octeon_ep: add Tx/Rx processing and interrupt support
authorVeerasenareddy Burru <vburru@marvell.com>
Wed, 13 Apr 2022 03:35:02 +0000 (20:35 -0700)
committerDavid S. Miller <davem@davemloft.net>
Wed, 13 Apr 2022 11:56:32 +0000 (12:56 +0100)
Add support to enable MSI-x and register interrupts.
Add support to process Tx and Rx traffic. Includes processing
Tx completions and Rx refill.

Signed-off-by: Veerasenareddy Burru <vburru@marvell.com>
Signed-off-by: Abhijit Ayarekar <aayarekar@marvell.com>
Signed-off-by: Satananda Burla <sburla@marvell.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/marvell/octeon_ep/octep_main.c
drivers/net/ethernet/marvell/octeon_ep/octep_rx.c
drivers/net/ethernet/marvell/octeon_ep/octep_tx.c

index dac1e94647942ebc84229e6809ec53e33ad3b974..2f6579d7c945160d17511db304e0530ac69812d7 100644 (file)
@@ -32,6 +32,431 @@ MODULE_DESCRIPTION(OCTEP_DRV_STRING);
 MODULE_LICENSE("GPL");
 MODULE_VERSION(OCTEP_DRV_VERSION_STR);
 
+/**
+ * octep_alloc_ioq_vectors() - Allocate Tx/Rx Queue interrupt info.
+ *
+ * @oct: Octeon device private data structure.
+ *
+ * Allocate resources to hold per Tx/Rx queue interrupt info.
+ * This is the information passed to interrupt handler, from which napi poll
+ * is scheduled and includes quick access to private data of Tx/Rx queue
+ * corresponding to the interrupt being handled.
+ *
+ * Return: 0, on successful allocation of resources for all queue interrupts.
+ *         -1, if failed to allocate any resource.
+ */
+static int octep_alloc_ioq_vectors(struct octep_device *oct)
+{
+       int i;
+       struct octep_ioq_vector *ioq_vector;
+
+       for (i = 0; i < oct->num_oqs; i++) {
+               oct->ioq_vector[i] = vzalloc(sizeof(*oct->ioq_vector[i]));
+               if (!oct->ioq_vector[i])
+                       goto free_ioq_vector;
+
+               ioq_vector = oct->ioq_vector[i];
+               ioq_vector->iq = oct->iq[i];
+               ioq_vector->oq = oct->oq[i];
+               ioq_vector->octep_dev = oct;
+       }
+
+       dev_info(&oct->pdev->dev, "Allocated %d IOQ vectors\n", oct->num_oqs);
+       return 0;
+
+free_ioq_vector:
+       while (i) {
+               i--;
+               vfree(oct->ioq_vector[i]);
+               oct->ioq_vector[i] = NULL;
+       }
+       return -1;
+}
+
+/**
+ * octep_free_ioq_vectors() - Free Tx/Rx Queue interrupt vector info.
+ *
+ * @oct: Octeon device private data structure.
+ */
+static void octep_free_ioq_vectors(struct octep_device *oct)
+{
+       int i;
+
+       for (i = 0; i < oct->num_oqs; i++) {
+               if (oct->ioq_vector[i]) {
+                       vfree(oct->ioq_vector[i]);
+                       oct->ioq_vector[i] = NULL;
+               }
+       }
+       netdev_info(oct->netdev, "Freed IOQ Vectors\n");
+}
+
+/**
+ * octep_enable_msix_range() - enable MSI-x interrupts.
+ *
+ * @oct: Octeon device private data structure.
+ *
+ * Allocate and enable all MSI-x interrupts (queue and non-queue interrupts)
+ * for the Octeon device.
+ *
+ * Return: 0, on successfully enabling all MSI-x interrupts.
+ *         -1, if failed to enable any MSI-x interrupt.
+ */
+static int octep_enable_msix_range(struct octep_device *oct)
+{
+       int num_msix, msix_allocated;
+       int i;
+
+       /* Generic interrupts apart from input/output queues */
+       num_msix = oct->num_oqs + CFG_GET_NON_IOQ_MSIX(oct->conf);
+       oct->msix_entries = kcalloc(num_msix,
+                                   sizeof(struct msix_entry), GFP_KERNEL);
+       if (!oct->msix_entries)
+               goto msix_alloc_err;
+
+       for (i = 0; i < num_msix; i++)
+               oct->msix_entries[i].entry = i;
+
+       msix_allocated = pci_enable_msix_range(oct->pdev, oct->msix_entries,
+                                              num_msix, num_msix);
+       if (msix_allocated != num_msix) {
+               dev_err(&oct->pdev->dev,
+                       "Failed to enable %d msix irqs; got only %d\n",
+                       num_msix, msix_allocated);
+               goto enable_msix_err;
+       }
+       oct->num_irqs = msix_allocated;
+       dev_info(&oct->pdev->dev, "MSI-X enabled successfully\n");
+
+       return 0;
+
+enable_msix_err:
+       if (msix_allocated > 0)
+               pci_disable_msix(oct->pdev);
+       kfree(oct->msix_entries);
+       oct->msix_entries = NULL;
+msix_alloc_err:
+       return -1;
+}
+
+/**
+ * octep_disable_msix() - disable MSI-x interrupts.
