net/nfp: support PF multiport
[dpdk.git] / drivers / net / nfp / nfp_net.c
1 /*
2  * Copyright (c) 2014, 2015 Netronome Systems, Inc.
3  * All rights reserved.
4  *
5  * Small portions derived from code Copyright(c) 2010-2015 Intel Corporation.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions are met:
9  *
10  * 1. Redistributions of source code must retain the above copyright notice,
11  *  this list of conditions and the following disclaimer.
12  *
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
16  *
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.
20  *
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.
32  */
33
34 /*
35  * vim:shiftwidth=8:noexpandtab
36  *
37  * @file dpdk/pmd/nfp_net.c
38  *
39  * Netronome vNIC DPDK Poll-Mode Driver: Main entry point
40  */
41
42 #include <rte_byteorder.h>
43 #include <rte_common.h>
44 #include <rte_log.h>
45 #include <rte_debug.h>
46 #include <rte_ethdev.h>
47 #include <rte_ethdev_pci.h>
48 #include <rte_dev.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>
57
58 #include "nfp_nfpu.h"
59 #include "nfp_net_pmd.h"
60 #include "nfp_net_logs.h"
61 #include "nfp_net_ctrl.h"
62
63 /* Prototypes */
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,
77                                        uint16_t queue_idx);
78 static uint16_t nfp_net_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
79                                   uint16_t nb_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,
96                                   uint16_t nb_pkts);
97
98 /*
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.
102  */
103 #define NFP_PCIE_QUEUE(_q)      (0x800 * ((_q) & 0xff))
104
105 /* Maximum value which can be added to a queue with one transaction */
106 #define NFP_QCP_MAX_ADD 0x7f
107
108 #define RTE_MBUF_DMA_ADDR_DEFAULT(mb) \
109         (uint64_t)((mb)->buf_physaddr + RTE_PKTMBUF_HEADROOM)
110
111 /* nfp_qcp_ptr - Read or Write Pointer of a queue */
112 enum nfp_qcp_ptr {
113         NFP_QCP_READ_PTR = 0,
114         NFP_QCP_WRITE_PTR
115 };
116
117 /*
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
122  *
123  * If @val is greater than @NFP_QCP_MAX_ADD multiple writes are performed.
124  */
125 static inline void
126 nfp_qcp_ptr_add(uint8_t *q, enum nfp_qcp_ptr ptr, uint32_t val)
127 {
128         uint32_t off;
129
130         if (ptr == NFP_QCP_READ_PTR)
131                 off = NFP_QCP_QUEUE_ADD_RPTR;
132         else
133                 off = NFP_QCP_QUEUE_ADD_WPTR;
134
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;
138         }
139
140         nn_writel(rte_cpu_to_le_32(val), q + off);
141 }
142
143 /*
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
147  */
148 static inline uint32_t
149 nfp_qcp_read(uint8_t *q, enum nfp_qcp_ptr ptr)
150 {
151         uint32_t off;
152         uint32_t val;
153
154         if (ptr == NFP_QCP_READ_PTR)
155                 off = NFP_QCP_QUEUE_STS_LO;
156         else
157                 off = NFP_QCP_QUEUE_STS_HI;
158
159         val = rte_cpu_to_le_32(nn_readl(q + off));
160
161         if (ptr == NFP_QCP_READ_PTR)
162                 return val & NFP_QCP_QUEUE_STS_LO_READPTR_mask;
163         else
164                 return val & NFP_QCP_QUEUE_STS_HI_WRITEPTR_mask;
165 }
166
167 /*
168  * Functions to read/write from/to Config BAR
169  * Performs any endian conversion necessary.
170  */
171 static inline uint8_t
172 nn_cfg_readb(struct nfp_net_hw *hw, int off)
173 {
174         return nn_readb(hw->ctrl_bar + off);
175 }
176
177 static inline void
178 nn_cfg_writeb(struct nfp_net_hw *hw, int off, uint8_t val)
179 {
180         nn_writeb(val, hw->ctrl_bar + off);
181 }
182
183 static inline uint32_t
184 nn_cfg_readl(struct nfp_net_hw *hw, int off)
185 {
186         return rte_le_to_cpu_32(nn_readl(hw->ctrl_bar + off));
187 }
188
189 static inline void
190 nn_cfg_writel(struct nfp_net_hw *hw, int off, uint32_t val)
191 {
192         nn_writel(rte_cpu_to_le_32(val), hw->ctrl_bar + off);
193 }
194
195 static inline uint64_t
196 nn_cfg_readq(struct nfp_net_hw *hw, int off)
197 {
198         return rte_le_to_cpu_64(nn_readq(hw->ctrl_bar + off));
199 }
200
201 static inline void
202 nn_cfg_writeq(struct nfp_net_hw *hw, int off, uint64_t val)
203 {
204         nn_writeq(rte_cpu_to_le_64(val), hw->ctrl_bar + off);
205 }
206
207 /*
208  * Atomically reads link status information from global structure rte_eth_dev.
209  *
210  * @param dev
211  *   - Pointer to the structure rte_eth_dev to read from.
212  *   - Pointer to the buffer to be saved with the link status.
213  *
214  * @return
215  *   - On success, zero.
216  *   - On failure, negative value.
217  */
218 static inline int
219 nfp_net_dev_atomic_read_link_status(struct rte_eth_dev *dev,
220                                     struct rte_eth_link *link)
221 {
222         struct rte_eth_link *dst = link;
223         struct rte_eth_link *src = &dev->data->dev_link;
224
225         if (rte_atomic64_cmpset((uint64_t *)dst, *(uint64_t *)dst,
226                                 *(uint64_t *)src) == 0)
227                 return -1;
228
229         return 0;
230 }
231
232 /*
233  * Atomically writes the link status information into global
234  * structure rte_eth_dev.
235  *
236  * @param dev
237  *   - Pointer to the structure rte_eth_dev to read from.
238  *   - Pointer to the buffer to be saved with the link status.
239  *
240  * @return
241  *   - On success, zero.
242  *   - On failure, negative value.
243  */
244 static inline int
245 nfp_net_dev_atomic_write_link_status(struct rte_eth_dev *dev,
246                                      struct rte_eth_link *link)
247 {
248         struct rte_eth_link *dst = &dev->data->dev_link;
249         struct rte_eth_link *src = link;
250
251         if (rte_atomic64_cmpset((uint64_t *)dst, *(uint64_t *)dst,
252                                 *(uint64_t *)src) == 0)
253                 return -1;
254
255         return 0;
256 }
257
258 static void
259 nfp_net_rx_queue_release_mbufs(struct nfp_net_rxq *rxq)
260 {
261         unsigned i;
262
263         if (rxq->rxbufs == NULL)
264                 return;
265
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;
270                 }
271         }
272 }
273
274 static void
275 nfp_net_rx_queue_release(void *rx_queue)
276 {
277         struct nfp_net_rxq *rxq = rx_queue;
278
279         if (rxq) {
280                 nfp_net_rx_queue_release_mbufs(rxq);
281                 rte_free(rxq->rxbufs);
282                 rte_free(rxq);
283         }
284 }
285
286 static void
287 nfp_net_reset_rx_queue(struct nfp_net_rxq *rxq)
288 {
289         nfp_net_rx_queue_release_mbufs(rxq);
290         rxq->rd_p = 0;
291         rxq->nb_rx_hold = 0;
292 }
293
294 static void
295 nfp_net_tx_queue_release_mbufs(struct nfp_net_txq *txq)
296 {
297         unsigned i;
298
299         if (txq->txbufs == NULL)
300                 return;
301
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;
306                 }
307         }
308 }
309
310 static void
311 nfp_net_tx_queue_release(void *tx_queue)
312 {
313         struct nfp_net_txq *txq = tx_queue;
314
315         if (txq) {
316                 nfp_net_tx_queue_release_mbufs(txq);
317                 rte_free(txq->txbufs);
318                 rte_free(txq);
319         }
320 }
321
322 static void
323 nfp_net_reset_tx_queue(struct nfp_net_txq *txq)
324 {
325         nfp_net_tx_queue_release_mbufs(txq);
326         txq->wr_p = 0;
327         txq->rd_p = 0;
328 }
329
330 static int
331 __nfp_net_reconfig(struct nfp_net_hw *hw, uint32_t update)
332 {
333         int cnt;
334         uint32_t new;
335         struct timespec wait;
336
337         PMD_DRV_LOG(DEBUG, "Writing to the configuration queue (%p)...\n",
338                     hw->qcp_cfg);
339
340         if (hw->qcp_cfg == NULL)
341                 rte_panic("Bad configuration queue pointer\n");
342
343         nfp_qcp_ptr_add(hw->qcp_cfg, NFP_QCP_WRITE_PTR, 1);
344
345         wait.tv_sec = 0;
346         wait.tv_nsec = 1000000;
347
348         PMD_DRV_LOG(DEBUG, "Polling for update ack...\n");
349
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);
353                 if (new == 0)
354                         break;
355                 if (new & NFP_NET_CFG_UPDATE_ERR) {
356                         PMD_INIT_LOG(ERR, "Reconfig error: 0x%08x", new);
357                         return -1;
358                 }
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");
363                 }
364                 nanosleep(&wait, 0); /* waiting for a 1ms */
365         }
366         PMD_DRV_LOG(DEBUG, "Ack DONE\n");
367         return 0;
368 }
369
370 /*
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
375  *
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.
378  */
379 static int
380 nfp_net_reconfig(struct nfp_net_hw *hw, uint32_t ctrl, uint32_t update)
381 {
382         uint32_t err;
383
384         PMD_DRV_LOG(DEBUG, "nfp_net_reconfig: ctrl=%08x update=%08x\n",
385                     ctrl, update);
386
387         rte_spinlock_lock(&hw->reconfig_lock);
388
389         nn_cfg_writel(hw, NFP_NET_CFG_CTRL, ctrl);
390         nn_cfg_writel(hw, NFP_NET_CFG_UPDATE, update);
391
392         rte_wmb();
393
394         err = __nfp_net_reconfig(hw, update);
395
396         rte_spinlock_unlock(&hw->reconfig_lock);
397
398         if (!err)
399                 return 0;
400
401         /*
402          * Reconfig errors imply situations where they can be handled.
403          * Otherwise, rte_panic is called inside __nfp_net_reconfig
404          */
405         PMD_INIT_LOG(ERR, "Error nfp_net reconfig for ctrl: %x update: %x",
406                      ctrl, update);
407         return -EIO;
408 }
409
410 /*
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.
