net/bnxt: support Rx queue count
[dpdk.git] / drivers / net / bnxt / bnxt_rxr.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) Broadcom Limited.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Broadcom Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33
34 #include <inttypes.h>
35 #include <stdbool.h>
36
37 #include <rte_bitmap.h>
38 #include <rte_byteorder.h>
39 #include <rte_malloc.h>
40 #include <rte_memory.h>
41
42 #include "bnxt.h"
43 #include "bnxt_cpr.h"
44 #include "bnxt_ring.h"
45 #include "bnxt_rxr.h"
46 #include "bnxt_rxq.h"
47 #include "hsi_struct_def_dpdk.h"
48
49 /*
50  * RX Ring handling
51  */
52
53 static inline struct rte_mbuf *__bnxt_alloc_rx_data(struct rte_mempool *mb)
54 {
55         struct rte_mbuf *data;
56
57         data = rte_mbuf_raw_alloc(mb);
58
59         return data;
60 }
61
62 static inline int bnxt_alloc_rx_data(struct bnxt_rx_queue *rxq,
63                                      struct bnxt_rx_ring_info *rxr,
64                                      uint16_t prod)
65 {
66         struct rx_prod_pkt_bd *rxbd = &rxr->rx_desc_ring[prod];
67         struct bnxt_sw_rx_bd *rx_buf = &rxr->rx_buf_ring[prod];
68         struct rte_mbuf *data;
69
70         data = __bnxt_alloc_rx_data(rxq->mb_pool);
71         if (!data) {
72                 rte_atomic64_inc(&rxq->bp->rx_mbuf_alloc_fail);
73                 return -ENOMEM;
74         }
75
76         rx_buf->mbuf = data;
77
78         rxbd->addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR(rx_buf->mbuf));
79
80         return 0;
81 }
82
83 static inline int bnxt_alloc_ag_data(struct bnxt_rx_queue *rxq,
84                                      struct bnxt_rx_ring_info *rxr,
85                                      uint16_t prod)
86 {
87         struct rx_prod_pkt_bd *rxbd = &rxr->ag_desc_ring[prod];
88         struct bnxt_sw_rx_bd *rx_buf = &rxr->ag_buf_ring[prod];
89         struct rte_mbuf *data;
90
91         data = __bnxt_alloc_rx_data(rxq->mb_pool);
92         if (!data) {
93                 rte_atomic64_inc(&rxq->bp->rx_mbuf_alloc_fail);
94                 return -ENOMEM;
95         }
96
97         if (rxbd == NULL)
98                 RTE_LOG(ERR, PMD, "Jumbo Frame. rxbd is NULL\n");
99         if (rx_buf == NULL)
100                 RTE_LOG(ERR, PMD, "Jumbo Frame. rx_buf is NULL\n");
101
102
103         rx_buf->mbuf = data;
104
105         rxbd->addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR(rx_buf->mbuf));
106
107         return 0;
108 }
109
110 static inline void bnxt_reuse_rx_mbuf(struct bnxt_rx_ring_info *rxr,
111                                struct rte_mbuf *mbuf)
112 {
113         uint16_t prod = RING_NEXT(rxr->rx_ring_struct, rxr->rx_prod);
114         struct bnxt_sw_rx_bd *prod_rx_buf;
115         struct rx_prod_pkt_bd *prod_bd;
116
117         prod_rx_buf = &rxr->rx_buf_ring[prod];
118
119         RTE_ASSERT(prod_rx_buf->mbuf == NULL);
120         RTE_ASSERT(mbuf != NULL);
121
122         prod_rx_buf->mbuf = mbuf;
123
124         prod_bd = &rxr->rx_desc_ring[prod];
125
126         prod_bd->addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR(mbuf));
127
128         rxr->rx_prod = prod;
129 }
130
131 #ifdef BNXT_DEBUG
132 static void bnxt_reuse_ag_mbuf(struct bnxt_rx_ring_info *rxr, uint16_t cons,
133                                struct rte_mbuf *mbuf)
134 {
135         uint16_t prod = rxr->ag_prod;
136         struct bnxt_sw_rx_bd *prod_rx_buf;
137         struct rx_prod_pkt_bd *prod_bd, *cons_bd;
138
139         prod_rx_buf = &rxr->ag_buf_ring[prod];
140
141         prod_rx_buf->mbuf = mbuf;
142
143         prod_bd = &rxr->ag_desc_ring[prod];
144         cons_bd = &rxr->ag_desc_ring[cons];
145
146         prod_bd->addr = cons_bd->addr;
147 }
148 #endif
149
150 static inline
151 struct rte_mbuf *bnxt_consume_rx_buf(struct bnxt_rx_ring_info *rxr,
152                                      uint16_t cons)
153 {
154         struct bnxt_sw_rx_bd *cons_rx_buf;
155         struct rte_mbuf *mbuf;
156
157         cons_rx_buf = &rxr->rx_buf_ring[cons];
158         RTE_ASSERT(cons_rx_buf->mbuf != NULL);
159         mbuf = cons_rx_buf->mbuf;
160         cons_rx_buf->mbuf = NULL;
161         return mbuf;
162 }
163
164 static void bnxt_tpa_start(struct bnxt_rx_queue *rxq,
165                            struct rx_tpa_start_cmpl *tpa_start,
166                            struct rx_tpa_start_cmpl_hi *tpa_start1)
167 {
168         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
169         uint8_t agg_id = rte_le_to_cpu_32(tpa_start->agg_id &
