ice: split ice_aq_wait_for_event() func into two

Mitigate race between registering on wait list and receiving
AQ Response from FW.

ice_aq_prep_for_event() should be called before sending AQ command,
ice_aq_wait_for_event() should be called after sending AQ command,
to wait for AQ Response.

Please note, that this was found by reading the code,
an actual race has not yet materialized.

Reviewed-by: Jacob Keller <[email protected]>
Signed-off-by: Przemek Kitszel <[email protected]>
Reviewed-by: Simon Horman <[email protected]>
Tested-by: Pucha Himasekhar Reddy <[email protected]> (A Contingent worker at Intel)
Reviewed-by: Leon Romanovsky <[email protected]>
Signed-off-by: Tony Nguyen <[email protected]>

Author: pkitszel

drivers/net/ethernet/intel/ice/ice.h

@@ -919,6 +919,7 @@ void ice_fdir_replay_fltrs(struct ice_pf *pf);
 int ice_fdir_create_dflt_rules(struct ice_pf *pf);
 
 enum ice_aq_task_state {
+	ICE_AQ_TASK_NOT_PREPARED,
 	ICE_AQ_TASK_WAITING,
 	ICE_AQ_TASK_COMPLETE,
 	ICE_AQ_TASK_CANCELED,
@@ -931,8 +932,10 @@ struct ice_aq_task {
 	u16 opcode;
 };
 
+void ice_aq_prep_for_event(struct ice_pf *pf, struct ice_aq_task *task,
+			   u16 opcode);
 int ice_aq_wait_for_event(struct ice_pf *pf, struct ice_aq_task *task,
-			  u16 opcode, unsigned long timeout);
+			  unsigned long timeout);
 int ice_open(struct net_device *netdev);
 int ice_open_internal(struct net_device *netdev);
 int ice_stop(struct net_device *netdev);

drivers/net/ethernet/intel/ice/ice_fw_update.c

@@ -302,6 +302,8 @@ ice_write_one_nvm_block(struct ice_pf *pf, u16 module, u32 offset,
 	dev_dbg(dev, "Writing block of %u bytes for module 0x%02x at offset %u\n",
 		block_size, module, offset);
 
+	ice_aq_prep_for_event(pf, &task, ice_aqc_opc_nvm_write);
+
 	err = ice_aq_update_nvm(hw, module, offset, block_size, block,
 				last_cmd, 0, NULL);
 	if (err) {
@@ -318,7 +320,7 @@ ice_write_one_nvm_block(struct ice_pf *pf, u16 module, u32 offset,
 	 * is conservative and is intended to prevent failure to update when
 	 * firmware is slow to respond.
 	 */
-	err = ice_aq_wait_for_event(pf, &task, ice_aqc_opc_nvm_write, 15 * HZ);
+	err = ice_aq_wait_for_event(pf, &task, 15 * HZ);
 	if (err) {
 		dev_err(dev, "Timed out while trying to flash module 0x%02x with block of size %u at offset %u, err %d\n",
 			module, block_size, offset, err);
@@ -491,6 +493,8 @@ ice_erase_nvm_module(struct ice_pf *pf, u16 module, const char *component,
 
 	devlink_flash_update_timeout_notify(devlink, "Erasing", component, ICE_FW_ERASE_TIMEOUT);
 
+	ice_aq_prep_for_event(pf, &task, ice_aqc_opc_nvm_erase);
+
 	err = ice_aq_erase_nvm(hw, module, NULL);
 	if (err) {
 		dev_err(dev, "Failed to erase %s (module 0x%02x), err %d aq_err %s\n",
@@ -501,7 +505,7 @@ ice_erase_nvm_module(struct ice_pf *pf, u16 module, const char *component,
 		goto out_notify_devlink;
 	}
 
-	err = ice_aq_wait_for_event(pf, &task, ice_aqc_opc_nvm_erase, ICE_FW_ERASE_TIMEOUT * HZ);
+	err = ice_aq_wait_for_event(pf, &task, ICE_FW_ERASE_TIMEOUT * HZ);
 	if (err) {
 		dev_err(dev, "Timed out waiting for firmware to respond with erase completion for %s (module 0x%02x), err %d\n",
 			component, module, err);
@@ -566,6 +570,8 @@ ice_switch_flash_banks(struct ice_pf *pf, u8 activate_flags,
 	u8 response_flags;
 	int err;
 
+	ice_aq_prep_for_event(pf, &task, ice_aqc_opc_nvm_write_activate);
+
 	err = ice_nvm_write_activate(hw, activate_flags, &response_flags);
 	if (err) {
 		dev_err(dev, "Failed to switch active flash banks, err %d aq_err %s\n",
@@ -590,8 +596,7 @@ ice_switch_flash_banks(struct ice_pf *pf, u8 activate_flags,
 		}
 	}
 
-	err = ice_aq_wait_for_event(pf, &task, ice_aqc_opc_nvm_write_activate,
-				    30 * HZ);
+	err = ice_aq_wait_for_event(pf, &task, 30 * HZ);
 	if (err) {
 		dev_err(dev, "Timed out waiting for firmware to switch active flash banks, err %d\n",
 			err);

drivers/net/ethernet/intel/ice/ice_main.c

@@ -1251,43 +1251,62 @@ ice_handle_link_event(struct ice_pf *pf, struct ice_rq_event_info *event)
 }
 