414  */
415 static int
416 nfp_net_configure(struct rte_eth_dev *dev)
417 {
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;
422         uint32_t update = 0;
423         struct nfp_net_hw *hw;
424
425         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
426
427         /*
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
433          */
434
435         PMD_INIT_LOG(DEBUG, "Configure");
436
437         dev_conf = &dev->data->dev_conf;
438         rxmode = &dev_conf->rxmode;
439         txmode = &dev_conf->txmode;
440
441         /* Checking TX mode */
442         if (txmode->mq_mode) {
443                 PMD_INIT_LOG(INFO, "TX mq_mode DCB and VMDq not supported");
444                 return -EINVAL;
445         }
446
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;
452                 } else {
453                         PMD_INIT_LOG(INFO, "RSS not supported");
454                         return -EINVAL;
455                 }
456         }
457
458         if (rxmode->split_hdr_size) {
459                 PMD_INIT_LOG(INFO, "rxmode does not support split header");
460                 return -EINVAL;
461         }
462
463         if (rxmode->hw_ip_checksum) {
464                 if (hw->cap & NFP_NET_CFG_CTRL_RXCSUM) {
465                         new_ctrl |= NFP_NET_CFG_CTRL_RXCSUM;
466                 } else {
467                         PMD_INIT_LOG(INFO, "RXCSUM not supported");
468                         return -EINVAL;
469                 }
470         }
471
472         if (rxmode->hw_vlan_filter) {
473                 PMD_INIT_LOG(INFO, "VLAN filter not supported");
474                 return -EINVAL;
475         }
476
477         if (rxmode->hw_vlan_strip) {
478                 if (hw->cap & NFP_NET_CFG_CTRL_RXVLAN) {
479                         new_ctrl |= NFP_NET_CFG_CTRL_RXVLAN;
480                 } else {
481                         PMD_INIT_LOG(INFO, "hw vlan strip not supported");
482                         return -EINVAL;
483                 }
484         }
485
486         if (rxmode->hw_vlan_extend) {
487                 PMD_INIT_LOG(INFO, "VLAN extended not supported");
488                 return -EINVAL;
489         }
490
491         /* Supporting VLAN insertion by default */
492         if (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)
493                 new_ctrl |= NFP_NET_CFG_CTRL_TXVLAN;
494
495         if (rxmode->jumbo_frame)
496                 /* this is handled in rte_eth_dev_configure */
497
498         if (rxmode->hw_strip_crc) {
499                 PMD_INIT_LOG(INFO, "strip CRC not supported");
500                 return -EINVAL;
501         }
502
503         if (rxmode->enable_scatter) {
504                 PMD_INIT_LOG(INFO, "Scatter not supported");
505                 return -EINVAL;
506         }
507
508         if (!new_ctrl)
509                 return 0;
510
511         update |= NFP_NET_CFG_UPDATE_GEN;
512
513         nn_cfg_writel(hw, NFP_NET_CFG_CTRL, new_ctrl);
514         if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
515                 return -EIO;
516
517         hw->ctrl = new_ctrl;
518
519         return 0;
520 }
521
522 static void
523 nfp_net_enable_queues(struct rte_eth_dev *dev)
524 {
525         struct nfp_net_hw *hw;
526         uint64_t enabled_queues = 0;
527         int i;
528
529         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
530
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);
534
535         nn_cfg_writeq(hw, NFP_NET_CFG_TXRS_ENABLE, enabled_queues);
536
537         enabled_queues = 0;
538
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);
542
543         nn_cfg_writeq(hw, NFP_NET_CFG_RXRS_ENABLE, enabled_queues);
544 }
545
546 static void
547 nfp_net_disable_queues(struct rte_eth_dev *dev)
548 {
549         struct nfp_net_hw *hw;
550         uint32_t new_ctrl, update = 0;
551
552         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
553
554         nn_cfg_writeq(hw, NFP_NET_CFG_TXRS_ENABLE, 0);
555         nn_cfg_writeq(hw, NFP_NET_CFG_RXRS_ENABLE, 0);
556
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;
560
561         if (hw->cap & NFP_NET_CFG_CTRL_RINGCFG)
562                 new_ctrl &= ~NFP_NET_CFG_CTRL_RINGCFG;
563
564         /* If an error when reconfig we avoid to change hw state */
565         if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
566                 return;
567
568         hw->ctrl = new_ctrl;
569 }
570
571 static int
572 nfp_net_rx_freelist_setup(struct rte_eth_dev *dev)
573 {
574         int i;
575
576         for (i = 0; i < dev->data->nb_rx_queues; i++) {
577                 if (nfp_net_rx_fill_freelist(dev->data->rx_queues[i]) < 0)
578                         return -1;
579         }
580         return 0;
581 }
582
583 static void
584 nfp_net_params_setup(struct nfp_net_hw *hw)
585 {
586         nn_cfg_writel(hw, NFP_NET_CFG_MTU, hw->mtu);
587         nn_cfg_writel(hw, NFP_NET_CFG_FLBUFSZ, hw->flbufsz);
588 }
589
590 static void
591 nfp_net_cfg_queue_setup(struct nfp_net_hw *hw)
592 {
593         hw->qcp_cfg = hw->tx_bar + NFP_QCP_QUEUE_ADDR_SZ;
594 }
595
596 static void nfp_net_read_mac(struct nfp_net_hw *hw)
597 {
598         uint32_t tmp;
599
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));
602
603         tmp = rte_be_to_cpu_32(nn_cfg_readl(hw, NFP_NET_CFG_MACADDR + 4));
604         memcpy(&hw->mac_addr[4], &tmp, 2);
605 }
606
607 static void
608 nfp_net_write_mac(struct nfp_net_hw *hw, uint8_t *mac)
609 {
610         uint32_t mac0 = *(uint32_t *)mac;
611         uint16_t mac1;
612
613         nn_writel(rte_cpu_to_be_32(mac0), hw->ctrl_bar + NFP_NET_CFG_MACADDR);
614
615         mac += 4;
616         mac1 = *(uint16_t *)mac;
617         nn_writew(rte_cpu_to_be_16(mac1),
618                   hw->ctrl_bar + NFP_NET_CFG_MACADDR + 6);
619 }
620
621 static int
622 nfp_configure_rx_interrupt(struct rte_eth_dev *dev,
623                            struct rte_intr_handle *intr_handle)
624 {
625         struct nfp_net_hw *hw;
626         int i;
627
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);
635                         return -ENOMEM;
636                 }
637         }
638
639         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
640
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;
646         } else {
647                 PMD_INIT_LOG(INFO, "VF: enabling RX interrupt with VFIO");
648                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
649                         /*
650                          * The first msix vector is reserved for non
651                          * efd interrupts
652                         */
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]);
657                 }
658         }
659
660         /* Avoiding TX interrupts */
661         hw->ctrl |= NFP_NET_CFG_CTRL_MSIX_TX_OFF;
662         return 0;
663 }
664
665 static int
666 nfp_net_start(struct rte_eth_dev *dev)
667 {
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;
673         int ret;
674
675         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
676
677         PMD_INIT_LOG(DEBUG, "Start");
678
679         /* Disabling queues just in case... */
680         nfp_net_disable_queues(dev);
681
682         /* Writing configuration parameters in the device */
683         nfp_net_params_setup(hw);
684
685         /* Enabling the required queues in the device */
686         nfp_net_enable_queues(dev);
687
688         /* check and configure queue intr-vector mapping */
689         if (dev->data->dev_conf.intr_conf.rxq != 0) {
690                 if (hw->pf_multiport_enabled) {
691                         PMD_INIT_LOG(ERR, "PMD rx interrupt is not supported "
692                                           "with NFP multiport PF");
693                                 return -EINVAL;
694                 }
695                 if (intr_handle->type == RTE_INTR_HANDLE_UIO) {
696                         /*
697                          * Better not to share LSC with RX interrupts.
698                          * Unregistering LSC interrupt handler
699                          */
700                         rte_intr_callback_unregister(&pci_dev->intr_handle,
701                                 nfp_net_dev_interrupt_handler, (void *)dev);
702
703                         if (dev->data->nb_rx_queues > 1) {
704                                 PMD_INIT_LOG(ERR, "PMD rx interrupt only "
705                                              "supports 1 queue with UIO");
706                                 return -EIO;
707                         }
708                 }
709                 intr_vector = dev->data->nb_rx_queues;
710                 if (rte_intr_efd_enable(intr_handle, intr_vector))
711                         return -1;
712
713                 nfp_configure_rx_interrupt(dev, intr_handle);
714                 update = NFP_NET_CFG_UPDATE_MSIX;
715         }
716
717         rte_intr_enable(intr_handle);
718
719         /* Enable device */
720         new_ctrl = hw->ctrl | NFP_NET_CFG_CTRL_ENABLE;
721
722         update |= NFP_NET_CFG_UPDATE_GEN | NFP_NET_CFG_UPDATE_RING;
723
724         if (hw->cap & NFP_NET_CFG_CTRL_RINGCFG)
725                 new_ctrl |= NFP_NET_CFG_CTRL_RINGCFG;
726
727         nn_cfg_writel(hw, NFP_NET_CFG_CTRL, new_ctrl);
728         if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
729                 return -EIO;
730
731         /*
732          * Allocating rte mbuffs for configured rx queues.
733          * This requires queues being enabled before
734          */
735         if (nfp_net_rx_freelist_setup(dev) < 0) {
736                 ret = -ENOMEM;
737                 goto error;
738         }
739
740         hw->ctrl = new_ctrl;
741
742         return 0;
743
744 error:
745         /*
746          * An error returned by this function should mean the app
747          * exiting and then the system releasing all the memory
748          * allocated even memory coming from hugepages.
749          *
750          * The device could be enabled at this point with some queues
751          * ready for getting packets. This is true if the call to
752          * nfp_net_rx_freelist_setup() succeeds for some queues but
753          * fails for subsequent queues.
754          *
755          * This should make the app exiting but better if we tell the
756          * device first.
757          */
758         nfp_net_disable_queues(dev);
759
760         return ret;
761 }
762
763 /* Stop device: disable rx and tx functions to allow for reconfiguring. */
764 static void
765 nfp_net_stop(struct rte_eth_dev *dev)
766 {
767         int i;
768
769         PMD_INIT_LOG(DEBUG, "Stop");
770
771         nfp_net_disable_queues(dev);
772
773         /* Clear queues */
774         for (i = 0; i < dev->data->nb_tx_queues; i++) {
775                 nfp_net_reset_tx_queue(
776                         (struct nfp_net_txq *)dev->data->tx_queues[i]);
777         }
778
779         for (i = 0; i < dev->data->nb_rx_queues; i++) {
780                 nfp_net_reset_rx_queue(
781                         (struct nfp_net_rxq *)dev->data->rx_queues[i]);
782         }
783 }
784
785 /* Reset and stop device. The device can not be restarted. */
786 static void
787 nfp_net_close(struct rte_eth_dev *dev)
788 {
789         struct nfp_net_hw *hw;
790         struct rte_pci_device *pci_dev;
791
792         PMD_INIT_LOG(DEBUG, "Close");
793
794         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
795         pci_dev = RTE_ETH_DEV_TO_PCI(dev);
796
797         /*
798          * We assume that the DPDK application is stopping all the
799          * threads/queues before calling the device close function.
800          */
801
802         nfp_net_stop(dev);
803
804         rte_intr_disable(&pci_dev->intr_handle);
805         nn_cfg_writeb(hw, NFP_NET_CFG_LSC, 0xff);
806
807         /* unregister callback func from eal lib */
808         rte_intr_callback_unregister(&pci_dev->intr_handle,
809                                      nfp_net_dev_interrupt_handler,
810                                      (void *)dev);
811
812         /*
813          * The ixgbe PMD driver disables the pcie master on the
814          * device. The i40e does not...
815          */
816 }
817
818 static void
819 nfp_net_promisc_enable(struct rte_eth_dev *dev)
820 {
821         uint32_t new_ctrl, update = 0;
822         struct nfp_net_hw *hw;
823
824         PMD_DRV_LOG(DEBUG, "Promiscuous mode enable\n");
825
826         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
827
828         if (!(hw->cap & NFP_NET_CFG_CTRL_PROMISC)) {
829                 PMD_INIT_LOG(INFO, "Promiscuous mode not supported");
830                 return;
831         }
832
833         if (hw->ctrl & NFP_NET_CFG_CTRL_PROMISC) {
834                 PMD_DRV_LOG(INFO, "Promiscuous mode already enabled\n");
835                 return;
836         }
837
838         new_ctrl = hw->ctrl | NFP_NET_CFG_CTRL_PROMISC;
839         update = NFP_NET_CFG_UPDATE_GEN;
840
841         /*
842          * DPDK sets promiscuous mode on just after this call assuming
843          * it can not fail ...
844          */
845         if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
846                 return;
847
848         hw->ctrl = new_ctrl;
849 }
850
851 static void
852 nfp_net_promisc_disable(struct rte_eth_dev *dev)
853 {
854         uint32_t new_ctrl, update = 0;
855         struct nfp_net_hw *hw;
856
857         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
858
859         if ((hw->ctrl & NFP_NET_CFG_CTRL_PROMISC) == 0) {
860                 PMD_DRV_LOG(INFO, "Promiscuous mode already disabled\n");
861                 return;
862         }
863
864         new_ctrl = hw->ctrl & ~NFP_NET_CFG_CTRL_PROMISC;
865         update = NFP_NET_CFG_UPDATE_GEN;
866
867         /*
868          * DPDK sets promiscuous mode off just before this call
869          * assuming it can not fail ...
870          */
871         if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
872                 return;
873
874         hw->ctrl = new_ctrl;
875 }
876
877 /*
878  * return 0 means link status changed, -1 means not changed
879  *
880  * Wait to complete is needed as it can take up to 9 seconds to get the Link
881  * status.
882  */
883 static int
884 nfp_net_link_update(struct rte_eth_dev *dev, __rte_unused int wait_to_complete)
885 {
886         struct nfp_net_hw *hw;
887         struct rte_eth_link link, old;
888         uint32_t nn_link_status;
889
890         static const uint32_t ls_to_ethtool[] = {
891                 [NFP_NET_CFG_STS_LINK_RATE_UNSUPPORTED] = ETH_SPEED_NUM_NONE,
892                 [NFP_NET_CFG_STS_LINK_RATE_UNKNOWN]     = ETH_SPEED_NUM_NONE,
893                 [NFP_NET_CFG_STS_LINK_RATE_1G]          = ETH_SPEED_NUM_1G,
894                 [NFP_NET_CFG_STS_LINK_RATE_10G]         = ETH_SPEED_NUM_10G,
895                 [NFP_NET_CFG_STS_LINK_RATE_25G]         = ETH_SPEED_NUM_25G,
896                 [NFP_NET_CFG_STS_LINK_RATE_40G]         = ETH_SPEED_NUM_40G,
897                 [NFP_NET_CFG_STS_LINK_RATE_50G]         = ETH_SPEED_NUM_50G,
898                 [NFP_NET_CFG_STS_LINK_RATE_100G]        = ETH_SPEED_NUM_100G,
899         };
900
901         PMD_DRV_LOG(DEBUG, "Link update\n");
902
903         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
904
905         memset(&old, 0, sizeof(old));
906         nfp_net_dev_atomic_read_link_status(dev, &old);
907
908         nn_link_status = nn_cfg_readl(hw, NFP_NET_CFG_STS);
909
910         memset(&link, 0, sizeof(struct rte_eth_link));
911
912         if (nn_link_status & NFP_NET_CFG_STS_LINK)
913                 link.link_status = ETH_LINK_UP;
914
915         link.link_duplex = ETH_LINK_FULL_DUPLEX;
916
917         nn_link_status = (nn_link_status >> NFP_NET_CFG_STS_LINK_RATE_SHIFT) &
918                          NFP_NET_CFG_STS_LINK_RATE_MASK;
919
920         if ((NFD_CFG_MAJOR_VERSION_of(hw->ver) < 4) ||
921             ((NFD_CFG_MINOR_VERSION_of(hw->ver) == 4) &&
922             (NFD_CFG_MINOR_VERSION_of(hw->ver) == 0)))
923                 /* We really do not know the speed wil old firmware */
924                 link.link_speed = ETH_SPEED_NUM_NONE;
925         else {
926                 if (nn_link_status >= RTE_DIM(ls_to_ethtool))
927                         link.link_speed = ETH_SPEED_NUM_NONE;
928                 else
929                         link.link_speed = ls_to_ethtool[nn_link_status];
930         }
931
932         if (old.link_status != link.link_status) {
933                 nfp_net_dev_atomic_write_link_status(dev, &link);
934                 if (link.link_status)
935                         PMD_DRV_LOG(INFO, "NIC Link is Up\n");
936                 else
937                         PMD_DRV_LOG(INFO, "NIC Link is Down\n");
938                 return 0;
939         }
940
941         return -1;
942 }
943
944 static void
945 nfp_net_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
946 {
947         int i;
948         struct nfp_net_hw *hw;
949         struct rte_eth_stats nfp_dev_stats;
950
951         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
952
953         /* RTE_ETHDEV_QUEUE_STAT_CNTRS default value is 16 */
954
955         /* reading per RX ring stats */
956         for (i = 0; i < dev->data->nb_rx_queues; i++) {
957                 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
958                         break;
959
960                 nfp_dev_stats.q_ipackets[i] =
961                         nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i));
962
963                 nfp_dev_stats.q_ipackets[i] -=
964                         hw->eth_stats_base.q_ipackets[i];
965
966                 nfp_dev_stats.q_ibytes[i] =
967                         nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i) + 0x8);
968
969                 nfp_dev_stats.q_ibytes[i] -=
970                         hw->eth_stats_base.q_ibytes[i];
971         }
972
973         /* reading per TX ring stats */
974         for (i = 0; i < dev->data->nb_tx_queues; i++) {
975                 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
976                         break;
977
978                 nfp_dev_stats.q_opackets[i] =
979                         nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i));
980
981                 nfp_dev_stats.q_opackets[i] -=
982                         hw->eth_stats_base.q_opackets[i];
983
984                 nfp_dev_stats.q_obytes[i] =
985                         nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i) + 0x8);
986
987                 nfp_dev_stats.q_obytes[i] -=
988                         hw->eth_stats_base.q_obytes[i];
989         }
990
991         nfp_dev_stats.ipackets =
992                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_FRAMES);
993
994         nfp_dev_stats.ipackets -= hw->eth_stats_base.ipackets;
995
996         nfp_dev_stats.ibytes =
997                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_OCTETS);
998
999         nfp_dev_stats.ibytes -= hw->eth_stats_base.ibytes;
1000
1001         nfp_dev_stats.opackets =
1002                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_FRAMES);
1003
1004         nfp_dev_stats.opackets -= hw->eth_stats_base.opackets;
1005
1006         nfp_dev_stats.obytes =
1007                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_OCTETS);
1008
1009         nfp_dev_stats.obytes -= hw->eth_stats_base.obytes;
1010
1011         /* reading general device stats */
1012         nfp_dev_stats.ierrors =
1013                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_ERRORS);
1014
1015         nfp_dev_stats.ierrors -= hw->eth_stats_base.ierrors;
1016
1017         nfp_dev_stats.oerrors =
1018                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_ERRORS);
1019
1020         nfp_dev_stats.oerrors -= hw->eth_stats_base.oerrors;
1021
1022         /* RX ring mbuf allocation failures */
1023         nfp_dev_stats.rx_nombuf = dev->data->rx_mbuf_alloc_failed;
1024
1025         nfp_dev_stats.imissed =
1026                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_DISCARDS);
1027
1028         nfp_dev_stats.imissed -= hw->eth_stats_base.imissed;
1029
1030         if (stats)
1031                 memcpy(stats, &nfp_dev_stats, sizeof(*stats));
1032 }
1033
1034 static void
1035 nfp_net_stats_reset(struct rte_eth_dev *dev)
1036 {
1037         int i;
1038         struct nfp_net_hw *hw;
1039
1040         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1041
1042         /*
1043          * hw->eth_stats_base records the per counter starting point.
