ixgbevf: allocate the rings as part of q_vector
authorEmil Tantilov <emil.s.tantilov@intel.com>
Wed, 31 Jan 2018 00:51:43 +0000 (16:51 -0800)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Mon, 26 Feb 2018 17:32:46 +0000 (09:32 -0800)
Make it so that all rings allocations are made as part of q_vector.
The advantage to this is that we can keep all of the memory related to
a single interrupt in one page.

The goal is to bring the logic of handling rings closer to ixgbe.

Signed-off-by: Emil Tantilov <emil.s.tantilov@intel.com>
Tested-by: Krishneil Singh <krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/ixgbevf/ixgbevf.h
drivers/net/ethernet/intel/ixgbevf/ixgbevf_main.c

index a5e9127a1156eae8aaa2878a001478b4b91cea65..f65ca156af2deacc0d6417d51cffd3f5d33aa4fe 100644 (file)
@@ -97,6 +97,7 @@ enum ixgbevf_ring_state_t {
 
 struct ixgbevf_ring {
        struct ixgbevf_ring *next;
+       struct ixgbevf_q_vector *q_vector;      /* backpointer to q_vector */
        struct net_device *netdev;
        struct device *dev;
        void *desc;                     /* descriptor ring memory */
@@ -128,7 +129,7 @@ struct ixgbevf_ring {
         */
        u16 reg_idx;
        int queue_index; /* needed for multiqueue queue management */
-};
+} ____cacheline_internodealigned_in_smp;
 
 /* How many Rx Buffers do we bundle into one write to the hardware ? */
 #define IXGBEVF_RX_BUFFER_WRITE        16      /* Must be power of 2 */
@@ -241,7 +242,11 @@ struct ixgbevf_q_vector {
        u16 itr; /* Interrupt throttle rate written to EITR */
        struct napi_struct napi;
        struct ixgbevf_ring_container rx, tx;
+       struct rcu_head rcu;    /* to avoid race with update stats on free */
        char name[IFNAMSIZ + 9];
+
+       /* for dynamic allocation of rings associated with this q_vector */
+       struct ixgbevf_ring ring[0] ____cacheline_internodealigned_in_smp;
 #ifdef CONFIG_NET_RX_BUSY_POLL
        unsigned int state;
 #define IXGBEVF_QV_STATE_IDLE          0
index a10b1bdc99e3049029aa9e53203407f2265b97df..6219ab2e3f52d1714633e3575ee147e269910024 100644 (file)
@@ -1270,85 +1270,6 @@ static irqreturn_t ixgbevf_msix_clean_rings(int irq, void *data)
        return IRQ_HANDLED;
 }
 
-static inline void map_vector_to_rxq(struct ixgbevf_adapter *a, int v_idx,
-                                    int r_idx)
-{
-       struct ixgbevf_q_vector *q_vector = a->q_vector[v_idx];
-
-       a->rx_ring[r_idx]->next = q_vector->rx.ring;
-       q_vector->rx.ring = a->rx_ring[r_idx];
-       q_vector->rx.count++;
-}
-
-static inline void map_vector_to_txq(struct ixgbevf_adapter *a, int v_idx,
-                                    int t_idx)
-{
-       struct ixgbevf_q_vector *q_vector = a->q_vector[v_idx];
-
-       a->tx_ring[t_idx]->next = q_vector->tx.ring;
-       q_vector->tx.ring = a->tx_ring[t_idx];
-       q_vector->tx.count++;
-}
-
-/**
- * ixgbevf_map_rings_to_vectors - Maps descriptor rings to vectors
- * @adapter: board private structure to initialize
- *
- * This function maps descriptor rings to the queue-specific vectors
- * we were allotted through the MSI-X enabling code.  Ideally, we'd have
- * one vector per ring/queue, but on a constrained vector budget, we
- * group the rings as "efficiently" as possible.  You would add new
- * mapping configurations in here.
- **/
-static int ixgbevf_map_rings_to_vectors(struct ixgbevf_adapter *adapter)
-{
-       int q_vectors;
-       int v_start = 0;
-       int rxr_idx = 0, txr_idx = 0;
-       int rxr_remaining = adapter->num_rx_queues;
-       int txr_remaining = adapter->num_tx_queues;
-       int i, j;
-       int rqpv, tqpv;
-
-       q_vectors = adapter->num_msix_vectors - NON_Q_VECTORS;
-
-       /* The ideal configuration...
-        * We have enough vectors to map one per queue.
-        */
-       if (q_vectors == adapter->num_rx_queues + adapter->num_tx_queues) {
-               for (; rxr_idx < rxr_remaining; v_start++, rxr_idx++)
-                       map_vector_to_rxq(adapter, v_start, rxr_idx);
-
-               for (; txr_idx < txr_remaining; v_start++, txr_idx++)
-                       map_vector_to_txq(adapter, v_start, txr_idx);
-               return 0;
-       }
-
-       /* If we don't have enough vectors for a 1-to-1
-        * mapping, we'll have to group them so there are
-        * multiple queues per vector.
-        */
-       /* Re-adjusting *qpv takes care of the remainder. */
-       for (i = v_start; i < q_vectors; i++) {
-               rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors - i);
-               for (j = 0; j < rqpv; j++) {
-                       map_vector_to_rxq(adapter, i, rxr_idx);
-                       rxr_idx++;
-                       rxr_remaining--;
-               }
-       }
-       for (i = v_start; i < q_vectors; i++) {
-               tqpv = DIV_ROUND_UP(txr_remaining, q_vectors - i);
-               for (j = 0; j < tqpv; j++) {
-                       map_vector_to_txq(adapter, i, txr_idx);
-                       txr_idx++;
-                       txr_remaining--;
-               }
-       }
-
-       return 0;
-}
-
 /**
  * ixgbevf_request_msix_irqs - Initialize MSI-X interrupts
  * @adapter: board private structure
@@ -1421,20 +1342,6 @@ free_queue_irqs:
        return err;
 }
 
