2 * Copyright (c) 2014, 2015 Netronome Systems, Inc.
5 * Small portions derived from code Copyright(c) 2010-2015 Intel Corporation.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions are met:
10 * 1. Redistributions of source code must retain the above copyright notice,
11 * this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution
17 * 3. Neither the name of the copyright holder nor the names of its
18 * contributors may be used to endorse or promote products derived from this
19 * software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
22 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
25 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
26 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
27 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
28 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
29 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
30 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31 * POSSIBILITY OF SUCH DAMAGE.
35 * vim:shiftwidth=8:noexpandtab
37 * @file dpdk/pmd/nfp_net.c
39 * Netronome vNIC DPDK Poll-Mode Driver: Main entry point
42 #include <rte_byteorder.h>
43 #include <rte_common.h>
45 #include <rte_debug.h>
46 #include <rte_ethdev.h>
47 #include <rte_ethdev_pci.h>
49 #include <rte_ether.h>
50 #include <rte_malloc.h>
51 #include <rte_memzone.h>
52 #include <rte_mempool.h>
53 #include <rte_version.h>
54 #include <rte_string_fns.h>
55 #include <rte_alarm.h>
56 #include <rte_spinlock.h>
59 #include "nfp_net_pmd.h"
60 #include "nfp_net_logs.h"
61 #include "nfp_net_ctrl.h"
64 static void nfp_net_close(struct rte_eth_dev *dev);
65 static int nfp_net_configure(struct rte_eth_dev *dev);
66 static void nfp_net_dev_interrupt_handler(void *param);
67 static void nfp_net_dev_interrupt_delayed_handler(void *param);
68 static int nfp_net_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu);
69 static void nfp_net_infos_get(struct rte_eth_dev *dev,
70 struct rte_eth_dev_info *dev_info);
71 static int nfp_net_init(struct rte_eth_dev *eth_dev);
72 static int nfp_net_link_update(struct rte_eth_dev *dev, int wait_to_complete);
73 static void nfp_net_promisc_enable(struct rte_eth_dev *dev);
74 static void nfp_net_promisc_disable(struct rte_eth_dev *dev);
75 static int nfp_net_rx_fill_freelist(struct nfp_net_rxq *rxq);
76 static uint32_t nfp_net_rx_queue_count(struct rte_eth_dev *dev,
78 static uint16_t nfp_net_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
80 static void nfp_net_rx_queue_release(void *rxq);
81 static int nfp_net_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
82 uint16_t nb_desc, unsigned int socket_id,
83 const struct rte_eth_rxconf *rx_conf,
84 struct rte_mempool *mp);
85 static int nfp_net_tx_free_bufs(struct nfp_net_txq *txq);
86 static void nfp_net_tx_queue_release(void *txq);
87 static int nfp_net_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
88 uint16_t nb_desc, unsigned int socket_id,
89 const struct rte_eth_txconf *tx_conf);
90 static int nfp_net_start(struct rte_eth_dev *dev);
91 static void nfp_net_stats_get(struct rte_eth_dev *dev,
92 struct rte_eth_stats *stats);
93 static void nfp_net_stats_reset(struct rte_eth_dev *dev);
94 static void nfp_net_stop(struct rte_eth_dev *dev);
95 static uint16_t nfp_net_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
99 * The offset of the queue controller queues in the PCIe Target. These
100 * happen to be at the same offset on the NFP6000 and the NFP3200 so
101 * we use a single macro here.
103 #define NFP_PCIE_QUEUE(_q) (0x800 * ((_q) & 0xff))
105 /* Maximum value which can be added to a queue with one transaction */
106 #define NFP_QCP_MAX_ADD 0x7f
108 #define RTE_MBUF_DMA_ADDR_DEFAULT(mb) \
109 (uint64_t)((mb)->buf_physaddr + RTE_PKTMBUF_HEADROOM)
111 /* nfp_qcp_ptr - Read or Write Pointer of a queue */
113 NFP_QCP_READ_PTR = 0,
118 * nfp_qcp_ptr_add - Add the value to the selected pointer of a queue
119 * @q: Base address for queue structure
120 * @ptr: Add to the Read or Write pointer
121 * @val: Value to add to the queue pointer
123 * If @val is greater than @NFP_QCP_MAX_ADD multiple writes are performed.
126 nfp_qcp_ptr_add(uint8_t *q, enum nfp_qcp_ptr ptr, uint32_t val)
130 if (ptr == NFP_QCP_READ_PTR)
131 off = NFP_QCP_QUEUE_ADD_RPTR;
133 off = NFP_QCP_QUEUE_ADD_WPTR;
135 while (val > NFP_QCP_MAX_ADD) {
136 nn_writel(rte_cpu_to_le_32(NFP_QCP_MAX_ADD), q + off);
137 val -= NFP_QCP_MAX_ADD;
140 nn_writel(rte_cpu_to_le_32(val), q + off);
144 * nfp_qcp_read - Read the current Read/Write pointer value for a queue
145 * @q: Base address for queue structure
146 * @ptr: Read or Write pointer
148 static inline uint32_t
149 nfp_qcp_read(uint8_t *q, enum nfp_qcp_ptr ptr)
154 if (ptr == NFP_QCP_READ_PTR)
155 off = NFP_QCP_QUEUE_STS_LO;
157 off = NFP_QCP_QUEUE_STS_HI;
159 val = rte_cpu_to_le_32(nn_readl(q + off));
161 if (ptr == NFP_QCP_READ_PTR)
162 return val & NFP_QCP_QUEUE_STS_LO_READPTR_mask;
164 return val & NFP_QCP_QUEUE_STS_HI_WRITEPTR_mask;
168 * Functions to read/write from/to Config BAR
169 * Performs any endian conversion necessary.
171 static inline uint8_t
172 nn_cfg_readb(struct nfp_net_hw *hw, int off)
174 return nn_readb(hw->ctrl_bar + off);
178 nn_cfg_writeb(struct nfp_net_hw *hw, int off, uint8_t val)
180 nn_writeb(val, hw->ctrl_bar + off);
183 static inline uint32_t
184 nn_cfg_readl(struct nfp_net_hw *hw, int off)
186 return rte_le_to_cpu_32(nn_readl(hw->ctrl_bar + off));
190 nn_cfg_writel(struct nfp_net_hw *hw, int off, uint32_t val)
192 nn_writel(rte_cpu_to_le_32(val), hw->ctrl_bar + off);
195 static inline uint64_t
196 nn_cfg_readq(struct nfp_net_hw *hw, int off)
198 return rte_le_to_cpu_64(nn_readq(hw->ctrl_bar + off));
202 nn_cfg_writeq(struct nfp_net_hw *hw, int off, uint64_t val)
204 nn_writeq(rte_cpu_to_le_64(val), hw->ctrl_bar + off);
208 * Atomically reads link status information from global structure rte_eth_dev.
211 * - Pointer to the structure rte_eth_dev to read from.
212 * - Pointer to the buffer to be saved with the link status.
215 * - On success, zero.
216 * - On failure, negative value.
219 nfp_net_dev_atomic_read_link_status(struct rte_eth_dev *dev,
220 struct rte_eth_link *link)
222 struct rte_eth_link *dst = link;
223 struct rte_eth_link *src = &dev->data->dev_link;
225 if (rte_atomic64_cmpset((uint64_t *)dst, *(uint64_t *)dst,
226 *(uint64_t *)src) == 0)
233 * Atomically writes the link status information into global
234 * structure rte_eth_dev.
237 * - Pointer to the structure rte_eth_dev to read from.
238 * - Pointer to the buffer to be saved with the link status.
241 * - On success, zero.
242 * - On failure, negative value.
245 nfp_net_dev_atomic_write_link_status(struct rte_eth_dev *dev,
246 struct rte_eth_link *link)
248 struct rte_eth_link *dst = &dev->data->dev_link;
249 struct rte_eth_link *src = link;
251 if (rte_atomic64_cmpset((uint64_t *)dst, *(uint64_t *)dst,
252 *(uint64_t *)src) == 0)
259 nfp_net_rx_queue_release_mbufs(struct nfp_net_rxq *rxq)
263 if (rxq->rxbufs == NULL)
266 for (i = 0; i < rxq->rx_count; i++) {
267 if (rxq->rxbufs[i].mbuf) {
268 rte_pktmbuf_free_seg(rxq->rxbufs[i].mbuf);
269 rxq->rxbufs[i].mbuf = NULL;
275 nfp_net_rx_queue_release(void *rx_queue)
277 struct nfp_net_rxq *rxq = rx_queue;
280 nfp_net_rx_queue_release_mbufs(rxq);
281 rte_free(rxq->rxbufs);
287 nfp_net_reset_rx_queue(struct nfp_net_rxq *rxq)
289 nfp_net_rx_queue_release_mbufs(rxq);
295 nfp_net_tx_queue_release_mbufs(struct nfp_net_txq *txq)
299 if (txq->txbufs == NULL)
302 for (i = 0; i < txq->tx_count; i++) {
303 if (txq->txbufs[i].mbuf) {
304 rte_pktmbuf_free(txq->txbufs[i].mbuf);
305 txq->txbufs[i].mbuf = NULL;
311 nfp_net_tx_queue_release(void *tx_queue)
313 struct nfp_net_txq *txq = tx_queue;
316 nfp_net_tx_queue_release_mbufs(txq);
317 rte_free(txq->txbufs);
323 nfp_net_reset_tx_queue(struct nfp_net_txq *txq)
325 nfp_net_tx_queue_release_mbufs(txq);
331 __nfp_net_reconfig(struct nfp_net_hw *hw, uint32_t update)
335 struct timespec wait;
337 PMD_DRV_LOG(DEBUG, "Writing to the configuration queue (%p)...\n",
340 if (hw->qcp_cfg == NULL)
341 rte_panic("Bad configuration queue pointer\n");
343 nfp_qcp_ptr_add(hw->qcp_cfg, NFP_QCP_WRITE_PTR, 1);
346 wait.tv_nsec = 1000000;
348 PMD_DRV_LOG(DEBUG, "Polling for update ack...\n");
350 /* Poll update field, waiting for NFP to ack the config */
351 for (cnt = 0; ; cnt++) {
352 new = nn_cfg_readl(hw, NFP_NET_CFG_UPDATE);
355 if (new & NFP_NET_CFG_UPDATE_ERR) {
356 PMD_INIT_LOG(ERR, "Reconfig error: 0x%08x", new);
359 if (cnt >= NFP_NET_POLL_TIMEOUT) {
360 PMD_INIT_LOG(ERR, "Reconfig timeout for 0x%08x after"
361 " %dms", update, cnt);
362 rte_panic("Exiting\n");
364 nanosleep(&wait, 0); /* waiting for a 1ms */
366 PMD_DRV_LOG(DEBUG, "Ack DONE\n");
371 * Reconfigure the NIC
372 * @nn: device to reconfigure
373 * @ctrl: The value for the ctrl field in the BAR config
374 * @update: The value for the update field in the BAR config
376 * Write the update word to the BAR and ping the reconfig queue. Then poll
377 * until the firmware has acknowledged the update by zeroing the update word.
380 nfp_net_reconfig(struct nfp_net_hw *hw, uint32_t ctrl, uint32_t update)
384 PMD_DRV_LOG(DEBUG, "nfp_net_reconfig: ctrl=%08x update=%08x\n",
387 rte_spinlock_lock(&hw->reconfig_lock);
389 nn_cfg_writel(hw, NFP_NET_CFG_CTRL, ctrl);
390 nn_cfg_writel(hw, NFP_NET_CFG_UPDATE, update);
394 err = __nfp_net_reconfig(hw, update);
396 rte_spinlock_unlock(&hw->reconfig_lock);
402 * Reconfig errors imply situations where they can be handled.
