X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fliquidio%2Flio_rxtx.c;h=2bbb893c27e423e49d12b027ae3c450d40c9b381;hb=f37dfab21c988d2d0ecb3c82be4ba9738c7e51c7;hp=14113e1bef05fb41d76ba199778e89f8517a79e9;hpb=1f720275d887edaf7491a89634668de2af78b53e;p=dpdk.git diff --git a/drivers/net/liquidio/lio_rxtx.c b/drivers/net/liquidio/lio_rxtx.c index 14113e1bef..2bbb893c27 100644 --- a/drivers/net/liquidio/lio_rxtx.c +++ b/drivers/net/liquidio/lio_rxtx.c @@ -41,6 +41,9 @@ #include "lio_rxtx.h" #define LIO_MAX_SG 12 +/* Flush iq if available tx_desc fall below LIO_FLUSH_WM */ +#define LIO_FLUSH_WM(_iq) ((_iq)->max_count / 2) +#define LIO_PKT_IN_DONE_CNT_MASK 0x00000000FFFFFFFFULL static void lio_droq_compute_max_packet_bufs(struct lio_droq *droq) @@ -78,28 +81,6 @@ lio_droq_destroy_ring_buffers(struct lio_droq *droq) lio_droq_reset_indices(droq); } -static void * -lio_recv_buffer_alloc(struct lio_device *lio_dev, int q_no) -{ - struct lio_droq *droq = lio_dev->droq[q_no]; - struct rte_mempool *mpool = droq->mpool; - struct rte_mbuf *m; - - m = rte_pktmbuf_alloc(mpool); - if (m == NULL) { - lio_dev_err(lio_dev, "Cannot allocate\n"); - return NULL; - } - - rte_mbuf_refcnt_set(m, 1); - m->next = NULL; - m->data_off = RTE_PKTMBUF_HEADROOM; - m->nb_segs = 1; - m->pool = mpool; - - return m; -} - static int lio_droq_setup_ring_buffers(struct lio_device *lio_dev, struct lio_droq *droq) @@ -109,9 +90,10 @@ lio_droq_setup_ring_buffers(struct lio_device *lio_dev, void *buf; for (i = 0; i < droq->max_count; i++) { - buf = lio_recv_buffer_alloc(lio_dev, droq->q_no); + buf = rte_pktmbuf_alloc(droq->mpool); if (buf == NULL) { lio_dev_err(lio_dev, "buffer alloc failed\n"); + droq->stats.rx_alloc_failure++; lio_droq_destroy_ring_buffers(droq); return -ENOMEM; } @@ -374,7 +356,6 @@ lio_droq_refill_pullup_descs(struct lio_droq *droq, /* lio_droq_refill * - * @param lio_dev - pointer to the lio device structure * @param droq - droq in which descriptors require new buffers. * * Description: @@ -390,7 +371,7 @@ lio_droq_refill_pullup_descs(struct lio_droq *droq, * This routine is called with droq->lock held. */ static uint32_t -lio_droq_refill(struct lio_device *lio_dev, struct lio_droq *droq) +lio_droq_refill(struct lio_droq *droq) { struct lio_droq_desc *desc_ring; uint32_t desc_refilled = 0; @@ -403,12 +384,14 @@ lio_droq_refill(struct lio_device *lio_dev, struct lio_droq *droq) * reuse the buffer, else allocate. */ if (droq->recv_buf_list[droq->refill_idx].buffer == NULL) { - buf = lio_recv_buffer_alloc(lio_dev, droq->q_no); + buf = rte_pktmbuf_alloc(droq->mpool); /* If a buffer could not be allocated, no point in * continuing */ - if (buf == NULL) + if (buf == NULL) { + droq->stats.rx_alloc_failure++; break; + } droq->recv_buf_list[droq->refill_idx].buffer = buf; } @@ -483,9 +466,6 @@ lio_droq_fast_process_packet(struct lio_device *lio_dev, droq->refill_count++; if (likely(nicbuf != NULL)) { - nicbuf->data_off = RTE_PKTMBUF_HEADROOM; - nicbuf->nb_segs = 1; - nicbuf->next = NULL; /* We don't have a way to pass flags yet */ nicbuf->ol_flags = 0; if (rh->r_dh.has_hash) { @@ -539,9 +519,6 @@ lio_droq_fast_process_packet(struct lio_device *lio_dev, if (!pkt_len) first_buf = nicbuf; - nicbuf->data_off = RTE_PKTMBUF_HEADROOM; - nicbuf->nb_segs = 1; - nicbuf->next = NULL; nicbuf->port = lio_dev->port_id; /* We don't have a way to pass * flags yet @@ -611,7 +588,7 @@ lio_droq_fast_process_packet(struct lio_device *lio_dev, } if (droq->refill_count >= droq->refill_threshold) { - int desc_refilled = lio_droq_refill(lio_dev, droq); + int desc_refilled = lio_droq_refill(droq); /* Flush the droq descriptor data to memory to be sure * that when we update the credits the data in memory is @@ -626,6 +603,11 @@ lio_droq_fast_process_packet(struct lio_device *lio_dev, info->length = 0; info->rh.rh64 = 0; + droq->stats.pkts_received++; + droq->stats.rx_pkts_received += data_pkts; + droq->stats.rx_bytes_received += data_total_len; + droq->stats.bytes_received += total_len; + return data_pkts; } @@ -907,6 +889,40 @@ release_lio_iq: return -1; } +int +lio_wait_for_instr_fetch(struct lio_device *lio_dev) +{ + int pending, instr_cnt; + int i, retry = 1000; + + do { + instr_cnt = 0; + + for (i = 0; i < LIO_MAX_INSTR_QUEUES(lio_dev); i++) { + if (!(lio_dev->io_qmask.iq & (1ULL << i))) + continue; + + if (lio_dev->instr_queue[i] == NULL) + break; + + pending = rte_atomic64_read( + &lio_dev->instr_queue[i]->instr_pending); + if (pending) + lio_flush_iq(lio_dev, lio_dev->instr_queue[i]); + + instr_cnt += pending; + } + + if (instr_cnt == 0) + break; + + rte_delay_ms(1); + + } while (retry-- && instr_cnt); + + return instr_cnt; +} + static inline void lio_ring_doorbell(struct lio_device *lio_dev, struct lio_instr_queue *iq) @@ -977,9 +993,151 @@ lio_add_to_request_list(struct lio_instr_queue *iq, iq->request_list[idx].reqtype = reqtype; } +static inline void +lio_free_netsgbuf(void *buf) +{ + struct lio_buf_free_info *finfo = buf; + struct lio_device *lio_dev = finfo->lio_dev; + struct rte_mbuf *m = finfo->mbuf; + struct lio_gather *g = finfo->g; + uint8_t iq = finfo->iq_no; + + /* This will take care of multiple segments also */ + rte_pktmbuf_free(m); + + rte_spinlock_lock(&lio_dev->glist_lock[iq]); + STAILQ_INSERT_TAIL(&lio_dev->glist_head[iq], &g->list, entries); + rte_spinlock_unlock(&lio_dev->glist_lock[iq]); + rte_free(finfo); +} + +/* Can only run in process context */ +static int +lio_process_iq_request_list(struct lio_device *lio_dev, + struct lio_instr_queue *iq) +{ + struct octeon_instr_irh *irh = NULL; + uint32_t old = iq->flush_index; + struct lio_soft_command *sc; + uint32_t inst_count = 0; + int reqtype; + void *buf; + + while (old != iq->lio_read_index) { + reqtype = iq->request_list[old].reqtype; + buf = iq->request_list[old].buf; + + if (reqtype == LIO_REQTYPE_NONE) + goto skip_this; + + switch (reqtype) { + case LIO_REQTYPE_NORESP_NET: + rte_pktmbuf_free((struct rte_mbuf *)buf); + break; + case LIO_REQTYPE_NORESP_NET_SG: + lio_free_netsgbuf(buf); + break; + case LIO_REQTYPE_SOFT_COMMAND: + sc = buf; + irh = (struct octeon_instr_irh *)&sc->cmd.cmd3.irh; + if (irh->rflag) { + /* We're expecting a response from Octeon. + * It's up to lio_process_ordered_list() to + * process sc. Add sc to the ordered soft + * command response list because we expect + * a response from Octeon. + */ + rte_spinlock_lock(&lio_dev->response_list.lock); + rte_atomic64_inc( + &lio_dev->response_list.pending_req_count); + STAILQ_INSERT_TAIL( + &lio_dev->response_list.head, + &sc->node, entries); + rte_spinlock_unlock( + &lio_dev->response_list.lock); + } else { + if (sc->callback) { + /* This callback must not sleep */ + sc->callback(LIO_REQUEST_DONE, + sc->callback_arg); + } + } + break; + default: + lio_dev_err(lio_dev, + "Unknown reqtype: %d buf: %p at idx %d\n", + reqtype, buf, old); + } + + iq->request_list[old].buf = NULL; + iq->request_list[old].reqtype = 0; + +skip_this: + inst_count++; + old = lio_incr_index(old, 1, iq->max_count); + } + + iq->flush_index = old; + + return inst_count; +} + +static void +lio_update_read_index(struct lio_instr_queue *iq) +{ + uint32_t pkt_in_done = rte_read32(iq->inst_cnt_reg); + uint32_t last_done; + + last_done = pkt_in_done - iq->pkt_in_done; + iq->pkt_in_done = pkt_in_done; + + /* Add last_done and modulo with the IQ size to get new index */ + iq->lio_read_index = (iq->lio_read_index + + (uint32_t)(last_done & LIO_PKT_IN_DONE_CNT_MASK)) % + iq->max_count; +} + +int +lio_flush_iq(struct lio_device *lio_dev, struct lio_instr_queue *iq) +{ + uint32_t tot_inst_processed = 0; + uint32_t inst_processed = 0; + int tx_done = 1; + + if (rte_atomic64_test_and_set(&iq->iq_flush_running) == 0) + return tx_done; + + rte_spinlock_lock(&iq->lock); + + lio_update_read_index(iq); + + do { + /* Process any outstanding IQ packets. */ + if (iq->flush_index == iq->lio_read_index) + break; + + inst_processed = lio_process_iq_request_list(lio_dev, iq); + + if (inst_processed) { + rte_atomic64_sub(&iq->instr_pending, inst_processed); + iq->stats.instr_processed += inst_processed; + } + + tot_inst_processed += inst_processed; + inst_processed = 0; + + } while (1); + + rte_spinlock_unlock(&iq->lock); + + rte_atomic64_clear(&iq->iq_flush_running); + + return tx_done; +} + static int lio_send_command(struct lio_device *lio_dev, uint32_t iq_no, void *cmd, - void *buf, uint32_t datasize __rte_unused, uint32_t reqtype) + void *buf, uint32_t datasize, uint32_t reqtype) { struct lio_instr_queue *iq = lio_dev->instr_queue[iq_no]; struct lio_iq_post_status st; @@ -990,7 +1148,13 @@ lio_send_command(struct lio_device *lio_dev, uint32_t iq_no, void *cmd, if (st.status != LIO_IQ_SEND_FAILED) { lio_add_to_request_list(iq, st.index, buf, reqtype); + LIO_INCR_INSTRQUEUE_PKT_COUNT(lio_dev, iq_no, bytes_sent, + datasize); + LIO_INCR_INSTRQUEUE_PKT_COUNT(lio_dev, iq_no, instr_posted, 1); + lio_ring_doorbell(lio_dev, iq); + } else { + LIO_INCR_INSTRQUEUE_PKT_COUNT(lio_dev, iq_no, instr_dropped, 1); } rte_spinlock_unlock(&iq->post_lock); @@ -1296,7 +1460,7 @@ list_delete_first_node(struct lio_stailq_head *head) return node; } -static void +void lio_delete_sglist(struct lio_instr_queue *txq) { struct lio_device *lio_dev = txq->lio_dev; @@ -1385,6 +1549,117 @@ lio_delete_instruction_queue(struct lio_device *lio_dev, int iq_no) lio_dev->num_iqs--; } +static inline uint32_t +lio_iq_get_available(struct lio_device *lio_dev, uint32_t q_no) +{ + return ((lio_dev->instr_queue[q_no]->max_count - 1) - + (uint32_t)rte_atomic64_read( + &lio_dev->instr_queue[q_no]->instr_pending)); +} + +static inline int +lio_iq_is_full(struct lio_device *lio_dev, uint32_t q_no) +{ + return ((uint32_t)rte_atomic64_read( + &lio_dev->instr_queue[q_no]->instr_pending) >= + (lio_dev->instr_queue[q_no]->max_count - 2)); +} + +static int +lio_dev_cleanup_iq(struct lio_device *lio_dev, int iq_no) +{ + struct lio_instr_queue *iq = lio_dev->instr_queue[iq_no]; + uint32_t count = 10000; + + while ((lio_iq_get_available(lio_dev, iq_no) < LIO_FLUSH_WM(iq)) && + --count) + lio_flush_iq(lio_dev, iq); + + return count ? 