bnx2x: Create workqueue for IOV related tasks

The bnx2x sriov mechanisms were done in the bnx2x slowpath workitem which
runs on the bnx2x's workqueue; This workitem is also responsible for the bottom
half of interrupt handling in the driver, and specifically it also receives
FW notifications of ramrod completions, allowing other flows to progress.

The original design of the sriov reltaed-flows was based on the notion such
flows must not sleep, since their context is the slowpath workitem.
Otherwise, we might reach timeouts - those flows may wait for ramrod completion
that will never arrive as the workitem wlll not be re-scheduled until that same
flow will be over.

In more recent time bnx2x started supporting features in which the VF interface
can be configured by the tools accessing the PF on the hypervisor.
This support created possible races on the VF-PF lock (which is taken either
when the PF is handling a VF message or when the PF is doing some slowpath work
on behalf of the VF) which may cause timeouts on the VF side and lags on the PF
side.

This patch changes the scheme - it creates a new workqueue for sriov related
tasks and moves all handling currently done in the slowpath task into the the
new workqueue.

Signed-off-by: Yuval Mintz <Yuval.Mintz@qlogic.com>
Signed-off-by: Ariel Elior <Ariel.Elior@qlogic.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/broadcom/bnx2x/bnx2x.h

@@ -1155,10 +1155,6 @@ struct bnx2x_port {
 			(offsetof(struct bnx2x_eth_stats, stat_name) / 4)
 
 /* slow path */
-
-/* slow path work-queue */
-extern struct workqueue_struct *bnx2x_wq;
-
 #define BNX2X_MAX_NUM_OF_VFS	64
 #define BNX2X_VF_CID_WND	4 /* log num of queues per VF. HW config. */
 #define BNX2X_CIDS_PER_VF	(1 << BNX2X_VF_CID_WND)
@@ -1416,6 +1412,12 @@ enum sp_rtnl_flag {
 	BNX2X_SP_RTNL_GET_DRV_VERSION,
 };
 
+enum bnx2x_iov_flag {
+	BNX2X_IOV_HANDLE_VF_MSG,
+	BNX2X_IOV_CONT_VFOP,
+	BNX2X_IOV_HANDLE_FLR,
+};
+
 struct bnx2x_prev_path_list {
 	struct list_head list;
 	u8 bus;
@@ -1614,6 +1616,8 @@ struct bnx2x {
 	int			mrrs;
 
 	struct delayed_work	sp_task;
+	struct delayed_work	iov_task;
+
 	atomic_t		interrupt_occurred;
 	struct delayed_work	sp_rtnl_task;
 
@@ -1897,6 +1901,9 @@ struct bnx2x {
 	/* operation indication for the sp_rtnl task */
 	unsigned long				sp_rtnl_state;
 
+	/* Indication of the IOV tasks */
+	unsigned long				iov_task_state;
+
 	/* DCBX Negotiation results */
 	struct dcbx_features			dcbx_local_feat;
 	u32					dcbx_error;

drivers/net/ethernet/broadcom/bnx2x/bnx2x_main.c

@@ -120,7 +120,8 @@ static int debug;
 module_param(debug, int, S_IRUGO);
 MODULE_PARM_DESC(debug, " Default debug msglevel");
 
-struct workqueue_struct *bnx2x_wq;
+static struct workqueue_struct *bnx2x_wq;
+struct workqueue_struct *bnx2x_iov_wq;
 
 struct bnx2x_mac_vals {
 	u32 xmac_addr;
@@ -1857,7 +1858,7 @@ void bnx2x_sp_event(struct bnx2x_fastpath *fp, union eth_rx_cqe *rr_cqe)
 		return;
 #endif
 	/* SRIOV: reschedule any 'in_progress' operations */
-	bnx2x_iov_sp_event(bp, cid, true);
+	bnx2x_iov_sp_event(bp, cid);
 