1044          * Lets update it now
1045          */
1046
1047         /* reading per RX ring stats */
1048         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1049                 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
1050                         break;
1051
1052                 hw->eth_stats_base.q_ipackets[i] =
1053                         nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i));
1054
1055                 hw->eth_stats_base.q_ibytes[i] =
1056                         nn_cfg_readq(hw, NFP_NET_CFG_RXR_STATS(i) + 0x8);
1057         }
1058
1059         /* reading per TX ring stats */
1060         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1061                 if (i == RTE_ETHDEV_QUEUE_STAT_CNTRS)
1062                         break;
1063
1064                 hw->eth_stats_base.q_opackets[i] =
1065                         nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i));
1066
1067                 hw->eth_stats_base.q_obytes[i] =
1068                         nn_cfg_readq(hw, NFP_NET_CFG_TXR_STATS(i) + 0x8);
1069         }
1070
1071         hw->eth_stats_base.ipackets =
1072                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_FRAMES);
1073
1074         hw->eth_stats_base.ibytes =
1075                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_OCTETS);
1076
1077         hw->eth_stats_base.opackets =
1078                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_FRAMES);
1079
1080         hw->eth_stats_base.obytes =
1081                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_OCTETS);
1082
1083         /* reading general device stats */
1084         hw->eth_stats_base.ierrors =
1085                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_ERRORS);
1086
1087         hw->eth_stats_base.oerrors =
1088                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_TX_ERRORS);
1089
1090         /* RX ring mbuf allocation failures */
1091         dev->data->rx_mbuf_alloc_failed = 0;
1092
1093         hw->eth_stats_base.imissed =
1094                 nn_cfg_readq(hw, NFP_NET_CFG_STATS_RX_DISCARDS);
1095 }
1096
1097 static void
1098 nfp_net_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
1099 {
1100         struct nfp_net_hw *hw;
1101
1102         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1103
1104         dev_info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1105         dev_info->max_rx_queues = (uint16_t)hw->max_rx_queues;
1106         dev_info->max_tx_queues = (uint16_t)hw->max_tx_queues;
1107         dev_info->min_rx_bufsize = ETHER_MIN_MTU;
1108         dev_info->max_rx_pktlen = hw->mtu;
1109         /* Next should change when PF support is implemented */
1110         dev_info->max_mac_addrs = 1;
1111
1112         if (hw->cap & NFP_NET_CFG_CTRL_RXVLAN)
1113                 dev_info->rx_offload_capa = DEV_RX_OFFLOAD_VLAN_STRIP;
1114
1115         if (hw->cap & NFP_NET_CFG_CTRL_RXCSUM)
1116                 dev_info->rx_offload_capa |= DEV_RX_OFFLOAD_IPV4_CKSUM |
1117                                              DEV_RX_OFFLOAD_UDP_CKSUM |
1118                                              DEV_RX_OFFLOAD_TCP_CKSUM;
1119
1120         if (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)
1121                 dev_info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
1122
1123         if (hw->cap & NFP_NET_CFG_CTRL_TXCSUM)
1124                 dev_info->tx_offload_capa |= DEV_TX_OFFLOAD_IPV4_CKSUM |
1125                                              DEV_TX_OFFLOAD_UDP_CKSUM |
1126                                              DEV_TX_OFFLOAD_TCP_CKSUM;
1127
1128         dev_info->default_rxconf = (struct rte_eth_rxconf) {
1129                 .rx_thresh = {
1130                         .pthresh = DEFAULT_RX_PTHRESH,
1131                         .hthresh = DEFAULT_RX_HTHRESH,
1132                         .wthresh = DEFAULT_RX_WTHRESH,
1133                 },
1134                 .rx_free_thresh = DEFAULT_RX_FREE_THRESH,
1135                 .rx_drop_en = 0,
1136         };
1137
1138         dev_info->default_txconf = (struct rte_eth_txconf) {
1139                 .tx_thresh = {
1140                         .pthresh = DEFAULT_TX_PTHRESH,
1141                         .hthresh = DEFAULT_TX_HTHRESH,
1142                         .wthresh = DEFAULT_TX_WTHRESH,
1143                 },
1144                 .tx_free_thresh = DEFAULT_TX_FREE_THRESH,
1145                 .tx_rs_thresh = DEFAULT_TX_RSBIT_THRESH,
1146                 .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS |
1147                              ETH_TXQ_FLAGS_NOOFFLOADS,
1148         };
1149
1150         dev_info->reta_size = NFP_NET_CFG_RSS_ITBL_SZ;
1151         dev_info->hash_key_size = NFP_NET_CFG_RSS_KEY_SZ;
1152
1153         dev_info->speed_capa = ETH_SPEED_NUM_1G | ETH_LINK_SPEED_10G |
1154                                ETH_SPEED_NUM_25G | ETH_SPEED_NUM_40G |
1155                                ETH_SPEED_NUM_50G | ETH_LINK_SPEED_100G;
1156
1157         if (hw->cap & NFP_NET_CFG_CTRL_LSO)
1158                 dev_info->tx_offload_capa |= DEV_TX_OFFLOAD_TCP_TSO;
1159 }
1160
1161 static const uint32_t *
1162 nfp_net_supported_ptypes_get(struct rte_eth_dev *dev)
1163 {
1164         static const uint32_t ptypes[] = {
1165                 /* refers to nfp_net_set_hash() */
1166                 RTE_PTYPE_INNER_L3_IPV4,
1167                 RTE_PTYPE_INNER_L3_IPV6,
1168                 RTE_PTYPE_INNER_L3_IPV6_EXT,
1169                 RTE_PTYPE_INNER_L4_MASK,
1170                 RTE_PTYPE_UNKNOWN
1171         };
1172
1173         if (dev->rx_pkt_burst == nfp_net_recv_pkts)
1174                 return ptypes;
1175         return NULL;
1176 }
1177
1178 static uint32_t
1179 nfp_net_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_idx)
1180 {
1181         struct nfp_net_rxq *rxq;
1182         struct nfp_net_rx_desc *rxds;
1183         uint32_t idx;
1184         uint32_t count;
1185
1186         rxq = (struct nfp_net_rxq *)dev->data->rx_queues[queue_idx];
1187
1188         idx = rxq->rd_p;
1189
1190         count = 0;
1191
1192         /*
1193          * Other PMDs are just checking the DD bit in intervals of 4
1194          * descriptors and counting all four if the first has the DD
1195          * bit on. Of course, this is not accurate but can be good for
1196          * performance. But ideally that should be done in descriptors
1197          * chunks belonging to the same cache line
1198          */
1199
1200         while (count < rxq->rx_count) {
1201                 rxds = &rxq->rxds[idx];
1202                 if ((rxds->rxd.meta_len_dd & PCIE_DESC_RX_DD) == 0)
1203                         break;
1204
1205                 count++;
1206                 idx++;
1207
1208                 /* Wrapping? */
1209                 if ((idx) == rxq->rx_count)
1210                         idx = 0;
1211         }
1212
1213         return count;
1214 }
1215
1216 static int
1217 nfp_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t queue_id)
1218 {
1219         struct rte_pci_device *pci_dev;
1220         struct nfp_net_hw *hw;
1221         int base = 0;
1222
1223         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1224         pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1225
1226         if (pci_dev->intr_handle.type != RTE_INTR_HANDLE_UIO)
1227                 base = 1;
1228
1229         /* Make sure all updates are written before un-masking */
1230         rte_wmb();
1231         nn_cfg_writeb(hw, NFP_NET_CFG_ICR(base + queue_id),
1232                       NFP_NET_CFG_ICR_UNMASKED);
1233         return 0;
1234 }
1235
1236 static int
1237 nfp_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t queue_id)
1238 {
1239         struct rte_pci_device *pci_dev;
1240         struct nfp_net_hw *hw;
1241         int base = 0;
1242
1243         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1244         pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1245
1246         if (pci_dev->intr_handle.type != RTE_INTR_HANDLE_UIO)
1247                 base = 1;
1248
1249         /* Make sure all updates are written before un-masking */
1250         rte_wmb();
1251         nn_cfg_writeb(hw, NFP_NET_CFG_ICR(base + queue_id), 0x1);
1252         return 0;
1253 }
1254
1255 static void
1256 nfp_net_dev_link_status_print(struct rte_eth_dev *dev)
1257 {
1258         struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1259         struct rte_eth_link link;
1260
1261         memset(&link, 0, sizeof(link));
1262         nfp_net_dev_atomic_read_link_status(dev, &link);
1263         if (link.link_status)
1264                 RTE_LOG(INFO, PMD, "Port %d: Link Up - speed %u Mbps - %s\n",
1265                         (int)(dev->data->port_id), (unsigned)link.link_speed,
1266                         link.link_duplex == ETH_LINK_FULL_DUPLEX
1267                         ? "full-duplex" : "half-duplex");
1268         else
1269                 RTE_LOG(INFO, PMD, " Port %d: Link Down\n",
1270                         (int)(dev->data->port_id));
1271
1272         RTE_LOG(INFO, PMD, "PCI Address: %04d:%02d:%02d:%d\n",
1273                 pci_dev->addr.domain, pci_dev->addr.bus,
1274                 pci_dev->addr.devid, pci_dev->addr.function);
1275 }
1276
1277 /* Interrupt configuration and handling */
1278
1279 /*
1280  * nfp_net_irq_unmask - Unmask an interrupt
1281  *
1282  * If MSI-X auto-masking is enabled clear the mask bit, otherwise
1283  * clear the ICR for the entry.
1284  */
1285 static void
1286 nfp_net_irq_unmask(struct rte_eth_dev *dev)
1287 {
1288         struct nfp_net_hw *hw;
1289         struct rte_pci_device *pci_dev;
1290
1291         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1292         pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1293
1294         if (hw->ctrl & NFP_NET_CFG_CTRL_MSIXAUTO) {
1295                 /* If MSI-X auto-masking is used, clear the entry */
1296                 rte_wmb();
1297                 rte_intr_enable(&pci_dev->intr_handle);
1298         } else {
1299                 /* Make sure all updates are written before un-masking */
1300                 rte_wmb();
1301                 nn_cfg_writeb(hw, NFP_NET_CFG_ICR(NFP_NET_IRQ_LSC_IDX),
1302                               NFP_NET_CFG_ICR_UNMASKED);
1303         }
1304 }
1305
1306 static void
1307 nfp_net_dev_interrupt_handler(void *param)
1308 {
1309         int64_t timeout;
1310         struct rte_eth_link link;
1311         struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
1312
1313         PMD_DRV_LOG(DEBUG, "We got a LSC interrupt!!!\n");
1314
1315         /* get the link status */
1316         memset(&link, 0, sizeof(link));
1317         nfp_net_dev_atomic_read_link_status(dev, &link);
1318
1319         nfp_net_link_update(dev, 0);
1320
1321         /* likely to up */
1322         if (!link.link_status) {
1323                 /* handle it 1 sec later, wait it being stable */
1324                 timeout = NFP_NET_LINK_UP_CHECK_TIMEOUT;
1325                 /* likely to down */
1326         } else {
1327                 /* handle it 4 sec later, wait it being stable */
1328                 timeout = NFP_NET_LINK_DOWN_CHECK_TIMEOUT;
1329         }
1330
1331         if (rte_eal_alarm_set(timeout * 1000,
1332                               nfp_net_dev_interrupt_delayed_handler,
1333                               (void *)dev) < 0) {
1334                 RTE_LOG(ERR, PMD, "Error setting alarm");
1335                 /* Unmasking */
1336                 nfp_net_irq_unmask(dev);
1337         }
1338 }
1339
1340 /*
1341  * Interrupt handler which shall be registered for alarm callback for delayed
1342  * handling specific interrupt to wait for the stable nic state. As the NIC
1343  * interrupt state is not stable for nfp after link is just down, it needs
1344  * to wait 4 seconds to get the stable status.