+ *
+ * @oct: Octeon device private data structure.
+ *
+ * Disable MSI-x on the Octeon device.
+ */
+static void octep_disable_msix(struct octep_device *oct)
+{
+       pci_disable_msix(oct->pdev);
+       kfree(oct->msix_entries);
+       oct->msix_entries = NULL;
+       dev_info(&oct->pdev->dev, "Disabled MSI-X\n");
+}
+
+/**
+ * octep_non_ioq_intr_handler() - common handler for all generic interrupts.
+ *
+ * @irq: Interrupt number.
+ * @data: interrupt data.
+ *
+ * this is common handler for all non-queue (generic) interrupts.
+ */
+static irqreturn_t octep_non_ioq_intr_handler(int irq, void *data)
+{
+       struct octep_device *oct = data;
+
+       return oct->hw_ops.non_ioq_intr_handler(oct);
+}
+
+/**
+ * octep_ioq_intr_handler() - handler for all Tx/Rx queue interrupts.
+ *
+ * @irq: Interrupt number.
+ * @data: interrupt data contains pointers to Tx/Rx queue private data
+ *         and correspong NAPI context.
+ *
+ * this is common handler for all non-queue (generic) interrupts.
+ */
+static irqreturn_t octep_ioq_intr_handler(int irq, void *data)
+{
+       struct octep_ioq_vector *ioq_vector = data;
+       struct octep_device *oct = ioq_vector->octep_dev;
+
+       return oct->hw_ops.ioq_intr_handler(ioq_vector);
+}
+
+/**
+ * octep_request_irqs() - Register interrupt handlers.
+ *
+ * @oct: Octeon device private data structure.
+ *
+ * Register handlers for all queue and non-queue interrupts.
+ *
+ * Return: 0, on successful registration of all interrupt handlers.
+ *         -1, on any error.
+ */
+static int octep_request_irqs(struct octep_device *oct)
+{
+       struct net_device *netdev = oct->netdev;
+       struct octep_ioq_vector *ioq_vector;
+       struct msix_entry *msix_entry;
+       char **non_ioq_msix_names;
+       int num_non_ioq_msix;
+       int ret, i;
+
+       num_non_ioq_msix = CFG_GET_NON_IOQ_MSIX(oct->conf);
+       non_ioq_msix_names = CFG_GET_NON_IOQ_MSIX_NAMES(oct->conf);
+
+       oct->non_ioq_irq_names = kcalloc(num_non_ioq_msix,
+                                        OCTEP_MSIX_NAME_SIZE, GFP_KERNEL);
+       if (!oct->non_ioq_irq_names)
+               goto alloc_err;
+
+       /* First few MSI-X interrupts are non-queue interrupts */
+       for (i = 0; i < num_non_ioq_msix; i++) {
+               char *irq_name;
+
+               irq_name = &oct->non_ioq_irq_names[i * OCTEP_MSIX_NAME_SIZE];
+               msix_entry = &oct->msix_entries[i];
+
+               snprintf(irq_name, OCTEP_MSIX_NAME_SIZE,
+                        "%s-%s", netdev->name, non_ioq_msix_names[i]);
+               ret = request_irq(msix_entry->vector,
+                                 octep_non_ioq_intr_handler, 0,
+                                 irq_name, oct);
+               if (ret) {
+                       netdev_err(netdev,
+                                  "request_irq failed for %s; err=%d",
+                                  irq_name, ret);
+                       goto non_ioq_irq_err;
+               }
+       }
+
+       /* Request IRQs for Tx/Rx queues */
+       for (i = 0; i < oct->num_oqs; i++) {
+               ioq_vector = oct->ioq_vector[i];
+               msix_entry = &oct->msix_entries[i + num_non_ioq_msix];
+
+               snprintf(ioq_vector->name, sizeof(ioq_vector->name),
+                        "%s-q%d", netdev->name, i);
+               ret = request_irq(msix_entry->vector,
+                                 octep_ioq_intr_handler, 0,
+                                 ioq_vector->name, ioq_vector);
+               if (ret) {
+                       netdev_err(netdev,
+                                  "request_irq failed for Q-%d; err=%d",
+                                  i, ret);
+                       goto ioq_irq_err;
+               }
+
+               cpumask_set_cpu(i % num_online_cpus(),
+                               &ioq_vector->affinity_mask);
+               irq_set_affinity_hint(msix_entry->vector,
+                                     &ioq_vector->affinity_mask);
+       }
+
+       return 0;
+ioq_irq_err:
+       while (i > num_non_ioq_msix) {
+               --i;
+               irq_set_affinity_hint(oct->msix_entries[i].vector, NULL);
+               free_irq(oct->msix_entries[i].vector, oct->ioq_vector[i]);
+       }
+non_ioq_irq_err:
+       while (i) {
+               --i;
+               free_irq(oct->msix_entries[i].vector, oct);
+       }
+alloc_err:
+       return -1;
+}
+
+/**
+ * octep_free_irqs() - free all registered interrupts.