403 * Otherwise, rte_panic is called inside __nfp_net_reconfig
405 PMD_INIT_LOG(ERR, "Error nfp_net reconfig for ctrl: %x update: %x",
411 * Configure an Ethernet device. This function must be invoked first
412 * before any other function in the Ethernet API. This function can
413 * also be re-invoked when a device is in the stopped state.
416 nfp_net_configure(struct rte_eth_dev *dev)
418 struct rte_eth_conf *dev_conf;
419 struct rte_eth_rxmode *rxmode;
420 struct rte_eth_txmode *txmode;
421 uint32_t new_ctrl = 0;
423 struct nfp_net_hw *hw;
425 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
428 * A DPDK app sends info about how many queues to use and how
429 * those queues need to be configured. This is used by the
430 * DPDK core and it makes sure no more queues than those
431 * advertised by the driver are requested. This function is
432 * called after that internal process
435 PMD_INIT_LOG(DEBUG, "Configure");
437 dev_conf = &dev->data->dev_conf;
438 rxmode = &dev_conf->rxmode;
439 txmode = &dev_conf->txmode;
441 /* Checking TX mode */
442 if (txmode->mq_mode) {
443 PMD_INIT_LOG(INFO, "TX mq_mode DCB and VMDq not supported");
447 /* Checking RX mode */
448 if (rxmode->mq_mode & ETH_MQ_RX_RSS) {
449 if (hw->cap & NFP_NET_CFG_CTRL_RSS) {
450 update = NFP_NET_CFG_UPDATE_RSS;
451 new_ctrl = NFP_NET_CFG_CTRL_RSS;
453 PMD_INIT_LOG(INFO, "RSS not supported");
458 if (rxmode->split_hdr_size) {
459 PMD_INIT_LOG(INFO, "rxmode does not support split header");
463 if (rxmode->hw_ip_checksum) {
464 if (hw->cap & NFP_NET_CFG_CTRL_RXCSUM) {
465 new_ctrl |= NFP_NET_CFG_CTRL_RXCSUM;
467 PMD_INIT_LOG(INFO, "RXCSUM not supported");
472 if (rxmode->hw_vlan_filter) {
473 PMD_INIT_LOG(INFO, "VLAN filter not supported");
477 if (rxmode->hw_vlan_strip) {
478 if (hw->cap & NFP_NET_CFG_CTRL_RXVLAN) {
479 new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
481 PMD_INIT_LOG(INFO, "hw vlan strip not supported");
486 if (rxmode->hw_vlan_extend) {
487 PMD_INIT_LOG(INFO, "VLAN extended not supported");
491 /* Supporting VLAN insertion by default */
492 if (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)
493 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
495 if (rxmode->jumbo_frame)
496 /* this is handled in rte_eth_dev_configure */
498 if (rxmode->hw_strip_crc) {
499 PMD_INIT_LOG(INFO, "strip CRC not supported");
503 if (rxmode->enable_scatter) {
504 PMD_INIT_LOG(INFO, "Scatter not supported");
511 update |= NFP_NET_CFG_UPDATE_GEN;
513 nn_cfg_writel(hw, NFP_NET_CFG_CTRL, new_ctrl);
514 if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
523 nfp_net_enable_queues(struct rte_eth_dev *dev)
525 struct nfp_net_hw *hw;
526 uint64_t enabled_queues = 0;
529 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
531 /* Enabling the required TX queues in the device */
532 for (i = 0; i < dev->data->nb_tx_queues; i++)
533 enabled_queues |= (1 << i);
535 nn_cfg_writeq(hw, NFP_NET_CFG_TXRS_ENABLE, enabled_queues);
539 /* Enabling the required RX queues in the device */
540 for (i = 0; i < dev->data->nb_rx_queues; i++)
541 enabled_queues |= (1 << i);
543 nn_cfg_writeq(hw, NFP_NET_CFG_RXRS_ENABLE, enabled_queues);
547 nfp_net_disable_queues(struct rte_eth_dev *dev)
549 struct nfp_net_hw *hw;
550 uint32_t new_ctrl, update = 0;
552 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
554 nn_cfg_writeq(hw, NFP_NET_CFG_TXRS_ENABLE, 0);
555 nn_cfg_writeq(hw, NFP_NET_CFG_RXRS_ENABLE, 0);
557 new_ctrl = hw->ctrl & ~NFP_NET_CFG_CTRL_ENABLE;
558 update = NFP_NET_CFG_UPDATE_GEN | NFP_NET_CFG_UPDATE_RING |
559 NFP_NET_CFG_UPDATE_MSIX;
561 if (hw->cap & NFP_NET_CFG_CTRL_RINGCFG)
562 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG;
564 /* If an error when reconfig we avoid to change hw state */
565 if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
572 nfp_net_rx_freelist_setup(struct rte_eth_dev *dev)
576 for (i = 0; i < dev->data->nb_rx_queues; i++) {
577 if (nfp_net_rx_fill_freelist(dev->data->rx_queues[i]) < 0)
584 nfp_net_params_setup(struct nfp_net_hw *hw)
586 nn_cfg_writel(hw, NFP_NET_CFG_MTU, hw->mtu);
587 nn_cfg_writel(hw, NFP_NET_CFG_FLBUFSZ, hw->flbufsz);
591 nfp_net_cfg_queue_setup(struct nfp_net_hw *hw)
593 hw->qcp_cfg = hw->tx_bar + NFP_QCP_QUEUE_ADDR_SZ;
596 static void nfp_net_read_mac(struct nfp_net_hw *hw)
600 tmp = rte_be_to_cpu_32(nn_cfg_readl(hw, NFP_NET_CFG_MACADDR));
601 memcpy(&hw->mac_addr[0], &tmp, sizeof(struct ether_addr));
603 tmp = rte_be_to_cpu_32(nn_cfg_readl(hw, NFP_NET_CFG_MACADDR + 4));
604 memcpy(&hw->mac_addr[4], &tmp, 2);
608 nfp_net_write_mac(struct nfp_net_hw *hw, uint8_t *mac)
610 uint32_t mac0 = *(uint32_t *)mac;
613 nn_writel(rte_cpu_to_be_32(mac0), hw->ctrl_bar + NFP_NET_CFG_MACADDR);
616 mac1 = *(uint16_t *)mac;
617 nn_writew(rte_cpu_to_be_16(mac1),
618 hw->ctrl_bar + NFP_NET_CFG_MACADDR + 6);
622 nfp_configure_rx_interrupt(struct rte_eth_dev *dev,
623 struct rte_intr_handle *intr_handle)
625 struct nfp_net_hw *hw;
628 if (!intr_handle->intr_vec) {
629 intr_handle->intr_vec =
630 rte_zmalloc("intr_vec",
631 dev->data->nb_rx_queues * sizeof(int), 0);
632 if (!intr_handle->intr_vec) {
633 PMD_INIT_LOG(ERR, "Failed to allocate %d rx_queues"
634 " intr_vec", dev->data->nb_rx_queues);
639 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
641 if (intr_handle->type == RTE_INTR_HANDLE_UIO) {
642 PMD_INIT_LOG(INFO, "VF: enabling RX interrupt with UIO");
643 /* UIO just supports one queue and no LSC*/
644 nn_cfg_writeb(hw, NFP_NET_CFG_RXR_VEC(0), 0);
645 intr_handle->intr_vec[0] = 0;
647 PMD_INIT_LOG(INFO, "VF: enabling RX interrupt with VFIO");
648 for (i = 0; i < dev->data->nb_rx_queues; i++) {
650 * The first msix vector is reserved for non
653 nn_cfg_writeb(hw, NFP_NET_CFG_RXR_VEC(i), i + 1);
654 intr_handle->intr_vec[i] = i + 1;
655 PMD_INIT_LOG(DEBUG, "intr_vec[%d]= %d\n", i,
656 intr_handle->intr_vec[i]);
660 /* Avoiding TX interrupts */
661 hw->ctrl |= NFP_NET_CFG_CTRL_MSIX_TX_OFF;
666 nfp_net_start(struct rte_eth_dev *dev)
668 struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
669 struct rte_intr_handle *intr_handle = &pci_dev->intr_handle;
670 uint32_t new_ctrl, update = 0;
671 struct nfp_net_hw *hw;
672 uint32_t intr_vector;
675 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
677 PMD_INIT_LOG(DEBUG, "Start");
679 /* Disabling queues just in case... */
680 nfp_net_disable_queues(dev);
682 /* Writing configuration parameters in the device */
683 nfp_net_params_setup(hw);
685 /* Enabling the required queues in the device */
686 nfp_net_enable_queues(dev);
688 /* check and configure queue intr-vector mapping */
689 if (dev->data->dev_conf.intr_conf.rxq != 0) {
690 if (intr_handle->type == RTE_INTR_HANDLE_UIO) {
692 * Better not to share LSC with RX interrupts.
693 * Unregistering LSC interrupt handler
695 rte_intr_callback_unregister(&pci_dev->intr_handle,
696 nfp_net_dev_interrupt_handler, (void *)dev);
698 if (dev->data->nb_rx_queues > 1) {
699 PMD_INIT_LOG(ERR, "PMD rx interrupt only "
700 "supports 1 queue with UIO");
704 intr_vector = dev->data->nb_rx_queues;
705 if (rte_intr_efd_enable(intr_handle, intr_vector))
708 nfp_configure_rx_interrupt(dev, intr_handle);
709 update = NFP_NET_CFG_UPDATE_MSIX;
712 rte_intr_enable(intr_handle);
715 new_ctrl = hw->ctrl | NFP_NET_CFG_CTRL_ENABLE;
717 update |= NFP_NET_CFG_UPDATE_GEN | NFP_NET_CFG_UPDATE_RING;
719 if (hw->cap & NFP_NET_CFG_CTRL_RINGCFG)
720 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG;
722 nn_cfg_writel(hw, NFP_NET_CFG_CTRL, new_ctrl);
723 if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
727 * Allocating rte mbuffs for configured rx queues.
728 * This requires queues being enabled before
730 if (nfp_net_rx_freelist_setup(dev) < 0) {
741 * An error returned by this function should mean the app
742 * exiting and then the system releasing all the memory
743 * allocated even memory coming from hugepages.
745 * The device could be enabled at this point with some queues
746 * ready for getting packets. This is true if the call to
747 * nfp_net_rx_freelist_setup() succeeds for some queues but
748 * fails for subsequent queues.
750 * This should make the app exiting but better if we tell the
753 nfp_net_disable_queues(dev);
758 /* Stop device: disable rx and tx functions to allow for reconfiguring. */
760 nfp_net_stop(struct rte_eth_dev *dev)
764 PMD_INIT_LOG(DEBUG, "Stop");
766 nfp_net_disable_queues(dev);
769 for (i = 0; i < dev->data->nb_tx_queues; i++) {
770 nfp_net_reset_tx_queue(
771 (struct nfp_net_txq *)dev->data->tx_queues[i]);
774 for (i = 0; i < dev->data->nb_rx_queues; i++) {
775 nfp_net_reset_rx_queue(
776 (struct nfp_net_rxq *)dev->data->rx_queues[i]);
780 /* Reset and stop device. The device can not be restarted. */
782 nfp_net_close(struct rte_eth_dev *dev)
784 struct nfp_net_hw *hw;
785 struct rte_pci_device *pci_dev;
787 PMD_INIT_LOG(DEBUG, "Close");
789 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
790 pci_dev = RTE_ETH_DEV_TO_PCI(dev);
793 * We assume that the DPDK application is stopping all the
794 * threads/queues before calling the device close function.