 /**
- * ice_aq_wait_for_event - Wait for an AdminQ event from firmware
+ * ice_aq_prep_for_event - Prepare to wait for an AdminQ event from firmware
  * @pf: pointer to the PF private structure
- * @task: ptr to task structure
+ * @task: intermediate helper storage and identifier for waiting
  * @opcode: the opcode to wait for
- * @timeout: how long to wait, in jiffies
  *
- * Waits for a specific AdminQ completion event on the ARQ for a given PF. The
- * current thread will be put to sleep until the specified event occurs or
- * until the given timeout is reached.
+ * Prepares to wait for a specific AdminQ completion event on the ARQ for
+ * a given PF. Actual wait would be done by a call to ice_aq_wait_for_event().
  *
- * To obtain only the descriptor contents, pass an event without an allocated
- * msg_buf. If the complete data buffer is desired, allocate the
- * event->msg_buf with enough space ahead of time.
+ * Calls are separated to allow caller registering for event before sending
+ * the command, which mitigates a race between registering and FW responding.
  *
- * Returns: zero on success, or a negative error code on failure.
+ * To obtain only the descriptor contents, pass an task->event with null
+ * msg_buf. If the complete data buffer is desired, allocate the
+ * task->event.msg_buf with enough space ahead of time.
  */
-int ice_aq_wait_for_event(struct ice_pf *pf, struct ice_aq_task *task,
-			  u16 opcode, unsigned long timeout)
+void ice_aq_prep_for_event(struct ice_pf *pf, struct ice_aq_task *task,
+			   u16 opcode)
 {
-	struct device *dev = ice_pf_to_dev(pf);
-	unsigned long start;
-	long ret;
-	int err;
-
 	INIT_HLIST_NODE(&task->entry);
 	task->opcode = opcode;
 	task->state = ICE_AQ_TASK_WAITING;
 
 	spin_lock_bh(&pf->aq_wait_lock);
 	hlist_add_head(&task->entry, &pf->aq_wait_list);
 	spin_unlock_bh(&pf->aq_wait_lock);
+}
 
-	start = jiffies;
+/**
+ * ice_aq_wait_for_event - Wait for an AdminQ event from firmware
+ * @pf: pointer to the PF private structure
+ * @task: ptr prepared by ice_aq_prep_for_event()
+ * @timeout: how long to wait, in jiffies
+ *
+ * Waits for a specific AdminQ completion event on the ARQ for a given PF. The
+ * current thread will be put to sleep until the specified event occurs or
+ * until the given timeout is reached.
+ *
+ * Returns: zero on success, or a negative error code on failure.
+ */
+int ice_aq_wait_for_event(struct ice_pf *pf, struct ice_aq_task *task,
+			  unsigned long timeout)
+{
+	enum ice_aq_task_state *state = &task->state;
+	struct device *dev = ice_pf_to_dev(pf);
+	unsigned long start = jiffies;
+	long ret;
+	int err;
 
-	ret = wait_event_interruptible_timeout(pf->aq_wait_queue, task->state,
+	ret = wait_event_interruptible_timeout(pf->aq_wait_queue,
+					       *state != ICE_AQ_TASK_WAITING,
 					       timeout);
-	switch (task->state) {
+	switch (*state) {
+	case ICE_AQ_TASK_NOT_PREPARED:
+		WARN(1, "call to %s without ice_aq_prep_for_event()", __func__);
+		err = -EINVAL;
+		break;
 	case ICE_AQ_TASK_WAITING:
 		err = ret < 0 ? ret : -ETIMEDOUT;
 		break;
@@ -1298,15 +1317,15 @@ int ice_aq_wait_for_event(struct ice_pf *pf, struct ice_aq_task *task,
 		err = ret < 0 ? ret : 0;
 		break;
 	default:
-		WARN(1, "Unexpected AdminQ wait task state %u", task->state);
+		WARN(1, "Unexpected AdminQ wait task state %u", *state);
 		err = -EINVAL;
 		break;
 	}
 
 	dev_dbg(dev, "Waited %u msecs (max %u msecs) for firmware response to op 0x%04x\n",
 		jiffies_to_msecs(jiffies - start),
 		jiffies_to_msecs(timeout),
-		opcode);
+		task->opcode);
 
 	spin_lock_bh(&pf->aq_wait_lock);
 	hlist_del(&task->entry);
@@ -1342,7 +1361,9 @@ static void ice_aq_check_events(struct ice_pf *pf, u16 opcode,
 
 	spin_lock_bh(&pf->aq_wait_lock);
 	hlist_for_each_entry(task, &pf->aq_wait_list, entry) {
-		if (task->state || task->opcode != opcode)
+		if (task->state != ICE_AQ_TASK_WAITING)
+			continue;
+		if (task->opcode != opcode)
 			continue;
 
 		task_ev = &task->event;