+ *
+ * @oct: Octeon device private data structure.
+ *
+ * Free all queue and non-queue interrupts of the Octeon device.
+ */
+static void octep_free_irqs(struct octep_device *oct)
+{
+       int i;
+
+       /* First few MSI-X interrupts are non queue interrupts; free them */
+       for (i = 0; i < CFG_GET_NON_IOQ_MSIX(oct->conf); i++)
+               free_irq(oct->msix_entries[i].vector, oct);
+       kfree(oct->non_ioq_irq_names);
+
+       /* Free IRQs for Input/Output (Tx/Rx) queues */
+       for (i = CFG_GET_NON_IOQ_MSIX(oct->conf); i < oct->num_irqs; i++) {
+               irq_set_affinity_hint(oct->msix_entries[i].vector, NULL);
+               free_irq(oct->msix_entries[i].vector,
+                        oct->ioq_vector[i - CFG_GET_NON_IOQ_MSIX(oct->conf)]);
+       }
+       netdev_info(oct->netdev, "IRQs freed\n");
+}
+
+/**
+ * octep_setup_irqs() - setup interrupts for the Octeon device.
+ *
+ * @oct: Octeon device private data structure.
+ *
+ * Allocate data structures to hold per interrupt information, allocate/enable
+ * MSI-x interrupt and register interrupt handlers.
+ *
+ * Return: 0, on successful allocation and registration of all interrupts.
+ *         -1, on any error.
+ */
+static int octep_setup_irqs(struct octep_device *oct)
+{
+       if (octep_alloc_ioq_vectors(oct))
+               goto ioq_vector_err;
+
+       if (octep_enable_msix_range(oct))
+               goto enable_msix_err;
+
+       if (octep_request_irqs(oct))
+               goto request_irq_err;
+
+       return 0;
+
+request_irq_err:
+       octep_disable_msix(oct);
+enable_msix_err:
+       octep_free_ioq_vectors(oct);
+ioq_vector_err:
+       return -1;
+}
+
+/**
+ * octep_clean_irqs() - free all interrupts and its resources.
+ *
+ * @oct: Octeon device private data structure.
+ */
+static void octep_clean_irqs(struct octep_device *oct)
+{
+       octep_free_irqs(oct);
+       octep_disable_msix(oct);
+       octep_free_ioq_vectors(oct);
+}
+
+/**
+ * octep_enable_ioq_irq() - Enable MSI-x interrupt of a Tx/Rx queue.
+ *
+ * @iq: Octeon Tx queue data structure.
+ * @oq: Octeon Rx queue data structure.
+ */
+static void octep_enable_ioq_irq(struct octep_iq *iq, struct octep_oq *oq)
+{
+       u32 pkts_pend = oq->pkts_pending;
+
+       netdev_dbg(iq->netdev, "enabling intr for Q-%u\n", iq->q_no);
+       if (iq->pkts_processed) {
+               writel(iq->pkts_processed, iq->inst_cnt_reg);
+               iq->pkt_in_done -= iq->pkts_processed;
+               iq->pkts_processed = 0;
+       }
+       if (oq->last_pkt_count - pkts_pend) {
+               writel(oq->last_pkt_count - pkts_pend, oq->pkts_sent_reg);
+               oq->last_pkt_count = pkts_pend;
+       }
+
+       /* Flush the previous wrties before writing to RESEND bit */
+       wmb();
+       writeq(1UL << OCTEP_OQ_INTR_RESEND_BIT, oq->pkts_sent_reg);
+       writeq(1UL << OCTEP_IQ_INTR_RESEND_BIT, iq->inst_cnt_reg);
+}
+
+/**
+ * octep_napi_poll() - NAPI poll function for Tx/Rx.
+ *
+ * @napi: pointer to napi context.
+ * @budget: max number of packets to be processed in single invocation.
+ */
+static int octep_napi_poll(struct napi_struct *napi, int budget)
+{
+       struct octep_ioq_vector *ioq_vector =
+               container_of(napi, struct octep_ioq_vector, napi);
+       u32 tx_pending, rx_done;
+
+       tx_pending = octep_iq_process_completions(ioq_vector->iq, budget);
+       rx_done = octep_oq_process_rx(ioq_vector->oq, budget);
+
+       /* need more polling if tx completion processing is still pending or
+        * processed at least 'budget' number of rx packets.
