 6 GHz channels.
https://wireless.docs.kernel.org/en/latest/en/users/drivers/ath12k/submittingpatches.html#tested-on-tag
Tested-on: tag missing. How was this tested?
>
> Fixes: acc152f9be20 ("wifi: ath12k: combine channel list for split-phy devices in single-wiphy")
> Signed-off-by: Linkai Gong <gonglinkai@kylinos.cn>
> ---
> drivers/net/wireless/ath/ath12k/mac.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/net/wireless/ath/ath12k/mac.c b/drivers/net/wireless/ath/ath12k/mac.c
> index a0928890671a..5468a8d2d5d5 100644
> --- a/drivers/net/wireless/ath/ath12k/mac.c
> +++ b/drivers/net/wireless/ath/ath12k/mac.c
> @@ -14275,6 +14275,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> sizeof(ath12k_6ghz_channels), GFP_KERNEL);
> if (!channels) {
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> return -ENOMEM;
> }
>
> @@ -14325,7 +14326,9 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> GFP_KERNEL);
> if (!channels) {
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
> + ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
> return -ENOMEM;
> }
>
> @@ -14365,7 +14368,7 @@ static int ath12k_mac_setup_channels_rates(struct ath12k *ar,
> kfree(ar->mac.sbands[NL80211_BAND_2GHZ].channels);
> ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> kfree(ar->mac.sbands[NL80211_BAND_6GHZ].channels);
> - ar->mac.sbands[NL80211_BAND_2GHZ].channels = NULL;
> + ar->mac.sbands[NL80211_BAND_6GHZ].channels = NULL;
> kfree(channels);
> band->channels = NULL;
> return ret;
On Thu, Aug 06, 2026 at 18:47:38 -0700, Jeff Johnson wrote: > https://wireless.docs.kernel.org/en/latest/en/users/drivers/ath12k/submittingpatches.html#tested-on-tag > Tested-on: tag missing. How was this tested? Thanks for the review. This issue was found by code review of the error paths in ath12k_mac_setup_channels_rates(). I do not have QCN9274/WCN7850 hardware, so I could not provide a Tested-on tag or reproduce the ENOMEM double-free on a device. The bug is: after kfree() of sbands[].channels on failure, the pointer is left non-NULL (or the wrong band is cleared), and ath12k_mac_cleanup_unregister() frees it again. I understand Tested-on is required for ath12k. I will follow up with a v2 including Tested-on once I can test on QCN9274/WCN7850 hardware. Meanwhile, any Tested-by from someone with the hardware would be appreciated. Thanks, Linkai
On 8/6/2026 7:56 PM, Linkai Gong wrote: > On Thu, Aug 06, 2026 at 18:47:38 -0700, Jeff Johnson wrote: >> https://wireless.docs.kernel.org/en/latest/en/users/drivers/ath12k/submittingpatches.html#tested-on-tag >> Tested-on: tag missing. How was this tested? > > Thanks for the review. > > This issue was found by code review of the error paths in > ath12k_mac_setup_channels_rates(). I do not have QCN9274/WCN7850 > hardware, so I could not provide a Tested-on tag or reproduce the > ENOMEM double-free on a device. > > The bug is: after kfree() of sbands[].channels on failure, the > pointer is left non-NULL (or the wrong band is cleared), and > ath12k_mac_cleanup_unregister() frees it again. > > I understand Tested-on is required for ath12k. I will follow up > with a v2 including Tested-on once I can test on QCN9274/WCN7850 > hardware. Meanwhile, any Tested-by from someone with the hardware > would be appreciated. It is OK to specify the patch was compiled tested only. I just need that as part of the commit text so that the level of testing is documented. The ath10k documentation was clear about this: For trivial patches it’s ok to skip Tested-on tag but then it should say “Compile tested only”. Not sure why that was dropped as part of the ath11k and ath12k documentation (that predates me becoming a maintainer). /jeff
© 2016 - 2026 Red Hat, Inc.