799 rte_intr_disable(&pci_dev->intr_handle);
800 nn_cfg_writeb(hw, NFP_NET_CFG_LSC, 0xff);
802 /* unregister callback func from eal lib */
803 rte_intr_callback_unregister(&pci_dev->intr_handle,
804 nfp_net_dev_interrupt_handler,
808 * The ixgbe PMD driver disables the pcie master on the
809 * device. The i40e does not...
814 nfp_net_promisc_enable(struct rte_eth_dev *dev)
816 uint32_t new_ctrl, update = 0;
817 struct nfp_net_hw *hw;
819 PMD_DRV_LOG(DEBUG, "Promiscuous mode enable\n");
821 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
823 if (!(hw->cap & NFP_NET_CFG_CTRL_PROMISC)) {
824 PMD_INIT_LOG(INFO, "Promiscuous mode not supported");
828 if (hw->ctrl & NFP_NET_CFG_CTRL_PROMISC) {
829 PMD_DRV_LOG(INFO, "Promiscuous mode already enabled\n");
833 new_ctrl = hw->ctrl | NFP_NET_CFG_CTRL_PROMISC;
834 update = NFP_NET_CFG_UPDATE_GEN;
837 * DPDK sets promiscuous mode on just after this call assuming
838 * it can not fail ...
840 if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
847 nfp_net_promisc_disable(struct rte_eth_dev *dev)
849 uint32_t new_ctrl, update = 0;
850 struct nfp_net_hw *hw;
852 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
854 if ((hw->ctrl & NFP_NET_CFG_CTRL_PROMISC) == 0) {
855 PMD_DRV_LOG(INFO, "Promiscuous mode already disabled\n");
859 new_ctrl = hw->ctrl & ~NFP_NET_CFG_CTRL_PROMISC;
860 update = NFP_NET_CFG_UPDATE_GEN;
863 * DPDK sets promiscuous mode off just before this call
864 * assuming it can not fail ...
866 if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
873 * return 0 means link status changed, -1 means not changed
875 * Wait to complete is needed as it can take up to 9 seconds to get the Link
879 nfp_net_link_update(struct rte_eth_dev *dev, __rte_unused int wait_to_complete)
881 struct nfp_net_hw *hw;
882 struct rte_eth_link link, old;
883 uint32_t nn_link_status;
885 static const uint32_t ls_to_ethtool[] = {
886 [NFP_NET_CFG_STS_LINK_RATE_UNSUPPORTED] = ETH_SPEED_NUM_NONE,
887 [NFP_NET_CFG_STS_LINK_RATE_UNKNOWN] = ETH_SPEED_NUM_NONE,
888 [NFP_NET_CFG_STS_LINK_RATE_1G] = ETH_SPEED_NUM_1G,
889 [NFP_NET_CFG_STS_LINK_RATE_10G] = ETH_SPEED_NUM_10G,
890 [NFP_NET_CFG_STS_LINK_RATE_25G] = ETH_SPEED_NUM_25G,
891 [NFP_NET_CFG_STS_LINK_RATE_40G] = ETH_SPEED_NUM_40G,
892 [NFP_NET_CFG_STS_LINK_RATE_50G] = ETH_SPEED_NUM_50G,
893 [NFP_NET_CFG_STS_LINK_RATE_100G] = ETH_SPEED_NUM_100G,
896 PMD_DRV_LOG(DEBUG, "Link update\n");
898 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
900 memset(&old, 0, sizeof(old));
901 nfp_net_dev_atomic_read_link_status(dev, &old);
903 nn_link_status = nn_cfg_readl(hw, NFP_NET_CFG_STS);
905 memset(&link, 0, sizeof(struct rte_eth_link));
907 if (nn_link_status & NFP_NET_CFG_STS_LINK)
908 link.link_status = ETH_LINK_UP;
910 link.link_duplex = ETH_LINK_FULL_DUPLEX;
912 nn_link_status = (nn_link_status >> NFP_NET_CFG_STS_LINK_RATE_SHIFT) &
913 NFP_NET_CFG_STS_LINK_RATE_MASK;
915 if ((NFD_CFG_MAJOR_VERSION_of(hw->ver) < 4) ||
916 ((NFD_CFG_MINOR_VERSION_of(hw->ver) == 4) &&
917 (NFD_CFG_MINOR_VERSION_of(hw->ver) == 0)))
918 /* We really do not know the speed wil old firmware */
919 link.link_speed = ETH_SPEED_NUM_NONE;
921 if (nn_link_status >= RTE_DIM(ls_to_ethtool))
922 link.link_speed = ETH_SPEED_NUM_NONE;
924 link.link_speed = ls_to_ethtool[nn_link_status];
927 if (old.link_status != link.link_status) {
928 nfp_net_dev_atomic_write_link_status(dev, &link);
929 if (link.link_status)
930 PMD_DRV_LOG(INFO, "NIC Link is Up\n");
932 PMD_DRV_LOG(INFO, "NIC Link is Down\n");
940 nfp_net_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
943 struct nfp_net_hw *hw;
944 struct rte_eth_stats nfp_dev_stats;
946 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
948 /* RTE_ETHDEV_QUEUE_STAT_CNTRS default value is 16 */
950 /* reading per RX ring stats */
951 for (i = 0; i < dev->data->nb_rx_queues; i++) {
952 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
955 nfp_dev_stats.q_ipackets[i] =
956 nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i));
958 nfp_dev_stats.q_ipackets[i] -=
959 hw->eth_stats_base.q_ipackets[i];
961 nfp_dev_stats.q_ibytes[i] =
962 nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i) + 0x8);
964 nfp_dev_stats.q_ibytes[i] -=
965 hw->eth_stats_base.q_ibytes[i];
968 /* reading per TX ring stats */
969 for (i = 0; i < dev->data->nb_tx_queues; i++) {
970 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
973 nfp_dev_stats.q_opackets[i] =
974 nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i));
976 nfp_dev_stats.q_opackets[i] -=
977 hw->eth_stats_base.q_opackets[i];
979 nfp_dev_stats.q_obytes[i] =
980 nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i) + 0x8);
982 nfp_dev_stats.q_obytes[i] -=
983 hw->eth_stats_base.q_obytes[i];
986 nfp_dev_stats.ipackets =
987 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_FRAMES);
989 nfp_dev_stats.ipackets -= hw->eth_stats_base.ipackets;
991 nfp_dev_stats.ibytes =
992 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_OCTETS);
994 nfp_dev_stats.ibytes -= hw->eth_stats_base.ibytes;
996 nfp_dev_stats.opackets =
997 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_FRAMES);
999 nfp_dev_stats.opackets -= hw->eth_stats_base.opackets;
1001 nfp_dev_stats.obytes =
1002 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_OCTETS);
1004 nfp_dev_stats.obytes -= hw->eth_stats_base.obytes;
1006 /* reading general device stats */
1007 nfp_dev_stats.ierrors =
1008 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_ERRORS);
1010 nfp_dev_stats.ierrors -= hw->eth_stats_base.ierrors;
1012 nfp_dev_stats.oerrors =
1013 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_ERRORS);
1015 nfp_dev_stats.oerrors -= hw->eth_stats_base.oerrors;
1017 /* RX ring mbuf allocation failures */
1018 nfp_dev_stats.rx_nombuf = dev->data->rx_mbuf_alloc_failed;
1020 nfp_dev_stats.imissed =
1021 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_DISCARDS);
1023 nfp_dev_stats.imissed -= hw->eth_stats_base.imissed;
1026 memcpy(stats, &nfp_dev_stats, sizeof(*stats));
1030 nfp_net_stats_reset(struct rte_eth_dev *dev)
1033 struct nfp_net_hw *hw;
1035 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1038 * hw->eth_stats_base records the per counter starting point.
1039 * Lets update it now
1042 /* reading per RX ring stats */
1043 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1044 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
1047 hw->eth_stats_base.q_ipackets[i] =
1048 nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i));
1050 hw->eth_stats_base.q_ibytes[i] =
1051 nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i) + 0x8);
1054 /* reading per TX ring stats */
1055 for (i = 0; i < dev->data->nb_tx_queues; i++) {
1056 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
1059 hw->eth_stats_base.q_opackets[i] =
1060 nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i));
1062 hw->eth_stats_base.q_obytes[i] =
1063 nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i) + 0x8);
1066 hw->eth_stats_base.ipackets =
1067 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_FRAMES);
1069 hw->eth_stats_base.ibytes =
1070 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_OCTETS);
1072 hw->eth_stats_base.opackets =
1073 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_FRAMES);
1075 hw->eth_stats_base.obytes =
1076 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_OCTETS);
1078 /* reading general device stats */
1079 hw->eth_stats_base.ierrors =
1080 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_ERRORS);
1082 hw->eth_stats_base.oerrors =
1083 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_ERRORS);
1085 /* RX ring mbuf allocation failures */
1086 dev->data->rx_mbuf_alloc_failed = 0;
1088 hw->eth_stats_base.imissed =
1089 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_DISCARDS);
1093 nfp_net_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
1095 struct nfp_net_hw *hw;
1097 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1099 dev_info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1100 dev_info->max_rx_queues = (uint16_t)hw->max_rx_queues;
1101 dev_info->max_tx_queues = (uint16_t)hw->max_tx_queues;
1102 dev_info->min_rx_bufsize = ETHER_MIN_MTU;
1103 dev_info->max_rx_pktlen = hw->mtu;
1104 /* Next should change when PF support is implemented */
1105 dev_info->max_mac_addrs = 1;
1107 if (hw->cap & NFP_NET_CFG_CTRL_RXVLAN)
1108 dev_info->rx_offload_capa = DEV_RX_OFFLOAD_VLAN_STRIP;
1110 if (hw->cap & NFP_NET_CFG_CTRL_RXCSUM)
1111 dev_info->rx_offload_capa |= DEV_RX_OFFLOAD_IPV4_CKSUM |
1112 DEV_RX_OFFLOAD_UDP_CKSUM |
1113 DEV_RX_OFFLOAD_TCP_CKSUM;
1115 if (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)