+        */
+       if (tx_pending || rx_done >= budget)
+               return budget;
+
+       napi_complete(napi);
+       octep_enable_ioq_irq(ioq_vector->iq, ioq_vector->oq);
+       return rx_done;
+}
+
+/**
+ * octep_napi_add() - Add NAPI poll for all Tx/Rx queues.
+ *
+ * @oct: Octeon device private data structure.
+ */
+static void octep_napi_add(struct octep_device *oct)
+{
+       int i;
+
+       for (i = 0; i < oct->num_oqs; i++) {
+               netdev_dbg(oct->netdev, "Adding NAPI on Q-%d\n", i);
+               netif_napi_add(oct->netdev, &oct->ioq_vector[i]->napi,
+                              octep_napi_poll, 64);
+               oct->oq[i]->napi = &oct->ioq_vector[i]->napi;
+       }
+}
+
+/**
+ * octep_napi_delete() - delete NAPI poll callback for all Tx/Rx queues.
+ *
+ * @oct: Octeon device private data structure.
+ */
+static void octep_napi_delete(struct octep_device *oct)
+{
+       int i;
+
+       for (i = 0; i < oct->num_oqs; i++) {
+               netdev_dbg(oct->netdev, "Deleting NAPI on Q-%d\n", i);
+               netif_napi_del(&oct->ioq_vector[i]->napi);
+               oct->oq[i]->napi = NULL;
+       }
+}
+
+/**
+ * octep_napi_enable() - enable NAPI for all Tx/Rx queues.
+ *
+ * @oct: Octeon device private data structure.
+ */
+static void octep_napi_enable(struct octep_device *oct)
+{
+       int i;
+
+       for (i = 0; i < oct->num_oqs; i++) {
+               netdev_dbg(oct->netdev, "Enabling NAPI on Q-%d\n", i);
+               napi_enable(&oct->ioq_vector[i]->napi);
+       }
+}
+
+/**
+ * octep_napi_disable() - disable NAPI for all Tx/Rx queues.
+ *
+ * @oct: Octeon device private data structure.
+ */
+static void octep_napi_disable(struct octep_device *oct)
+{
+       int i;
+
+       for (i = 0; i < oct->num_oqs; i++) {
+               netdev_dbg(oct->netdev, "Disabling NAPI on Q-%d\n", i);
+               napi_disable(&oct->ioq_vector[i]->napi);
+       }
+}
+
 static void octep_link_up(struct net_device *netdev)
 {
        netif_carrier_on(netdev);
@@ -63,6 +488,8 @@ static int octep_open(struct net_device *netdev)
                goto setup_iq_err;
        if (octep_setup_oqs(oct))
                goto setup_oq_err;
+       if (octep_setup_irqs(oct))
+               goto setup_irq_err;
 
        err = netif_set_real_num_tx_queues(netdev, oct->num_oqs);
        if (err)
@@ -71,6 +498,9 @@ static int octep_open(struct net_device *netdev)
        if (err)
                goto set_queues_err;
 
+       octep_napi_add(oct);
+       octep_napi_enable(oct);
+
        oct->link_info.admin_up = 1;
        octep_set_rx_state(oct, true);
 
@@ -93,6 +523,10 @@ static int octep_open(struct net_device *netdev)
        return 0;
 
 set_queues_err:
+       octep_napi_disable(oct);
+       octep_napi_delete(oct);
+       octep_clean_irqs(oct);
+setup_irq_err:
        octep_free_oqs(oct);
 setup_oq_err:
        octep_free_iqs(oct);
@@ -126,7 +560,10 @@ static int octep_stop(struct net_device *netdev)
        oct->link_info.oper_up = 0;
 
        oct->hw_ops.disable_interrupts(oct);
+       octep_napi_disable(oct);
+       octep_napi_delete(oct);
 
+       octep_clean_irqs(oct);
        octep_clean_iqs(oct);
 
        oct->hw_ops.disable_io_queues(oct);
@@ -137,6 +574,36 @@ static int octep_stop(struct net_device *netdev)
        return 0;
 }
 
+/**
+ * octep_iq_full_check() - check if a Tx queue is full.
+ *
+ * @iq: Octeon Tx queue data structure.
+ *
+ * Return: 0, if the Tx queue is not full.
+ *         1, if the Tx queue is full.
+ */
+static inline int octep_iq_full_check(struct octep_iq *iq)
+{
+       if (likely((iq->max_count - atomic_read(&iq->instr_pending)) >=
+                  OCTEP_WAKE_QUEUE_THRESHOLD))
+               return 0;
+
+       /* Stop the queue if unable to send */
+       netif_stop_subqueue(iq->netdev, iq->q_no);
+
+       /* check again and restart the queue, in case NAPI has just freed
+        * enough Tx ring entries.