 	smp_mb__before_atomic_inc();
 	atomic_inc(&bp->cq_spq_left);
@@ -4160,7 +4161,8 @@ static void bnx2x_attn_int_deasserted3(struct bnx2x *bp, u32 attn)
 				bnx2x_handle_drv_info_req(bp);
 
 			if (val & DRV_STATUS_VF_DISABLED)
-				bnx2x_vf_handle_flr_event(bp);
+				bnx2x_schedule_iov_task(bp,
+							BNX2X_IOV_HANDLE_FLR);
 
 			if ((bp->port.pmf == 0) && (val & DRV_STATUS_PMF))
 				bnx2x_pmf_update(bp);
@@ -5351,8 +5353,8 @@ static void bnx2x_eq_int(struct bnx2x *bp)
 		/* handle eq element */
 		switch (opcode) {
 		case EVENT_RING_OPCODE_VF_PF_CHANNEL:
-			DP(BNX2X_MSG_IOV, "vf pf channel element on eq\n");
-			bnx2x_vf_mbx(bp, &elem->message.data.vf_pf_event);
+			bnx2x_vf_mbx_schedule(bp,
+					      &elem->message.data.vf_pf_event);
 			continue;
 
 		case EVENT_RING_OPCODE_STAT_QUERY:
@@ -5567,13 +5569,6 @@ static void bnx2x_sp_task(struct work_struct *work)
 			     le16_to_cpu(bp->def_att_idx), IGU_INT_ENABLE, 1);
 	}
 
-	/* must be called after the EQ processing (since eq leads to sriov
-	 * ramrod completion flows).
-	 * This flow may have been scheduled by the arrival of a ramrod
-	 * completion, or by the sriov code rescheduling itself.
-	 */
-	bnx2x_iov_sp_task(bp);
-
 	/* afex - poll to check if VIFSET_ACK should be sent to MFW */
 	if (test_and_clear_bit(BNX2X_AFEX_PENDING_VIFSET_MCP_ACK,
 			       &bp->sp_state)) {
@@ -8990,6 +8985,7 @@ static int bnx2x_func_wait_started(struct bnx2x *bp)
 		synchronize_irq(bp->pdev->irq);
 
 	flush_workqueue(bnx2x_wq);
+	flush_workqueue(bnx2x_iov_wq);
 
 	while (bnx2x_func_get_state(bp, &bp->func_obj) !=
 				BNX2X_F_STATE_STARTED && tout--)
@@ -11877,6 +11873,7 @@ static int bnx2x_init_bp(struct bnx2x *bp)
 	INIT_DELAYED_WORK(&bp->sp_task, bnx2x_sp_task);
 	INIT_DELAYED_WORK(&bp->sp_rtnl_task, bnx2x_sp_rtnl_task);
 	INIT_DELAYED_WORK(&bp->period_task, bnx2x_period_task);
+	INIT_DELAYED_WORK(&bp->iov_task, bnx2x_iov_task);
 	if (IS_PF(bp)) {
 		rc = bnx2x_get_hwinfo(bp);
 		if (rc)
@@ -13499,11 +13496,18 @@ static int __init bnx2x_init(void)
 		pr_err("Cannot create workqueue\n");
 		return -ENOMEM;
 	}
+	bnx2x_iov_wq = create_singlethread_workqueue("bnx2x_iov");
+	if (!bnx2x_iov_wq) {
+		pr_err("Cannot create iov workqueue\n");
+		destroy_workqueue(bnx2x_wq);
+		return -ENOMEM;
+	}
 
 	ret = pci_register_driver(&bnx2x_pci_driver);
 	if (ret) {
 		pr_err("Cannot register driver\n");
 		destroy_workqueue(bnx2x_wq);
+		destroy_workqueue(bnx2x_iov_wq);
 	}
 	return ret;
 }
@@ -13515,6 +13519,7 @@ static void __exit bnx2x_cleanup(void)
 	pci_unregister_driver(&bnx2x_pci_driver);
 
 	destroy_workqueue(bnx2x_wq);
+	destroy_workqueue(bnx2x_iov_wq);
 