1116 dev_info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
1118 if (hw->cap & NFP_NET_CFG_CTRL_TXCSUM)
1119 dev_info->tx_offload_capa |= DEV_TX_OFFLOAD_IPV4_CKSUM |
1120 DEV_TX_OFFLOAD_UDP_CKSUM |
1121 DEV_TX_OFFLOAD_TCP_CKSUM;
1123 dev_info->default_rxconf = (struct rte_eth_rxconf) {
1125 .pthresh = DEFAULT_RX_PTHRESH,
1126 .hthresh = DEFAULT_RX_HTHRESH,
1127 .wthresh = DEFAULT_RX_WTHRESH,
1129 .rx_free_thresh = DEFAULT_RX_FREE_THRESH,
1133 dev_info->default_txconf = (struct rte_eth_txconf) {
1135 .pthresh = DEFAULT_TX_PTHRESH,
1136 .hthresh = DEFAULT_TX_HTHRESH,
1137 .wthresh = DEFAULT_TX_WTHRESH,
1139 .tx_free_thresh = DEFAULT_TX_FREE_THRESH,
1140 .tx_rs_thresh = DEFAULT_TX_RSBIT_THRESH,
1141 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS |
1142 ETH_TXQ_FLAGS_NOOFFLOADS,
1145 dev_info->reta_size = NFP_NET_CFG_RSS_ITBL_SZ;
1146 dev_info->hash_key_size = NFP_NET_CFG_RSS_KEY_SZ;
1148 dev_info->speed_capa = ETH_SPEED_NUM_1G | ETH_LINK_SPEED_10G |
1149 ETH_SPEED_NUM_25G | ETH_SPEED_NUM_40G |
1150 ETH_SPEED_NUM_50G | ETH_LINK_SPEED_100G;
1152 if (hw->cap & NFP_NET_CFG_CTRL_LSO)
1153 dev_info->tx_offload_capa |= DEV_TX_OFFLOAD_TCP_TSO;
1156 static const uint32_t *
1157 nfp_net_supported_ptypes_get(struct rte_eth_dev *dev)
1159 static const uint32_t ptypes[] = {
1160 /* refers to nfp_net_set_hash() */
1161 RTE_PTYPE_INNER_L3_IPV4,
1162 RTE_PTYPE_INNER_L3_IPV6,
1163 RTE_PTYPE_INNER_L3_IPV6_EXT,
1164 RTE_PTYPE_INNER_L4_MASK,
1168 if (dev->rx_pkt_burst == nfp_net_recv_pkts)
1174 nfp_net_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_idx)
1176 struct nfp_net_rxq *rxq;
1177 struct nfp_net_rx_desc *rxds;
1181 rxq = (struct nfp_net_rxq *)dev->data->rx_queues[queue_idx];
1188 * Other PMDs are just checking the DD bit in intervals of 4
1189 * descriptors and counting all four if the first has the DD
1190 * bit on. Of course, this is not accurate but can be good for
1191 * performance. But ideally that should be done in descriptors
1192 * chunks belonging to the same cache line
1195 while (count < rxq->rx_count) {
1196 rxds = &rxq->rxds[idx];
1197 if ((rxds->rxd.meta_len_dd & PCIE_DESC_RX_DD) == 0)
1204 if ((idx) == rxq->rx_count)
1212 nfp_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t queue_id)
1214 struct rte_pci_device *pci_dev;
1215 struct nfp_net_hw *hw;
1218 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1219 pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1221 if (pci_dev->intr_handle.type != RTE_INTR_HANDLE_UIO)
1224 /* Make sure all updates are written before un-masking */
1226 nn_cfg_writeb(hw, NFP_NET_CFG_ICR(base + queue_id),
1227 NFP_NET_CFG_ICR_UNMASKED);
1232 nfp_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t queue_id)
1234 struct rte_pci_device *pci_dev;
1235 struct nfp_net_hw *hw;
1238 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1239 pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1241 if (pci_dev->intr_handle.type != RTE_INTR_HANDLE_UIO)
1244 /* Make sure all updates are written before un-masking */
1246 nn_cfg_writeb(hw, NFP_NET_CFG_ICR(base + queue_id), 0x1);
1251 nfp_net_dev_link_status_print(struct rte_eth_dev *dev)
1253 struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1254 struct rte_eth_link link;
1256 memset(&link, 0, sizeof(link));
1257 nfp_net_dev_atomic_read_link_status(dev, &link);
1258 if (link.link_status)
1259 RTE_LOG(INFO, PMD, "Port %d: Link Up - speed %u Mbps - %s\n",
1260 (int)(dev->data->port_id), (unsigned)link.link_speed,
1261 link.link_duplex == ETH_LINK_FULL_DUPLEX
1262 ? "full-duplex" : "half-duplex");
1264 RTE_LOG(INFO, PMD, " Port %d: Link Down\n",
1265 (int)(dev->data->port_id));
1267 RTE_LOG(INFO, PMD, "PCI Address: %04d:%02d:%02d:%d\n",
1268 pci_dev->addr.domain, pci_dev->addr.bus,
1269 pci_dev->addr.devid, pci_dev->addr.function);
1272 /* Interrupt configuration and handling */
1275 * nfp_net_irq_unmask - Unmask an interrupt
1277 * If MSI-X auto-masking is enabled clear the mask bit, otherwise
1278 * clear the ICR for the entry.
1281 nfp_net_irq_unmask(struct rte_eth_dev *dev)
1283 struct nfp_net_hw *hw;
1284 struct rte_pci_device *pci_dev;
1286 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1287 pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1289 if (hw->ctrl & NFP_NET_CFG_CTRL_MSIXAUTO) {
1290 /* If MSI-X auto-masking is used, clear the entry */
1292 rte_intr_enable(&pci_dev->intr_handle);
1294 /* Make sure all updates are written before un-masking */
1296 nn_cfg_writeb(hw, NFP_NET_CFG_ICR(NFP_NET_IRQ_LSC_IDX),
1297 NFP_NET_CFG_ICR_UNMASKED);
1302 nfp_net_dev_interrupt_handler(void *param)
1305 struct rte_eth_link link;
1306 struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
1308 PMD_DRV_LOG(DEBUG, "We got a LSC interrupt!!!\n");
1310 /* get the link status */
1311 memset(&link, 0, sizeof(link));
1312 nfp_net_dev_atomic_read_link_status(dev, &link);
1314 nfp_net_link_update(dev, 0);
1317 if (!link.link_status) {
1318 /* handle it 1 sec later, wait it being stable */
1319 timeout = NFP_NET_LINK_UP_CHECK_TIMEOUT;
1320 /* likely to down */
1322 /* handle it 4 sec later, wait it being stable */
1323 timeout = NFP_NET_LINK_DOWN_CHECK_TIMEOUT;
1326 if (rte_eal_alarm_set(timeout * 1000,
1327 nfp_net_dev_interrupt_delayed_handler,
1329 RTE_LOG(ERR, PMD, "Error setting alarm");
1331 nfp_net_irq_unmask(dev);
1336 * Interrupt handler which shall be registered for alarm callback for delayed
1337 * handling specific interrupt to wait for the stable nic state. As the NIC
1338 * interrupt state is not stable for nfp after link is just down, it needs
1339 * to wait 4 seconds to get the stable status.
1341 * @param handle Pointer to interrupt handle.
1342 * @param param The address of parameter (struct rte_eth_dev *)
1347 nfp_net_dev_interrupt_delayed_handler(void *param)
1349 struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
1351 nfp_net_link_update(dev, 0);
1352 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL, NULL);
1354 nfp_net_dev_link_status_print(dev);
1357 nfp_net_irq_unmask(dev);
1361 nfp_net_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1363 struct nfp_net_hw *hw;
1365 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1367 /* check that mtu is within the allowed range */
1368 if ((mtu < ETHER_MIN_MTU) || ((uint32_t)mtu > hw->max_mtu))
1371 /* switch to jumbo mode if needed */
1372 if ((uint32_t)mtu > ETHER_MAX_LEN)
1373 dev->data->dev_conf.rxmode.jumbo_frame = 1;
1375 dev->data->dev_conf.rxmode.jumbo_frame = 0;
1377 /* update max frame size */
1378 dev->data->dev_conf.rxmode.max_rx_pkt_len = (uint32_t)mtu;
1380 /* writing to configuration space */
1381 nn_cfg_writel(hw, NFP_NET_CFG_MTU, (uint32_t)mtu);
1389 nfp_net_rx_queue_setup(struct rte_eth_dev *dev,
1390 uint16_t queue_idx, uint16_t nb_desc,
1391 unsigned int socket_id,
1392 const struct rte_eth_rxconf *rx_conf,
1393 struct rte_mempool *mp)
1395 const struct rte_memzone *tz;
1396 struct nfp_net_rxq *rxq;
1397 struct nfp_net_hw *hw;
1399 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1401 PMD_INIT_FUNC_TRACE();
1403 /* Validating number of descriptors */
1404 if (((nb_desc * sizeof(struct nfp_net_rx_desc)) % 128) != 0 ||
1405 (nb_desc > NFP_NET_MAX_RX_DESC) ||
1406 (nb_desc < NFP_NET_MIN_RX_DESC)) {
1407 RTE_LOG(ERR, PMD, "Wrong nb_desc value\n");
1412 * Free memory prior to re-allocation if needed. This is the case after
1413 * calling nfp_net_stop
1415 if (dev->data->rx_queues[queue_idx]) {
1416 nfp_net_rx_queue_release(dev->data->rx_queues[queue_idx]);
1417 dev->data->rx_queues[queue_idx] = NULL;
1420 /* Allocating rx queue data structure */
1421 rxq = rte_zmalloc_socket("ethdev RX queue", sizeof(struct nfp_net_rxq),
1422 RTE_CACHE_LINE_SIZE, socket_id);
1426 /* Hw queues mapping based on firmware confifguration */
1427 rxq->qidx = queue_idx;
1428 rxq->fl_qcidx = queue_idx * hw->stride_rx;
1429 rxq->rx_qcidx = rxq->fl_qcidx + (hw->stride_rx - 1);
1430 rxq->qcp_fl = hw->rx_bar + NFP_QCP_QUEUE_OFF(rxq->fl_qcidx);
1431 rxq->qcp_rx = hw->rx_bar + NFP_QCP_QUEUE_OFF(rxq->rx_qcidx);
1434 * Tracking mbuf size for detecting a potential mbuf overflow due to
1438 rxq->mbuf_size = rxq->mem_pool->elt_size;
1439 rxq->mbuf_size -= (sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM);
1440 hw->flbufsz = rxq->mbuf_size;
1442 rxq->rx_count = nb_desc;
1443 rxq->port_id = dev->data->port_id;
1444 rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1445 rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ? 0
1447 rxq->drop_en = rx_conf->rx_drop_en;
1450 * Allocate RX ring hardware descriptors. A memzone large enough to
1451 * handle the maximum ring size is allocated in order to allow for
1452 * resizing in later calls to the queue setup function.