+        */
+       if (unlikely((iq->max_count - atomic_read(&iq->instr_pending)) >=
+                    OCTEP_WAKE_QUEUE_THRESHOLD)) {
+               netif_start_subqueue(iq->netdev, iq->q_no);
+               iq->stats.restart_cnt++;
+               return 0;
+       }
+
+       return 1;
+}
+
 /**
  * octep_start_xmit() - Enqueue packet to Octoen hardware Tx Queue.
  *
@@ -149,6 +616,126 @@ static int octep_stop(struct net_device *netdev)
 static netdev_tx_t octep_start_xmit(struct sk_buff *skb,
                                    struct net_device *netdev)
 {
+       struct octep_device *oct = netdev_priv(netdev);
+       struct octep_tx_sglist_desc *sglist;
+       struct octep_tx_buffer *tx_buffer;
+       struct octep_tx_desc_hw *hw_desc;
+       struct skb_shared_info *shinfo;
+       struct octep_instr_hdr *ih;
+       struct octep_iq *iq;
+       skb_frag_t *frag;
+       u16 nr_frags, si;
+       u16 q_no, wi;
+
+       q_no = skb_get_queue_mapping(skb);
+       if (q_no >= oct->num_iqs) {
+               netdev_err(netdev, "Invalid Tx skb->queue_mapping=%d\n", q_no);
+               q_no = q_no % oct->num_iqs;
+       }
+
+       iq = oct->iq[q_no];
+       if (octep_iq_full_check(iq)) {
+               iq->stats.tx_busy++;
+               return NETDEV_TX_BUSY;
+       }
+
+       shinfo = skb_shinfo(skb);
+       nr_frags = shinfo->nr_frags;
+
+       wi = iq->host_write_index;
+       hw_desc = &iq->desc_ring[wi];
+       hw_desc->ih64 = 0;
+
+       tx_buffer = iq->buff_info + wi;
+       tx_buffer->skb = skb;
+
+       ih = &hw_desc->ih;
+       ih->tlen = skb->len;
+       ih->pkind = oct->pkind;
+
+       if (!nr_frags) {
+               tx_buffer->gather = 0;
+               tx_buffer->dma = dma_map_single(iq->dev, skb->data,
+                                               skb->len, DMA_TO_DEVICE);
+               if (dma_mapping_error(iq->dev, tx_buffer->dma))
+                       goto dma_map_err;
+               hw_desc->dptr = tx_buffer->dma;
+       } else {
+               /* Scatter/Gather */
+               dma_addr_t dma;
+               u16 len;
+
+               sglist = tx_buffer->sglist;
+
+               ih->gsz = nr_frags + 1;
+               ih->gather = 1;
+               tx_buffer->gather = 1;
+
+               len = skb_headlen(skb);
+               dma = dma_map_single(iq->dev, skb->data, len, DMA_TO_DEVICE);
+               if (dma_mapping_error(iq->dev, dma))
+                       goto dma_map_err;
+
+               dma_sync_single_for_cpu(iq->dev, tx_buffer->sglist_dma,
+                                       OCTEP_SGLIST_SIZE_PER_PKT,
+                                       DMA_TO_DEVICE);
+               memset(sglist, 0, OCTEP_SGLIST_SIZE_PER_PKT);
+               sglist[0].len[3] = len;
+               sglist[0].dma_ptr[0] = dma;
+
+               si = 1; /* entry 0 is main skb, mapped above */
+               frag = &shinfo->frags[0];
+               while (nr_frags--) {
+                       len = skb_frag_size(frag);
+                       dma = skb_frag_dma_map(iq->dev, frag, 0,
+                                              len, DMA_TO_DEVICE);
+                       if (dma_mapping_error(iq->dev, dma))
+                               goto dma_map_sg_err;
+
+                       sglist[si >> 2].len[3 - (si & 3)] = len;
+                       sglist[si >> 2].dma_ptr[si & 3] = dma;
+
+                       frag++;
+                       si++;
+               }
+               dma_sync_single_for_device(iq->dev, tx_buffer->sglist_dma,
+                                          OCTEP_SGLIST_SIZE_PER_PKT,
+                                          DMA_TO_DEVICE);
+
+               hw_desc->dptr = tx_buffer->sglist_dma;
+       }
+
+       /* Flush the hw descriptor before writing to doorbell */
+       wmb();
+
+       /* Ring Doorbell to notify the NIC there is a new packet */
+       writel(1, iq->doorbell_reg);
+       atomic_inc(&iq->instr_pending);
+       wi++;
+       if (wi == iq->max_count)
+               wi = 0;
+       iq->host_write_index = wi;
+
+       netdev_tx_sent_queue(iq->netdev_q, skb->len);
+       iq->stats.instr_posted++;
+       skb_tx_timestamp(skb);
+       return NETDEV_TX_OK;
+
+dma_map_sg_err:
+       if (si > 0) {
+               dma_unmap_single(iq->dev, sglist[0].dma_ptr[0],
+                                sglist[0].len[0], DMA_TO_DEVICE);
+               sglist[0].len[0] = 0;
+       }
+       while (si > 1) {
+               dma_unmap_page(iq->dev, sglist[si >> 2].dma_ptr[si & 3],
+                              sglist[si >> 2].len[si & 3], DMA_TO_DEVICE);
+               sglist[si >> 2].len[si & 3] = 0;
+               si--;
+       }
+       tx_buffer->gather = 0;
+dma_map_err:
+       dev_kfree_skb_any(skb);
        return NETDEV_TX_OK;
 }
 
index 7086ede58475fc1f8275c6919661cb6916b5085d..945947ec77239b057dccd9766190dc91b558ca84 100644 (file)
@@ -68,6 +68,50 @@ rx_buf_alloc_err:
        return -1;
 }
 
+/**
+ * octep_oq_refill() - refill buffers for used Rx ring descriptors.
