liquidio: Napi rx/tx traffic

This Patch adds tx buffer handling  to Napi along with RX
traffic. Also separate spinlocks are introduced for handling
iq posting and buffer reclaim so that tx path and tx interrupt
do not compete against each other.

Signed-off-by: Derek Chickles <derek.chickles@caviumnetworks.com>
Signed-off-by: Satanand Burla <satananda.burla@caviumnetworks.com>
Signed-off-by: Felix Manlunas <felix.manlunas@caviumnetworks.com>
Signed-off-by: Raghu Vatsavayi <raghu.vatsavayi@caviumnetworks.com>
Signed-off-by: Raghu Vatsavayi <rvatsavayi@caviumnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 70 insertions, 40 deletions

diff --git a/drivers/net/ethernet/cavium/liquidio/request_manager.c b/drivers/net/ethernet/cavium/liquidio/request_manager.c
index 51031010c307..7eafa75ac095 100644
--- a/drivers/net/ethernet/cavium/liquidio/request_manager.c
+++ b/drivers/net/ethernet/cavium/liquidio/request_manager.c
@@ -51,7 +51,7 @@ struct iq_post_status {
 };
 
 static void check_db_timeout(struct work_struct *work);
-static void  __check_db_timeout(struct octeon_device *oct, unsigned long iq_no);
+static void  __check_db_timeout(struct octeon_device *oct, u64 iq_no);
 
 static void (*reqtype_free_fn[MAX_OCTEON_DEVICES][REQTYPE_LAST + 1]) (void *);
 
@@ -149,6 +149,9 @@ int octeon_init_instr_queue(struct octeon_device *oct,
 
 	/* Initialize the spinlock for this instruction queue */
 	spin_lock_init(&iq->lock);
+	spin_lock_init(&iq->post_lock);
+
+	spin_lock_init(&iq->iq_flush_running_lock);
 
 	oct->io_qmask.iq |= (1ULL << iq_no);
 
@@ -391,13 +394,13 @@ __add_to_request_list(struct octeon_instr_queue *iq,
 
 int
 lio_process_iq_request_list(struct octeon_device *oct,
-			    struct octeon_instr_queue *iq)
+			    struct octeon_instr_queue *iq, u32 napi_budget)
 {
 	int reqtype;
 	void *buf;
 	u32 old = iq->flush_index;
 	u32 inst_count = 0;
-	unsigned pkts_compl = 0, bytes_compl = 0;
+	unsigned int pkts_compl = 0, bytes_compl = 0;
 	struct octeon_soft_command *sc;
 	struct octeon_instr_irh *irh;
 
@@ -457,6 +460,9 @@ lio_process_iq_request_list(struct octeon_device *oct,
  skip_this:
 		inst_count++;
 		INCR_INDEX_BY1(old, iq->max_count);
+
+		if ((napi_budget) && (inst_count >= napi_budget))
+			break;
 	}
 	if (bytes_compl)
 		octeon_report_tx_completion_to_bql(iq->app_ctx, pkts_compl,
@@ -466,38 +472,63 @@ lio_process_iq_request_list(struct octeon_device *oct,
 	return inst_count;
 }
 
-static inline void
-update_iq_indices(struct octeon_device *oct, struct octeon_instr_queue *iq)
+/* Can only be called from process context */
+int
+octeon_flush_iq(struct octeon_device *oct, struct octeon_instr_queue *iq,
+		u32 pending_thresh, u32 napi_budget)
 {
 	u32 inst_processed = 0;
+	u32 tot_inst_processed = 0;
+	int tx_done = 1;
 
-	/* Calculate how many commands Octeon has read and move the read index
-	 * accordingly.
-	 */
-	iq->octeon_read_index = oct->fn_list.update_iq_read_idx(oct, iq);
+	if (!spin_trylock(&iq->iq_flush_running_lock))
+		return tx_done;
 
-	/* Move the NORESPONSE requests to the per-device completion list. */
-	if (iq->flush_index != iq->octeon_read_index)
-		inst_processed = lio_process_iq_request_list(oct, iq);
+	spin_lock_bh(&iq->lock);
 