 	/* Free globally allocated resources */
 	list_for_each_safe(pos, q, &bnx2x_prev_list) {

drivers/net/ethernet/broadcom/bnx2x/bnx2x_sriov.c

@@ -2042,6 +2042,9 @@ int bnx2x_iov_init_one(struct bnx2x *bp, int int_mode_param,
 		goto failed;
 	}
 
+	/* Prepare the VFs event synchronization mechanism */
+	mutex_init(&bp->vfdb->event_mutex);
+
 	return 0;
 failed:
 	DP(BNX2X_MSG_IOV, "Failed err=%d\n", err);
@@ -2469,7 +2472,7 @@ get_vf:
 		return 0;
 	}
 	/* SRIOV: reschedule any 'in_progress' operations */
-	bnx2x_iov_sp_event(bp, cid, false);
+	bnx2x_iov_sp_event(bp, cid);
 
 	return 0;
 }
@@ -2506,7 +2509,7 @@ void bnx2x_iov_set_queue_sp_obj(struct bnx2x *bp, int vf_cid,
 	}
 }
 
-void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid, bool queue_work)
+void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid)
 {
 	struct bnx2x_virtf *vf;
 
@@ -2518,8 +2521,7 @@ void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid, bool queue_work)
 	if (vf) {
 		/* set in_progress flag */
 		atomic_set(&vf->op_in_progress, 1);
-		if (queue_work)
-			queue_delayed_work(bnx2x_wq, &bp->sp_task, 0);
+		bnx2x_schedule_iov_task(bp, BNX2X_IOV_CONT_VFOP);
 	}
 }
 
@@ -2604,7 +2606,7 @@ void bnx2x_iov_adjust_stats_req(struct bnx2x *bp)
 	bp->fw_stats_req->hdr.cmd_num = bp->fw_stats_num + stats_count;
 }
 
-void bnx2x_iov_sp_task(struct bnx2x *bp)
+void bnx2x_iov_vfop_cont(struct bnx2x *bp)
 {
 	int i;
 
@@ -3875,3 +3877,32 @@ void bnx2x_iov_channel_down(struct bnx2x *bp)
 		bnx2x_post_vf_bulletin(bp, vf_idx);
 	}
 }
+
+void bnx2x_iov_task(struct work_struct *work)
+{
+	struct bnx2x *bp = container_of(work, struct bnx2x, iov_task.work);
+
+	if (!netif_running(bp->dev))
+		return;
+
+	if (test_and_clear_bit(BNX2X_IOV_HANDLE_FLR,
+			       &bp->iov_task_state))
+		bnx2x_vf_handle_flr_event(bp);
+
+	if (test_and_clear_bit(BNX2X_IOV_CONT_VFOP,
+			       &bp->iov_task_state))
+		bnx2x_iov_vfop_cont(bp);
+
+	if (test_and_clear_bit(BNX2X_IOV_HANDLE_VF_MSG,
+			       &bp->iov_task_state))
+		bnx2x_vf_mbx(bp);
+}
+
+void bnx2x_schedule_iov_task(struct bnx2x *bp, enum bnx2x_iov_flag flag)
+{
+	smp_mb__before_clear_bit();
+	set_bit(flag, &bp->iov_task_state);
+	smp_mb__after_clear_bit();
+	DP(BNX2X_MSG_IOV, "Scheduling iov task [Flag: %d]\n", flag);
+	queue_delayed_work(bnx2x_iov_wq, &bp->iov_task, 0);
+}