+ *
+ * @oct: Octeon device private data structure.
+ * @oq: Octeon Rx queue data structure.
+ *
+ * Return: number of descriptors successfully refilled with receive buffers.
+ */
+static int octep_oq_refill(struct octep_device *oct, struct octep_oq *oq)
+{
+       struct octep_oq_desc_hw *desc_ring = oq->desc_ring;
+       struct page *page;
+       u32 refill_idx, i;
+
+       refill_idx = oq->host_refill_idx;
+       for (i = 0; i < oq->refill_count; i++) {
+               page = dev_alloc_page();
+               if (unlikely(!page)) {
+                       dev_err(oq->dev, "refill: rx buffer alloc failed\n");
+                       oq->stats.alloc_failures++;
+                       break;
+               }
+
+               desc_ring[refill_idx].buffer_ptr = dma_map_page(oq->dev, page, 0,
+                                                               PAGE_SIZE, DMA_FROM_DEVICE);
+               if (dma_mapping_error(oq->dev, desc_ring[refill_idx].buffer_ptr)) {
+                       dev_err(oq->dev,
+                               "OQ-%d buffer refill: DMA mapping error!\n",
+                               oq->q_no);
+                       put_page(page);
+                       oq->stats.alloc_failures++;
+                       break;
+               }
+               oq->buff_info[refill_idx].page = page;
+               refill_idx++;
+               if (refill_idx == oq->max_count)
+                       refill_idx = 0;
+       }
+       oq->host_refill_idx = refill_idx;
+       oq->refill_count -= i;
+
+       return i;
+}
+
 /**
  * octep_setup_oq() - Setup a Rx queue.
  *
@@ -262,3 +306,203 @@ void octep_free_oqs(struct octep_device *oct)
                        "Successfully freed OQ(RxQ)-%d.\n", i);
        }
 }
+
+/**
+ * octep_oq_check_hw_for_pkts() - Check for new Rx packets.
+ *
+ * @oct: Octeon device private data structure.
+ * @oq: Octeon Rx queue data structure.
+ *
+ * Return: packets received after previous check.
+ */
+static int octep_oq_check_hw_for_pkts(struct octep_device *oct,
+                                     struct octep_oq *oq)
+{
+       u32 pkt_count, new_pkts;
+
+       pkt_count = readl(oq->pkts_sent_reg);
+       new_pkts = pkt_count - oq->last_pkt_count;
+
+       /* Clear the hardware packets counter register if the rx queue is
+        * being processed continuously with-in a single interrupt and
+        * reached half its max value.
+        * this counter is not cleared every time read, to save write cycles.
+        */
+       if (unlikely(pkt_count > 0xF0000000U)) {
+               writel(pkt_count, oq->pkts_sent_reg);
+               pkt_count = readl(oq->pkts_sent_reg);
+               new_pkts += pkt_count;
+       }
+       oq->last_pkt_count = pkt_count;
+       oq->pkts_pending += new_pkts;
+       return new_pkts;
+}
+
+/**
+ * __octep_oq_process_rx() - Process hardware Rx queue and push to stack.
+ *
+ * @oct: Octeon device private data structure.
+ * @oq: Octeon Rx queue data structure.
+ * @pkts_to_process: number of packets to be processed.
+ *
+ * Process the new packets in Rx queue.
+ * Packets larger than single Rx buffer arrive in consecutive descriptors.
+ * But, count returned by the API only accounts full packets, not fragments.
+ *
+ * Return: number of packets processed and pushed to stack.