1345  *
1346  * @param handle   Pointer to interrupt handle.
1347  * @param param    The address of parameter (struct rte_eth_dev *)
1348  *
1349  * @return  void
1350  */
1351 static void
1352 nfp_net_dev_interrupt_delayed_handler(void *param)
1353 {
1354         struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
1355
1356         nfp_net_link_update(dev, 0);
1357         _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL, NULL);
1358
1359         nfp_net_dev_link_status_print(dev);
1360
1361         /* Unmasking */
1362         nfp_net_irq_unmask(dev);
1363 }
1364
1365 static int
1366 nfp_net_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1367 {
1368         struct nfp_net_hw *hw;
1369
1370         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1371
1372         /* check that mtu is within the allowed range */
1373         if ((mtu < ETHER_MIN_MTU) || ((uint32_t)mtu > hw->max_mtu))
1374                 return -EINVAL;
1375
1376         /* switch to jumbo mode if needed */
1377         if ((uint32_t)mtu > ETHER_MAX_LEN)
1378                 dev->data->dev_conf.rxmode.jumbo_frame = 1;
1379         else
1380                 dev->data->dev_conf.rxmode.jumbo_frame = 0;
1381
1382         /* update max frame size */
1383         dev->data->dev_conf.rxmode.max_rx_pkt_len = (uint32_t)mtu;
1384
1385         /* writing to configuration space */
1386         nn_cfg_writel(hw, NFP_NET_CFG_MTU, (uint32_t)mtu);
1387
1388         hw->mtu = mtu;
1389
1390         return 0;
1391 }
1392
1393 static int
1394 nfp_net_rx_queue_setup(struct rte_eth_dev *dev,
1395                        uint16_t queue_idx, uint16_t nb_desc,
1396                        unsigned int socket_id,
1397                        const struct rte_eth_rxconf *rx_conf,
1398                        struct rte_mempool *mp)
1399 {
1400         const struct rte_memzone *tz;
1401         struct nfp_net_rxq *rxq;
1402         struct nfp_net_hw *hw;
1403
1404         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1405
1406         PMD_INIT_FUNC_TRACE();
1407
1408         /* Validating number of descriptors */
1409         if (((nb_desc * sizeof(struct nfp_net_rx_desc)) % 128) != 0 ||
1410             (nb_desc > NFP_NET_MAX_RX_DESC) ||
1411             (nb_desc < NFP_NET_MIN_RX_DESC)) {
1412                 RTE_LOG(ERR, PMD, "Wrong nb_desc value\n");
1413                 return -EINVAL;
1414         }
1415
1416         /*
1417          * Free memory prior to re-allocation if needed. This is the case after
1418          * calling nfp_net_stop
1419          */
1420         if (dev->data->rx_queues[queue_idx]) {
1421                 nfp_net_rx_queue_release(dev->data->rx_queues[queue_idx]);
1422                 dev->data->rx_queues[queue_idx] = NULL;
1423         }
1424
1425         /* Allocating rx queue data structure */
1426         rxq = rte_zmalloc_socket("ethdev RX queue", sizeof(struct nfp_net_rxq),
1427                                  RTE_CACHE_LINE_SIZE, socket_id);
1428         if (rxq == NULL)
1429                 return -ENOMEM;
1430
1431         /* Hw queues mapping based on firmware confifguration */
1432         rxq->qidx = queue_idx;
1433         rxq->fl_qcidx = queue_idx * hw->stride_rx;
1434         rxq->rx_qcidx = rxq->fl_qcidx + (hw->stride_rx - 1);
1435         rxq->qcp_fl = hw->rx_bar + NFP_QCP_QUEUE_OFF(rxq->fl_qcidx);
1436         rxq->qcp_rx = hw->rx_bar + NFP_QCP_QUEUE_OFF(rxq->rx_qcidx);
1437
1438         /*
1439          * Tracking mbuf size for detecting a potential mbuf overflow due to
1440          * RX offset
1441          */
1442         rxq->mem_pool = mp;
1443         rxq->mbuf_size = rxq->mem_pool->elt_size;
1444         rxq->mbuf_size -= (sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM);
1445         hw->flbufsz = rxq->mbuf_size;
1446
1447         rxq->rx_count = nb_desc;
1448         rxq->port_id = dev->data->port_id;
1449         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1450         rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ? 0
1451                                   : ETHER_CRC_LEN);
1452         rxq->drop_en = rx_conf->rx_drop_en;
1453
1454         /*
1455          * Allocate RX ring hardware descriptors. A memzone large enough to
1456          * handle the maximum ring size is allocated in order to allow for
1457          * resizing in later calls to the queue setup function.
1458          */
1459         tz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1460                                    sizeof(struct nfp_net_rx_desc) *
1461                                    NFP_NET_MAX_RX_DESC, NFP_MEMZONE_ALIGN,
1462                                    socket_id);
1463
1464         if (tz == NULL) {
1465                 RTE_LOG(ERR, PMD, "Error allocatig rx dma\n");
1466                 nfp_net_rx_queue_release(rxq);
1467                 return -ENOMEM;
1468         }
1469
1470         /* Saving physical and virtual addresses for the RX ring */
1471         rxq->dma = (uint64_t)tz->phys_addr;
1472         rxq->rxds = (struct nfp_net_rx_desc *)tz->addr;
1473
1474         /* mbuf pointers array for referencing mbufs linked to RX descriptors */
1475         rxq->rxbufs = rte_zmalloc_socket("rxq->rxbufs",
1476                                          sizeof(*rxq->rxbufs) * nb_desc,
1477                                          RTE_CACHE_LINE_SIZE, socket_id);
1478         if (rxq->rxbufs == NULL) {
1479                 nfp_net_rx_queue_release(rxq);
1480                 return -ENOMEM;
1481         }
1482
1483         PMD_RX_LOG(DEBUG, "rxbufs=%p hw_ring=%p dma_addr=0x%" PRIx64 "\n",
1484                    rxq->rxbufs, rxq->rxds, (unsigned long int)rxq->dma);
1485
1486         nfp_net_reset_rx_queue(rxq);
1487
1488         dev->data->rx_queues[queue_idx] = rxq;
1489         rxq->hw = hw;
1490
1491         /*
1492          * Telling the HW about the physical address of the RX ring and number
1493          * of descriptors in log2 format
1494          */
1495         nn_cfg_writeq(hw, NFP_NET_CFG_RXR_ADDR(queue_idx), rxq->dma);
1496         nn_cfg_writeb(hw, NFP_NET_CFG_RXR_SZ(queue_idx), rte_log2_u32(nb_desc));
1497
1498         return 0;
1499 }
1500
1501 static int
1502 nfp_net_rx_fill_freelist(struct nfp_net_rxq *rxq)
1503 {
1504         struct nfp_net_rx_buff *rxe = rxq->rxbufs;
1505         uint64_t dma_addr;
1506         unsigned i;
1507
1508         PMD_RX_LOG(DEBUG, "nfp_net_rx_fill_freelist for %u descriptors\n",
1509                    rxq->rx_count);
1510
1511         for (i = 0; i < rxq->rx_count; i++) {
1512                 struct nfp_net_rx_desc *rxd;
1513                 struct rte_mbuf *mbuf = rte_pktmbuf_alloc(rxq->mem_pool);
1514
1515                 if (mbuf == NULL) {
1516                         RTE_LOG(ERR, PMD, "RX mbuf alloc failed queue_id=%u\n",
1517                                 (unsigned)rxq->qidx);
1518                         return -ENOMEM;
1519                 }
1520
1521                 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DMA_ADDR_DEFAULT(mbuf));
1522
1523                 rxd = &rxq->rxds[i];
1524                 rxd->fld.dd = 0;
1525                 rxd->fld.dma_addr_hi = (dma_addr >> 32) & 0xff;
1526                 rxd->fld.dma_addr_lo = dma_addr & 0xffffffff;
1527                 rxe[i].mbuf = mbuf;
1528                 PMD_RX_LOG(DEBUG, "[%d]: %" PRIx64 "\n", i, dma_addr);
1529         }
1530
1531         /* Make sure all writes are flushed before telling the hardware */
1532         rte_wmb();
1533
1534         /* Not advertising the whole ring as the firmware gets confused if so */
1535         PMD_RX_LOG(DEBUG, "Increment FL write pointer in %u\n",
1536                    rxq->rx_count - 1);
1537
1538         nfp_qcp_ptr_add(rxq->qcp_fl, NFP_QCP_WRITE_PTR, rxq->rx_count - 1);
1539
1540         return 0;
1541 }
1542
1543 static int
1544 nfp_net_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
1545                        uint16_t nb_desc, unsigned int socket_id,
1546                        const struct rte_eth_txconf *tx_conf)
1547 {
1548         const struct rte_memzone *tz;
1549         struct nfp_net_txq *txq;
1550         uint16_t tx_free_thresh;
1551         struct nfp_net_hw *hw;
1552
1553         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1554
1555         PMD_INIT_FUNC_TRACE();
1556
1557         /* Validating number of descriptors */
1558         if (((nb_desc * sizeof(struct nfp_net_tx_desc)) % 128) != 0 ||
1559             (nb_desc > NFP_NET_MAX_TX_DESC) ||
1560             (nb_desc < NFP_NET_MIN_TX_DESC)) {
1561                 RTE_LOG(ERR, PMD, "Wrong nb_desc value\n");
1562                 return -EINVAL;
1563         }
1564
1565         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
1566                                     tx_conf->tx_free_thresh :
1567                                     DEFAULT_TX_FREE_THRESH);
1568
1569         if (tx_free_thresh > (nb_desc)) {
1570                 RTE_LOG(ERR, PMD,
1571                         "tx_free_thresh must be less than the number of TX "
1572                         "descriptors. (tx_free_thresh=%u port=%d "
1573                         "queue=%d)\n", (unsigned int)tx_free_thresh,
1574                         (int)dev->data->port_id, (int)queue_idx);
1575                 return -(EINVAL);
1576         }
1577
1578         /*
1579          * Free memory prior to re-allocation if needed. This is the case after
1580          * calling nfp_net_stop
1581          */
1582         if (dev->data->tx_queues[queue_idx]) {
1583                 PMD_TX_LOG(DEBUG, "Freeing memory prior to re-allocation %d\n",
1584                            queue_idx);
1585                 nfp_net_tx_queue_release(dev->data->tx_queues[queue_idx]);
1586                 dev->data->tx_queues[queue_idx] = NULL;
1587         }
1588
1589         /* Allocating tx queue data structure */
1590         txq = rte_zmalloc_socket("ethdev TX queue", sizeof(struct nfp_net_txq),
1591                                  RTE_CACHE_LINE_SIZE, socket_id);
1592         if (txq == NULL) {
1593                 RTE_LOG(ERR, PMD, "Error allocating tx dma\n");
1594                 return -ENOMEM;
1595         }
1596
1597         /*
1598          * Allocate TX ring hardware descriptors. A memzone large enough to
1599          * handle the maximum ring size is allocated in order to allow for
1600          * resizing in later calls to the queue setup function.