-static inline void ixgbevf_reset_q_vectors(struct ixgbevf_adapter *adapter)
-{
-       int i, q_vectors = adapter->num_msix_vectors - NON_Q_VECTORS;
-
-       for (i = 0; i < q_vectors; i++) {
-               struct ixgbevf_q_vector *q_vector = adapter->q_vector[i];
-
-               q_vector->rx.ring = NULL;
-               q_vector->tx.ring = NULL;
-               q_vector->rx.count = 0;
-               q_vector->tx.count = 0;
-       }
-}
-
 /**
  * ixgbevf_request_irq - initialize interrupts
  * @adapter: board private structure
@@ -1474,8 +1381,6 @@ static void ixgbevf_free_irq(struct ixgbevf_adapter *adapter)
                free_irq(adapter->msix_entries[i].vector,
                         adapter->q_vector[i]);
        }
-
-       ixgbevf_reset_q_vectors(adapter);
 }
 
 /**
@@ -2457,105 +2362,171 @@ static void ixgbevf_set_num_queues(struct ixgbevf_adapter *adapter)
 }
 
 /**
- * ixgbevf_alloc_queues - Allocate memory for all rings
+ * ixgbevf_set_interrupt_capability - set MSI-X or FAIL if not supported
+ * @adapter: board private structure to initialize
+ *
+ * Attempt to configure the interrupts using the best available
+ * capabilities of the hardware and the kernel.
+ **/
+static int ixgbevf_set_interrupt_capability(struct ixgbevf_adapter *adapter)
+{
+       int vector, v_budget;
+
+       /* It's easy to be greedy for MSI-X vectors, but it really
+        * doesn't do us much good if we have a lot more vectors
+        * than CPU's.  So let's be conservative and only ask for
+        * (roughly) the same number of vectors as there are CPU's.
+        * The default is to use pairs of vectors.
+        */
+       v_budget = max(adapter->num_rx_queues, adapter->num_tx_queues);
+       v_budget = min_t(int, v_budget, num_online_cpus());
+       v_budget += NON_Q_VECTORS;
+
+       adapter->msix_entries = kcalloc(v_budget,
+                                       sizeof(struct msix_entry), GFP_KERNEL);
+       if (!adapter->msix_entries)
+               return -ENOMEM;
+
+       for (vector = 0; vector < v_budget; vector++)
+               adapter->msix_entries[vector].entry = vector;
+
+       /* A failure in MSI-X entry allocation isn't fatal, but the VF driver
+        * does not support any other modes, so we will simply fail here. Note
+        * that we clean up the msix_entries pointer else-where.
+        */
+       return ixgbevf_acquire_msix_vectors(adapter, v_budget);
+}
+
+static void ixgbevf_add_ring(struct ixgbevf_ring *ring,
+                            struct ixgbevf_ring_container *head)
+{
+       ring->next = head->ring;
+       head->ring = ring;
+       head->count++;
+}
+
+/**
+ * ixgbevf_alloc_q_vector - Allocate memory for a single interrupt vector
  * @adapter: board private structure to initialize
+ * @v_idx: index of vector in adapter struct
+ * @txr_count: number of Tx rings for q vector
+ * @txr_idx: index of first Tx ring to assign
+ * @rxr_count: number of Rx rings for q vector
+ * @rxr_idx: index of first Rx ring to assign
  *
- * We allocate one ring per queue at run-time since we don't know the
- * number of queues at compile-time.  The polling_netdev array is
- * intended for Multiqueue, but should work fine with a single queue.
+ * We allocate one q_vector.  If allocation fails we return -ENOMEM.
  **/
-static int ixgbevf_alloc_queues(struct ixgbevf_adapter *adapter)
+static int ixgbevf_alloc_q_vector(struct ixgbevf_adapter *adapter, int v_idx,
+                                 int txr_count, int txr_idx,
+                                 int rxr_count, int rxr_idx)
 {
+       struct ixgbevf_q_vector *q_vector;
        struct ixgbevf_ring *ring;
-       int rx = 0, tx = 0;
+       int ring_count, size;
+
+       ring_count = txr_count + rxr_count;
+       size = sizeof(*q_vector) + (sizeof(*ring) * ring_count);
+
+       /* allocate q_vector and rings */
+       q_vector = kzalloc(size, GFP_KERNEL);
+       if (!q_vector)
+               return -ENOMEM;
+
+       /* initialize NAPI */
+       netif_napi_add(adapter->netdev, &q_vector->napi, ixgbevf_poll, 64);
+
+       /* tie q_vector and adapter together */
+       adapter->q_vector[v_idx] = q_vector;
+       q_vector->adapter = adapter;
+       q_vector->v_idx = v_idx;
 
