cxgb3: Allocate multiqueues at init time
Divy Le Ray [Thu, 9 Oct 2008 00:37:33 +0000 (17:37 -0700)]
Allocate a queue set per core, up to the maximum of available qsets.
Share the queue sets on multi port adapters.
Rename MSI-X interrupt vectors ethX-N, N being the queue set number.

Signed-off-by: Divy Le Ray <divy@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/cxgb3/common.h
drivers/net/cxgb3/cxgb3_ioctl.h
drivers/net/cxgb3/cxgb3_main.c

index d6dbcd4..ace6b58 100644 (file)
@@ -358,6 +358,7 @@ struct qset_params {                /* SGE queue set parameters */
        unsigned int jumbo_size;        /* # of entries in jumbo free list */
        unsigned int txq_size[SGE_TXQ_PER_SET]; /* Tx queue sizes */
        unsigned int cong_thres;        /* FL congestion threshold */
+       unsigned int vector;            /* Interrupt (line or vector) number */
 };
 
 struct sge_params {
index 68200a1..3e8d5fa 100644 (file)
@@ -92,6 +92,8 @@ struct ch_qset_params {
        int32_t polling;
        int32_t lro;
        int32_t cong_thres;
+       int32_t  vector;
+       int32_t  qnum;
 };
 
 struct ch_pktsched_params {
index 0e51d49..fea2c71 100644 (file)
@@ -274,10 +274,10 @@ static void name_msix_vecs(struct adapter *adap)
 
                for (i = 0; i < pi->nqsets; i++, msi_idx++) {
                        snprintf(adap->msix_info[msi_idx].desc, n,
-                                "%s (queue %d)", d->name, i);
+                                "%s-%d", d->name, pi->first_qset + i);
                        adap->msix_info[msi_idx].desc[n] = 0;
                }
-       }
+       }
 }
 
 static int request_msix_data_irqs(struct adapter *adap)
@@ -306,6 +306,22 @@ static int request_msix_data_irqs(struct adapter *adap)
        return 0;
 }
 
+static void free_irq_resources(struct adapter *adapter)
+{
+       if (adapter->flags & USING_MSIX) {
+               int i, n = 0;
+
+               free_irq(adapter->msix_info[0].vec, adapter);
+               for_each_port(adapter, i)
+                   n += adap2pinfo(adapter, i)->nqsets;
+
+               for (i = 0; i < n; ++i)
+                       free_irq(adapter->msix_info[i + 1].vec,
+                                &adapter->sge.qs[i]);
+       } else
+               free_irq(adapter->pdev->irq, adapter);
+}
+
 static int await_mgmt_replies(struct adapter *adap, unsigned long init_cnt,
                              unsigned long n)
 {
@@ -473,7 +489,10 @@ static int setup_sge_qsets(struct adapter *adap)
                struct port_info *pi = netdev_priv(dev);
 
                pi->qs = &adap->sge.qs[pi->first_qset];
-               for (j = 0; j < pi->nqsets; ++j, ++qset_idx) {
+               for (j = pi->first_qset; j < pi->first_qset + pi->nqsets;
+                    ++j, ++qset_idx) {
+                       if (!pi->rx_csum_offload)
+                               adap->params.sge.qset[qset_idx].lro = 0;
                        err = t3_sge_alloc_qset(adap, qset_idx, 1,
                                (adap->flags & USING_MSIX) ? qset_idx + 1 :
                                                             irq_idx,
@@ -740,11 +759,12 @@ static void init_port_mtus(struct adapter *adapter)
        t3_write_reg(adapter, A_TP_MTU_PORT_TABLE, mtus);
 }
 
-static void send_pktsched_cmd(struct adapter *adap, int sched, int qidx, int lo,
+static int send_pktsched_cmd(struct adapter *adap, int sched, int qidx, int lo,
                              int hi, int port)
 {
        struct sk_buff *skb;
        struct mngt_pktsched_wr *req;
+       int ret;
 
        skb = alloc_skb(sizeof(*req), GFP_KERNEL | __GFP_NOFAIL);
        req = (struct mngt_pktsched_wr *)skb_put(skb, sizeof(*req));
@@ -755,20 +775,28 @@ static void send_pktsched_cmd(struct adapter *adap, int sched, int qidx, int lo,
        req->min = lo;
        req->max = hi;
        req->binding = port;
-       t3_mgmt_tx(adap, skb);
+       ret = t3_mgmt_tx(adap, skb);
+
+       return ret;
 }
 
-static void bind_qsets(struct adapter *adap)
+static int bind_qsets(struct adapter *adap)
 {
-       int i, j;
+       int i, j, err = 0;
 
        for_each_port(adap, i) {
                const struct port_info *pi = adap2pinfo(adap, i);
 
-               for (j = 0; j < pi->nqsets; ++j)
-                       send_pktsched_cmd(adap, 1, pi->first_qset + j, -1,
-                                         -1, i);
+               for (j = 0; j < pi->nqsets; ++j) {
+                       int ret = send_pktsched_cmd(adap, 1,
+                                                   pi->first_qset + j, -1,
+                                                   -1, i);
+                       if (ret)
+                               err = ret;
+               }
        }
+
+       return err;
 }
 