1601          */
1602         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
1603                                    sizeof(struct nfp_net_tx_desc) *
1604                                    NFP_NET_MAX_TX_DESC, NFP_MEMZONE_ALIGN,
1605                                    socket_id);
1606         if (tz == NULL) {
1607                 RTE_LOG(ERR, PMD, "Error allocating tx dma\n");
1608                 nfp_net_tx_queue_release(txq);
1609                 return -ENOMEM;
1610         }
1611
1612         txq->tx_count = nb_desc;
1613         txq->tx_free_thresh = tx_free_thresh;
1614         txq->tx_pthresh = tx_conf->tx_thresh.pthresh;
1615         txq->tx_hthresh = tx_conf->tx_thresh.hthresh;
1616         txq->tx_wthresh = tx_conf->tx_thresh.wthresh;
1617
1618         /* queue mapping based on firmware configuration */
1619         txq->qidx = queue_idx;
1620         txq->tx_qcidx = queue_idx * hw->stride_tx;
1621         txq->qcp_q = hw->tx_bar + NFP_QCP_QUEUE_OFF(txq->tx_qcidx);
1622
1623         txq->port_id = dev->data->port_id;
1624         txq->txq_flags = tx_conf->txq_flags;
1625
1626         /* Saving physical and virtual addresses for the TX ring */
1627         txq->dma = (uint64_t)tz->phys_addr;
1628         txq->txds = (struct nfp_net_tx_desc *)tz->addr;
1629
1630         /* mbuf pointers array for referencing mbufs linked to TX descriptors */
1631         txq->txbufs = rte_zmalloc_socket("txq->txbufs",
1632                                          sizeof(*txq->txbufs) * nb_desc,
1633                                          RTE_CACHE_LINE_SIZE, socket_id);
1634         if (txq->txbufs == NULL) {
1635                 nfp_net_tx_queue_release(txq);
1636                 return -ENOMEM;
1637         }
1638         PMD_TX_LOG(DEBUG, "txbufs=%p hw_ring=%p dma_addr=0x%" PRIx64 "\n",
1639                    txq->txbufs, txq->txds, (unsigned long int)txq->dma);
1640
1641         nfp_net_reset_tx_queue(txq);
1642
1643         dev->data->tx_queues[queue_idx] = txq;
1644         txq->hw = hw;
1645
1646         /*
1647          * Telling the HW about the physical address of the TX ring and number
1648          * of descriptors in log2 format
1649          */
1650         nn_cfg_writeq(hw, NFP_NET_CFG_TXR_ADDR(queue_idx), txq->dma);
1651         nn_cfg_writeb(hw, NFP_NET_CFG_TXR_SZ(queue_idx), rte_log2_u32(nb_desc));
1652
1653         return 0;
1654 }
1655
1656 /* nfp_net_tx_tso - Set TX descriptor for TSO */
1657 static inline void
1658 nfp_net_tx_tso(struct nfp_net_txq *txq, struct nfp_net_tx_desc *txd,
1659                struct rte_mbuf *mb)
1660 {
1661         uint64_t ol_flags;
1662         struct nfp_net_hw *hw = txq->hw;
1663
1664         if (!(hw->cap & NFP_NET_CFG_CTRL_LSO))
1665                 goto clean_txd;
1666
1667         ol_flags = mb->ol_flags;
1668
1669         if (!(ol_flags & PKT_TX_TCP_SEG))
1670                 goto clean_txd;
1671
1672         txd->l4_offset = mb->l2_len + mb->l3_len + mb->l4_len;
1673         txd->lso = rte_cpu_to_le_16(mb->tso_segsz);
1674         txd->flags = PCIE_DESC_TX_LSO;
1675         return;
1676
1677 clean_txd:
1678         txd->flags = 0;
1679         txd->l4_offset = 0;
1680         txd->lso = 0;
1681 }
1682
1683 /* nfp_net_tx_cksum - Set TX CSUM offload flags in TX descriptor */
1684 static inline void
1685 nfp_net_tx_cksum(struct nfp_net_txq *txq, struct nfp_net_tx_desc *txd,
1686                  struct rte_mbuf *mb)
1687 {
1688         uint64_t ol_flags;
1689         struct nfp_net_hw *hw = txq->hw;
1690
1691         if (!(hw->cap & NFP_NET_CFG_CTRL_TXCSUM))
1692                 return;
1693
1694         ol_flags = mb->ol_flags;
1695
1696         /* IPv6 does not need checksum */
1697         if (ol_flags & PKT_TX_IP_CKSUM)
1698                 txd->flags |= PCIE_DESC_TX_IP4_CSUM;
1699
1700         switch (ol_flags & PKT_TX_L4_MASK) {
1701         case PKT_TX_UDP_CKSUM:
1702                 txd->flags |= PCIE_DESC_TX_UDP_CSUM;
1703                 break;
1704         case PKT_TX_TCP_CKSUM:
1705                 txd->flags |= PCIE_DESC_TX_TCP_CSUM;
1706                 break;
1707         }
1708
1709         if (ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK))
1710                 txd->flags |= PCIE_DESC_TX_CSUM;
1711 }
1712
1713 /* nfp_net_rx_cksum - set mbuf checksum flags based on RX descriptor flags */
1714 static inline void
1715 nfp_net_rx_cksum(struct nfp_net_rxq *rxq, struct nfp_net_rx_desc *rxd,
1716                  struct rte_mbuf *mb)
1717 {
1718         struct nfp_net_hw *hw = rxq->hw;
1719
1720         if (!(hw->ctrl & NFP_NET_CFG_CTRL_RXCSUM))
1721                 return;
1722
1723         /* If IPv4 and IP checksum error, fail */
1724         if ((rxd->rxd.flags & PCIE_DESC_RX_IP4_CSUM) &&
1725             !(rxd->rxd.flags & PCIE_DESC_RX_IP4_CSUM_OK))
1726                 mb->ol_flags |= PKT_RX_IP_CKSUM_BAD;
1727
1728         /* If neither UDP nor TCP return */
1729         if (!(rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM) &&
1730             !(rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM))
1731                 return;
1732
1733         if ((rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM) &&
1734             !(rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK))
1735                 mb->ol_flags |= PKT_RX_L4_CKSUM_BAD;
1736
1737         if ((rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM) &&
1738             !(rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK))
1739                 mb->ol_flags |= PKT_RX_L4_CKSUM_BAD;
1740 }
1741
1742 #define NFP_HASH_OFFSET      ((uint8_t *)mbuf->buf_addr + mbuf->data_off - 4)
1743 #define NFP_HASH_TYPE_OFFSET ((uint8_t *)mbuf->buf_addr + mbuf->data_off - 8)
1744
1745 #define NFP_DESC_META_LEN(d) (d->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK)
1746
1747 /*
1748  * nfp_net_set_hash - Set mbuf hash data
1749  *
1750  * The RSS hash and hash-type are pre-pended to the packet data.
1751  * Extract and decode it and set the mbuf fields.
1752  */
1753 static inline void
1754 nfp_net_set_hash(struct nfp_net_rxq *rxq, struct nfp_net_rx_desc *rxd,
1755                  struct rte_mbuf *mbuf)
1756 {
1757         struct nfp_net_hw *hw = rxq->hw;
1758         uint8_t *meta_offset;
1759         uint32_t meta_info;
1760         uint32_t hash = 0;
1761         uint32_t hash_type = 0;
1762
1763         if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
1764                 return;
1765
1766         if (NFD_CFG_MAJOR_VERSION_of(hw->ver) <= 3) {
1767                 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
1768                         return;
1769
1770                 hash = rte_be_to_cpu_32(*(uint32_t *)NFP_HASH_OFFSET);
1771                 hash_type = rte_be_to_cpu_32(*(uint32_t *)NFP_HASH_TYPE_OFFSET);
1772
1773         } else if (NFP_DESC_META_LEN(rxd)) {
1774                 /*
1775                  * new metadata api:
1776                  * <----  32 bit  ----->
1777                  * m    field type word
1778                  * e     data field #2
1779                  * t     data field #1
1780                  * a     data field #0
1781                  * ====================
1782                  *    packet data
1783                  *
1784                  * Field type word contains up to 8 4bit field types
1785                  * A 4bit field type refers to a data field word
1786                  * A data field word can have several 4bit field types
1787                  */
1788                 meta_offset = rte_pktmbuf_mtod(mbuf, uint8_t *);
1789                 meta_offset -= NFP_DESC_META_LEN(rxd);
1790                 meta_info = rte_be_to_cpu_32(*(uint32_t *)meta_offset);
1791                 meta_offset += 4;
1792                 /* NFP PMD just supports metadata for hashing */
1793                 switch (meta_info & NFP_NET_META_FIELD_MASK) {
1794                 case NFP_NET_META_HASH:
1795                         /* next field type is about the hash type */
1796                         meta_info >>= NFP_NET_META_FIELD_SIZE;
1797                         /* hash value is in the data field */
1798                         hash = rte_be_to_cpu_32(*(uint32_t *)meta_offset);
1799                         hash_type = meta_info & NFP_NET_META_FIELD_MASK;
1800                         break;
1801                 default:
1802                         /* Unsupported metadata can be a performance issue */
1803                         return;
1804                 }
1805         } else {
1806                 return;
1807         }
1808
1809         mbuf->hash.rss = hash;
1810         mbuf->ol_flags |= PKT_RX_RSS_HASH;
1811
1812         switch (hash_type) {
1813         case NFP_NET_RSS_IPV4:
1814                 mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV4;
1815                 break;
1816         case NFP_NET_RSS_IPV6:
1817                 mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV6;
1818                 break;
1819         case NFP_NET_RSS_IPV6_EX:
1820                 mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV6_EXT;
1821                 break;
1822         default:
1823                 mbuf->packet_type |= RTE_PTYPE_INNER_L4_MASK;
1824         }
1825 }
1826
1827 static inline void
1828 nfp_net_mbuf_alloc_failed(struct nfp_net_rxq *rxq)
1829 {
1830         rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
1831 }
1832
1833 #define NFP_DESC_META_LEN(d) (d->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK)
1834
1835 /*
1836  * RX path design:
1837  *
1838  * There are some decissions to take:
1839  * 1) How to check DD RX descriptors bit
1840  * 2) How and when to allocate new mbufs
1841  *
1842  * Current implementation checks just one single DD bit each loop. As each
1843  * descriptor is 8 bytes, it is likely a good idea to check descriptors in
1844  * a single cache line instead. Tests with this change have not shown any
1845  * performance improvement but it requires further investigation. For example,
1846  * depending on which descriptor is next, the number of descriptors could be
1847  * less than 8 for just checking those in the same cache line. This implies
1848  * extra work which could be counterproductive by itself. Indeed, last firmware
1849  * changes are just doing this: writing several descriptors with the DD bit
1850  * for saving PCIe bandwidth and DMA operations from the NFP.
1851  *
1852  * Mbuf allocation is done when a new packet is received. Then the descriptor
1853  * is automatically linked with the new mbuf and the old one is given to the
1854  * user. The main drawback with this design is mbuf allocation is heavier than
1855  * using bulk allocations allowed by DPDK with rte_mempool_get_bulk. From the
1856  * cache point of view it does not seem allocating the mbuf early on as we are
1857  * doing now have any benefit at all. Again, tests with this change have not
1858  * shown any improvement. Also, rte_mempool_get_bulk returns all or nothing
1859  * so looking at the implications of this type of allocation should be studied
1860  * deeply
1861  */
1862
1863 static uint16_t
1864 nfp_net_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
1865 {
1866         struct nfp_net_rxq *rxq;
1867         struct nfp_net_rx_desc *rxds;
1868         struct nfp_net_rx_buff *rxb;
1869         struct nfp_net_hw *hw;
1870         struct rte_mbuf *mb;
1871         struct rte_mbuf *new_mb;
1872         uint16_t nb_hold;
1873         uint64_t dma_addr;
1874         int avail;
1875
1876         rxq = rx_queue;
1877         if (unlikely(rxq == NULL)) {
1878                 /*
1879                  * DPDK just checks the queue is lower than max queues
1880                  * enabled. But the queue needs to be configured
1881                  */
1882                 RTE_LOG_DP(ERR, PMD, "RX Bad queue\n");
1883                 return -EINVAL;
1884         }
1885
1886         hw = rxq->hw;
1887         avail = 0;
1888         nb_hold = 0;
1889
1890         while (avail < nb_pkts) {
1891                 rxb = &rxq->rxbufs[rxq->rd_p];
1892                 if (unlikely(rxb == NULL)) {
1893                         RTE_LOG_DP(ERR, PMD, "rxb does not exist!\n");
1894                         break;
1895                 }
1896
1897                 /*
1898                  * Memory barrier to ensure that we won't do other
1899                  * reads before the DD bit.
