iwlagn: fix the race in the unmapping of the HCMD

As Stanislaw pointed out, my patch

	iwlagn: fix a race in the unmapping of the TFDs

solved only part of the problem. The race still exists for TFDs of
the host commands. Fix that too.

Reported-by: Stanislaw Gruszka <sgruszka@redhat.com>
Signed-off-by: Emmanuel Grumbach <emmanuel.grumbach@intel.com>
Signed-off-by: Wey-Yi Guy <wey-yi.w.guy@intel.com>
Signed-off-by: John W. Linville <linville@tuxdriver.com>

diff --git a/drivers/net/wireless/iwlwifi/iwl-trans-pcie.c b/drivers/net/wireless/iwlwifi/iwl-trans-pcie.c
index 8e8c75c..da34110 100644 (file)
--- a/drivers/net/wireless/iwlwifi/iwl-trans-pcie.c
+++ b/drivers/net/wireless/iwlwifi/iwl-trans-pcie.c
@@ -407,6 +407,7 @@ static void iwl_tx_queue_unmap(struct iwl_trans *trans, int txq_id)
        struct iwl_queue *q = &txq->q;
        enum dma_data_direction dma_dir;
        unsigned long flags;
+       spinlock_t *lock;
 
        if (!q->n_bd)
                return;
@@ -414,19 +415,22 @@ static void iwl_tx_queue_unmap(struct iwl_trans *trans, int txq_id)
        /* In the command queue, all the TBs are mapped as BIDI
         * so unmap them as such.
         */
-       if (txq_id == trans->shrd->cmd_queue)
+       if (txq_id == trans->shrd->cmd_queue) {
                dma_dir = DMA_BIDIRECTIONAL;
-       else
+               lock = &trans->hcmd_lock;
+       } else {
                dma_dir = DMA_TO_DEVICE;
+               lock = &trans->shrd->sta_lock;
+       }
 
-       spin_lock_irqsave(&trans->shrd->sta_lock, flags);
+       spin_lock_irqsave(lock, flags);
        while (q->write_ptr != q->read_ptr) {
                /* The read_ptr needs to bound by q->n_window */
                iwlagn_txq_free_tfd(trans, txq, get_cmd_index(q, q->read_ptr),
                                    dma_dir);
                q->read_ptr = iwl_queue_inc_wrap(q->read_ptr, q->n_bd);
        }
-       spin_unlock_irqrestore(&trans->shrd->sta_lock, flags);
+       spin_unlock_irqrestore(lock, flags);
 }
 
 /**