 #define FW_FNAME "t3fw-%d.%d.%d.bin"
@@ -954,9 +982,16 @@ static int cxgb_up(struct adapter *adap)
                t3_write_reg(adap, A_TP_INT_ENABLE, 0x7fbfffff);
        }
 
-       if ((adap->flags & (USING_MSIX | QUEUES_BOUND)) == USING_MSIX)
-               bind_qsets(adap);
-       adap->flags |= QUEUES_BOUND;
+       if (!(adap->flags & QUEUES_BOUND)) {
+               err = bind_qsets(adap);
+               if (err) {
+                       CH_ERR(adap, "failed to bind qsets, err %d\n", err);
+                       t3_intr_disable(adap);
+                       free_irq_resources(adap);
+                       goto out;
+               }
+               adap->flags |= QUEUES_BOUND;
+       }
 
 out:
        return err;
@@ -975,19 +1010,7 @@ static void cxgb_down(struct adapter *adapter)
        t3_intr_disable(adapter);
        spin_unlock_irq(&adapter->work_lock);
 
-       if (adapter->flags & USING_MSIX) {
-               int i, n = 0;
-
-               free_irq(adapter->msix_info[0].vec, adapter);
-               for_each_port(adapter, i)
-                   n += adap2pinfo(adapter, i)->nqsets;
-
-               for (i = 0; i < n; ++i)
-                       free_irq(adapter->msix_info[i + 1].vec,
-                                &adapter->sge.qs[i]);
-       } else
-               free_irq(adapter->pdev->irq, adapter);
-
+       free_irq_resources(adapter);
        flush_workqueue(cxgb3_wq);      /* wait for external IRQ handler */
        quiesce_rx(adapter);
 }
@@ -1292,8 +1315,8 @@ static unsigned long collect_sge_port_stats(struct adapter *adapter,
        int i;
        unsigned long tot = 0;
 
-       for (i = 0; i < p->nqsets; ++i)
-               tot += adapter->sge.qs[i + p->first_qset].port_stats[idx];
+       for (i = p->first_qset; i < p->first_qset + p->nqsets; ++i)
+               tot += adapter->sge.qs[i].port_stats[idx];
        return tot;
 }
 
@@ -1497,7 +1520,7 @@ static int set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
        struct link_config *lc = &p->link_config;
 
        if (!(lc->supported & SUPPORTED_Autoneg))
-               return -EOPNOTSUPP;     /* can't change speed/duplex */
+               return -EOPNOTSUPP;     /* can't change speed/duplex */
 
        if (cmd->autoneg == AUTONEG_DISABLE) {
                int cap = speed_duplex_to_caps(cmd->speed, cmd->duplex);
@@ -1576,8 +1599,10 @@ static int set_rx_csum(struct net_device *dev, u32 data)
                struct adapter *adap = p->adapter;
                int i;
 
-               for (i = p->first_qset; i < p->first_qset + p->nqsets; i++)
+               for (i = p->first_qset; i < p->first_qset + p->nqsets; i++) {
+                       adap->params.sge.qset[i].lro = 0;
                        adap->sge.qs[i].lro_enabled = 0;
+               }
        }
        return 0;
 }
@@ -1783,6 +1808,8 @@ static int cxgb_extension_ioctl(struct net_device *dev, void __user *useraddr)
                int i;
                struct qset_params *q;
                struct ch_qset_params t;
+               int q1 = pi->first_qset;
+               int nqsets = pi->nqsets;
 
                if (!capable(CAP_NET_ADMIN))
                        return -EPERM;
@@ -1805,6 +1832,16 @@ static int cxgb_extension_ioctl(struct net_device *dev, void __user *useraddr)
                        || !in_range(t.rspq_size, MIN_RSPQ_ENTRIES,
                                        MAX_RSPQ_ENTRIES))
                        return -EINVAL;
+
+               if ((adapter->flags & FULL_INIT_DONE) && t.lro > 0)
+                       for_each_port(adapter, i) {
+                               pi = adap2pinfo(adapter, i);
+                               if (t.qset_idx >= pi->first_qset &&
+                                   t.qset_idx < pi->first_qset + pi->nqsets &&
+                                   !pi->rx_csum_offload)
+                                       return -EINVAL;
+                       }
+
                if ((adapter->flags & FULL_INIT_DONE) &&
                        (t.rspq_size >= 0 || t.fl_size[0] >= 0 ||
                        t.fl_size[1] >= 0 || t.txq_size[0] >= 0 ||
@@ -1812,6 +1849,20 @@ static int cxgb_extension_ioctl(struct net_device *dev, void __user *useraddr)
                        t.polling >= 0 || t.cong_thres >= 0))
                        return -EBUSY;
 