1900                  */
1901                 rte_rmb();
1902
1903                 rxds = &rxq->rxds[rxq->rd_p];
1904                 if ((rxds->rxd.meta_len_dd & PCIE_DESC_RX_DD) == 0)
1905                         break;
1906
1907                 /*
1908                  * We got a packet. Let's alloc a new mbuff for refilling the
1909                  * free descriptor ring as soon as possible
1910                  */
1911                 new_mb = rte_pktmbuf_alloc(rxq->mem_pool);
1912                 if (unlikely(new_mb == NULL)) {
1913                         RTE_LOG_DP(DEBUG, PMD, "RX mbuf alloc failed port_id=%u "
1914                                 "queue_id=%u\n", (unsigned)rxq->port_id,
1915                                 (unsigned)rxq->qidx);
1916                         nfp_net_mbuf_alloc_failed(rxq);
1917                         break;
1918                 }
1919
1920                 nb_hold++;
1921
1922                 /*
1923                  * Grab the mbuff and refill the descriptor with the
1924                  * previously allocated mbuff
1925                  */
1926                 mb = rxb->mbuf;
1927                 rxb->mbuf = new_mb;
1928
1929                 PMD_RX_LOG(DEBUG, "Packet len: %u, mbuf_size: %u\n",
1930                            rxds->rxd.data_len, rxq->mbuf_size);
1931
1932                 /* Size of this segment */
1933                 mb->data_len = rxds->rxd.data_len - NFP_DESC_META_LEN(rxds);
1934                 /* Size of the whole packet. We just support 1 segment */
1935                 mb->pkt_len = rxds->rxd.data_len - NFP_DESC_META_LEN(rxds);
1936
1937                 if (unlikely((mb->data_len + hw->rx_offset) >
1938                              rxq->mbuf_size)) {
1939                         /*
1940                          * This should not happen and the user has the
1941                          * responsibility of avoiding it. But we have
1942                          * to give some info about the error
1943                          */
1944                         RTE_LOG_DP(ERR, PMD,
1945                                 "mbuf overflow likely due to the RX offset.\n"
1946                                 "\t\tYour mbuf size should have extra space for"
1947                                 " RX offset=%u bytes.\n"
1948                                 "\t\tCurrently you just have %u bytes available"
1949                                 " but the received packet is %u bytes long",
1950                                 hw->rx_offset,
1951                                 rxq->mbuf_size - hw->rx_offset,
1952                                 mb->data_len);
1953                         return -EINVAL;
1954                 }
1955
1956                 /* Filling the received mbuff with packet info */
1957                 if (hw->rx_offset)
1958                         mb->data_off = RTE_PKTMBUF_HEADROOM + hw->rx_offset;
1959                 else
1960                         mb->data_off = RTE_PKTMBUF_HEADROOM +
1961                                        NFP_DESC_META_LEN(rxds);
1962
1963                 /* No scatter mode supported */
1964                 mb->nb_segs = 1;
1965                 mb->next = NULL;
1966
1967                 /* Checking the RSS flag */
1968                 nfp_net_set_hash(rxq, rxds, mb);
1969
1970                 /* Checking the checksum flag */
1971                 nfp_net_rx_cksum(rxq, rxds, mb);
1972
1973                 if ((rxds->rxd.flags & PCIE_DESC_RX_VLAN) &&
1974                     (hw->ctrl & NFP_NET_CFG_CTRL_RXVLAN)) {
1975                         mb->vlan_tci = rte_cpu_to_le_32(rxds->rxd.vlan);
1976                         mb->ol_flags |= PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED;
1977                 }
1978
1979                 /* Adding the mbuff to the mbuff array passed by the app */
1980                 rx_pkts[avail++] = mb;
1981
1982                 /* Now resetting and updating the descriptor */
1983                 rxds->vals[0] = 0;
1984                 rxds->vals[1] = 0;
1985                 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DMA_ADDR_DEFAULT(new_mb));
1986                 rxds->fld.dd = 0;
1987                 rxds->fld.dma_addr_hi = (dma_addr >> 32) & 0xff;
1988                 rxds->fld.dma_addr_lo = dma_addr & 0xffffffff;
1989
1990                 rxq->rd_p++;
1991                 if (unlikely(rxq->rd_p == rxq->rx_count)) /* wrapping?*/
1992                         rxq->rd_p = 0;
1993         }
1994
1995         if (nb_hold == 0)
1996                 return nb_hold;
1997
1998         PMD_RX_LOG(DEBUG, "RX  port_id=%u queue_id=%u, %d packets received\n",
1999                    (unsigned)rxq->port_id, (unsigned)rxq->qidx, nb_hold);
2000
2001         nb_hold += rxq->nb_rx_hold;
2002
2003         /*
2004          * FL descriptors needs to be written before incrementing the
2005          * FL queue WR pointer
2006          */
2007         rte_wmb();
2008         if (nb_hold > rxq->rx_free_thresh) {
2009                 PMD_RX_LOG(DEBUG, "port=%u queue=%u nb_hold=%u avail=%u\n",
2010                            (unsigned)rxq->port_id, (unsigned)rxq->qidx,
2011                            (unsigned)nb_hold, (unsigned)avail);
2012                 nfp_qcp_ptr_add(rxq->qcp_fl, NFP_QCP_WRITE_PTR, nb_hold);
2013                 nb_hold = 0;
2014         }
2015         rxq->nb_rx_hold = nb_hold;
2016
2017         return avail;
2018 }
2019
2020 /*
2021  * nfp_net_tx_free_bufs - Check for descriptors with a complete
2022  * status
2023  * @txq: TX queue to work with
2024  * Returns number of descriptors freed
2025  */
2026 int
2027 nfp_net_tx_free_bufs(struct nfp_net_txq *txq)
2028 {
2029         uint32_t qcp_rd_p;
2030         int todo;
2031
2032         PMD_TX_LOG(DEBUG, "queue %u. Check for descriptor with a complete"
2033                    " status\n", txq->qidx);
2034
2035         /* Work out how many packets have been sent */
2036         qcp_rd_p = nfp_qcp_read(txq->qcp_q, NFP_QCP_READ_PTR);
2037
2038         if (qcp_rd_p == txq->rd_p) {
2039                 PMD_TX_LOG(DEBUG, "queue %u: It seems harrier is not sending "
2040                            "packets (%u, %u)\n", txq->qidx,
2041                            qcp_rd_p, txq->rd_p);
2042                 return 0;
2043         }
2044
2045         if (qcp_rd_p > txq->rd_p)
2046                 todo = qcp_rd_p - txq->rd_p;
2047         else
2048                 todo = qcp_rd_p + txq->tx_count - txq->rd_p;
2049
2050         PMD_TX_LOG(DEBUG, "qcp_rd_p %u, txq->rd_p: %u, qcp->rd_p: %u\n",
2051                    qcp_rd_p, txq->rd_p, txq->rd_p);
2052
2053         if (todo == 0)
2054                 return todo;
2055
2056         txq->rd_p += todo;
2057         if (unlikely(txq->rd_p >= txq->tx_count))
2058                 txq->rd_p -= txq->tx_count;
2059
2060         return todo;
2061 }
2062
2063 /* Leaving always free descriptors for avoiding wrapping confusion */
2064 static inline
2065 uint32_t nfp_free_tx_desc(struct nfp_net_txq *txq)
2066 {
2067         if (txq->wr_p >= txq->rd_p)
2068                 return txq->tx_count - (txq->wr_p - txq->rd_p) - 8;
2069         else
2070                 return txq->rd_p - txq->wr_p - 8;
2071 }
2072
2073 /*
2074  * nfp_net_txq_full - Check if the TX queue free descriptors
2075  * is below tx_free_threshold
2076  *
2077  * @txq: TX queue to check
2078  *
2079  * This function uses the host copy* of read/write pointers
2080  */
2081 static inline
2082 uint32_t nfp_net_txq_full(struct nfp_net_txq *txq)
2083 {
2084         return (nfp_free_tx_desc(txq) < txq->tx_free_thresh);
2085 }
2086
2087 static uint16_t
2088 nfp_net_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
2089 {
2090         struct nfp_net_txq *txq;
2091         struct nfp_net_hw *hw;
2092         struct nfp_net_tx_desc *txds, txd;
2093         struct rte_mbuf *pkt;
2094         uint64_t dma_addr;
2095         int pkt_size, dma_size;
2096         uint16_t free_descs, issued_descs;
2097         struct rte_mbuf **lmbuf;
2098         int i;
2099
2100         txq = tx_queue;
2101         hw = txq->hw;
2102         txds = &txq->txds[txq->wr_p];
2103
2104         PMD_TX_LOG(DEBUG, "working for queue %u at pos %d and %u packets\n",
2105                    txq->qidx, txq->wr_p, nb_pkts);
2106
2107         if ((nfp_free_tx_desc(txq) < nb_pkts) || (nfp_net_txq_full(txq)))
2108                 nfp_net_tx_free_bufs(txq);
2109
2110         free_descs = (uint16_t)nfp_free_tx_desc(txq);
2111         if (unlikely(free_descs == 0))
2112                 return 0;
2113
2114         pkt = *tx_pkts;
2115
2116         i = 0;
2117         issued_descs = 0;
2118         PMD_TX_LOG(DEBUG, "queue: %u. Sending %u packets\n",
2119                    txq->qidx, nb_pkts);
2120         /* Sending packets */
2121         while ((i < nb_pkts) && free_descs) {
2122                 /* Grabbing the mbuf linked to the current descriptor */
2123                 lmbuf = &txq->txbufs[txq->wr_p].mbuf;
2124                 /* Warming the cache for releasing the mbuf later on */
2125                 RTE_MBUF_PREFETCH_TO_FREE(*lmbuf);
2126
2127                 pkt = *(tx_pkts + i);
2128
2129                 if (unlikely((pkt->nb_segs > 1) &&
2130                              !(hw->cap & NFP_NET_CFG_CTRL_GATHER))) {
2131                         PMD_INIT_LOG(INFO, "NFP_NET_CFG_CTRL_GATHER not set");
2132                         rte_panic("Multisegment packet unsupported\n");
2133                 }
2134
2135                 /* Checking if we have enough descriptors */
2136                 if (unlikely(pkt->nb_segs > free_descs))
2137                         goto xmit_end;
2138
2139                 /*
2140                  * Checksum and VLAN flags just in the first descriptor for a
2141                  * multisegment packet, but TSO info needs to be in all of them.
2142                  */
2143                 txd.data_len = pkt->pkt_len;
2144                 nfp_net_tx_tso(txq, &txd, pkt);
2145                 nfp_net_tx_cksum(txq, &txd, pkt);
2146
2147                 if ((pkt->ol_flags & PKT_TX_VLAN_PKT) &&
2148                     (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)) {
2149                         txd.flags |= PCIE_DESC_TX_VLAN;
2150                         txd.vlan = pkt->vlan_tci;
2151                 }
2152
2153                 /*
2154                  * mbuf data_len is the data in one segment and pkt_len data
2155                  * in the whole packet. When the packet is just one segment,
2156                  * then data_len = pkt_len
2157                  */
2158                 pkt_size = pkt->pkt_len;
2159
2160                 while (pkt) {
2161                         /* Copying TSO, VLAN and cksum info */
2162                         *txds = txd;
2163
2164                         /* Releasing mbuf used by this descriptor previously*/
2165                         if (*lmbuf)
2166                                 rte_pktmbuf_free_seg(*lmbuf);
2167
2168                         /*
2169                          * Linking mbuf with descriptor for being released
2170                          * next time descriptor is used
2171                          */
2172                         *lmbuf = pkt;
2173
2174                         dma_size = pkt->data_len;
2175                         dma_addr = rte_mbuf_data_dma_addr(pkt);
2176                         PMD_TX_LOG(DEBUG, "Working with mbuf at dma address:"
2177                                    "%" PRIx64 "\n", dma_addr);
2178
2179                         /* Filling descriptors fields */
2180                         txds->dma_len = dma_size;
2181                         txds->data_len = txd.data_len;
2182                         txds->dma_addr_hi = (dma_addr >> 32) & 0xff;
2183                         txds->dma_addr_lo = (dma_addr & 0xffffffff);
2184                         ASSERT(free_descs > 0);
2185                         free_descs--;
2186
2187                         txq->wr_p++;
2188                         if (unlikely(txq->wr_p == txq->tx_count)) /* wrapping?*/
2189                                 txq->wr_p = 0;
2190
2191                         pkt_size -= dma_size;
2192                         if (!pkt_size)
2193                                 /* End of packet */
2194                                 txds->offset_eop |= PCIE_DESC_TX_EOP;
2195                         else
2196                                 txds->offset_eop &= PCIE_DESC_TX_OFFSET_MASK;
2197
2198                         pkt = pkt->next;
2199                         /* Referencing next free TX descriptor */
2200                         txds = &txq->txds[txq->wr_p];
2201                         lmbuf = &txq->txbufs[txq->wr_p].mbuf;
2202                         issued_descs++;
2203                 }
2204                 i++;
2205         }
2206
2207 xmit_end:
2208         /* Increment write pointers. Force memory write before we let HW know */
2209         rte_wmb();
2210         nfp_qcp_ptr_add(txq->qcp_q, NFP_QCP_WRITE_PTR, issued_descs);
2211
2212         return i;
2213 }
2214
2215 static void
2216 nfp_net_vlan_offload_set(struct rte_eth_dev *dev, int mask)
2217 {
2218         uint32_t new_ctrl, update;
2219         struct nfp_net_hw *hw;
2220
2221         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2222         new_ctrl = 0;
2223
2224         if ((mask & ETH_VLAN_FILTER_OFFLOAD) ||
2225             (mask & ETH_VLAN_EXTEND_OFFLOAD))
2226                 RTE_LOG(INFO, PMD, "No support for ETH_VLAN_FILTER_OFFLOAD or"
2227                         " ETH_VLAN_EXTEND_OFFLOAD");
2228
2229         /* Enable vlan strip if it is not configured yet */
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;
2233
2234         /* Disable vlan strip just if it is configured */
2235         if (!(mask & ETH_VLAN_STRIP_OFFLOAD) &&
2236             (hw->ctrl & NFP_NET_CFG_CTRL_RXVLAN))
2237                 new_ctrl = hw->ctrl & ~NFP_NET_CFG_CTRL_RXVLAN;
2238
2239         if (new_ctrl == 0)
2240                 return;
2241
2242         update = NFP_NET_CFG_UPDATE_GEN;
2243
2244         if (nfp_net_reconfig(hw, new_ctrl, update) < 0)
2245                 return;
2246
2247         hw->ctrl = new_ctrl;
2248 }
2249
2250 /* Update Redirection Table(RETA) of Receive Side Scaling of Ethernet device */
2251 static int
2252 nfp_net_reta_update(struct rte_eth_dev *dev,
2253                     struct rte_eth_rss_reta_entry64 *reta_conf,
2254                     uint16_t reta_size)