drivers/net/ethernet/broadcom/bnx2x/bnx2x_sriov.h

@@ -30,6 +30,8 @@ enum sample_bulletin_result {
 
 #ifdef CONFIG_BNX2X_SRIOV
 
+extern struct workqueue_struct *bnx2x_iov_wq;
+
 /* The bnx2x device structure holds vfdb structure described below.
  * The VF array is indexed by the relative vfid.
  */
@@ -346,11 +348,6 @@ struct bnx2x_vf_mbx {
 	u32 vf_addr_hi;
 
 	struct vfpf_first_tlv first_tlv;	/* saved VF request header */
-
-	u8 flags;
-#define VF_MSG_INPROCESS	0x1	/* failsafe - the FW should prevent
-					 * more then one pending msg
-					 */
 };
 
 struct bnx2x_vf_sp {
@@ -427,6 +424,10 @@ struct bnx2x_vfdb {
 	/* the number of msix vectors belonging to this PF designated for VFs */
 	u16 vf_sbs_pool;
 	u16 first_vf_igu_entry;
+
+	/* sp_rtnl synchronization */
+	struct mutex			event_mutex;
+	u64				event_occur;
 };
 
 /* queue access */
@@ -476,13 +477,14 @@ void bnx2x_iov_init_dq(struct bnx2x *bp);
 void bnx2x_iov_init_dmae(struct bnx2x *bp);
 void bnx2x_iov_set_queue_sp_obj(struct bnx2x *bp, int vf_cid,
 				struct bnx2x_queue_sp_obj **q_obj);
-void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid, bool queue_work);
+void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid);
 int bnx2x_iov_eq_sp_event(struct bnx2x *bp, union event_ring_elem *elem);
 void bnx2x_iov_adjust_stats_req(struct bnx2x *bp);
 void bnx2x_iov_storm_stats_update(struct bnx2x *bp);
-void bnx2x_iov_sp_task(struct bnx2x *bp);
 /* global vf mailbox routines */
-void bnx2x_vf_mbx(struct bnx2x *bp, struct vf_pf_event_data *vfpf_event);
+void bnx2x_vf_mbx(struct bnx2x *bp);
+void bnx2x_vf_mbx_schedule(struct bnx2x *bp,
+			   struct vf_pf_event_data *vfpf_event);
 void bnx2x_vf_enable_mbx(struct bnx2x *bp, u8 abs_vfid);
 
 /* CORE VF API */
@@ -520,7 +522,8 @@ enum {
 		else {							\
 			DP(BNX2X_MSG_IOV, "no ramrod. Scheduling\n");	\
 			atomic_set(&vf->op_in_progress, 1);		\
-			queue_delayed_work(bnx2x_wq, &bp->sp_task, 0);  \
+			bnx2x_schedule_iov_task(bp,			\
+						BNX2X_IOV_CONT_VFOP);	\
 			return;						\
 		}							\
 	} while (0)
@@ -785,18 +788,21 @@ void bnx2x_pf_set_vfs_vlan(struct bnx2x *bp);
 int bnx2x_sriov_configure(struct pci_dev *dev, int num_vfs);
 void bnx2x_iov_channel_down(struct bnx2x *bp);
 
+void bnx2x_iov_task(struct work_struct *work);
+
+void bnx2x_schedule_iov_task(struct bnx2x *bp, enum bnx2x_iov_flag flag);
+
 #else /* CONFIG_BNX2X_SRIOV */
 
