ethdev: add namespace
[dpdk.git] / drivers / net / cxgbe / sge.c
index b3c4ec2..21b8fe6 100644 (file)
@@ -20,7 +20,6 @@
 #include <rte_log.h>
 #include <rte_debug.h>
 #include <rte_pci.h>
-#include <rte_atomic.h>
 #include <rte_branch_prediction.h>
 #include <rte_memory.h>
 #include <rte_memzone.h>
@@ -28,7 +27,7 @@
 #include <rte_eal.h>
 #include <rte_alarm.h>
 #include <rte_ether.h>
-#include <rte_ethdev_driver.h>
+#include <ethdev_driver.h>
 #include <rte_malloc.h>
 #include <rte_random.h>
 #include <rte_dev.h>
@@ -366,13 +365,10 @@ static unsigned int refill_fl_usembufs(struct adapter *adap, struct sge_fl *q,
        struct rte_mbuf *buf_bulk[n];
        int ret, i;
        struct rte_pktmbuf_pool_private *mbp_priv;
-       u8 jumbo_en = rxq->rspq.eth_dev->data->dev_conf.rxmode.offloads &
-               DEV_RX_OFFLOAD_JUMBO_FRAME;
 
        /* Use jumbo mtu buffers if mbuf data room size can fit jumbo data. */
        mbp_priv = rte_mempool_get_priv(rxq->rspq.mb_pool);
-       if (jumbo_en &&
-           ((mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM) >= 9000))
+       if ((mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM) >= 9000)
                buf_size_idx = RX_LARGE_MTU_BUF;
 
        ret = rte_mempool_get_bulk(rxq->rspq.mb_pool, (void *)buf_bulk, n);
@@ -1004,12 +1000,6 @@ static inline int tx_do_packet_coalesce(struct sge_eth_txq *txq,
        struct cpl_tx_pkt_core *cpl;
        struct tx_sw_desc *sd;
        unsigned int idx = q->coalesce.idx, len = mbuf->pkt_len;
-       unsigned int max_coal_pkt_num = is_pf4(adap) ? ETH_COALESCE_PKT_NUM :
-                                                      ETH_COALESCE_VF_PKT_NUM;
-
-#ifdef RTE_LIBRTE_CXGBE_TPUT
-       RTE_SET_USED(nb_pkts);
-#endif
 
        if (q->coalesce.type == 0) {
                mc = (struct ulp_txpkt *)q->coalesce.ptr;
@@ -1082,13 +1072,15 @@ static inline int tx_do_packet_coalesce(struct sge_eth_txq *txq,
        sd->coalesce.sgl[idx & 1] = (struct ulptx_sgl *)(cpl + 1);
        sd->coalesce.idx = (idx & 1) + 1;
 
-       /* send the coaelsced work request if max reached */
-       if (++q->coalesce.idx == max_coal_pkt_num
-#ifndef RTE_LIBRTE_CXGBE_TPUT
-           || q->coalesce.idx >= nb_pkts
-#endif
-           )
+       /* Send the coalesced work request, only if max reached. However,
+        * if lower latency is preferred over throughput, then don't wait
+        * for coalescing the next Tx burst and send the packets now.
+        */
+       q->coalesce.idx++;
+       if (q->coalesce.idx == adap->params.max_tx_coalesce_num ||
+           (adap->devargs.tx_mode_latency && q->coalesce.idx >= nb_pkts))
                ship_tx_pkt_coalesce_wr(adap, txq);
+
        return 0;
 }
 
@@ -1118,7 +1110,7 @@ int t4_eth_xmit(struct sge_eth_txq *txq, struct rte_mbuf *mbuf,
        u32 wr_mid;
        u64 cntrl, *end;
        bool v6;
-       u32 max_pkt_len = txq->data->dev_conf.rxmode.max_rx_pkt_len;
+       u32 max_pkt_len;
 
        /* Reject xmit if queue is stopped */
        if (unlikely(txq->flags & EQ_STOPPED))
@@ -1134,11 +1126,12 @@ out_free:
                return 0;
        }
 
+       max_pkt_len = txq->data->mtu + RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN;
        if ((!(m->ol_flags & PKT_TX_TCP_SEG)) &&
            (unlikely(m->pkt_len > max_pkt_len)))
                goto out_free;
 