1454 tz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1455 sizeof(struct nfp_net_rx_desc) *
1456 NFP_NET_MAX_RX_DESC, NFP_MEMZONE_ALIGN,
1460 RTE_LOG(ERR, PMD, "Error allocatig rx dma\n");
1461 nfp_net_rx_queue_release(rxq);
1465 /* Saving physical and virtual addresses for the RX ring */
1466 rxq->dma = (uint64_t)tz->phys_addr;
1467 rxq->rxds = (struct nfp_net_rx_desc *)tz->addr;
1469 /* mbuf pointers array for referencing mbufs linked to RX descriptors */
1470 rxq->rxbufs = rte_zmalloc_socket("rxq->rxbufs",
1471 sizeof(*rxq->rxbufs) * nb_desc,
1472 RTE_CACHE_LINE_SIZE, socket_id);
1473 if (rxq->rxbufs == NULL) {
1474 nfp_net_rx_queue_release(rxq);
1478 PMD_RX_LOG(DEBUG, "rxbufs=%p hw_ring=%p dma_addr=0x%" PRIx64 "\n",
1479 rxq->rxbufs, rxq->rxds, (unsigned long int)rxq->dma);
1481 nfp_net_reset_rx_queue(rxq);
1483 dev->data->rx_queues[queue_idx] = rxq;
1487 * Telling the HW about the physical address of the RX ring and number
1488 * of descriptors in log2 format
1490 nn_cfg_writeq(hw, NFP_NET_CFG_RXR_ADDR(queue_idx), rxq->dma);
1491 nn_cfg_writeb(hw, NFP_NET_CFG_RXR_SZ(queue_idx), rte_log2_u32(nb_desc));
1497 nfp_net_rx_fill_freelist(struct nfp_net_rxq *rxq)
1499 struct nfp_net_rx_buff *rxe = rxq->rxbufs;
1503 PMD_RX_LOG(DEBUG, "nfp_net_rx_fill_freelist for %u descriptors\n",
1506 for (i = 0; i < rxq->rx_count; i++) {
1507 struct nfp_net_rx_desc *rxd;
1508 struct rte_mbuf *mbuf = rte_pktmbuf_alloc(rxq->mem_pool);
1511 RTE_LOG(ERR, PMD, "RX mbuf alloc failed queue_id=%u\n",
1512 (unsigned)rxq->qidx);
1516 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DMA_ADDR_DEFAULT(mbuf));
1518 rxd = &rxq->rxds[i];
1520 rxd->fld.dma_addr_hi = (dma_addr >> 32) & 0xff;
1521 rxd->fld.dma_addr_lo = dma_addr & 0xffffffff;
1523 PMD_RX_LOG(DEBUG, "[%d]: %" PRIx64 "\n", i, dma_addr);
1526 /* Make sure all writes are flushed before telling the hardware */
1529 /* Not advertising the whole ring as the firmware gets confused if so */
1530 PMD_RX_LOG(DEBUG, "Increment FL write pointer in %u\n",
1533 nfp_qcp_ptr_add(rxq->qcp_fl, NFP_QCP_WRITE_PTR, rxq->rx_count - 1);
1539 nfp_net_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
1540 uint16_t nb_desc, unsigned int socket_id,
1541 const struct rte_eth_txconf *tx_conf)
1543 const struct rte_memzone *tz;
1544 struct nfp_net_txq *txq;
1545 uint16_t tx_free_thresh;
1546 struct nfp_net_hw *hw;
1548 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1550 PMD_INIT_FUNC_TRACE();
1552 /* Validating number of descriptors */
1553 if (((nb_desc * sizeof(struct nfp_net_tx_desc)) % 128) != 0 ||
1554 (nb_desc > NFP_NET_MAX_TX_DESC) ||
1555 (nb_desc < NFP_NET_MIN_TX_DESC)) {
1556 RTE_LOG(ERR, PMD, "Wrong nb_desc value\n");
1560 tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
1561 tx_conf->tx_free_thresh :
1562 DEFAULT_TX_FREE_THRESH);
1564 if (tx_free_thresh > (nb_desc)) {
1566 "tx_free_thresh must be less than the number of TX "
1567 "descriptors. (tx_free_thresh=%u port=%d "
1568 "queue=%d)\n", (unsigned int)tx_free_thresh,
1569 (int)dev->data->port_id, (int)queue_idx);
1574 * Free memory prior to re-allocation if needed. This is the case after
1575 * calling nfp_net_stop
1577 if (dev->data->tx_queues[queue_idx]) {
1578 PMD_TX_LOG(DEBUG, "Freeing memory prior to re-allocation %d\n",
1580 nfp_net_tx_queue_release(dev->data->tx_queues[queue_idx]);
1581 dev->data->tx_queues[queue_idx] = NULL;
1584 /* Allocating tx queue data structure */
1585 txq = rte_zmalloc_socket("ethdev TX queue", sizeof(struct nfp_net_txq),
1586 RTE_CACHE_LINE_SIZE, socket_id);
1588 RTE_LOG(ERR, PMD, "Error allocating tx dma\n");
1593 * Allocate TX ring hardware descriptors. A memzone large enough to
1594 * handle the maximum ring size is allocated in order to allow for
1595 * resizing in later calls to the queue setup function.
1597 tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
1598 sizeof(struct nfp_net_tx_desc) *
1599 NFP_NET_MAX_TX_DESC, NFP_MEMZONE_ALIGN,
1602 RTE_LOG(ERR, PMD, "Error allocating tx dma\n");
1603 nfp_net_tx_queue_release(txq);
1607 txq->tx_count = nb_desc;
1608 txq->tx_free_thresh = tx_free_thresh;
1609 txq->tx_pthresh = tx_conf->tx_thresh.pthresh;
1610 txq->tx_hthresh = tx_conf->tx_thresh.hthresh;
1611 txq->tx_wthresh = tx_conf->tx_thresh.wthresh;
1613 /* queue mapping based on firmware configuration */
1614 txq->qidx = queue_idx;
1615 txq->tx_qcidx = queue_idx * hw->stride_tx;
1616 txq->qcp_q = hw->tx_bar + NFP_QCP_QUEUE_OFF(txq->tx_qcidx);
1618 txq->port_id = dev->data->port_id;
1619 txq->txq_flags = tx_conf->txq_flags;
1621 /* Saving physical and virtual addresses for the TX ring */
1622 txq->dma = (uint64_t)tz->phys_addr;
1623 txq->txds = (struct nfp_net_tx_desc *)tz->addr;
1625 /* mbuf pointers array for referencing mbufs linked to TX descriptors */
1626 txq->txbufs = rte_zmalloc_socket("txq->txbufs",
1627 sizeof(*txq->txbufs) * nb_desc,
1628 RTE_CACHE_LINE_SIZE, socket_id);
1629 if (txq->txbufs == NULL) {
1630 nfp_net_tx_queue_release(txq);
1633 PMD_TX_LOG(DEBUG, "txbufs=%p hw_ring=%p dma_addr=0x%" PRIx64 "\n",
1634 txq->txbufs, txq->txds, (unsigned long int)txq->dma);
1636 nfp_net_reset_tx_queue(txq);
1638 dev->data->tx_queues[queue_idx] = txq;
1642 * Telling the HW about the physical address of the TX ring and number
1643 * of descriptors in log2 format
1645 nn_cfg_writeq(hw, NFP_NET_CFG_TXR_ADDR(queue_idx), txq->dma);
1646 nn_cfg_writeb(hw, NFP_NET_CFG_TXR_SZ(queue_idx), rte_log2_u32(nb_desc));
1651 /* nfp_net_tx_tso - Set TX descriptor for TSO */
1653 nfp_net_tx_tso(struct nfp_net_txq *txq, struct nfp_net_tx_desc *txd,
1654 struct rte_mbuf *mb)
1657 struct nfp_net_hw *hw = txq->hw;
1659 if (!(hw->cap & NFP_NET_CFG_CTRL_LSO))
1662 ol_flags = mb->ol_flags;
1664 if (!(ol_flags & PKT_TX_TCP_SEG))
1667 txd->l4_offset = mb->l2_len + mb->l3_len + mb->l4_len;
1668 txd->lso = rte_cpu_to_le_16(mb->tso_segsz);
1669 txd->flags = PCIE_DESC_TX_LSO;
1678 /* nfp_net_tx_cksum - Set TX CSUM offload flags in TX descriptor */
1680 nfp_net_tx_cksum(struct nfp_net_txq *txq, struct nfp_net_tx_desc *txd,
1681 struct rte_mbuf *mb)
1684 struct nfp_net_hw *hw = txq->hw;
1686 if (!(hw->cap & NFP_NET_CFG_CTRL_TXCSUM))
1689 ol_flags = mb->ol_flags;
1691 /* IPv6 does not need checksum */
1692 if (ol_flags & PKT_TX_IP_CKSUM)
1693 txd->flags |= PCIE_DESC_TX_IP4_CSUM;
1695 switch (ol_flags & PKT_TX_L4_MASK) {
1696 case PKT_TX_UDP_CKSUM:
1697 txd->flags |= PCIE_DESC_TX_UDP_CSUM;
1699 case PKT_TX_TCP_CKSUM:
1700 txd->flags |= PCIE_DESC_TX_TCP_CSUM;
1704 if (ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK))
1705 txd->flags |= PCIE_DESC_TX_CSUM;
1708 /* nfp_net_rx_cksum - set mbuf checksum flags based on RX descriptor flags */
1710 nfp_net_rx_cksum(struct nfp_net_rxq *rxq, struct nfp_net_rx_desc *rxd,
1711 struct rte_mbuf *mb)
1713 struct nfp_net_hw *hw = rxq->hw;
1715 if (!(hw->ctrl & NFP_NET_CFG_CTRL_RXCSUM))
1718 /* If IPv4 and IP checksum error, fail */
1719 if ((rxd->rxd.flags & PCIE_DESC_RX_IP4_CSUM) &&
1720 !(rxd->rxd.flags & PCIE_DESC_RX_IP4_CSUM_OK))
1721 mb->ol_flags |= PKT_RX_IP_CKSUM_BAD;
1723 /* If neither UDP nor TCP return */
1724 if (!(rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM) &&
1725 !(rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM))
1728 if ((rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM) &&
1729 !(rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK))
1730 mb->ol_flags |= PKT_RX_L4_CKSUM_BAD;
1732 if ((rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM) &&
1733 !(rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK))
1734 mb->ol_flags |= PKT_RX_L4_CKSUM_BAD;
1737 #define NFP_HASH_OFFSET ((uint8_t *)mbuf->buf_addr + mbuf->data_off - 4)
1738 #define NFP_HASH_TYPE_OFFSET ((uint8_t *)mbuf->buf_addr + mbuf->data_off - 8)
1740 #define NFP_DESC_META_LEN(d) (d->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK)
1743 * nfp_net_set_hash - Set mbuf hash data
1745 * The RSS hash and hash-type are pre-pended to the packet data.
1746 * Extract and decode it and set the mbuf fields.
1749 nfp_net_set_hash(struct nfp_net_rxq *rxq, struct nfp_net_rx_desc *rxd,
1750 struct rte_mbuf *mbuf)
1752 struct nfp_net_hw *hw = rxq->hw;
1753 uint8_t *meta_offset;
1756 uint32_t hash_type = 0;
1758 if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
1761 if (NFD_CFG_MAJOR_VERSION_of(hw->ver) <= 3) {
1762 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
1765 hash = rte_be_to_cpu_32(*(uint32_t *)NFP_HASH_OFFSET);
1766 hash_type = rte_be_to_cpu_32(*(uint32_t *)NFP_HASH_TYPE_OFFSET);
1768 } else if (NFP_DESC_META_LEN(rxd)) {
1771 * <---- 32 bit ----->
1776 * ====================
1779 * Field type word contains up to 8 4bit field types
1780 * A 4bit field type refers to a data field word
1781 * A data field word can have several 4bit field types
1783 meta_offset = rte_pktmbuf_mtod(mbuf, uint8_t *);
1784 meta_offset -= NFP_DESC_META_LEN(rxd);
1785 meta_info = rte_be_to_cpu_32(*(uint32_t *)meta_offset);
1787 /* NFP PMD just supports metadata for hashing */
1788 switch (meta_info & NFP_NET_META_FIELD_MASK) {
1789 case NFP_NET_META_HASH:
1790 /* next field type is about the hash type */
1791 meta_info >>= NFP_NET_META_FIELD_SIZE;
1792 /* hash value is in the data field */
1793 hash = rte_be_to_cpu_32(*(uint32_t *)meta_offset);
1794 hash_type = meta_info & NFP_NET_META_FIELD_MASK;
1797 /* Unsupported metadata can be a performance issue */
1804 mbuf->hash.rss = hash;
1805 mbuf->ol_flags |= PKT_RX_RSS_HASH;
1807 switch (hash_type) {
1808 case NFP_NET_RSS_IPV4:
1809 mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV4;
1811 case NFP_NET_RSS_IPV6:
1812 mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV6;
1814 case NFP_NET_RSS_IPV6_EX:
1815 mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV6_EXT;
1818 mbuf->packet_type |= RTE_PTYPE_INNER_L4_MASK;
1823 nfp_net_mbuf_alloc_failed(struct nfp_net_rxq *rxq)
1825 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
1828 #define NFP_DESC_META_LEN(d) (d->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK)
1833 * There are some decissions to take:
1834 * 1) How to check DD RX descriptors bit
1835 * 2) How and when to allocate new mbufs
1837 * Current implementation checks just one single DD bit each loop. As each
1838 * descriptor is 8 bytes, it is likely a good idea to check descriptors in
1839 * a single cache line instead. Tests with this change have not shown any
1840 * performance improvement but it requires further investigation. For example,
1841 * depending on which descriptor is next, the number of descriptors could be
1842 * less than 8 for just checking those in the same cache line. This implies
1843 * extra work which could be counterproductive by itself. Indeed, last firmware
1844 * changes are just doing this: writing several descriptors with the DD bit
1845 * for saving PCIe bandwidth and DMA operations from the NFP.