 static inline void bnx2x_iov_set_queue_sp_obj(struct bnx2x *bp, int vf_cid,
 				struct bnx2x_queue_sp_obj **q_obj) {}
-static inline void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid,
-				      bool queue_work) {}
+static inline void bnx2x_iov_sp_event(struct bnx2x *bp, int vf_cid) {}
 static inline void bnx2x_vf_handle_flr_event(struct bnx2x *bp) {}
 static inline int bnx2x_iov_eq_sp_event(struct bnx2x *bp,
 					union event_ring_elem *elem) {return 1; }
-static inline void bnx2x_iov_sp_task(struct bnx2x *bp) {}
-static inline void bnx2x_vf_mbx(struct bnx2x *bp,
-				struct vf_pf_event_data *vfpf_event) {}
+static inline void bnx2x_vf_mbx(struct bnx2x *bp) {}
+static inline void bnx2x_vf_mbx_schedule(struct bnx2x *bp,
+					 struct vf_pf_event_data *vfpf_event) {}
 static inline int bnx2x_iov_init_ilt(struct bnx2x *bp, u16 line) {return line; }
 static inline void bnx2x_iov_init_dq(struct bnx2x *bp) {}
 static inline int bnx2x_iov_alloc_mem(struct bnx2x *bp) {return 0; }
@@ -843,5 +849,8 @@ static inline void bnx2x_pf_set_vfs_vlan(struct bnx2x *bp) {}
 static inline int bnx2x_sriov_configure(struct pci_dev *dev, int num_vfs) {return 0; }
 static inline void bnx2x_iov_channel_down(struct bnx2x *bp) {}
 
+static inline void bnx2x_iov_task(struct work_struct *work) {}
+void bnx2x_schedule_iov_task(struct bnx2x *bp, enum bnx2x_iov_flag flag) {}
+
 #endif /* CONFIG_BNX2X_SRIOV */
 #endif /* bnx2x_sriov.h */

drivers/net/ethernet/broadcom/bnx2x/bnx2x_vfpf.c

@@ -1089,9 +1089,6 @@ static void bnx2x_vf_mbx_resp_send_msg(struct bnx2x *bp,
 	storm_memset_vf_mbx_ack(bp, vf->abs_vfid);
 	mmiowb();
 
-	/* initiate dmae to send the response */
-	mbx->flags &= ~VF_MSG_INPROCESS;
-
 	/* copy the response header including status-done field,
 	 * must be last dmae, must be after FW is acked
 	 */
@@ -2059,13 +2056,10 @@ static void bnx2x_vf_mbx_request(struct bnx2x *bp, struct bnx2x_virtf *vf,
 	}
 }
 
-/* handle new vf-pf message */
-void bnx2x_vf_mbx(struct bnx2x *bp, struct vf_pf_event_data *vfpf_event)
+void bnx2x_vf_mbx_schedule(struct bnx2x *bp,
+			   struct vf_pf_event_data *vfpf_event)
 {
-	struct bnx2x_virtf *vf;
-	struct bnx2x_vf_mbx *mbx;
 	u8 vf_idx;
-	int rc;
 
 	DP(BNX2X_MSG_IOV,
 	   "vf pf event received: vfid %d, address_hi %x, address lo %x",
@@ -2077,50 +2071,73 @@ void bnx2x_vf_mbx(struct bnx2x *bp, struct vf_pf_event_data *vfpf_event)
 	    BNX2X_NR_VIRTFN(bp)) {
 		BNX2X_ERR("Illegal vf_id %d max allowed: %d\n",
 			  vfpf_event->vf_id, BNX2X_NR_VIRTFN(bp));
-		goto mbx_done;
+		return;
 	}
+
 	vf_idx = bnx2x_vf_idx_by_abs_fid(bp, vfpf_event->vf_id);
-	mbx = BP_VF_MBX(bp, vf_idx);
 
-	/* verify an event is not currently being processed -
-	 * debug failsafe only
-	 */
-	if (mbx->flags & VF_MSG_INPROCESS) {
-		BNX2X_ERR("Previous message is still being processed, vf_id %d\n",
-			  vfpf_event->vf_id);
-		goto mbx_done;
-	}
-	vf = BP_VF(bp, vf_idx);
+	/* Update VFDB with current message and schedule its handling */
+	mutex_lock(&BP_VFDB(bp)->event_mutex);
+	BP_VF_MBX(bp, vf_idx)->vf_addr_hi = vfpf_event->msg_addr_hi;
+	BP_VF_MBX(bp, vf_idx)->vf_addr_lo = vfpf_event->msg_addr_lo;
+	BP_VFDB(bp)->event_occur |= (1ULL << vf_idx);
+	mutex_unlock(&BP_VFDB(bp)->event_mutex);
 