-       pi = (struct port_info *)txq->data->dev_private;
+       pi = txq->data->dev_private;
        adap = pi->adapter;
 
        cntrl = F_TXPKT_L4CSUM_DIS | F_TXPKT_IPCSUM_DIS;
@@ -1154,7 +1147,6 @@ out_free:
                                txq->stats.mapping_err++;
                                goto out_free;
                        }
-                       rte_prefetch0((volatile void *)addr);
                        return tx_do_packet_coalesce(txq, mbuf, cflits, adap,
                                                     pi, addr, nb_pkts);
                } else {
@@ -1428,14 +1420,16 @@ int t4_mgmt_tx(struct sge_ctrl_txq *q, struct rte_mbuf *mbuf)
 
 /**
  * alloc_ring - allocate resources for an SGE descriptor ring
- * @dev: the PCI device's core device
+ * @dev: the port associated with the queue
+ * @z_name: memzone's name
+ * @queue_id: queue index
+ * @socket_id: preferred socket id for memory allocations
  * @nelem: the number of descriptors
  * @elem_size: the size of each descriptor
+ * @stat_size: extra space in HW ring for status information
  * @sw_size: the size of the SW state associated with each ring element
  * @phys: the physical address of the allocated ring
  * @metadata: address of the array holding the SW state for the ring
- * @stat_size: extra space in HW ring for status information
- * @node: preferred node for memory allocations
  *
  * Allocates resources for an SGE descriptor ring, such as Tx queues,
  * free buffer lists, or response queues.  Each SGE ring requires
@@ -1445,39 +1439,34 @@ int t4_mgmt_tx(struct sge_ctrl_txq *q, struct rte_mbuf *mbuf)
  * of the function), the bus address of the HW ring, and the address
  * of the SW ring.
  */
-static void *alloc_ring(size_t nelem, size_t elem_size,
-                       size_t sw_size, dma_addr_t *phys, void *metadata,
-                       size_t stat_size, __rte_unused uint16_t queue_id,
-                       int socket_id, const char *z_name,
-                       const char *z_name_sw)
+static void *alloc_ring(struct rte_eth_dev *dev, const char *z_name,
+                       uint16_t queue_id, int socket_id, size_t nelem,
+                       size_t elem_size, size_t stat_size, size_t sw_size,
+                       dma_addr_t *phys, void *metadata)
 {
        size_t len = CXGBE_MAX_RING_DESC_SIZE * elem_size + stat_size;
+       char z_name_sw[RTE_MEMZONE_NAMESIZE];
        const struct rte_memzone *tz;
        void *s = NULL;
 
+       snprintf(z_name_sw, sizeof(z_name_sw), "eth_p%d_q%d_%s_sw_ring",
+                dev->data->port_id, queue_id, z_name);
+
        dev_debug(adapter, "%s: nelem = %zu; elem_size = %zu; sw_size = %zu; "
                  "stat_size = %zu; queue_id = %u; socket_id = %d; z_name = %s;"
                  " z_name_sw = %s\n", __func__, nelem, elem_size, sw_size,
                  stat_size, queue_id, socket_id, z_name, z_name_sw);
 
-       tz = rte_memzone_lookup(z_name);
-       if (tz) {
-               dev_debug(adapter, "%s: tz exists...returning existing..\n",
-                         __func__);
-               goto alloc_sw_ring;
-       }
-
        /*
         * Allocate TX/RX ring hardware descriptors. A memzone large enough to
         * handle the maximum ring size is allocated in order to allow for
         * resizing in later calls to the queue setup function.
         */
-       tz = rte_memzone_reserve_aligned(z_name, len, socket_id,
-                       RTE_MEMZONE_IOVA_CONTIG, 4096);
+       tz = rte_eth_dma_zone_reserve(dev, z_name, queue_id, len, 4096,
+                                     socket_id);
        if (!tz)
                return NULL;
 
-alloc_sw_ring:
        memset(tz->addr, 0, len);
        if (sw_size) {
                s = rte_zmalloc_socket(z_name_sw, nelem * sw_size,
@@ -1702,6 +1691,11 @@ int cxgbe_poll(struct sge_rspq *q, struct rte_mbuf **rx_pkts,
        unsigned int params;
        u32 val;
 
+       if (unlikely(rxq->flags & IQ_STOPPED)) {
+               *work_done = 0;
+               return 0;
+       }
+
        *work_done = process_responses(q, budget, rx_pkts);
 
        if (*work_done) {
@@ -1762,22 +1756,22 @@ static void __iomem *bar2_address(struct adapter *adapter, unsigned int qid,
        return adapter->bar2 + bar2_qoffset;
 }
 
