Re: [PATCH ath-next v2 1/8] wifi: ath10k: fix SDIO aux workqueue starvation in recovery detection

[Kang Yang <kang.yang@oss.qualcomm.com>]

On 10/21/2026 9:33 PM, Loic Poulain wrote:
> On Wed, Oct 21, 2026 at 11:09 AM Kang Yang <kang.yang@oss.qualcomm.com> wrote:
>>
>> Recovery monitoring runs on workqueue_aux, which is also
>> used by SDIO RX dispatch and WMI SERVICE_READY handling.
>>
>> A single firmware crash may generate multiple recovery
>> triggers. Recovery detection currently handles each one by
>> waiting for recovery completion on workqueue_aux. On SDIO,
>> this blocks the same workqueue needed to deliver recovery
>> traffic, preventing recovery from completing.
>>
>> As a result, otherwise recoverable firmware crashes fail
>> with:
>>    ctl_resp never came in (-110)
>>
>> Treat duplicate recovery triggers as part of a single
>> recovery attempt by gating recovery startup in
>> ath10k_core_start_recovery() with an atomic_cmpxchg() on
>> pending_recovery, and by making the monitor path check the
>> elapsed time since the attempt started instead of blocking
>> on a completion.
>>
>> Tested-on: QCA6174 hw3.2 PCI WLAN.RM.4.4.1-00288-QCARMSWPZ-1
>> Tested-on: QCA6174 hw3.2 SDIO WLAN.RMH.4.4.1-00189
>>
>> Fixes: f35a07a4842a ("wifi: ath10k: move recovery check logic into a new work")
>> Signed-off-by: Kang Yang <kang.yang@oss.qualcomm.com>
>> ---
>>   drivers/net/wireless/ath/ath10k/core.c | 64 +++++++++++++++++++---------------
>>   drivers/net/wireless/ath/ath10k/core.h |  2 ++
>>   2 files changed, 37 insertions(+), 29 deletions(-)
>>
>> diff --git a/drivers/net/wireless/ath/ath10k/core.c b/drivers/net/wireless/ath/ath10k/core.c
>> index 7c2939cbde5f..07beefa06c4c 100644
>> --- a/drivers/net/wireless/ath/ath10k/core.c
>> +++ b/drivers/net/wireless/ath/ath10k/core.c
>> @@ -2494,46 +2494,52 @@ static int ath10k_init_hw_params(struct ath10k *ar)
>>   static void ath10k_core_recovery_check_work(struct work_struct *work)
>>   {
>>          struct ath10k *ar = container_of(work, struct ath10k, recovery_check_work);
>> -       long time_left;
>> +       unsigned long recovery_start = READ_ONCE(ar->recovery_start_jiffies);
>> +       int fail_count;
>>
>> -       /* Sometimes the recovery will fail and then the next all recovery fail,
>> -        * so avoid infinite recovery.
>> -        */
>> -       if (atomic_read(&ar->fail_cont_count) >= ATH10K_RECOVERY_MAX_FAIL_COUNT) {
>> -               ath10k_err(ar, "consecutive fail %d times, will shutdown driver!",
>> -                          atomic_read(&ar->fail_cont_count));
>> -               ar->state = ATH10K_STATE_WEDGED;
>> +       if (ar->state == ATH10K_STATE_WEDGED)
>>                  return;
>> -       }
>>
>> -       ath10k_dbg(ar, ATH10K_DBG_BOOT, "total recovery count: %d", ++ar->recovery_count);
>> +       /*
>> +        * Pure monitor: only ath10k_core_start_recovery() starts a fresh
>> +        * cycle. Bail if there is nothing in flight.
>> +        */
>> +       if (!atomic_read(&ar->pending_recovery))
>> +               return;
>>
>> -       if (atomic_read(&ar->pending_recovery)) {
>> -               /* Sometimes it happened another recovery work before the previous one
>> -                * completed, then the second recovery work will destroy the previous
>> -                * one, thus below is to avoid that.
>> -                */
>> -               time_left = wait_for_completion_timeout(&ar->driver_recovery,