-	/* save the VF message address */
-	mbx->vf_addr_hi = vfpf_event->msg_addr_hi;
-	mbx->vf_addr_lo = vfpf_event->msg_addr_lo;
-	DP(BNX2X_MSG_IOV, "mailbox vf address hi 0x%x, lo 0x%x, offset 0x%x\n",
-	   mbx->vf_addr_hi, mbx->vf_addr_lo, mbx->first_tlv.resp_msg_offset);
+	bnx2x_schedule_iov_task(bp, BNX2X_IOV_HANDLE_VF_MSG);
+}
 
-	/* dmae to get the VF request */
-	rc = bnx2x_copy32_vf_dmae(bp, true, mbx->msg_mapping, vf->abs_vfid,
-				  mbx->vf_addr_hi, mbx->vf_addr_lo,
-				  sizeof(union vfpf_tlvs)/4);
-	if (rc) {
-		BNX2X_ERR("Failed to copy request VF %d\n", vf->abs_vfid);
-		goto mbx_error;
-	}
+/* handle new vf-pf messages */
+void bnx2x_vf_mbx(struct bnx2x *bp)
+{
+	struct bnx2x_vfdb *vfdb = BP_VFDB(bp);
+	u64 events;
+	u8 vf_idx;
+	int rc;
 
-	/* process the VF message header */
-	mbx->first_tlv = mbx->msg->req.first_tlv;
+	if (!vfdb)
+		return;
 
-	/* Clean response buffer to refrain from falsely seeing chains */
-	memset(&mbx->msg->resp, 0, sizeof(union pfvf_tlvs));
+	mutex_lock(&vfdb->event_mutex);
+	events = vfdb->event_occur;
+	vfdb->event_occur = 0;
+	mutex_unlock(&vfdb->event_mutex);
 
-	/* dispatch the request (will prepare the response) */
-	bnx2x_vf_mbx_request(bp, vf, mbx);
-	goto mbx_done;
+	for_each_vf(bp, vf_idx) {
+		struct bnx2x_vf_mbx *mbx = BP_VF_MBX(bp, vf_idx);
+		struct bnx2x_virtf *vf = BP_VF(bp, vf_idx);
 
-mbx_error:
-	bnx2x_vf_release(bp, vf, false); /* non blocking */
-mbx_done:
-	return;
+		/* Handle VFs which have pending events */
+		if (!(events & (1ULL << vf_idx)))
+			continue;
+
+		DP(BNX2X_MSG_IOV,
+		   "Handling vf pf event vfid %d, address: [%x:%x], resp_offset 0x%x\n",
+		   vf_idx, mbx->vf_addr_hi, mbx->vf_addr_lo,
+		   mbx->first_tlv.resp_msg_offset);
+
+		/* dmae to get the VF request */
+		rc = bnx2x_copy32_vf_dmae(bp, true, mbx->msg_mapping,
+					  vf->abs_vfid, mbx->vf_addr_hi,
+					  mbx->vf_addr_lo,
+					  sizeof(union vfpf_tlvs)/4);
+		if (rc) {
+			BNX2X_ERR("Failed to copy request VF %d\n",
+				  vf->abs_vfid);
+			bnx2x_vf_release(bp, vf, false); /* non blocking */
+			return;
+		}
+
+		/* process the VF message header */
+		mbx->first_tlv = mbx->msg->req.first_tlv;
+
+		/* Clean response buffer to refrain from falsely
+		 * seeing chains.
+		 */
+		memset(&mbx->msg->resp, 0, sizeof(union pfvf_tlvs));
+
+		/* dispatch the request (will prepare the response) */
+		bnx2x_vf_mbx_request(bp, vf, mbx);
+	}
 }
 
 /* propagate local bulletin board to vf */