1847 * Mbuf allocation is done when a new packet is received. Then the descriptor
1848 * is automatically linked with the new mbuf and the old one is given to the
1849 * user. The main drawback with this design is mbuf allocation is heavier than
1850 * using bulk allocations allowed by DPDK with rte_mempool_get_bulk. From the
1851 * cache point of view it does not seem allocating the mbuf early on as we are
1852 * doing now have any benefit at all. Again, tests with this change have not
1853 * shown any improvement. Also, rte_mempool_get_bulk returns all or nothing
1854 * so looking at the implications of this type of allocation should be studied
1859 nfp_net_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1861 struct nfp_net_rxq *rxq;
1862 struct nfp_net_rx_desc *rxds;
1863 struct nfp_net_rx_buff *rxb;
1864 struct nfp_net_hw *hw;
1865 struct rte_mbuf *mb;
1866 struct rte_mbuf *new_mb;
1872 if (unlikely(rxq == NULL)) {
1874 * DPDK just checks the queue is lower than max queues
1875 * enabled. But the queue needs to be configured
1877 RTE_LOG_DP(ERR, PMD, "RX Bad queue\n");
1885 while (avail < nb_pkts) {
1886 rxb = &rxq->rxbufs[rxq->rd_p];
1887 if (unlikely(rxb == NULL)) {
1888 RTE_LOG_DP(ERR, PMD, "rxb does not exist!\n");
1893 * Memory barrier to ensure that we won't do other
1894 * reads before the DD bit.
1898 rxds = &rxq->rxds[rxq->rd_p];
1899 if ((rxds->rxd.meta_len_dd & PCIE_DESC_RX_DD) == 0)
1903 * We got a packet. Let's alloc a new mbuff for refilling the
1904 * free descriptor ring as soon as possible
1906 new_mb = rte_pktmbuf_alloc(rxq->mem_pool);
1907 if (unlikely(new_mb == NULL)) {
1908 RTE_LOG_DP(DEBUG, PMD, "RX mbuf alloc failed port_id=%u "
1909 "queue_id=%u\n", (unsigned)rxq->port_id,
1910 (unsigned)rxq->qidx);
1911 nfp_net_mbuf_alloc_failed(rxq);
1918 * Grab the mbuff and refill the descriptor with the
1919 * previously allocated mbuff
1924 PMD_RX_LOG(DEBUG, "Packet len: %u, mbuf_size: %u\n",
1925 rxds->rxd.data_len, rxq->mbuf_size);
1927 /* Size of this segment */
1928 mb->data_len = rxds->rxd.data_len - NFP_DESC_META_LEN(rxds);
1929 /* Size of the whole packet. We just support 1 segment */
1930 mb->pkt_len = rxds->rxd.data_len - NFP_DESC_META_LEN(rxds);
1932 if (unlikely((mb->data_len + hw->rx_offset) >
1935 * This should not happen and the user has the
1936 * responsibility of avoiding it. But we have
1937 * to give some info about the error
1939 RTE_LOG_DP(ERR, PMD,
1940 "mbuf overflow likely due to the RX offset.\n"
1941 "\t\tYour mbuf size should have extra space for"
1942 " RX offset=%u bytes.\n"
1943 "\t\tCurrently you just have %u bytes available"
1944 " but the received packet is %u bytes long",
1946 rxq->mbuf_size - hw->rx_offset,
1951 /* Filling the received mbuff with packet info */
1953 mb->data_off = RTE_PKTMBUF_HEADROOM + hw->rx_offset;
1955 mb->data_off = RTE_PKTMBUF_HEADROOM +
1956 NFP_DESC_META_LEN(rxds);
1958 /* No scatter mode supported */
1962 /* Checking the RSS flag */
1963 nfp_net_set_hash(rxq, rxds, mb);
1965 /* Checking the checksum flag */
1966 nfp_net_rx_cksum(rxq, rxds, mb);
1968 if ((rxds->rxd.flags & PCIE_DESC_RX_VLAN) &&
1969 (hw->ctrl & NFP_NET_CFG_CTRL_RXVLAN)) {
1970 mb->vlan_tci = rte_cpu_to_le_32(rxds->rxd.vlan);
1971 mb->ol_flags |= PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED;
1974 /* Adding the mbuff to the mbuff array passed by the app */
1975 rx_pkts[avail++] = mb;
1977 /* Now resetting and updating the descriptor */
1980 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DMA_ADDR_DEFAULT(new_mb));
1982 rxds->fld.dma_addr_hi = (dma_addr >> 32) & 0xff;
1983 rxds->fld.dma_addr_lo = dma_addr & 0xffffffff;
1986 if (unlikely(rxq->rd_p == rxq->rx_count)) /* wrapping?*/
1993 PMD_RX_LOG(DEBUG, "RX port_id=%u queue_id=%u, %d packets received\n",
1994 (unsigned)rxq->port_id, (unsigned)rxq->qidx, nb_hold);
1996 nb_hold += rxq->nb_rx_hold;
1999 * FL descriptors needs to be written before incrementing the
2000 * FL queue WR pointer
2003 if (nb_hold > rxq->rx_free_thresh) {
2004 PMD_RX_LOG(DEBUG, "port=%u queue=%u nb_hold=%u avail=%u\n",
2005 (unsigned)rxq->port_id, (unsigned)rxq->qidx,
2006 (unsigned)nb_hold, (unsigned)avail);
2007 nfp_qcp_ptr_add(rxq->qcp_fl, NFP_QCP_WRITE_PTR, nb_hold);
2010 rxq->nb_rx_hold = nb_hold;
2016 * nfp_net_tx_free_bufs - Check for descriptors with a complete
2018 * @txq: TX queue to work with
2019 * Returns number of descriptors freed
2022 nfp_net_tx_free_bufs(struct nfp_net_txq *txq)
2027 PMD_TX_LOG(DEBUG, "queue %u. Check for descriptor with a complete"
2028 " status\n", txq->qidx);
2030 /* Work out how many packets have been sent */
2031 qcp_rd_p = nfp_qcp_read(txq->qcp_q, NFP_QCP_READ_PTR);
2033 if (qcp_rd_p == txq->rd_p) {
2034 PMD_TX_LOG(DEBUG, "queue %u: It seems harrier is not sending "
2035 "packets (%u, %u)\n", txq->qidx,
2036 qcp_rd_p, txq->rd_p);
2040 if (qcp_rd_p > txq->rd_p)
2041 todo = qcp_rd_p - txq->rd_p;
2043 todo = qcp_rd_p + txq->tx_count - txq->rd_p;
2045 PMD_TX_LOG(DEBUG, "qcp_rd_p %u, txq->rd_p: %u, qcp->rd_p: %u\n",
2046 qcp_rd_p, txq->rd_p, txq->rd_p);
2052 if (unlikely(txq->rd_p >= txq->tx_count))
2053 txq->rd_p -= txq->tx_count;
2058 /* Leaving always free descriptors for avoiding wrapping confusion */
2060 uint32_t nfp_free_tx_desc(struct nfp_net_txq *txq)
2062 if (txq->wr_p >= txq->rd_p)
2063 return txq->tx_count - (txq->wr_p - txq->rd_p) - 8;
2065 return txq->rd_p - txq->wr_p - 8;
2069 * nfp_net_txq_full - Check if the TX queue free descriptors
2070 * is below tx_free_threshold
2072 * @txq: TX queue to check
2074 * This function uses the host copy* of read/write pointers
2077 uint32_t nfp_net_txq_full(struct nfp_net_txq *txq)
2079 return (nfp_free_tx_desc(txq) < txq->tx_free_thresh);
2083 nfp_net_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2085 struct nfp_net_txq *txq;
2086 struct nfp_net_hw *hw;
2087 struct nfp_net_tx_desc *txds, txd;
2088 struct rte_mbuf *pkt;
2090 int pkt_size, dma_size;
2091 uint16_t free_descs, issued_descs;
2092 struct rte_mbuf **lmbuf;
2097 txds = &txq->txds[txq->wr_p];
2099 PMD_TX_LOG(DEBUG, "working for queue %u at pos %d and %u packets\n",
2100 txq->qidx, txq->wr_p, nb_pkts);
2102 if ((nfp_free_tx_desc(txq) < nb_pkts) || (nfp_net_txq_full(txq)))
2103 nfp_net_tx_free_bufs(txq);
2105 free_descs = (uint16_t)nfp_free_tx_desc(txq);
2106 if (unlikely(free_descs == 0))
2113 PMD_TX_LOG(DEBUG, "queue: %u. Sending %u packets\n",
2114 txq->qidx, nb_pkts);
2115 /* Sending packets */
2116 while ((i < nb_pkts) && free_descs) {
2117 /* Grabbing the mbuf linked to the current descriptor */
2118 lmbuf = &txq->txbufs[txq->wr_p].mbuf;
2119 /* Warming the cache for releasing the mbuf later on */
2120 RTE_MBUF_PREFETCH_TO_FREE(*lmbuf);
2122 pkt = *(tx_pkts + i);
2124 if (unlikely((pkt->nb_segs > 1) &&
2125 !(hw->cap & NFP_NET_CFG_CTRL_GATHER))) {
2126 PMD_INIT_LOG(INFO, "NFP_NET_CFG_CTRL_GATHER not set");
2127 rte_panic("Multisegment packet unsupported\n");
2130 /* Checking if we have enough descriptors */
2131 if (unlikely(pkt->nb_segs > free_descs))
2135 * Checksum and VLAN flags just in the first descriptor for a
2136 * multisegment packet, but TSO info needs to be in all of them.