-       for (; tx < adapter->num_tx_queues; tx++) {
-               ring = kzalloc(sizeof(*ring), GFP_KERNEL);
-               if (!ring)
-                       goto err_allocation;
+       /* initialize pointer to rings */
+       ring = q_vector->ring;
 
+       while (txr_count) {
+               /* assign generic ring traits */
                ring->dev = &adapter->pdev->dev;
                ring->netdev = adapter->netdev;
+
+               /* configure backlink on ring */
+               ring->q_vector = q_vector;
+
+               /* update q_vector Tx values */
+               ixgbevf_add_ring(ring, &q_vector->tx);
+
+               /* apply Tx specific ring traits */
                ring->count = adapter->tx_ring_count;
-               ring->queue_index = tx;
-               ring->reg_idx = tx;
+               ring->queue_index = txr_idx;
+               ring->reg_idx = txr_idx;
 
-               adapter->tx_ring[tx] = ring;
-       }
+               /* assign ring to adapter */
+                adapter->tx_ring[txr_idx] = ring;
+
+               /* update count and index */
+               txr_count--;
+               txr_idx++;
 
-       for (; rx < adapter->num_rx_queues; rx++) {
-               ring = kzalloc(sizeof(*ring), GFP_KERNEL);
-               if (!ring)
-                       goto err_allocation;
+               /* push pointer to next ring */
+               ring++;
+       }
 
+       while (rxr_count) {
+               /* assign generic ring traits */
                ring->dev = &adapter->pdev->dev;
                ring->netdev = adapter->netdev;
 
+               /* configure backlink on ring */
+               ring->q_vector = q_vector;
+
+               /* update q_vector Rx values */
+               ixgbevf_add_ring(ring, &q_vector->rx);
+
+               /* apply Rx specific ring traits */
                ring->count = adapter->rx_ring_count;
-               ring->queue_index = rx;
-               ring->reg_idx = rx;
+               ring->queue_index = rxr_idx;
+               ring->reg_idx = rxr_idx;
 
-               adapter->rx_ring[rx] = ring;
-       }
+               /* assign ring to adapter */
+               adapter->rx_ring[rxr_idx] = ring;
 
-       return 0;
+               /* update count and index */
+               rxr_count--;
+               rxr_idx++;
 
-err_allocation:
-       while (tx) {
-               kfree(adapter->tx_ring[--tx]);
-               adapter->tx_ring[tx] = NULL;
+               /* push pointer to next ring */
+               ring++;
        }
 
-       while (rx) {
-               kfree(adapter->rx_ring[--rx]);
-               adapter->rx_ring[rx] = NULL;
-       }
-       return -ENOMEM;
+       return 0;
 }
 