+ */
+static int __octep_oq_process_rx(struct octep_device *oct,
+                                struct octep_oq *oq, u16 pkts_to_process)
+{
+       struct octep_oq_resp_hw_ext *resp_hw_ext = NULL;
+       struct octep_rx_buffer *buff_info;
+       struct octep_oq_resp_hw *resp_hw;
+       u32 pkt, rx_bytes, desc_used;
+       struct sk_buff *skb;
+       u16 data_offset;
+       u32 read_idx;
+
+       read_idx = oq->host_read_idx;
+       rx_bytes = 0;
+       desc_used = 0;
+       for (pkt = 0; pkt < pkts_to_process; pkt++) {
+               buff_info = (struct octep_rx_buffer *)&oq->buff_info[read_idx];
+               dma_unmap_page(oq->dev, oq->desc_ring[read_idx].buffer_ptr,
+                              PAGE_SIZE, DMA_FROM_DEVICE);
+               resp_hw = page_address(buff_info->page);
+               buff_info->page = NULL;
+
+               /* Swap the length field that is in Big-Endian to CPU */
+               buff_info->len = be64_to_cpu(resp_hw->length);
+               if (oct->caps_enabled & OCTEP_CAP_RX_CHECKSUM) {
+                       /* Extended response header is immediately after
+                        * response header (resp_hw)
+                        */
+                       resp_hw_ext = (struct octep_oq_resp_hw_ext *)
+                                     (resp_hw + 1);
+                       buff_info->len -= OCTEP_OQ_RESP_HW_EXT_SIZE;
+                       /* Packet Data is immediately after
+                        * extended response header.
+                        */
+                       data_offset = OCTEP_OQ_RESP_HW_SIZE +
+                                     OCTEP_OQ_RESP_HW_EXT_SIZE;
+               } else {
+                       /* Data is immediately after
+                        * Hardware Rx response header.
+                        */
+                       data_offset = OCTEP_OQ_RESP_HW_SIZE;
+               }
+               rx_bytes += buff_info->len;
+
+               if (buff_info->len <= oq->max_single_buffer_size) {
+                       skb = build_skb((void *)resp_hw, PAGE_SIZE);
+                       skb_reserve(skb, data_offset);
+                       skb_put(skb, buff_info->len);
+                       read_idx++;
+                       desc_used++;
+                       if (read_idx == oq->max_count)
+                               read_idx = 0;
+               } else {
+                       struct skb_shared_info *shinfo;
+                       u16 data_len;
+
+                       skb = build_skb((void *)resp_hw, PAGE_SIZE);
+                       skb_reserve(skb, data_offset);
+                       /* Head fragment includes response header(s);
+                        * subsequent fragments contains only data.
+                        */
+                       skb_put(skb, oq->max_single_buffer_size);
+                       read_idx++;
+                       desc_used++;
+                       if (read_idx == oq->max_count)
+                               read_idx = 0;
+
+                       shinfo = skb_shinfo(skb);
+                       data_len = buff_info->len - oq->max_single_buffer_size;
+                       while (data_len) {
+                               dma_unmap_page(oq->dev, oq->desc_ring[read_idx].buffer_ptr,
+                                              PAGE_SIZE, DMA_FROM_DEVICE);
+                               buff_info = (struct octep_rx_buffer *)
+                                           &oq->buff_info[read_idx];
+                               if (data_len < oq->buffer_size) {
+                                       buff_info->len = data_len;
+                                       data_len = 0;
+                               } else {
+                                       buff_info->len = oq->buffer_size;
+                                       data_len -= oq->buffer_size;
+                               }
+
+                               skb_add_rx_frag(skb, shinfo->nr_frags,
+                                               buff_info->page, 0,
+                                               buff_info->len,
+                                               buff_info->len);
+                               buff_info->page = NULL;
+                               read_idx++;
+                               desc_used++;
+                               if (read_idx == oq->max_count)
+                                       read_idx = 0;
+                       }
+               }
+
+               skb->dev = oq->netdev;
+               skb->protocol =  eth_type_trans(skb, skb->dev);
+               if (resp_hw_ext &&
+                   resp_hw_ext->csum_verified == OCTEP_CSUM_VERIFIED)
+                       skb->ip_summed = CHECKSUM_UNNECESSARY;
+               else
+                       skb->ip_summed = CHECKSUM_NONE;
+               napi_gro_receive(oq->napi, skb);
+       }
+
+       oq->host_read_idx = read_idx;
+       oq->refill_count += desc_used;
+       oq->stats.packets += pkt;
+       oq->stats.bytes += rx_bytes;
+
+       return pkt;
+}
+
+/**
+ * octep_oq_process_rx() - Process Rx queue.
+ *
+ * @oq: Octeon Rx queue data structure.