2138 txd.data_len = pkt->pkt_len;
2139 nfp_net_tx_tso(txq, &txd, pkt);
2140 nfp_net_tx_cksum(txq, &txd, pkt);
2142 if ((pkt->ol_flags & PKT_TX_VLAN_PKT) &&
2143 (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)) {
2144 txd.flags |= PCIE_DESC_TX_VLAN;
2145 txd.vlan = pkt->vlan_tci;
2149 * mbuf data_len is the data in one segment and pkt_len data
2150 * in the whole packet. When the packet is just one segment,
2151 * then data_len = pkt_len
2153 pkt_size = pkt->pkt_len;
2156 /* Copying TSO, VLAN and cksum info */
2159 /* Releasing mbuf used by this descriptor previously*/
2161 rte_pktmbuf_free_seg(*lmbuf);
2164 * Linking mbuf with descriptor for being released
2165 * next time descriptor is used
2169 dma_size = pkt->data_len;
2170 dma_addr = rte_mbuf_data_dma_addr(pkt);
2171 PMD_TX_LOG(DEBUG, "Working with mbuf at dma address:"
2172 "%" PRIx64 "\n", dma_addr);
2174 /* Filling descriptors fields */
2175 txds->dma_len = dma_size;
2176 txds->data_len = txd.data_len;
2177 txds->dma_addr_hi = (dma_addr >> 32) & 0xff;
2178 txds->dma_addr_lo = (dma_addr & 0xffffffff);
2179 ASSERT(free_descs > 0);
2183 if (unlikely(txq->wr_p == txq->tx_count)) /* wrapping?*/
2186 pkt_size -= dma_size;
2189 txds->offset_eop |= PCIE_DESC_TX_EOP;
2191 txds->offset_eop &= PCIE_DESC_TX_OFFSET_MASK;
2194 /* Referencing next free TX descriptor */
2195 txds = &txq->txds[txq->wr_p];
2196 lmbuf = &txq->txbufs[txq->wr_p].mbuf;
2203 /* Increment write pointers. Force memory write before we let HW know */
2205 nfp_qcp_ptr_add(txq->qcp_q, NFP_QCP_WRITE_PTR, issued_descs);
2211 nfp_net_vlan_offload_set(struct rte_eth_dev *dev, int mask)
2213 uint32_t new_ctrl, update;
2214 struct nfp_net_hw *hw;
2216 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2219 if ((mask & ETH_VLAN_FILTER_OFFLOAD) ||
2220 (mask & ETH_VLAN_EXTEND_OFFLOAD))
2221 RTE_LOG(INFO, PMD, "No support for ETH_VLAN_FILTER_OFFLOAD or"
2222 " ETH_VLAN_EXTEND_OFFLOAD");
2224 /* Enable vlan strip if it is not configured yet */
2225 if ((mask & ETH_VLAN_STRIP_OFFLOAD) &&
2226 !(hw->ctrl & NFP_NET_CFG_CTRL_RXVLAN))
2227 new_ctrl = hw->ctrl | NFP_NET_CFG_CTRL_RXVLAN;
2229 /* Disable vlan strip just if it is configured */
2230 if (!(mask & ETH_VLAN_STRIP_OFFLOAD) &&
2231 (hw->ctrl & NFP_NET_CFG_CTRL_RXVLAN))
2232 new_ctrl = hw->ctrl & ~NFP_NET_CFG_CTRL_RXVLAN;
2237 update = NFP_NET_CFG_UPDATE_GEN;
2239 if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
2242 hw->ctrl = new_ctrl;
2245 /* Update Redirection Table(RETA) of Receive Side Scaling of Ethernet device */
2247 nfp_net_reta_update(struct rte_eth_dev *dev,
2248 struct rte_eth_rss_reta_entry64 *reta_conf,
2251 uint32_t reta, mask;
2255 struct nfp_net_hw *hw =
2256 NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2258 if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
2261 if (reta_size != NFP_NET_CFG_RSS_ITBL_SZ) {
2262 RTE_LOG(ERR, PMD, "The size of hash lookup table configured "
2263 "(%d) doesn't match the number hardware can supported "
2264 "(%d)\n", reta_size, NFP_NET_CFG_RSS_ITBL_SZ);
2269 * Update Redirection Table. There are 128 8bit-entries which can be
2270 * manage as 32 32bit-entries
2272 for (i = 0; i < reta_size; i += 4) {
2273 /* Handling 4 RSS entries per loop */
2274 idx = i / RTE_RETA_GROUP_SIZE;
2275 shift = i % RTE_RETA_GROUP_SIZE;
2276 mask = (uint8_t)((reta_conf[idx].mask >> shift) & 0xF);
2282 /* If all 4 entries were set, don't need read RETA register */
2284 reta = nn_cfg_readl(hw, NFP_NET_CFG_RSS_ITBL + i);
2286 for (j = 0; j < 4; j++) {
2287 if (!(mask & (0x1 << j)))
2290 /* Clearing the entry bits */
2291 reta &= ~(0xFF << (8 * j));
2292 reta |= reta_conf[idx].reta[shift + j] << (8 * j);
2294 nn_cfg_writel(hw, NFP_NET_CFG_RSS_ITBL + (idx * 64) + shift,
2298 update = NFP_NET_CFG_UPDATE_RSS;
2300 if (nfp_net_reconfig(hw, hw->ctrl, update) < 0)
2306 /* Query Redirection Table(RETA) of Receive Side Scaling of Ethernet device. */
2308 nfp_net_reta_query(struct rte_eth_dev *dev,
2309 struct rte_eth_rss_reta_entry64 *reta_conf,
2315 struct nfp_net_hw *hw;
2317 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2319 if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
2322 if (reta_size != NFP_NET_CFG_RSS_ITBL_SZ) {
2323 RTE_LOG(ERR, PMD, "The size of hash lookup table configured "
2324 "(%d) doesn't match the number hardware can supported "
2325 "(%d)\n", reta_size, NFP_NET_CFG_RSS_ITBL_SZ);
2330 * Reading Redirection Table. There are 128 8bit-entries which can be
2331 * manage as 32 32bit-entries
2333 for (i = 0; i < reta_size; i += 4) {
2334 /* Handling 4 RSS entries per loop */
2335 idx = i / RTE_RETA_GROUP_SIZE;
2336 shift = i % RTE_RETA_GROUP_SIZE;
2337 mask = (uint8_t)((reta_conf[idx].mask >> shift) & 0xF);
2342 reta = nn_cfg_readl(hw, NFP_NET_CFG_RSS_ITBL + (idx * 64) +
2344 for (j = 0; j < 4; j++) {
2345 if (!(mask & (0x1 << j)))
2347 reta_conf->reta[shift + j] =
2348 (uint8_t)((reta >> (8 * j)) & 0xF);
2355 nfp_net_rss_hash_update(struct rte_eth_dev *dev,
2356 struct rte_eth_rss_conf *rss_conf)
2359 uint32_t cfg_rss_ctrl = 0;
2363 struct nfp_net_hw *hw;
2365 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2367 rss_hf = rss_conf->rss_hf;
2369 /* Checking if RSS is enabled */
2370 if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS)) {
2371 if (rss_hf != 0) { /* Enable RSS? */
2372 RTE_LOG(ERR, PMD, "RSS unsupported\n");
2375 return 0; /* Nothing to do */
2378 if (rss_conf->rss_key_len > NFP_NET_CFG_RSS_KEY_SZ) {
2379 RTE_LOG(ERR, PMD, "hash key too long\n");
2383 if (rss_hf & ETH_RSS_IPV4)
2384 cfg_rss_ctrl |= NFP_NET_CFG_RSS_IPV4 |
2385 NFP_NET_CFG_RSS_IPV4_TCP |
2386 NFP_NET_CFG_RSS_IPV4_UDP;
2388 if (rss_hf & ETH_RSS_IPV6)
2389 cfg_rss_ctrl |= NFP_NET_CFG_RSS_IPV6 |
2390 NFP_NET_CFG_RSS_IPV6_TCP |
2391 NFP_NET_CFG_RSS_IPV6_UDP;
2393 cfg_rss_ctrl |= NFP_NET_CFG_RSS_MASK;
2394 cfg_rss_ctrl |= NFP_NET_CFG_RSS_TOEPLITZ;
2396 /* configuring where to apply the RSS hash */
2397 nn_cfg_writel(hw, NFP_NET_CFG_RSS_CTRL, cfg_rss_ctrl);
2399 /* Writing the key byte a byte */
2400 for (i = 0; i < rss_conf->rss_key_len; i++) {
2401 memcpy(&key, &rss_conf->rss_key[i], 1);
2402 nn_cfg_writeb(hw, NFP_NET_CFG_RSS_KEY + i, key);
2405 /* Writing the key size */
2406 nn_cfg_writeb(hw, NFP_NET_CFG_RSS_KEY_SZ, rss_conf->rss_key_len);
2408 update = NFP_NET_CFG_UPDATE_RSS;
2410 if (nfp_net_reconfig(hw, hw->ctrl, update) < 0)
2417 nfp_net_rss_hash_conf_get(struct rte_eth_dev *dev,
2418 struct rte_eth_rss_conf *rss_conf)
2421 uint32_t cfg_rss_ctrl;
2424 struct nfp_net_hw *hw;
2426 hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2428 if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
2431 rss_hf = rss_conf->rss_hf;
2432 cfg_rss_ctrl = nn_cfg_readl(hw, NFP_NET_CFG_RSS_CTRL);
2434 if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV4)
2435 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP | ETH_RSS_NONFRAG_IPV4_UDP;
2437 if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV4_TCP)
2438 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
2440 if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV6_TCP)
2441 rss_hf |= ETH_RSS_NONFRAG_IPV6_TCP;
2443 if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV4_UDP)
2444 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP;
2446 if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV6_UDP)
2447 rss_hf |= ETH_RSS_NONFRAG_IPV6_UDP;
2449 if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV6)
2450 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP | ETH_RSS_NONFRAG_IPV6_UDP;
2452 /* Reading the key size */
2453 rss_conf->rss_key_len = nn_cfg_readl(hw, NFP_NET_CFG_RSS_KEY_SZ);
2455 /* Reading the key byte a byte */
2456 for (i = 0; i < rss_conf->rss_key_len; i++) {
2457 key = nn_cfg_readb(hw, NFP_NET_CFG_RSS_KEY + i);
2458 memcpy(&rss_conf->rss_key[i], &key, 1);
2464 /* Initialise and register driver with DPDK Application */
2465 static const struct eth_dev_ops nfp_net_eth_dev_ops = {
2466 .dev_configure = nfp_net_configure,
2467 .dev_start = nfp_net_start,
2468 .dev_stop = nfp_net_stop,
2469 .dev_close = nfp_net_close,
2470 .promiscuous_enable = nfp_net_promisc_enable,
2471 .promiscuous_disable = nfp_net_promisc_disable,
2472 .link_update = nfp_net_link_update,
2473 .stats_get = nfp_net_stats_get,
2474 .stats_reset = nfp_net_stats_reset,
2475 .dev_infos_get = nfp_net_infos_get,
2476 .dev_supported_ptypes_get = nfp_net_supported_ptypes_get,
2477 .mtu_set = nfp_net_dev_mtu_set,
2478 .vlan_offload_set = nfp_net_vlan_offload_set,
2479 .reta_update = nfp_net_reta_update,
2480 .reta_query = nfp_net_reta_query,
2481 .rss_hash_update = nfp_net_rss_hash_update,
2482 .rss_hash_conf_get = nfp_net_rss_hash_conf_get,
2483 .rx_queue_setup = nfp_net_rx_queue_setup,
2484 .rx_queue_release = nfp_net_rx_queue_release,
2485 .rx_queue_count = nfp_net_rx_queue_count,
2486 .tx_queue_setup = nfp_net_tx_queue_setup,
2487 .tx_queue_release = nfp_net_tx_queue_release,
2488 .rx_queue_intr_enable = nfp_rx_queue_intr_enable,
2489 .rx_queue_intr_disable = nfp_rx_queue_intr_disable,
2493 nfp_net_init(struct rte_eth_dev *eth_dev)
2495 struct rte_pci_device *pci_dev;
2496 struct nfp_net_hw *hw;
2498 uint64_t tx_bar_off = 0, rx_bar_off = 0;
2502 nspu_desc_t *nspu_desc = NULL;
2503 uint64_t bar_offset;
2505 PMD_INIT_FUNC_TRACE();
2507 hw = NFP_NET_DEV_PRIVATE_TO_HW(eth_dev->data->dev_private);
2509 eth_dev->dev_ops = &nfp_net_eth_dev_ops;
2510 eth_dev->rx_pkt_burst = &nfp_net_recv_pkts;
2511 eth_dev->tx_pkt_burst = &nfp_net_xmit_pkts;
2513 /* For secondary processes, the primary has done all the work */
2514 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
2517 pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
2518 rte_eth_copy_pci_info(eth_dev, pci_dev);
2519 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
2521 hw->device_id = pci_dev->id.device_id;
2522 hw->vendor_id = pci_dev->id.vendor_id;
2523 hw->subsystem_device_id = pci_dev->id.subsystem_device_id;
2524 hw->subsystem_vendor_id = pci_dev->id.subsystem_vendor_id;
2526 PMD_INIT_LOG(DEBUG, "nfp_net: device (%u:%u) %u:%u:%u:%u",
2527 pci_dev->id.vendor_id, pci_dev->id.device_id,
2528 pci_dev->addr.domain, pci_dev->addr.bus,
2529 pci_dev->addr.devid, pci_dev->addr.function);
2531 hw->ctrl_bar = (uint8_t *)pci_dev->mem_resource[0].addr;
2532 if (hw->ctrl_bar == NULL) {
2534 "hw->ctrl_bar is NULL. BAR0 not configured\n");
2538 /* Is this a PF device? */
2539 if ((pci_dev->id.device_id == PCI_DEVICE_ID_NFP4000_PF_NIC) ||
2540 (pci_dev->id.device_id == PCI_DEVICE_ID_NFP6000_PF_NIC)) {
2541 nspu_desc = hw->nspu_desc;
2543 if (nfp_nsp_map_ctrl_bar(nspu_desc, &bar_offset) != 0) {
2545 * A firmware should be there after PF probe so this
2546 * should not happen.