>> -                                                       ATH10K_RECOVERY_TIMEOUT_HZ);
>> -               if (time_left) {
>> -                       ath10k_warn(ar, "previous recovery succeeded, skip this!\n");
>> -                       return;
>> -               }
>> +       if (!time_after(jiffies, recovery_start + ATH10K_RECOVERY_TIMEOUT_HZ))
> 
> You can use time_is_after_jiffies() for this.
> 
>> +               return;
>>
>> -               /* Record the continuous recovery fail count when recovery failed. */
>> -               atomic_inc(&ar->fail_cont_count);
>> +       /* Record the continuous recovery fail count when recovery is late. */
>> +       fail_count = atomic_inc_return(&ar->fail_cont_count);
>>
>> -               /* Avoid having multiple recoveries at the same time. */
>> -               return;
>> +       if (fail_count >= ATH10K_RECOVERY_MAX_FAIL_COUNT) {
>> +               ath10k_err(ar, "consecutive fail %d times, will shutdown driver!",
>> +                          fail_count);
>> +               ar->state = ATH10K_STATE_WEDGED;
>>          }
>> -
>> -       atomic_inc(&ar->pending_recovery);
>> -       queue_work(ar->workqueue, &ar->restart_work);
> 
> So we are not re-arming anymore for additional tries? or I miss something.
> 
>>   }
>>
>>   void ath10k_core_start_recovery(struct ath10k *ar)
>>   {
>> -       /* Use workqueue_aux to avoid blocking recovery tracking */
>> +       if (ar->state == ATH10K_STATE_WEDGED)
>> +               return;
>> +
>> +       /*
>> +        * SDIO raises several triggers within microseconds of one crash;
>> +        * only the first winner starts a fresh cycle and stamps the
>> +        * deadline. Losers do nothing while the owner runs recovery.
>> +        */
>> +       if (atomic_cmpxchg(&ar->pending_recovery, 0, 1) != 0)
>> +               return;
>> +       ath10k_dbg(ar, ATH10K_DBG_BOOT, "total recovery count: %d",
>> +                  ++ar->recovery_count);
>> +       /* Start a fresh failure budget; the last cycle may not have completed. */
>> +       atomic_set(&ar->fail_cont_count, 0);
>> +       WRITE_ONCE(ar->recovery_start_jiffies, jiffies);
>> +       queue_work(ar->workqueue, &ar->restart_work);
>> +
>> +       /* Start monitoring after publishing the recovery timestamp. */
>>          queue_work(ar->workqueue_aux, &ar->recovery_check_work);
> 
> So now that the recovery check is no longer blocking, it seems to me
> that this could be implemented as a simple timer (alternatively, a
> delayed work) as it would avoid having to perform the timeout
> calculation...
> 

Patch 5 will change it into delayed work, let me think if i need to 
merge patch 5 into patch 1.


>>   }
>>   EXPORT_SYMBOL(ath10k_core_start_recovery);
>> diff --git a/drivers/net/wireless/ath/ath10k/core.h b/drivers/net/wireless/ath/ath10k/core.h
>> index dfee432615eb..c739fbadc8bb 100644
>> --- a/drivers/net/wireless/ath/ath10k/core.h
>> +++ b/drivers/net/wireless/ath/ath10k/core.h
>> @@ -1215,6 +1215,8 @@ struct ath10k {
>>          unsigned int recovery_count;
>>          /* continuous recovery fail count */
>>          atomic_t fail_cont_count;
>> +       /* jiffies when the current recovery attempt started */
>> +       unsigned long recovery_start_jiffies;
>>
>>          /* cycle count is reported twice for each visited channel during scan.
>>           * access protected by data_lock
>>
>> --
>> 2.34.1
>>