-int t4_sge_eth_rxq_start(struct adapter *adap, struct sge_rspq *rq)
+int t4_sge_eth_rxq_start(struct adapter *adap, struct sge_eth_rxq *rxq)
 {
-       struct sge_eth_rxq *rxq = container_of(rq, struct sge_eth_rxq, rspq);
        unsigned int fl_id = rxq->fl.size ? rxq->fl.cntxt_id : 0xffff;
 
+       rxq->flags &= ~IQ_STOPPED;
        return t4_iq_start_stop(adap, adap->mbox, true, adap->pf, 0,
-                               rq->cntxt_id, fl_id, 0xffff);
+                               rxq->rspq.cntxt_id, fl_id, 0xffff);
 }
 
-int t4_sge_eth_rxq_stop(struct adapter *adap, struct sge_rspq *rq)
+int t4_sge_eth_rxq_stop(struct adapter *adap, struct sge_eth_rxq *rxq)
 {
-       struct sge_eth_rxq *rxq = container_of(rq, struct sge_eth_rxq, rspq);
        unsigned int fl_id = rxq->fl.size ? rxq->fl.cntxt_id : 0xffff;
 
+       rxq->flags |= IQ_STOPPED;
        return t4_iq_start_stop(adap, adap->mbox, false, adap->pf, 0,
-                               rq->cntxt_id, fl_id, 0xffff);
+                               rxq->rspq.cntxt_id, fl_id, 0xffff);
 }
 
 /*
@@ -1792,22 +1786,16 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
        int ret, flsz = 0;
        struct fw_iq_cmd c;
        struct sge *s = &adap->sge;
-       struct port_info *pi = (struct port_info *)(eth_dev->data->dev_private);
-       char z_name[RTE_MEMZONE_NAMESIZE];
-       char z_name_sw[RTE_MEMZONE_NAMESIZE];
+       struct port_info *pi = eth_dev->data->dev_private;
        unsigned int nb_refill;
        u8 pciechan;
 
        /* Size needs to be multiple of 16, including status entry. */
        iq->size = cxgbe_roundup(iq->size, 16);
 
-       snprintf(z_name, sizeof(z_name), "eth_p%d_q%d_%s",
-                       eth_dev->data->port_id, queue_id,
-                       fwevtq ? "fwq_ring" : "rx_ring");
-       snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
-
-       iq->desc = alloc_ring(iq->size, iq->iqe_len, 0, &iq->phys_addr, NULL, 0,
-                             queue_id, socket_id, z_name, z_name_sw);
+       iq->desc = alloc_ring(eth_dev, fwevtq ? "fwq_ring" : "rx_ring",
+                             queue_id, socket_id, iq->size, iq->iqe_len,
+                             0, 0, &iq->phys_addr, NULL);
        if (!iq->desc)
                return -ENOMEM;
 
@@ -1865,18 +1853,14 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
                        fl->size = s->fl_starve_thres - 1 + 2 * 8;
                fl->size = cxgbe_roundup(fl->size, 8);
 
-               snprintf(z_name, sizeof(z_name), "eth_p%d_q%d_%s",
-                               eth_dev->data->port_id, queue_id,
-                               fwevtq ? "fwq_ring" : "fl_ring");
-               snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
-
-               fl->desc = alloc_ring(fl->size, sizeof(__be64),
+               fl->desc = alloc_ring(eth_dev, "fl_ring", queue_id, socket_id,
+                                     fl->size, sizeof(__be64), s->stat_len,
                                      sizeof(struct rx_sw_desc),
-                                     &fl->addr, &fl->sdesc, s->stat_len,
-                                     queue_id, socket_id, z_name, z_name_sw);
-
-               if (!fl->desc)
-                       goto fl_nomem;
+                                     &fl->addr, &fl->sdesc);
+               if (!fl->desc) {
+                       ret = -ENOMEM;
+                       goto err;
+               }
 
                flsz = fl->size / 8 + s->stat_len / sizeof(struct tx_desc);
                c.iqns_to_fl0congen |=
@@ -1967,7 +1951,8 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
         * simple (and hopefully less wrong).
         */
        if (is_pf4(adap) && !is_t4(adap->params.chip) && cong >= 0) {
-               u32 param, val;
+               u8 cng_ch_bits_log = adap->params.arch.cng_ch_bits_log;
+               u32 param, val, ch_map = 0;
                int i;
 