2548 RTE_LOG(ERR, PMD, "PF BAR symbol resolution failed\n");
2552 /* vNIC PF control BAR is a subset of PF PCI device BAR */
2553 hw->ctrl_bar += bar_offset;
2554 PMD_INIT_LOG(DEBUG, "ctrl bar: %p\n", hw->ctrl_bar);
2557 hw->max_rx_queues = nn_cfg_readl(hw, NFP_NET_CFG_MAX_RXRINGS);
2558 hw->max_tx_queues = nn_cfg_readl(hw, NFP_NET_CFG_MAX_TXRINGS);
2560 /* Work out where in the BAR the queues start. */
2561 switch (pci_dev->id.device_id) {
2562 case PCI_DEVICE_ID_NFP4000_PF_NIC:
2563 case PCI_DEVICE_ID_NFP6000_PF_NIC:
2564 case PCI_DEVICE_ID_NFP6000_VF_NIC:
2565 start_q = nn_cfg_readl(hw, NFP_NET_CFG_START_TXQ);
2566 tx_bar_off = NFP_PCIE_QUEUE(start_q);
2567 start_q = nn_cfg_readl(hw, NFP_NET_CFG_START_RXQ);
2568 rx_bar_off = NFP_PCIE_QUEUE(start_q);
2571 RTE_LOG(ERR, PMD, "nfp_net: no device ID matching\n");
2575 PMD_INIT_LOG(DEBUG, "tx_bar_off: 0x%" PRIx64 "\n", tx_bar_off);
2576 PMD_INIT_LOG(DEBUG, "rx_bar_off: 0x%" PRIx64 "\n", rx_bar_off);
2578 if ((pci_dev->id.device_id == PCI_DEVICE_ID_NFP4000_PF_NIC) ||
2579 (pci_dev->id.device_id == PCI_DEVICE_ID_NFP6000_PF_NIC)) {
2580 /* configure access to tx/rx vNIC BARs */
2581 nfp_nsp_map_queues_bar(nspu_desc, &bar_offset);
2582 PMD_INIT_LOG(DEBUG, "tx/rx bar_offset: %" PRIx64 "\n",
2584 hw->hw_queues = (uint8_t *)pci_dev->mem_resource[0].addr;
2586 /* vNIC PF tx/rx BARs are a subset of PF PCI device */
2587 hw->hw_queues += bar_offset;
2588 hw->tx_bar = hw->hw_queues + tx_bar_off;
2589 hw->rx_bar = hw->hw_queues + rx_bar_off;
2591 hw->tx_bar = (uint8_t *)pci_dev->mem_resource[2].addr +
2593 hw->rx_bar = (uint8_t *)pci_dev->mem_resource[2].addr +
2597 PMD_INIT_LOG(DEBUG, "ctrl_bar: %p, tx_bar: %p, rx_bar: %p",
2598 hw->ctrl_bar, hw->tx_bar, hw->rx_bar);
2600 nfp_net_cfg_queue_setup(hw);
2602 /* Get some of the read-only fields from the config BAR */
2603 hw->ver = nn_cfg_readl(hw, NFP_NET_CFG_VERSION);
2604 hw->cap = nn_cfg_readl(hw, NFP_NET_CFG_CAP);
2605 hw->max_mtu = nn_cfg_readl(hw, NFP_NET_CFG_MAX_MTU);
2606 hw->mtu = hw->max_mtu;
2608 if (NFD_CFG_MAJOR_VERSION_of(hw->ver) < 2)
2609 hw->rx_offset = NFP_NET_RX_OFFSET;
2611 hw->rx_offset = nn_cfg_readl(hw, NFP_NET_CFG_RX_OFFSET_ADDR);
2613 PMD_INIT_LOG(INFO, "VER: %#x, Maximum supported MTU: %d",
2614 hw->ver, hw->max_mtu);
2615 PMD_INIT_LOG(INFO, "CAP: %#x, %s%s%s%s%s%s%s%s%s", hw->cap,
2616 hw->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "",
2617 hw->cap & NFP_NET_CFG_CTRL_RXCSUM ? "RXCSUM " : "",
2618 hw->cap & NFP_NET_CFG_CTRL_TXCSUM ? "TXCSUM " : "",
2619 hw->cap & NFP_NET_CFG_CTRL_RXVLAN ? "RXVLAN " : "",
2620 hw->cap & NFP_NET_CFG_CTRL_TXVLAN ? "TXVLAN " : "",
2621 hw->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "",
2622 hw->cap & NFP_NET_CFG_CTRL_GATHER ? "GATHER " : "",
2623 hw->cap & NFP_NET_CFG_CTRL_LSO ? "TSO " : "",
2624 hw->cap & NFP_NET_CFG_CTRL_RSS ? "RSS " : "");
2628 hw->stride_rx = stride;
2629 hw->stride_tx = stride;
2631 PMD_INIT_LOG(INFO, "max_rx_queues: %u, max_tx_queues: %u",
2632 hw->max_rx_queues, hw->max_tx_queues);
2634 /* Initializing spinlock for reconfigs */
2635 rte_spinlock_init(&hw->reconfig_lock);
2637 /* Allocating memory for mac addr */
2638 eth_dev->data->mac_addrs = rte_zmalloc("mac_addr", ETHER_ADDR_LEN, 0);
2639 if (eth_dev->data->mac_addrs == NULL) {
2640 PMD_INIT_LOG(ERR, "Failed to space for MAC address");
2644 nfp_net_read_mac(hw);
2646 if (!is_valid_assigned_ether_addr((struct ether_addr *)&hw->mac_addr)) {
2647 /* Using random mac addresses for VFs */
2648 eth_random_addr(&hw->mac_addr[0]);
2649 nfp_net_write_mac(hw, (uint8_t *)&hw->mac_addr);
2652 /* Copying mac address to DPDK eth_dev struct */
2653 ether_addr_copy((struct ether_addr *)hw->mac_addr,
2654 ð_dev->data->mac_addrs[0]);
2656 PMD_INIT_LOG(INFO, "port %d VendorID=0x%x DeviceID=0x%x "
2657 "mac=%02x:%02x:%02x:%02x:%02x:%02x",
2658 eth_dev->data->port_id, pci_dev->id.vendor_id,
2659 pci_dev->id.device_id,
2660 hw->mac_addr[0], hw->mac_addr[1], hw->mac_addr[2],
2661 hw->mac_addr[3], hw->mac_addr[4], hw->mac_addr[5]);
2663 /* Registering LSC interrupt handler */
2664 rte_intr_callback_register(&pci_dev->intr_handle,
2665 nfp_net_dev_interrupt_handler,
2668 /* Telling the firmware about the LSC interrupt entry */
2669 nn_cfg_writeb(hw, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2671 /* Recording current stats counters values */
2672 nfp_net_stats_reset(eth_dev);
2677 static int nfp_pf_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2678 struct rte_pci_device *dev)
2680 struct rte_eth_dev *eth_dev;
2681 struct nfp_net_hw *hw;
2682 nfpu_desc_t *nfpu_desc;
2683 nspu_desc_t *nspu_desc;
2684 uint64_t offset_symbol;
2691 nfpu_desc = rte_malloc("nfp nfpu", sizeof(nfpu_desc_t), 0);
2695 if (nfpu_open(dev, nfpu_desc, 0) < 0) {
2697 "nfpu_open failed\n");
2701 nspu_desc = nfpu_desc->nspu;
2704 /* Check NSP ABI version */
2705 if (nfp_nsp_get_abi_version(nspu_desc, &major, &minor) < 0) {
2706 RTE_LOG(INFO, PMD, "NFP NSP not present\n");
2709 PMD_INIT_LOG(INFO, "nspu ABI version: %d.%d\n", major, minor);
2711 if ((major == 0) && (minor < 20)) {
2712 RTE_LOG(INFO, PMD, "NFP NSP ABI version too old. Required 0.20 or higher\n");
2716 ret = nfp_nsp_fw_setup(nspu_desc, "nfd_cfg_pf0_num_ports",
2721 eth_dev = rte_eth_dev_allocate(dev->device.name);
2727 eth_dev->data->dev_private = rte_zmalloc("nfp_pf_port",
2728 sizeof(struct nfp_net_adapter),
2729 RTE_CACHE_LINE_SIZE);
2730 if (!eth_dev->data->dev_private) {
2731 rte_eth_dev_release_port(eth_dev);
2736 hw = (struct nfp_net_hw *)(eth_dev->data->dev_private);
2737 hw->nspu_desc = nspu_desc;
2738 hw->nfpu_desc = nfpu_desc;
2741 eth_dev->device = &dev->device;
2742 rte_eth_copy_pci_info(eth_dev, dev);
2744 ret = nfp_net_init(eth_dev);
2749 /* something went wrong */
2750 rte_eth_dev_release_port(eth_dev);
2753 nfpu_close(nfpu_desc);
2755 rte_free(nfpu_desc);
2760 static const struct rte_pci_id pci_id_nfp_pf_net_map[] = {
2762 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETRONOME,
2763 PCI_DEVICE_ID_NFP4000_PF_NIC)
2766 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETRONOME,
2767 PCI_DEVICE_ID_NFP6000_PF_NIC)
2774 static const struct rte_pci_id pci_id_nfp_vf_net_map[] = {
2776 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETRONOME,
2777 PCI_DEVICE_ID_NFP6000_VF_NIC)
2784 static int eth_nfp_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2785 struct rte_pci_device *pci_dev)
2787 return rte_eth_dev_pci_generic_probe(pci_dev,
2788 sizeof(struct nfp_net_adapter), nfp_net_init);
2791 static int eth_nfp_pci_remove(struct rte_pci_device *pci_dev)
2793 return rte_eth_dev_pci_generic_remove(pci_dev, NULL);
2796 static struct rte_pci_driver rte_nfp_net_pf_pmd = {
2797 .id_table = pci_id_nfp_pf_net_map,
2798 .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
2799 .probe = nfp_pf_pci_probe,
2800 .remove = eth_nfp_pci_remove,
2803 static struct rte_pci_driver rte_nfp_net_vf_pmd = {
2804 .id_table = pci_id_nfp_vf_net_map,
2805 .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
2806 .probe = eth_nfp_pci_probe,
2807 .remove = eth_nfp_pci_remove,
2810 RTE_PMD_REGISTER_PCI(net_nfp_pf, rte_nfp_net_pf_pmd);
2811 RTE_PMD_REGISTER_PCI(net_nfp_vf, rte_nfp_net_vf_pmd);
2812 RTE_PMD_REGISTER_PCI_TABLE(net_nfp_pf, pci_id_nfp_pf_net_map);
2813 RTE_PMD_REGISTER_PCI_TABLE(net_nfp_vf, pci_id_nfp_vf_net_map);
2814 RTE_PMD_REGISTER_KMOD_DEP(net_nfp_pf, "* igb_uio | uio_pci_generic | vfio");
2815 RTE_PMD_REGISTER_KMOD_DEP(net_nfp_vf, "* igb_uio | uio_pci_generic | vfio");
2819 * c-file-style: "Linux"
2820 * indent-tabs-mode: t