2255 {
2256         uint32_t reta, mask;
2257         int i, j;
2258         int idx, shift;
2259         uint32_t update;
2260         struct nfp_net_hw *hw =
2261                 NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2262
2263         if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
2264                 return -EINVAL;
2265
2266         if (reta_size != NFP_NET_CFG_RSS_ITBL_SZ) {
2267                 RTE_LOG(ERR, PMD, "The size of hash lookup table configured "
2268                         "(%d) doesn't match the number hardware can supported "
2269                         "(%d)\n", reta_size, NFP_NET_CFG_RSS_ITBL_SZ);
2270                 return -EINVAL;
2271         }
2272
2273         /*
2274          * Update Redirection Table. There are 128 8bit-entries which can be
2275          * manage as 32 32bit-entries
2276          */
2277         for (i = 0; i < reta_size; i += 4) {
2278                 /* Handling 4 RSS entries per loop */
2279                 idx = i / RTE_RETA_GROUP_SIZE;
2280                 shift = i % RTE_RETA_GROUP_SIZE;
2281                 mask = (uint8_t)((reta_conf[idx].mask >> shift) & 0xF);
2282
2283                 if (!mask)
2284                         continue;
2285
2286                 reta = 0;
2287                 /* If all 4 entries were set, don't need read RETA register */
2288                 if (mask != 0xF)
2289                         reta = nn_cfg_readl(hw, NFP_NET_CFG_RSS_ITBL + i);
2290
2291                 for (j = 0; j < 4; j++) {
2292                         if (!(mask & (0x1 << j)))
2293                                 continue;
2294                         if (mask != 0xF)
2295                                 /* Clearing the entry bits */
2296                                 reta &= ~(0xFF << (8 * j));
2297                         reta |= reta_conf[idx].reta[shift + j] << (8 * j);
2298                 }
2299                 nn_cfg_writel(hw, NFP_NET_CFG_RSS_ITBL + (idx * 64) + shift,
2300                               reta);
2301         }
2302
2303         update = NFP_NET_CFG_UPDATE_RSS;
2304
2305         if (nfp_net_reconfig(hw, hw->ctrl, update) < 0)
2306                 return -EIO;
2307
2308         return 0;
2309 }
2310
2311  /* Query Redirection Table(RETA) of Receive Side Scaling of Ethernet device. */
2312 static int
2313 nfp_net_reta_query(struct rte_eth_dev *dev,
2314                    struct rte_eth_rss_reta_entry64 *reta_conf,
2315                    uint16_t reta_size)
2316 {
2317         uint8_t i, j, mask;
2318         int idx, shift;
2319         uint32_t reta;
2320         struct nfp_net_hw *hw;
2321
2322         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2323
2324         if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
2325                 return -EINVAL;
2326
2327         if (reta_size != NFP_NET_CFG_RSS_ITBL_SZ) {
2328                 RTE_LOG(ERR, PMD, "The size of hash lookup table configured "
2329                         "(%d) doesn't match the number hardware can supported "
2330                         "(%d)\n", reta_size, NFP_NET_CFG_RSS_ITBL_SZ);
2331                 return -EINVAL;
2332         }
2333
2334         /*
2335          * Reading Redirection Table. There are 128 8bit-entries which can be
2336          * manage as 32 32bit-entries
2337          */
2338         for (i = 0; i < reta_size; i += 4) {
2339                 /* Handling 4 RSS entries per loop */
2340                 idx = i / RTE_RETA_GROUP_SIZE;
2341                 shift = i % RTE_RETA_GROUP_SIZE;
2342                 mask = (uint8_t)((reta_conf[idx].mask >> shift) & 0xF);
2343
2344                 if (!mask)
2345                         continue;
2346
2347                 reta = nn_cfg_readl(hw, NFP_NET_CFG_RSS_ITBL + (idx * 64) +
2348                                     shift);
2349                 for (j = 0; j < 4; j++) {
2350                         if (!(mask & (0x1 << j)))
2351                                 continue;
2352                         reta_conf->reta[shift + j] =
2353                                 (uint8_t)((reta >> (8 * j)) & 0xF);
2354                 }
2355         }
2356         return 0;
2357 }
2358
2359 static int
2360 nfp_net_rss_hash_update(struct rte_eth_dev *dev,
2361                         struct rte_eth_rss_conf *rss_conf)
2362 {
2363         uint32_t update;
2364         uint32_t cfg_rss_ctrl = 0;
2365         uint8_t key;
2366         uint64_t rss_hf;
2367         int i;
2368         struct nfp_net_hw *hw;
2369
2370         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2371
2372         rss_hf = rss_conf->rss_hf;
2373
2374         /* Checking if RSS is enabled */
2375         if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS)) {
2376                 if (rss_hf != 0) { /* Enable RSS? */
2377                         RTE_LOG(ERR, PMD, "RSS unsupported\n");
2378                         return -EINVAL;
2379                 }
2380                 return 0; /* Nothing to do */
2381         }
2382
2383         if (rss_conf->rss_key_len > NFP_NET_CFG_RSS_KEY_SZ) {
2384                 RTE_LOG(ERR, PMD, "hash key too long\n");
2385                 return -EINVAL;
2386         }
2387
2388         if (rss_hf & ETH_RSS_IPV4)
2389                 cfg_rss_ctrl |= NFP_NET_CFG_RSS_IPV4 |
2390                                 NFP_NET_CFG_RSS_IPV4_TCP |
2391                                 NFP_NET_CFG_RSS_IPV4_UDP;
2392
2393         if (rss_hf & ETH_RSS_IPV6)
2394                 cfg_rss_ctrl |= NFP_NET_CFG_RSS_IPV6 |
2395                                 NFP_NET_CFG_RSS_IPV6_TCP |
2396                                 NFP_NET_CFG_RSS_IPV6_UDP;
2397
2398         cfg_rss_ctrl |= NFP_NET_CFG_RSS_MASK;
2399         cfg_rss_ctrl |= NFP_NET_CFG_RSS_TOEPLITZ;
2400
2401         /* configuring where to apply the RSS hash */
2402         nn_cfg_writel(hw, NFP_NET_CFG_RSS_CTRL, cfg_rss_ctrl);
2403
2404         /* Writing the key byte a byte */
2405         for (i = 0; i < rss_conf->rss_key_len; i++) {
2406                 memcpy(&key, &rss_conf->rss_key[i], 1);
2407                 nn_cfg_writeb(hw, NFP_NET_CFG_RSS_KEY + i, key);
2408         }
2409
2410         /* Writing the key size */
2411         nn_cfg_writeb(hw, NFP_NET_CFG_RSS_KEY_SZ, rss_conf->rss_key_len);
2412
2413         update = NFP_NET_CFG_UPDATE_RSS;
2414
2415         if (nfp_net_reconfig(hw, hw->ctrl, update) < 0)
2416                 return -EIO;
2417
2418         return 0;
2419 }
2420
2421 static int
2422 nfp_net_rss_hash_conf_get(struct rte_eth_dev *dev,
2423                           struct rte_eth_rss_conf *rss_conf)
2424 {
2425         uint64_t rss_hf;
2426         uint32_t cfg_rss_ctrl;
2427         uint8_t key;
2428         int i;
2429         struct nfp_net_hw *hw;
2430
2431         hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2432
2433         if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
2434                 return -EINVAL;
2435
2436         rss_hf = rss_conf->rss_hf;
2437         cfg_rss_ctrl = nn_cfg_readl(hw, NFP_NET_CFG_RSS_CTRL);
2438
2439         if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV4)
2440                 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP | ETH_RSS_NONFRAG_IPV4_UDP;
2441
2442         if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV4_TCP)
2443                 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
2444
2445         if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV6_TCP)
2446                 rss_hf |= ETH_RSS_NONFRAG_IPV6_TCP;
2447
2448         if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV4_UDP)
2449                 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP;
2450
2451         if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV6_UDP)
2452                 rss_hf |= ETH_RSS_NONFRAG_IPV6_UDP;
2453
2454         if (cfg_rss_ctrl & NFP_NET_CFG_RSS_IPV6)
2455                 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP | ETH_RSS_NONFRAG_IPV6_UDP;
2456
2457         /* Reading the key size */
2458         rss_conf->rss_key_len = nn_cfg_readl(hw, NFP_NET_CFG_RSS_KEY_SZ);
2459
2460         /* Reading the key byte a byte */
2461         for (i = 0; i < rss_conf->rss_key_len; i++) {
2462                 key = nn_cfg_readb(hw, NFP_NET_CFG_RSS_KEY + i);
2463                 memcpy(&rss_conf->rss_key[i], &key, 1);
2464         }
2465
2466         return 0;
2467 }
2468
2469 /* Initialise and register driver with DPDK Application */
2470 static const struct eth_dev_ops nfp_net_eth_dev_ops = {
2471         .dev_configure          = nfp_net_configure,
2472         .dev_start              = nfp_net_start,
2473         .dev_stop               = nfp_net_stop,
2474         .dev_close              = nfp_net_close,
2475         .promiscuous_enable     = nfp_net_promisc_enable,
2476         .promiscuous_disable    = nfp_net_promisc_disable,
2477         .link_update            = nfp_net_link_update,
2478         .stats_get              = nfp_net_stats_get,
2479         .stats_reset            = nfp_net_stats_reset,
2480         .dev_infos_get          = nfp_net_infos_get,
2481         .dev_supported_ptypes_get = nfp_net_supported_ptypes_get,
2482         .mtu_set                = nfp_net_dev_mtu_set,
2483         .vlan_offload_set       = nfp_net_vlan_offload_set,
2484         .reta_update            = nfp_net_reta_update,
2485         .reta_query             = nfp_net_reta_query,
2486         .rss_hash_update        = nfp_net_rss_hash_update,
2487         .rss_hash_conf_get      = nfp_net_rss_hash_conf_get,
2488         .rx_queue_setup         = nfp_net_rx_queue_setup,
2489         .rx_queue_release       = nfp_net_rx_queue_release,
2490         .rx_queue_count         = nfp_net_rx_queue_count,
2491         .tx_queue_setup         = nfp_net_tx_queue_setup,
2492         .tx_queue_release       = nfp_net_tx_queue_release,
2493         .rx_queue_intr_enable   = nfp_rx_queue_intr_enable,
2494         .rx_queue_intr_disable  = nfp_rx_queue_intr_disable,
2495 };
2496
2497 /*
2498  * All eth_dev created got its private data, but before nfp_net_init, that
2499  * private data is referencing private data for all the PF ports. This is due
2500  * to how the vNIC bars are mapped based on first port, so all ports need info
2501  * about port 0 private data. Inside nfp_net_init the private data pointer is
2502  * changed to the right address for each port once the bars have been mapped.
2503  *
2504  * This functions helps to find out which port and therefore which offset
2505  * inside the private data array to use.
2506  */
2507 static int
2508 get_pf_port_number(char *name)
2509 {
2510         char *pf_str = name;
2511         int size = 0;
2512
2513         while ((*pf_str != '_') && (*pf_str != '\0') && (size++ < 30))
2514                 pf_str++;
2515
2516         if (size == 30)
2517                 /*
2518                  * This should not happen at all and it would mean major
2519                  * implementation fault.
2520                  */
2521                 rte_panic("nfp_net: problem with pf device name\n");
2522
2523         /* Expecting _portX with X within [0,7] */
2524         pf_str += 5;
2525
2526         return (int)strtol(pf_str, NULL, 10);
2527 }
2528
2529 static int
2530 nfp_net_init(struct rte_eth_dev *eth_dev)
2531 {
2532         struct rte_pci_device *pci_dev;
2533         struct nfp_net_hw *hw, *hwport0;
2534
2535         uint64_t tx_bar_off = 0, rx_bar_off = 0;
2536         uint32_t start_q;
2537         int stride = 4;
2538
2539         nspu_desc_t *nspu_desc = NULL;
2540         uint64_t bar_offset;
2541         int port = 0;
2542
2543         PMD_INIT_FUNC_TRACE();
2544
2545         pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
2546
2547         if ((pci_dev->id.device_id == PCI_DEVICE_ID_NFP4000_PF_NIC) ||
2548             (pci_dev->id.device_id == PCI_DEVICE_ID_NFP6000_PF_NIC)) {
2549                 port = get_pf_port_number(eth_dev->data->name);
2550                 if (port < 0 || port > 7) {
2551                         RTE_LOG(ERR, PMD, "Port value is wrong\n");
2552                         return -ENODEV;
2553                 }
2554
2555                 PMD_INIT_LOG(DEBUG, "Working with PF port value %d\n", port);
2556
2557                 /* This points to port 0 private data */
2558                 hwport0 = NFP_NET_DEV_PRIVATE_TO_HW(eth_dev->data->dev_private);
2559
2560                 /* This points to the specific port private data */
2561                 hw = &hwport0[port];
2562                 hw->pf_port_idx = port;
2563         } else {
2564                 hw = NFP_NET_DEV_PRIVATE_TO_HW(eth_dev->data->dev_private);
2565                 hwport0 = 0;
2566         }
2567
2568         eth_dev->dev_ops = &nfp_net_eth_dev_ops;
2569         eth_dev->rx_pkt_burst = &nfp_net_recv_pkts;
2570         eth_dev->tx_pkt_burst = &nfp_net_xmit_pkts;
2571
2572         /* For secondary processes, the primary has done all the work */
2573         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
2574                 return 0;
2575
2576         rte_eth_copy_pci_info(eth_dev, pci_dev);
2577         /* hotplug is not possible with multiport PF */
2578         if (!hw->pf_multiport_enabled)
2579                 eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
2580
2581         hw->device_id = pci_dev->id.device_id;
2582         hw->vendor_id = pci_dev->id.vendor_id;
2583         hw->subsystem_device_id = pci_dev->id.subsystem_device_id;
2584         hw->subsystem_vendor_id = pci_dev->id.subsystem_vendor_id;
2585
2586         PMD_INIT_LOG(DEBUG, "nfp_net: device (%u:%u) %u:%u:%u:%u",
2587                      pci_dev->id.vendor_id, pci_dev->id.device_id,
2588                      pci_dev->addr.domain, pci_dev->addr.bus,
2589                      pci_dev->addr.devid, pci_dev->addr.function);
2590
2591         hw->ctrl_bar = (uint8_t *)pci_dev->mem_resource[0].addr;
2592         if (hw->ctrl_bar == NULL) {
2593                 RTE_LOG(ERR, PMD,
2594                         "hw->ctrl_bar is NULL. BAR0 not configured\n");
2595                 return -ENODEV;
2596         }
2597
2598         if (hw->is_pf && port == 0) {
2599                 nspu_desc = hw->nspu_desc;
2600
2601                 if (nfp_nsp_map_ctrl_bar(nspu_desc, &bar_offset) != 0) {
2602                         /*
2603                          * A firmware should be there after PF probe so this
2604                          * should not happen.