+ * @budget: max number of packets can be processed in one invocation.
+ *
+ * Check for newly received packets and process them.
+ * Keeps checking for new packets until budget is used or no new packets seen.
+ *
+ * Return: number of packets processed.
+ */
+int octep_oq_process_rx(struct octep_oq *oq, int budget)
+{
+       u32 pkts_available, pkts_processed, total_pkts_processed;
+       struct octep_device *oct = oq->octep_dev;
+
+       pkts_available = 0;
+       pkts_processed = 0;
+       total_pkts_processed = 0;
+       while (total_pkts_processed < budget) {
+                /* update pending count only when current one exhausted */
+               if (oq->pkts_pending == 0)
+                       octep_oq_check_hw_for_pkts(oct, oq);
+               pkts_available = min(budget - total_pkts_processed,
+                                    oq->pkts_pending);
+               if (!pkts_available)
+                       break;
+
+               pkts_processed = __octep_oq_process_rx(oct, oq,
+                                                      pkts_available);
+               oq->pkts_pending -= pkts_processed;
+               total_pkts_processed += pkts_processed;
+       }
+
+       if (oq->refill_count >= oq->refill_threshold) {
+               u32 desc_refilled = octep_oq_refill(oct, oq);
+
+               /* flush pending writes before updating credits */
+               wmb();
+               writel(desc_refilled, oq->pkts_credit_reg);
+       }
+
+       return total_pkts_processed;
+}
index 185522d724758fbaa1ea4fdbdc1793427b97215f..511552bc3e87a094131acb3e7c313d24ba0f830e 100644 (file)
@@ -24,6 +24,79 @@ static void octep_iq_reset_indices(struct octep_iq *iq)
        atomic_set(&iq->instr_pending, 0);
 }
 
+/**
+ * octep_iq_process_completions() - Process Tx queue completions.
+ *
+ * @iq: Octeon Tx queue data structure.
+ * @budget: max number of completions to be processed in one invocation.
+ */
+int octep_iq_process_completions(struct octep_iq *iq, u16 budget)
+{
+       u32 compl_pkts, compl_bytes, compl_sg;
+       struct octep_device *oct = iq->octep_dev;
+       struct octep_tx_buffer *tx_buffer;
+       struct skb_shared_info *shinfo;
+       u32 fi = iq->flush_index;
+       struct sk_buff *skb;
+       u8 frags, i;
+
+       compl_pkts = 0;
+       compl_sg = 0;
+       compl_bytes = 0;
+       iq->octep_read_index = oct->hw_ops.update_iq_read_idx(iq);
+
+       while (likely(budget && (fi != iq->octep_read_index))) {
+               tx_buffer = iq->buff_info + fi;
+               skb = tx_buffer->skb;
+
+               fi++;
+               if (unlikely(fi == iq->max_count))
+                       fi = 0;
+               compl_bytes += skb->len;
+               compl_pkts++;
+               budget--;
+
+               if (!tx_buffer->gather) {
+                       dma_unmap_single(iq->dev, tx_buffer->dma,
+                                        tx_buffer->skb->len, DMA_TO_DEVICE);
+                       dev_kfree_skb_any(skb);
+                       continue;
+               }
+
+               /* Scatter/Gather */
+               shinfo = skb_shinfo(skb);
+               frags = shinfo->nr_frags;
+               compl_sg++;
+
+               dma_unmap_single(iq->dev, tx_buffer->sglist[0].dma_ptr[0],
+                                tx_buffer->sglist[0].len[0], DMA_TO_DEVICE);
+
+               i = 1; /* entry 0 is main skb, unmapped above */
+               while (frags--) {
+                       dma_unmap_page(iq->dev, tx_buffer->sglist[i >> 2].dma_ptr[i & 3],
+                                      tx_buffer->sglist[i >> 2].len[i & 3], DMA_TO_DEVICE);
+                       i++;
+               }
+
+               dev_kfree_skb_any(skb);
+       }
+
+       iq->pkts_processed += compl_pkts;
+       atomic_sub(compl_pkts, &iq->instr_pending);
+       iq->stats.instr_completed += compl_pkts;
+       iq->stats.bytes_sent += compl_bytes;
+       iq->stats.sgentry_sent += compl_sg;
+       iq->flush_index = fi;
+
+       netdev_tx_completed_queue(iq->netdev_q, compl_pkts, compl_bytes);
+
+       if (unlikely(__netif_subqueue_stopped(iq->netdev, iq->q_no)) &&
+           ((iq->max_count - atomic_read(&iq->instr_pending)) >
+            OCTEP_WAKE_QUEUE_THRESHOLD))
+               netif_wake_subqueue(iq->netdev, iq->q_no);
+       return !budget;
+}
+
 /**
  * octep_iq_free_pending() - Free Tx buffers for pending completions.
  *