 /**
- * ixgbevf_set_interrupt_capability - set MSI-X or FAIL if not supported
+ * ixgbevf_free_q_vector - Free memory allocated for specific interrupt vector
  * @adapter: board private structure to initialize
+ * @v_idx: index of vector in adapter struct
  *
- * Attempt to configure the interrupts using the best available
- * capabilities of the hardware and the kernel.
+ * This function frees the memory allocated to the q_vector.  In addition if
+ * NAPI is enabled it will delete any references to the NAPI struct prior
+ * to freeing the q_vector.
  **/
-static int ixgbevf_set_interrupt_capability(struct ixgbevf_adapter *adapter)
+static void ixgbevf_free_q_vector(struct ixgbevf_adapter *adapter, int v_idx)
 {
-       struct net_device *netdev = adapter->netdev;
-       int err;
-       int vector, v_budget;
+       struct ixgbevf_q_vector *q_vector = adapter->q_vector[v_idx];
+       struct ixgbevf_ring *ring;
 
-       /* It's easy to be greedy for MSI-X vectors, but it really
-        * doesn't do us much good if we have a lot more vectors
-        * than CPU's.  So let's be conservative and only ask for
-        * (roughly) the same number of vectors as there are CPU's.
-        * The default is to use pairs of vectors.
-        */
-       v_budget = max(adapter->num_rx_queues, adapter->num_tx_queues);
-       v_budget = min_t(int, v_budget, num_online_cpus());
-       v_budget += NON_Q_VECTORS;
+       ixgbevf_for_each_ring(ring, q_vector->tx)
+               adapter->tx_ring[ring->queue_index] = NULL;
 
-       /* A failure in MSI-X entry allocation isn't fatal, but it does
-        * mean we disable MSI-X capabilities of the adapter.
-        */
-       adapter->msix_entries = kcalloc(v_budget,
-                                       sizeof(struct msix_entry), GFP_KERNEL);
-       if (!adapter->msix_entries)
-               return -ENOMEM;
+       ixgbevf_for_each_ring(ring, q_vector->rx)
+               adapter->rx_ring[ring->queue_index] = NULL;
 
-       for (vector = 0; vector < v_budget; vector++)
-               adapter->msix_entries[vector].entry = vector;
+       adapter->q_vector[v_idx] = NULL;
+       netif_napi_del(&q_vector->napi);
 
-       err = ixgbevf_acquire_msix_vectors(adapter, v_budget);
-       if (err)
-               return err;
-
-       err = netif_set_real_num_tx_queues(netdev, adapter->num_tx_queues);
-       if (err)
-               return err;
-
-       return netif_set_real_num_rx_queues(netdev, adapter->num_rx_queues);
+       /* ixgbevf_get_stats() might access the rings on this vector,
+        * we must wait a grace period before freeing it.
+        */
+       kfree_rcu(q_vector, rcu);
 }
 
 /**
@@ -2567,35 +2538,53 @@ static int ixgbevf_set_interrupt_capability(struct ixgbevf_adapter *adapter)
  **/
 static int ixgbevf_alloc_q_vectors(struct ixgbevf_adapter *adapter)
 {
-       int q_idx, num_q_vectors;
-       struct ixgbevf_q_vector *q_vector;
+       int q_vectors = adapter->num_msix_vectors - NON_Q_VECTORS;
+       int rxr_remaining = adapter->num_rx_queues;
+       int txr_remaining = adapter->num_tx_queues;
+       int rxr_idx = 0, txr_idx = 0, v_idx = 0;
+       int err;
+
+       if (q_vectors >= (rxr_remaining + txr_remaining)) {
+               for (; rxr_remaining; v_idx++, q_vectors--) {
+                       int rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors);
+
+                       err = ixgbevf_alloc_q_vector(adapter, v_idx,
+                                                    0, 0, rqpv, rxr_idx);
+                       if (err)
+                               goto err_out;
+
+                       /* update counts and index */
+                       rxr_remaining -= rqpv;
+                       rxr_idx += rqpv;
+               }
+       }
+
+       for (; q_vectors; v_idx++, q_vectors--) {
+               int rqpv = DIV_ROUND_UP(rxr_remaining, q_vectors);
+               int tqpv = DIV_ROUND_UP(txr_remaining, q_vectors);
 
-       num_q_vectors = adapter->num_msix_vectors - NON_Q_VECTORS;
+               err = ixgbevf_alloc_q_vector(adapter, v_idx,
+                                            tqpv, txr_idx,
+                                            rqpv, rxr_idx);
 