                param = (V_FW_PARAMS_MNEM(FW_PARAMS_MNEM_DMAQ) |
@@ -1980,9 +1965,9 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
                                        X_CONMCTXT_CNGTPMODE_CHANNEL);
                        for (i = 0; i < 4; i++) {
                                if (cong & (1 << i))
-                                       val |= V_CONMCTXT_CNGCHMAP(1 <<
-                                                                  (i << 2));
+                                       ch_map |= 1 << (i << cng_ch_bits_log);
                        }
+                       val |= V_CONMCTXT_CNGCHMAP(ch_map);
                }
                ret = t4_set_params(adap, adap->mbox, adap->pf, 0, 1,
                                    &param, &val);
@@ -1996,8 +1981,6 @@ int t4_sge_alloc_rxq(struct adapter *adap, struct sge_rspq *iq, bool fwevtq,
 refill_fl_err:
        t4_iq_free(adap, adap->mbox, adap->pf, 0, FW_IQ_TYPE_FL_INT_CAP,
                   iq->cntxt_id, fl->cntxt_id, 0xffff);
-fl_nomem:
-       ret = -ENOMEM;
 err:
        iq->cntxt_id = 0;
        iq->abs_id = 0;
@@ -2062,22 +2045,16 @@ int t4_sge_alloc_eth_txq(struct adapter *adap, struct sge_eth_txq *txq,
        int ret, nentries;
        struct fw_eq_eth_cmd c;
        struct sge *s = &adap->sge;
-       struct port_info *pi = (struct port_info *)(eth_dev->data->dev_private);
-       char z_name[RTE_MEMZONE_NAMESIZE];
-       char z_name_sw[RTE_MEMZONE_NAMESIZE];
+       struct port_info *pi = eth_dev->data->dev_private;
        u8 pciechan;
 
        /* Add status entries */
        nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
 
-       snprintf(z_name, sizeof(z_name), "eth_p%d_q%d_%s",
-                       eth_dev->data->port_id, queue_id, "tx_ring");
-       snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
-
-       txq->q.desc = alloc_ring(txq->q.size, sizeof(struct tx_desc),
-                                sizeof(struct tx_sw_desc), &txq->q.phys_addr,
-                                &txq->q.sdesc, s->stat_len, queue_id,
-                                socket_id, z_name, z_name_sw);
+       txq->q.desc = alloc_ring(eth_dev, "tx_ring", queue_id, socket_id,
+                                txq->q.size, sizeof(struct tx_desc),
+                                s->stat_len, sizeof(struct tx_sw_desc),
+                                &txq->q.phys_addr, &txq->q.sdesc);
        if (!txq->q.desc)
                return -ENOMEM;
 
@@ -2141,21 +2118,14 @@ int t4_sge_alloc_ctrl_txq(struct adapter *adap, struct sge_ctrl_txq *txq,
        int ret, nentries;
        struct fw_eq_ctrl_cmd c;
        struct sge *s = &adap->sge;
-       struct port_info *pi = (struct port_info *)(eth_dev->data->dev_private);
-       char z_name[RTE_MEMZONE_NAMESIZE];
-       char z_name_sw[RTE_MEMZONE_NAMESIZE];
+       struct port_info *pi = eth_dev->data->dev_private;
 
        /* Add status entries */
        nentries = txq->q.size + s->stat_len / sizeof(struct tx_desc);
 
-       snprintf(z_name, sizeof(z_name), "eth_p%d_q%d_%s",
-                       eth_dev->data->port_id, queue_id, "ctrl_tx_ring");
-       snprintf(z_name_sw, sizeof(z_name_sw), "%s_sw_ring", z_name);
-
-       txq->q.desc = alloc_ring(txq->q.size, sizeof(struct tx_desc),
-                                0, &txq->q.phys_addr,
-                                NULL, 0, queue_id,
-                                socket_id, z_name, z_name_sw);
+       txq->q.desc = alloc_ring(eth_dev, "ctrl_tx_ring", queue_id,
+                                socket_id, txq->q.size, sizeof(struct tx_desc),
+                                0, 0, &txq->q.phys_addr, NULL);
        if (!txq->q.desc)
                return -ENOMEM;
 
@@ -2226,15 +2196,18 @@ static void free_rspq_fl(struct adapter *adap, struct sge_rspq *rq,
  */
 void t4_sge_eth_clear_queues(struct port_info *pi)
 {
-       int i;
        struct adapter *adap = pi->adapter;
-       struct sge_eth_rxq *rxq = &adap->sge.ethrxq[pi->first_qset];
-       struct sge_eth_txq *txq = &adap->sge.ethtxq[pi->first_qset];
+       struct sge_eth_rxq *rxq;
+       struct sge_eth_txq *txq;
+       int i;
 