-	if (inst_processed) {
-		atomic_sub(inst_processed, &iq->instr_pending);
-		iq->stats.instr_processed += inst_processed;
-	}
-}
+	iq->octeon_read_index = oct->fn_list.update_iq_read_idx(iq);
 
-static void
-octeon_flush_iq(struct octeon_device *oct, struct octeon_instr_queue *iq,
-		u32 pending_thresh)
-{
 	if (atomic_read(&iq->instr_pending) >= (s32)pending_thresh) {
-		spin_lock_bh(&iq->lock);
-		update_iq_indices(oct, iq);
-		spin_unlock_bh(&iq->lock);
+		do {
+			/* Process any outstanding IQ packets. */
+			if (iq->flush_index == iq->octeon_read_index)
+				break;
+
+			if (napi_budget)
+				inst_processed = lio_process_iq_request_list
+					(oct, iq,
+					 napi_budget - tot_inst_processed);
+			else
+				inst_processed =
+					lio_process_iq_request_list(oct, iq, 0);
+
+			if (inst_processed) {
+				atomic_sub(inst_processed, &iq->instr_pending);
+				iq->stats.instr_processed += inst_processed;
+			}
+
+			tot_inst_processed += inst_processed;
+			inst_processed = 0;
+
+		} while (tot_inst_processed < napi_budget);
+
+		if (napi_budget && (tot_inst_processed >= napi_budget))
+			tx_done = 0;
 	}
+
+	iq->last_db_time = jiffies;
+
+	spin_unlock_bh(&iq->lock);
+
+	spin_unlock(&iq->iq_flush_running_lock);
+
+	return tx_done;
 }
 
-static void __check_db_timeout(struct octeon_device *oct, unsigned long iq_no)
+/* Process instruction queue after timeout.
+ * This routine gets called from a workqueue or when removing the module.
+ */
+static void __check_db_timeout(struct octeon_device *oct, u64 iq_no)
 {
 	struct octeon_instr_queue *iq;
 	u64 next_time;
@@ -508,24 +539,17 @@ static void __check_db_timeout(struct octeon_device *oct, unsigned long iq_no)
 	if (!iq)
 		return;
 
+	/* return immediately, if no work pending */
+	if (!atomic_read(&iq->instr_pending))
+		return;
 	/* If jiffies - last_db_time < db_timeout do nothing  */
 	next_time = iq->last_db_time + iq->db_timeout;
 	if (!time_after(jiffies, (unsigned long)next_time))
 		return;
 	iq->last_db_time = jiffies;
 
-	/* Get the lock and prevent tasklets. This routine gets called from
-	 * the poll thread. Instructions can now be posted in tasklet context
-	 */
-	spin_lock_bh(&iq->lock);
-	if (iq->fill_cnt != 0)
-		ring_doorbell(oct, iq);
-
-	spin_unlock_bh(&iq->lock);
-
 	/* Flush the instruction queue */
-	if (iq->do_auto_flush)
-		octeon_flush_iq(oct, iq, 1);
+	octeon_flush_iq(oct, iq, 1, 0);
 }
 
 /* Called by the Poll thread at regular intervals to check the instruction
@@ -550,7 +574,10 @@ octeon_send_command(struct octeon_device *oct, u32 iq_no,
 	struct iq_post_status st;
 	struct octeon_instr_queue *iq = oct->instr_queue[iq_no];
 
-	spin_lock_bh(&iq->lock);
+	/* Get the lock and prevent other tasks and tx interrupt handler from
+	 * running.
+	 */
+	spin_lock_bh(&iq->post_lock);
 
 	st = __post_command2(oct, iq, force_db, cmd);
 
@@ -566,10 +593,13 @@ octeon_send_command(struct octeon_device *oct, u32 iq_no,
 		INCR_INSTRQUEUE_PKT_COUNT(oct, iq_no, instr_dropped, 1);
 	}
 
-	spin_unlock_bh(&iq->lock);
+	spin_unlock_bh(&iq->post_lock);
 
-	if (iq->do_auto_flush)
-		octeon_flush_iq(oct, iq, 2);
+	/* This is only done here to expedite packets being flushed
+	 * for cases where there are no IQ completion interrupts.
+	 */
+	/*if (iq->do_auto_flush)*/
+	/*	octeon_flush_iq(oct, iq, 2, 0);*/
 
 	return st.status;
 }