-       for (q_idx = 0; q_idx < num_q_vectors; q_idx++) {
-               q_vector = kzalloc(sizeof(struct ixgbevf_q_vector), GFP_KERNEL);
-               if (!q_vector)
+               if (err)
                        goto err_out;
-               q_vector->adapter = adapter;
-               q_vector->v_idx = q_idx;
-               netif_napi_add(adapter->netdev, &q_vector->napi,
-                              ixgbevf_poll, 64);
-               adapter->q_vector[q_idx] = q_vector;
+
+               /* update counts and index */
+               rxr_remaining -= rqpv;
+               rxr_idx += rqpv;
+               txr_remaining -= tqpv;
+               txr_idx += tqpv;
        }
 
        return 0;
 
 err_out:
-       while (q_idx) {
-               q_idx--;
-               q_vector = adapter->q_vector[q_idx];
-#ifdef CONFIG_NET_RX_BUSY_POLL
-               napi_hash_del(&q_vector->napi);
-#endif
-               netif_napi_del(&q_vector->napi);
-               kfree(q_vector);
-               adapter->q_vector[q_idx] = NULL;
+       while (v_idx) {
+               v_idx--;
+               ixgbevf_free_q_vector(adapter, v_idx);
        }
+
        return -ENOMEM;
 }
 
@@ -2609,17 +2598,11 @@ err_out:
  **/
 static void ixgbevf_free_q_vectors(struct ixgbevf_adapter *adapter)
 {
-       int q_idx, num_q_vectors = adapter->num_msix_vectors - NON_Q_VECTORS;
-
-       for (q_idx = 0; q_idx < num_q_vectors; q_idx++) {
-               struct ixgbevf_q_vector *q_vector = adapter->q_vector[q_idx];
+       int q_vectors = adapter->num_msix_vectors - NON_Q_VECTORS;
 
-               adapter->q_vector[q_idx] = NULL;
-#ifdef CONFIG_NET_RX_BUSY_POLL
-               napi_hash_del(&q_vector->napi);
-#endif
-               netif_napi_del(&q_vector->napi);
-               kfree(q_vector);
+       while (q_vectors) {
+               q_vectors--;
+               ixgbevf_free_q_vector(adapter, q_vectors);
        }
 }
 
@@ -2663,12 +2646,6 @@ static int ixgbevf_init_interrupt_scheme(struct ixgbevf_adapter *adapter)
                goto err_alloc_q_vectors;
        }
 
-       err = ixgbevf_alloc_queues(adapter);
-       if (err) {
-               pr_err("Unable to allocate memory for queues\n");
-               goto err_alloc_queues;
-       }
-
        hw_dbg(&adapter->hw, "Multiqueue %s: Rx Queue count = %u, Tx Queue count = %u\n",
               (adapter->num_rx_queues > 1) ? "Enabled" :
               "Disabled", adapter->num_rx_queues, adapter->num_tx_queues);
@@ -2676,8 +2653,6 @@ static int ixgbevf_init_interrupt_scheme(struct ixgbevf_adapter *adapter)
        set_bit(__IXGBEVF_DOWN, &adapter->state);
 
        return 0;
-err_alloc_queues:
-       ixgbevf_free_q_vectors(adapter);
 err_alloc_q_vectors:
        ixgbevf_reset_interrupt_capability(adapter);
 err_set_interrupt:
@@ -2693,17 +2668,6 @@ err_set_interrupt:
  **/
 static void ixgbevf_clear_interrupt_scheme(struct ixgbevf_adapter *adapter)
 {
-       int i;
-
-       for (i = 0; i < adapter->num_tx_queues; i++) {
-               kfree(adapter->tx_ring[i]);
-               adapter->tx_ring[i] = NULL;
-       }
-       for (i = 0; i < adapter->num_rx_queues; i++) {
-               kfree(adapter->rx_ring[i]);
-               adapter->rx_ring[i] = NULL;
-       }
-
        adapter->num_tx_queues = 0;
        adapter->num_rx_queues = 0;
 
@@ -3307,12 +3271,6 @@ int ixgbevf_open(struct net_device *netdev)
 
        ixgbevf_configure(adapter);
 
-       /* Map the Tx/Rx rings to the vectors we were allotted.
-        * if request_irq will be called in this function map_rings
-        * must be called *before* up_complete
-        */
-       ixgbevf_map_rings_to_vectors(adapter);
-
        err = ixgbevf_request_irq(adapter);
        if (err)
                goto err_req_irq;
@@ -4042,6 +4000,7 @@ static void ixgbevf_get_stats(struct net_device *netdev,
 
        stats->multicast = adapter->stats.vfmprc - adapter->stats.base_vfmprc;
 
+       rcu_read_lock();
        for (i = 0; i < adapter->num_rx_queues; i++) {
                ring = adapter->rx_ring[i];
                do {
@@ -4063,6 +4022,7 @@ static void ixgbevf_get_stats(struct net_device *netdev,
                stats->tx_bytes += bytes;
                stats->tx_packets += packets;
        }
+       rcu_read_unlock();
 }
 
 #define IXGBEVF_MAX_MAC_HDR_LEN                127