+       rxq = &adap->sge.ethrxq[pi->first_rxqset];
        for (i = 0; i < pi->n_rx_qsets; i++, rxq++) {
                if (rxq->rspq.desc)
-                       t4_sge_eth_rxq_stop(adap, &rxq->rspq);
+                       t4_sge_eth_rxq_stop(adap, rxq);
        }
+
+       txq = &adap->sge.ethtxq[pi->first_txqset];
        for (i = 0; i < pi->n_tx_qsets; i++, txq++) {
                if (txq->q.desc) {
                        struct sge_txq *q = &txq->q;
@@ -2250,7 +2223,7 @@ void t4_sge_eth_clear_queues(struct port_info *pi)
 void t4_sge_eth_rxq_release(struct adapter *adap, struct sge_eth_rxq *rxq)
 {
        if (rxq->rspq.desc) {
-               t4_sge_eth_rxq_stop(adap, &rxq->rspq);
+               t4_sge_eth_rxq_stop(adap, rxq);
                free_rspq_fl(adap, &rxq->rspq, rxq->fl.size ? &rxq->fl : NULL);
        }
 }
@@ -2267,6 +2240,36 @@ void t4_sge_eth_txq_release(struct adapter *adap, struct sge_eth_txq *txq)
        }
 }
 
+void t4_sge_eth_release_queues(struct port_info *pi)
+{
+       struct adapter *adap = pi->adapter;
+       struct sge_eth_rxq *rxq;
+       struct sge_eth_txq *txq;
+       unsigned int i;
+
+       rxq = &adap->sge.ethrxq[pi->first_rxqset];
+       /* clean up Ethernet Tx/Rx queues */
+       for (i = 0; i < pi->n_rx_qsets; i++, rxq++) {
+               /* Free only the queues allocated */
+               if (rxq->rspq.desc) {
+                       t4_sge_eth_rxq_release(adap, rxq);
+                       rte_eth_dma_zone_free(rxq->rspq.eth_dev, "fl_ring", i);
+                       rte_eth_dma_zone_free(rxq->rspq.eth_dev, "rx_ring", i);
+                       rxq->rspq.eth_dev = NULL;
+               }
+       }
+
+       txq = &adap->sge.ethtxq[pi->first_txqset];
+       for (i = 0; i < pi->n_tx_qsets; i++, txq++) {
+               /* Free only the queues allocated */
+               if (txq->q.desc) {
+                       t4_sge_eth_txq_release(adap, txq);
+                       rte_eth_dma_zone_free(txq->eth_dev, "tx_ring", i);
+                       txq->eth_dev = NULL;
+               }
+       }
+}
+
 void t4_sge_tx_monitor_start(struct adapter *adap)
 {
        rte_eal_alarm_set(50, tx_timer_cb, (void *)adap);
@@ -2286,21 +2289,6 @@ void t4_sge_tx_monitor_stop(struct adapter *adap)
 void t4_free_sge_resources(struct adapter *adap)
 {
        unsigned int i;
-       struct sge_eth_rxq *rxq = &adap->sge.ethrxq[0];
-       struct sge_eth_txq *txq = &adap->sge.ethtxq[0];
-
-       /* clean up Ethernet Tx/Rx queues */
-       for (i = 0; i < adap->sge.max_ethqsets; i++, rxq++, txq++) {
-               /* Free only the queues allocated */
-               if (rxq->rspq.desc) {
-                       t4_sge_eth_rxq_release(adap, rxq);
-                       rxq->rspq.eth_dev = NULL;
-               }
-               if (txq->q.desc) {
-                       t4_sge_eth_txq_release(adap, txq);
-                       txq->eth_dev = NULL;
-               }
-       }
 
        /* clean up control Tx queues */
        for (i = 0; i < ARRAY_SIZE(adap->sge.ctrlq); i++) {
@@ -2310,12 +2298,17 @@ void t4_free_sge_resources(struct adapter *adap)
                        reclaim_completed_tx_imm(&cq->q);
                        t4_ctrl_eq_free(adap, adap->mbox, adap->pf, 0,
                                        cq->q.cntxt_id);
+                       rte_eth_dma_zone_free(adap->eth_dev, "ctrl_tx_ring", i);
+                       rte_mempool_free(cq->mb_pool);
                        free_txq(&cq->q);
                }
        }
 
-       if (adap->sge.fw_evtq.desc)
+       /* clean up firmware event queue */
+       if (adap->sge.fw_evtq.desc) {
                free_rspq_fl(adap, &adap->sge.fw_evtq, NULL);
+               rte_eth_dma_zone_free(adap->eth_dev, "fwq_ring", 0);
+       }
 }
 
 /**