2605                          */
2606                         RTE_LOG(ERR, PMD, "PF BAR symbol resolution failed\n");
2607                         return -ENODEV;
2608                 }
2609
2610                 /* vNIC PF control BAR is a subset of PF PCI device BAR */
2611                 hw->ctrl_bar += bar_offset;
2612                 PMD_INIT_LOG(DEBUG, "ctrl bar: %p\n", hw->ctrl_bar);
2613         }
2614
2615         if (port > 0) {
2616                 if (!hwport0->ctrl_bar)
2617                         return -ENODEV;
2618
2619                 /* address based on port0 offset */
2620                 hw->ctrl_bar = hwport0->ctrl_bar +
2621                                (port * NFP_PF_CSR_SLICE_SIZE);
2622         }
2623
2624         PMD_INIT_LOG(DEBUG, "ctrl bar: %p\n", hw->ctrl_bar);
2625
2626         hw->max_rx_queues = nn_cfg_readl(hw, NFP_NET_CFG_MAX_RXRINGS);
2627         hw->max_tx_queues = nn_cfg_readl(hw, NFP_NET_CFG_MAX_TXRINGS);
2628
2629         /* Work out where in the BAR the queues start. */
2630         switch (pci_dev->id.device_id) {
2631         case PCI_DEVICE_ID_NFP4000_PF_NIC:
2632         case PCI_DEVICE_ID_NFP6000_PF_NIC:
2633         case PCI_DEVICE_ID_NFP6000_VF_NIC:
2634                 start_q = nn_cfg_readl(hw, NFP_NET_CFG_START_TXQ);
2635                 tx_bar_off = NFP_PCIE_QUEUE(start_q);
2636                 start_q = nn_cfg_readl(hw, NFP_NET_CFG_START_RXQ);
2637                 rx_bar_off = NFP_PCIE_QUEUE(start_q);
2638                 break;
2639         default:
2640                 RTE_LOG(ERR, PMD, "nfp_net: no device ID matching\n");
2641                 return -ENODEV;
2642         }
2643
2644         PMD_INIT_LOG(DEBUG, "tx_bar_off: 0x%" PRIx64 "\n", tx_bar_off);
2645         PMD_INIT_LOG(DEBUG, "rx_bar_off: 0x%" PRIx64 "\n", rx_bar_off);
2646
2647         if (hw->is_pf && port == 0) {
2648                 /* configure access to tx/rx vNIC BARs */
2649                 nfp_nsp_map_queues_bar(nspu_desc, &bar_offset);
2650                 PMD_INIT_LOG(DEBUG, "tx/rx bar_offset: %" PRIx64 "\n",
2651                                     bar_offset);
2652                 hwport0->hw_queues = (uint8_t *)pci_dev->mem_resource[0].addr;
2653
2654                 /* vNIC PF tx/rx BARs are a subset of PF PCI device */
2655                 hwport0->hw_queues += bar_offset;
2656         }
2657
2658         if (hw->is_pf) {
2659                 hw->tx_bar = hwport0->hw_queues + tx_bar_off;
2660                 hw->rx_bar = hwport0->hw_queues + rx_bar_off;
2661                 eth_dev->data->dev_private = hw;
2662         } else {
2663                 hw->tx_bar = (uint8_t *)pci_dev->mem_resource[2].addr +
2664                              tx_bar_off;
2665                 hw->rx_bar = (uint8_t *)pci_dev->mem_resource[2].addr +
2666                              rx_bar_off;
2667         }
2668
2669         PMD_INIT_LOG(DEBUG, "ctrl_bar: %p, tx_bar: %p, rx_bar: %p",
2670                      hw->ctrl_bar, hw->tx_bar, hw->rx_bar);
2671
2672         nfp_net_cfg_queue_setup(hw);
2673
2674         /* Get some of the read-only fields from the config BAR */
2675         hw->ver = nn_cfg_readl(hw, NFP_NET_CFG_VERSION);
2676         hw->cap = nn_cfg_readl(hw, NFP_NET_CFG_CAP);
2677         hw->max_mtu = nn_cfg_readl(hw, NFP_NET_CFG_MAX_MTU);
2678         hw->mtu = hw->max_mtu;
2679
2680         if (NFD_CFG_MAJOR_VERSION_of(hw->ver) < 2)
2681                 hw->rx_offset = NFP_NET_RX_OFFSET;
2682         else
2683                 hw->rx_offset = nn_cfg_readl(hw, NFP_NET_CFG_RX_OFFSET_ADDR);
2684
2685         PMD_INIT_LOG(INFO, "VER: %#x, Maximum supported MTU: %d",
2686                      hw->ver, hw->max_mtu);
2687         PMD_INIT_LOG(INFO, "CAP: %#x, %s%s%s%s%s%s%s%s%s", hw->cap,
2688                      hw->cap & NFP_NET_CFG_CTRL_PROMISC ? "PROMISC " : "",
2689                      hw->cap & NFP_NET_CFG_CTRL_RXCSUM  ? "RXCSUM "  : "",
2690                      hw->cap & NFP_NET_CFG_CTRL_TXCSUM  ? "TXCSUM "  : "",
2691                      hw->cap & NFP_NET_CFG_CTRL_RXVLAN  ? "RXVLAN "  : "",
2692                      hw->cap & NFP_NET_CFG_CTRL_TXVLAN  ? "TXVLAN "  : "",
2693                      hw->cap & NFP_NET_CFG_CTRL_SCATTER ? "SCATTER " : "",
2694                      hw->cap & NFP_NET_CFG_CTRL_GATHER  ? "GATHER "  : "",
2695                      hw->cap & NFP_NET_CFG_CTRL_LSO     ? "TSO "     : "",
2696                      hw->cap & NFP_NET_CFG_CTRL_RSS     ? "RSS "     : "");
2697
2698         hw->ctrl = 0;
2699
2700         hw->stride_rx = stride;
2701         hw->stride_tx = stride;
2702
2703         PMD_INIT_LOG(INFO, "max_rx_queues: %u, max_tx_queues: %u",
2704                      hw->max_rx_queues, hw->max_tx_queues);
2705
2706         /* Initializing spinlock for reconfigs */
2707         rte_spinlock_init(&hw->reconfig_lock);
2708
2709         /* Allocating memory for mac addr */
2710         eth_dev->data->mac_addrs = rte_zmalloc("mac_addr", ETHER_ADDR_LEN, 0);
2711         if (eth_dev->data->mac_addrs == NULL) {
2712                 PMD_INIT_LOG(ERR, "Failed to space for MAC address");
2713                 return -ENOMEM;
2714         }
2715
2716         nfp_net_read_mac(hw);
2717
2718         if (!is_valid_assigned_ether_addr((struct ether_addr *)&hw->mac_addr)) {
2719                 /* Using random mac addresses for VFs */
2720                 eth_random_addr(&hw->mac_addr[0]);
2721                 nfp_net_write_mac(hw, (uint8_t *)&hw->mac_addr);
2722         }
2723
2724         /* Copying mac address to DPDK eth_dev struct */
2725         ether_addr_copy((struct ether_addr *)hw->mac_addr,
2726                         &eth_dev->data->mac_addrs[0]);
2727
2728         PMD_INIT_LOG(INFO, "port %d VendorID=0x%x DeviceID=0x%x "
2729                      "mac=%02x:%02x:%02x:%02x:%02x:%02x",
2730                      eth_dev->data->port_id, pci_dev->id.vendor_id,
2731                      pci_dev->id.device_id,
2732                      hw->mac_addr[0], hw->mac_addr[1], hw->mac_addr[2],
2733                      hw->mac_addr[3], hw->mac_addr[4], hw->mac_addr[5]);
2734
2735         /* Registering LSC interrupt handler */
2736         rte_intr_callback_register(&pci_dev->intr_handle,
2737                                    nfp_net_dev_interrupt_handler,
2738                                    (void *)eth_dev);
2739
2740         /* Telling the firmware about the LSC interrupt entry */
2741         nn_cfg_writeb(hw, NFP_NET_CFG_LSC, NFP_NET_IRQ_LSC_IDX);
2742
2743         /* Recording current stats counters values */
2744         nfp_net_stats_reset(eth_dev);
2745
2746         return 0;
2747 }
2748
2749 static int
2750 nfp_pf_create_dev(struct rte_pci_device *dev, int port, int ports,
2751                   nfpu_desc_t *nfpu_desc, void **priv)
2752 {
2753         struct rte_eth_dev *eth_dev;
2754         struct nfp_net_hw *hw;
2755         char *port_name;
2756         int ret;
2757
2758         port_name = rte_zmalloc("nfp_pf_port_name", 100, 0);
2759         if (!port_name)
2760                 return -ENOMEM;
2761
2762         if (ports > 1)
2763                 sprintf(port_name, "%s_port%d", dev->device.name, port);
2764         else
2765                 sprintf(port_name, "%s", dev->device.name);
2766
2767         eth_dev = rte_eth_dev_allocate(port_name);
2768         if (!eth_dev)
2769                 return -ENOMEM;
2770
2771         if (port == 0) {
2772                 *priv = rte_zmalloc(port_name,
2773                                     sizeof(struct nfp_net_adapter) * ports,
2774                                     RTE_CACHE_LINE_SIZE);
2775                 if (!*priv) {
2776                         rte_eth_dev_release_port(eth_dev);
2777                         return -ENOMEM;
2778                 }
2779         }
2780
2781         eth_dev->data->dev_private = *priv;
2782
2783         /*
2784          * dev_private pointing to port0 dev_private because we need
2785          * to configure vNIC bars based on port0 at nfp_net_init.
2786          * Then dev_private is adjusted per port.
2787          */
2788         hw = (struct nfp_net_hw *)(eth_dev->data->dev_private) + port;
2789         hw->nspu_desc = nfpu_desc->nspu;
2790         hw->nfpu_desc = nfpu_desc;
2791         hw->is_pf = 1;
2792         if (ports > 1)
2793                 hw->pf_multiport_enabled = 1;
2794
2795         eth_dev->device = &dev->device;
2796         rte_eth_copy_pci_info(eth_dev, dev);
2797
2798         ret = nfp_net_init(eth_dev);
2799
2800         if (ret)
2801                 rte_eth_dev_release_port(eth_dev);
2802
2803         rte_free(port_name);
2804
2805         return ret;
2806 }
2807
2808 static int nfp_pf_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2809                             struct rte_pci_device *dev)
2810 {
2811         nfpu_desc_t *nfpu_desc;
2812         nspu_desc_t *nspu_desc;
2813         uint64_t offset_symbol;
2814         uint8_t *bar_offset;
2815         int major, minor;
2816         int total_ports;
2817         void *priv = 0;
2818         int ret = -ENODEV;
2819         int i;
2820
2821         if (!dev)
2822                 return ret;
2823
2824         nfpu_desc = rte_malloc("nfp nfpu", sizeof(nfpu_desc_t), 0);
2825         if (!nfpu_desc)
2826                 return -ENOMEM;
2827
2828         if (nfpu_open(dev, nfpu_desc, 0) < 0) {
2829                 RTE_LOG(ERR, PMD,
2830                         "nfpu_open failed\n");
2831                 goto nfpu_error;
2832         }
2833
2834         nspu_desc = nfpu_desc->nspu;
2835
2836
2837         /* Check NSP ABI version */
2838         if (nfp_nsp_get_abi_version(nspu_desc, &major, &minor) < 0) {
2839                 RTE_LOG(INFO, PMD, "NFP NSP not present\n");
2840                 goto error;
2841         }
2842         PMD_INIT_LOG(INFO, "nspu ABI version: %d.%d\n", major, minor);
2843
2844         if ((major == 0) && (minor < 20)) {
2845                 RTE_LOG(INFO, PMD, "NFP NSP ABI version too old. Required 0.20 or higher\n");
2846                 goto error;
2847         }
2848
2849         ret = nfp_nsp_fw_setup(nspu_desc, "nfd_cfg_pf0_num_ports",
2850                                &offset_symbol);
2851         if (ret)
2852                 goto error;
2853
2854         bar_offset = (uint8_t *)dev->mem_resource[0].addr;
2855         bar_offset += offset_symbol;
2856         total_ports = (uint32_t)*bar_offset;
2857         PMD_INIT_LOG(INFO, "Total pf ports: %d\n", total_ports);
2858
2859         if (total_ports <= 0 || total_ports > 8) {
2860                 RTE_LOG(ERR, PMD, "nfd_cfg_pf0_num_ports symbol with wrong value");
2861                 ret = -ENODEV;
2862                 goto error;
2863         }
2864
2865         for (i = 0; i < total_ports; i++) {
2866                 ret = nfp_pf_create_dev(dev, i, total_ports, nfpu_desc, &priv);
2867                 if (ret)
2868                         goto error;
2869         }
2870
2871         return 0;
2872
2873 error:
2874         nfpu_close(nfpu_desc);
2875 nfpu_error:
2876         rte_free(nfpu_desc);
2877
2878         return ret;
2879 }
2880
2881 static const struct rte_pci_id pci_id_nfp_pf_net_map[] = {
2882         {
2883                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETRONOME,
2884                                PCI_DEVICE_ID_NFP4000_PF_NIC)
2885         },
2886         {
2887                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETRONOME,
2888                                PCI_DEVICE_ID_NFP6000_PF_NIC)
2889         },
2890         {
2891                 .vendor_id = 0,
2892         },
2893 };
2894
2895 static const struct rte_pci_id pci_id_nfp_vf_net_map[] = {
2896         {
2897                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETRONOME,
2898                                PCI_DEVICE_ID_NFP6000_VF_NIC)
2899         },
2900         {
2901                 .vendor_id = 0,
2902         },
2903 };
2904
2905 static int eth_nfp_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
2906         struct rte_pci_device *pci_dev)
2907 {
2908         return rte_eth_dev_pci_generic_probe(pci_dev,
2909                 sizeof(struct nfp_net_adapter), nfp_net_init);
2910 }
2911
2912 static int eth_nfp_pci_remove(struct rte_pci_device *pci_dev)
2913 {
2914         return rte_eth_dev_pci_generic_remove(pci_dev, NULL);
2915 }
2916
2917 static struct rte_pci_driver rte_nfp_net_pf_pmd = {
2918         .id_table = pci_id_nfp_pf_net_map,
2919         .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
2920         .probe = nfp_pf_pci_probe,
2921         .remove = eth_nfp_pci_remove,
2922 };
2923
2924 static struct rte_pci_driver rte_nfp_net_vf_pmd = {
2925         .id_table = pci_id_nfp_vf_net_map,
2926         .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
2927         .probe = eth_nfp_pci_probe,
2928         .remove = eth_nfp_pci_remove,
2929 };
2930
2931 RTE_PMD_REGISTER_PCI(net_nfp_pf, rte_nfp_net_pf_pmd);
2932 RTE_PMD_REGISTER_PCI(net_nfp_vf, rte_nfp_net_vf_pmd);
2933 RTE_PMD_REGISTER_PCI_TABLE(net_nfp_pf, pci_id_nfp_pf_net_map);
2934 RTE_PMD_REGISTER_PCI_TABLE(net_nfp_vf, pci_id_nfp_vf_net_map);
2935 RTE_PMD_REGISTER_KMOD_DEP(net_nfp_pf, "* igb_uio | uio_pci_generic | vfio");
2936 RTE_PMD_REGISTER_KMOD_DEP(net_nfp_vf, "* igb_uio | uio_pci_generic | vfio");
2937
2938 /*
2939  * Local variables:
2940  * c-file-style: "Linux"
2941  * indent-tabs-mode: t
2942  * End:
2943  */