0 : 1; +} + +static void +lio_ctrl_cmd_callback(uint32_t status __rte_unused, void *sc_ptr) +{ + struct lio_soft_command *sc = sc_ptr; + struct lio_dev_ctrl_cmd *ctrl_cmd; + struct lio_ctrl_pkt *ctrl_pkt; + + ctrl_pkt = (struct lio_ctrl_pkt *)sc->ctxptr; + ctrl_cmd = ctrl_pkt->ctrl_cmd; + ctrl_cmd->cond = 1; + + lio_free_soft_command(sc); +} + +static inline struct lio_soft_command * +lio_alloc_ctrl_pkt_sc(struct lio_device *lio_dev, + struct lio_ctrl_pkt *ctrl_pkt) +{ + struct lio_soft_command *sc = NULL; + uint32_t uddsize, datasize; + uint32_t rdatasize; + uint8_t *data; + + uddsize = (uint32_t)(ctrl_pkt->ncmd.s.more * 8); + + datasize = OCTEON_CMD_SIZE + uddsize; + rdatasize = (ctrl_pkt->wait_time) ? 16 : 0; + + sc = lio_alloc_soft_command(lio_dev, datasize, + rdatasize, sizeof(struct lio_ctrl_pkt)); + if (sc == NULL) + return NULL; + + rte_memcpy(sc->ctxptr, ctrl_pkt, sizeof(struct lio_ctrl_pkt)); + + data = (uint8_t *)sc->virtdptr; + + rte_memcpy(data, &ctrl_pkt->ncmd, OCTEON_CMD_SIZE); + + lio_swap_8B_data((uint64_t *)data, OCTEON_CMD_SIZE >> 3); + + if (uddsize) { + /* Endian-Swap for UDD should have been done by caller. */ + rte_memcpy(data + OCTEON_CMD_SIZE, ctrl_pkt->udd, uddsize); + } + + sc->iq_no = (uint32_t)ctrl_pkt->iq_no; + + lio_prepare_soft_command(lio_dev, sc, + LIO_OPCODE, LIO_OPCODE_CMD, + 0, 0, 0); + + sc->callback = lio_ctrl_cmd_callback; + sc->callback_arg = sc; + sc->wait_time = ctrl_pkt->wait_time; + + return sc; +} + +int +lio_send_ctrl_pkt(struct lio_device *lio_dev, struct lio_ctrl_pkt *ctrl_pkt) +{ + struct lio_soft_command *sc = NULL; + int retval; + + sc = lio_alloc_ctrl_pkt_sc(lio_dev, ctrl_pkt); + if (sc == NULL) { + lio_dev_err(lio_dev, "soft command allocation failed\n"); + return -1; + } + + retval = lio_send_soft_command(lio_dev, sc); + if (retval == LIO_IQ_SEND_FAILED) { + lio_free_soft_command(sc); + lio_dev_err(lio_dev, "Port: %d soft command: %d send failed status: %x\n", + lio_dev->port_id, ctrl_pkt->ncmd.s.cmd, retval); + return -1; + } + + return retval; +} + /** Send data packet to the device * @param lio_dev - lio device pointer * @param ndata - control structure with queueing, and buffer information @@ -1405,6 +1680,7 @@ lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts) struct lio_instr_queue *txq = tx_queue; union lio_cmd_setup cmdsetup; struct lio_device *lio_dev; + struct lio_iq_stats *stats; struct lio_data_pkt ndata; int i, processed = 0; struct rte_mbuf *m; @@ -1414,13 +1690,16 @@ lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts) lio_dev = txq->lio_dev; iq_no = txq->txpciq.s.q_no; + stats = &lio_dev->instr_queue[iq_no]->stats; - if (!lio_dev->linfo.link.s.link_up) { + if (!lio_dev->intf_open || !lio_dev->linfo.link.s.link_up) { PMD_TX_LOG(lio_dev, ERR, "Transmit failed link_status : %d\n", lio_dev->linfo.link.s.link_up); goto xmit_failed; } + lio_dev_cleanup_iq(lio_dev, iq_no); + for (i = 0; i < nb_pkts; i++) { uint32_t pkt_len = 0; @@ -1432,6 +1711,15 @@ lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts) ndata.buf = m; ndata.q_no = iq_no; + if (lio_iq_is_full(lio_dev, ndata.q_no)) { + stats->tx_iq_busy++; + if (lio_dev_cleanup_iq(lio_dev, iq_no)) { + PMD_TX_LOG(lio_dev, ERR, + "Transmit failed iq:%d full\n", + ndata.q_no); + break; + } + } cmdsetup.cmd_setup64 = 0; cmdsetup.s.iq_no = iq_no; @@ -1440,7 +1728,9 @@ lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts) if (m->ol_flags & PKT_TX_IP_CKSUM) cmdsetup.s.ip_csum = 1; - if ((m->ol_flags & PKT_TX_TCP_CKSUM) || + if (m->ol_flags & PKT_TX_OUTER_IP_CKSUM) + cmdsetup.s.tnl_csum = 1; + else if ((m->ol_flags & PKT_TX_TCP_CKSUM) || (m->ol_flags & PKT_TX_UDP_CKSUM)) cmdsetup.s.transport_csum = 1; @@ -1451,6 +1741,68 @@ lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts) &cmdsetup, tag); ndata.cmd.cmd3.dptr = rte_mbuf_data_dma_addr(m); ndata.reqtype = LIO_REQTYPE_NORESP_NET; + } else { + struct lio_buf_free_info *finfo; + struct lio_gather *g; + phys_addr_t phyaddr; + int i, frags; + + finfo = (struct lio_buf_free_info *)rte_malloc(NULL, + sizeof(*finfo), 0); + if (finfo == NULL) { + PMD_TX_LOG(lio_dev, ERR, + "free buffer alloc failed\n"); + goto xmit_failed; + } + + rte_spinlock_lock(&lio_dev->glist_lock[iq_no]); + g = (struct lio_gather *)list_delete_first_node( + &lio_dev->glist_head[iq_no]); + rte_spinlock_unlock(&lio_dev->glist_lock[iq_no]); + if (g == NULL) { + PMD_TX_LOG(lio_dev, ERR, + "Transmit scatter gather: glist null!\n"); + goto xmit_failed; + } + + cmdsetup.s.gather = 1; + cmdsetup.s.u.gatherptrs = m->nb_segs; + lio_prepare_pci_cmd(lio_dev, &ndata.cmd, + &cmdsetup, tag); + + memset(g->sg, 0, g->sg_size); + g->sg[0].ptr[0] = rte_mbuf_data_dma_addr(m); + lio_add_sg_size(&g->sg[0], m->data_len, 0); + pkt_len = m->data_len; + finfo->mbuf = m; + + /* First seg taken care above */ + frags = m->nb_segs - 1; + i = 1; + m = m->next; + while (frags--) { + g->sg[(i >> 2)].ptr[(i & 3)] = + rte_mbuf_data_dma_addr(m); + lio_add_sg_size(&g->sg[(i >> 2)], + m->data_len, (i & 3)); + pkt_len += m->data_len; + i++; + m = m->next; + } + + phyaddr = rte_mem_virt2phy(g->sg); + if (phyaddr == RTE_BAD_PHYS_ADDR) { + PMD_TX_LOG(lio_dev, ERR, "bad phys addr\n"); + goto xmit_failed; + } + + ndata.cmd.cmd3.dptr = phyaddr; + ndata.reqtype = LIO_REQTYPE_NORESP_NET_SG; + + finfo->g = g; + finfo->lio_dev = lio_dev; + finfo->iq_no = (uint64_t)iq_no; + ndata.buf = finfo; } ndata.datasize = pkt_len; @@ -1462,14 +1814,43 @@ lio_dev_xmit_pkts(void *tx_queue, struct rte_mbuf **pkts, uint16_t nb_pkts) break; } - if (unlikely(status == LIO_IQ_SEND_STOP)) + if (unlikely(status == LIO_IQ_SEND_STOP)) { PMD_TX_LOG(lio_dev, DEBUG, "iq full\n"); + /* create space as iq is full */ + lio_dev_cleanup_iq(lio_dev, iq_no); + } + stats->tx_done++; + stats->tx_tot_bytes += pkt_len; processed++; } xmit_failed: + stats->tx_dropped += (nb_pkts - processed); return processed; } +void +lio_dev_clear_queues(struct rte_eth_dev *eth_dev) +{ + struct lio_instr_queue *txq; + struct lio_droq *rxq; + uint16_t i; + + for (i = 0; i < eth_dev->data->nb_tx_queues; i++) { + txq = eth_dev->data->tx_queues[i]; + if (txq != NULL) { + lio_dev_tx_queue_release(txq); + eth_dev->data->tx_queues[i] = NULL; + } + } + + for (i = 0; i < eth_dev->data->nb_rx_queues; i++) { + rxq = eth_dev->data->rx_queues[i]; + if (rxq != NULL) { + lio_dev_rx_queue_release(rxq); + eth_dev->data->rx_queues[i] = NULL; + } + } +}