+               /* Allow setting of any available qset when offload enabled */
+               if (test_bit(OFFLOAD_DEVMAP_BIT, &adapter->open_device_map)) {
+                       q1 = 0;
+                       for_each_port(adapter, i) {
+                               pi = adap2pinfo(adapter, i);
+                               nqsets += pi->first_qset + pi->nqsets;
+                       }
+               }
+
+               if (t.qset_idx < q1)
+                       return -EINVAL;
+               if (t.qset_idx > q1 + nqsets - 1)
+                       return -EINVAL;
+
                q = &adapter->params.sge.qset[t.qset_idx];
 
                if (t.rspq_size >= 0)
@@ -1861,13 +1912,26 @@ static int cxgb_extension_ioctl(struct net_device *dev, void __user *useraddr)
        case CHELSIO_GET_QSET_PARAMS:{
                struct qset_params *q;
                struct ch_qset_params t;
+               int q1 = pi->first_qset;
+               int nqsets = pi->nqsets;
+               int i;
 
                if (copy_from_user(&t, useraddr, sizeof(t)))
                        return -EFAULT;
-               if (t.qset_idx >= SGE_QSETS)
+
+               /* Display qsets for all ports when offload enabled */
+               if (test_bit(OFFLOAD_DEVMAP_BIT, &adapter->open_device_map)) {
+                       q1 = 0;
+                       for_each_port(adapter, i) {
+                               pi = adap2pinfo(adapter, i);
+                               nqsets = pi->first_qset + pi->nqsets;
+                       }
+               }
+
+               if (t.qset_idx >= nqsets)
                        return -EINVAL;
 
-               q = &adapter->params.sge.qset[t.qset_idx];
+               q = &adapter->params.sge.qset[q1 + t.qset_idx];
                t.rspq_size = q->rspq_size;
                t.txq_size[0] = q->txq_size[0];
                t.txq_size[1] = q->txq_size[1];
@@ -1878,6 +1942,12 @@ static int cxgb_extension_ioctl(struct net_device *dev, void __user *useraddr)
                t.lro = q->lro;
                t.intr_lat = q->coalesce_usecs;
                t.cong_thres = q->cong_thres;
+               t.qnum = q1;
+
+               if (adapter->flags & USING_MSIX)
+                       t.vector = adapter->msix_info[q1 + t.qset_idx + 1].vec;
+               else
+                       t.vector = adapter->pdev->irq;
 
                if (copy_to_user(useraddr, &t, sizeof(t)))
                        return -EFAULT;
@@ -2223,8 +2293,8 @@ static void t3_synchronize_rx(struct adapter *adap, const struct port_info *p)
 {
        int i;
 
-       for (i = 0; i < p->nqsets; i++) {
-               struct sge_rspq *q = &adap->sge.qs[i + p->first_qset].rspq;
+       for (i = p->first_qset; i < p->first_qset + p->nqsets; i++) {
+               struct sge_rspq *q = &adap->sge.qs[i].rspq;
 
                spin_lock_irq(&q->lock);
                spin_unlock_irq(&q->lock);
@@ -2581,6 +2651,42 @@ static struct pci_error_handlers t3_err_handler = {
        .resume = t3_io_resume,
 };
 
+/*
+ * Set the number of qsets based on the number of CPUs and the number of ports,
+ * not to exceed the number of available qsets, assuming there are enough qsets
+ * per port in HW.
+ */
+static void set_nqsets(struct adapter *adap)
+{
+       int i, j = 0;
+       int num_cpus = num_online_cpus();
+       int hwports = adap->params.nports;
+       int nqsets = SGE_QSETS;
+
+       if (adap->params.rev > 0) {
+               if (hwports == 2 &&
+                   (hwports * nqsets > SGE_QSETS ||
+                    num_cpus >= nqsets / hwports))
+                       nqsets /= hwports;
+               if (nqsets > num_cpus)
+                       nqsets = num_cpus;
+               if (nqsets < 1 || hwports == 4)
+                       nqsets = 1;
+       } else
+               nqsets = 1;
+
+       for_each_port(adap, i) {
+               struct port_info *pi = adap2pinfo(adap, i);
+
+               pi->first_qset = j;
+               pi->nqsets = nqsets;
+               j = pi->first_qset + nqsets;
+
+               dev_info(&adap->pdev->dev,
+                        "Port %d using %d queue sets.\n", i, nqsets);
+       }
+}
+
 static int __devinit cxgb_enable_msix(struct adapter *adap)
 {
        struct msix_entry entries[SGE_QSETS + 1];
@@ -2739,9 +2845,6 @@ static int __devinit init_one(struct pci_dev *pdev,
                pi = netdev_priv(netdev);
                pi->adapter = adapter;
                pi->rx_csum_offload = 1;
-               pi->nqsets = 1;
-               pi->first_qset = i;
-               pi->activity = 0;
                pi->port_id = i;
                netif_carrier_off(netdev);
                netdev->irq = pdev->irq;
@@ -2818,6 +2921,8 @@ static int __devinit init_one(struct pci_dev *pdev,
        else if (msi > 0 && pci_enable_msi(pdev) == 0)
                adapter->flags |= USING_MSI;
 
+       set_nqsets(adapter);
+
        err = sysfs_create_group(&adapter->port[0]->dev.kobj,
                                 &cxgb3_attr_group);