170                 RX_TPA_START_CMPL_AGG_ID_MASK) >> RX_TPA_START_CMPL_AGG_ID_SFT;
171         uint16_t data_cons;
172         struct bnxt_tpa_info *tpa_info;
173         struct rte_mbuf *mbuf;
174
175         data_cons = tpa_start->opaque;
176         tpa_info = &rxr->tpa_info[agg_id];
177
178         mbuf = bnxt_consume_rx_buf(rxr, data_cons);
179
180         bnxt_reuse_rx_mbuf(rxr, tpa_info->mbuf);
181
182         tpa_info->mbuf = mbuf;
183         tpa_info->len = rte_le_to_cpu_32(tpa_start->len);
184
185         mbuf->nb_segs = 1;
186         mbuf->next = NULL;
187         mbuf->pkt_len = rte_le_to_cpu_32(tpa_start->len);
188         mbuf->data_len = mbuf->pkt_len;
189         mbuf->port = rxq->port_id;
190         mbuf->ol_flags = PKT_RX_LRO;
191         if (likely(tpa_start->flags_type &
192                    rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS_RSS_VALID))) {
193                 mbuf->hash.rss = rte_le_to_cpu_32(tpa_start->rss_hash);
194                 mbuf->ol_flags |= PKT_RX_RSS_HASH;
195         } else {
196                 mbuf->hash.fdir.id = rte_le_to_cpu_16(tpa_start1->cfa_code);
197                 mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
198         }
199         if (tpa_start1->flags2 &
200             rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_META_FORMAT_VLAN)) {
201                 mbuf->vlan_tci = rte_le_to_cpu_32(tpa_start1->metadata);
202                 mbuf->ol_flags |= PKT_RX_VLAN_PKT;
203         }
204         if (likely(tpa_start1->flags2 &
205                    rte_cpu_to_le_32(RX_TPA_START_CMPL_FLAGS2_L4_CS_CALC)))
206                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
207
208         /* recycle next mbuf */
209         data_cons = RING_NEXT(rxr->rx_ring_struct, data_cons);
210         bnxt_reuse_rx_mbuf(rxr, bnxt_consume_rx_buf(rxr, data_cons));
211 }
212
213 static int bnxt_agg_bufs_valid(struct bnxt_cp_ring_info *cpr,
214                 uint8_t agg_bufs, uint32_t raw_cp_cons)
215 {
216         uint16_t last_cp_cons;
217         struct rx_pkt_cmpl *agg_cmpl;
218
219         raw_cp_cons = ADV_RAW_CMP(raw_cp_cons, agg_bufs);
220         last_cp_cons = RING_CMP(cpr->cp_ring_struct, raw_cp_cons);
221         agg_cmpl = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[last_cp_cons];
222         cpr->valid = FLIP_VALID(raw_cp_cons,
223                                 cpr->cp_ring_struct->ring_mask,
224                                 cpr->valid);
225         return CMP_VALID(agg_cmpl, raw_cp_cons, cpr->cp_ring_struct);
226 }
227
228 /* TPA consume agg buffer out of order, allocate connected data only */
229 static int bnxt_prod_ag_mbuf(struct bnxt_rx_queue *rxq)
230 {
231         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
232         uint16_t next = RING_NEXT(rxr->ag_ring_struct, rxr->ag_prod);
233
234         /* TODO batch allocation for better performance */
235         while (rte_bitmap_get(rxr->ag_bitmap, next)) {
236                 if (unlikely(bnxt_alloc_ag_data(rxq, rxr, next))) {
237                         RTE_LOG(ERR, PMD,
238                                 "agg mbuf alloc failed: prod=0x%x\n", next);
239                         break;
240                 }
241                 rte_bitmap_clear(rxr->ag_bitmap, next);
242                 rxr->ag_prod = next;
243                 next = RING_NEXT(rxr->ag_ring_struct, next);
244         }
245
246         return 0;
247 }
248
249 static int bnxt_rx_pages(struct bnxt_rx_queue *rxq,
250                          struct rte_mbuf *mbuf, uint32_t *tmp_raw_cons,
251                          uint8_t agg_buf)
252 {
253         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
254         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
255         int i;
256         uint16_t cp_cons, ag_cons;
257         struct rx_pkt_cmpl *rxcmp;
258         struct rte_mbuf *last = mbuf;
259
260         for (i = 0; i < agg_buf; i++) {
261                 struct bnxt_sw_rx_bd *ag_buf;
262                 struct rte_mbuf *ag_mbuf;
263                 *tmp_raw_cons = NEXT_RAW_CMP(*tmp_raw_cons);
264                 cp_cons = RING_CMP(cpr->cp_ring_struct, *tmp_raw_cons);
265                 rxcmp = (struct rx_pkt_cmpl *)
266                                         &cpr->cp_desc_ring[cp_cons];
267
268 #ifdef BNXT_DEBUG
269                 bnxt_dump_cmpl(cp_cons, rxcmp);
270 #endif
271
272                 ag_cons = rxcmp->opaque;
273                 RTE_ASSERT(ag_cons <= rxr->ag_ring_struct->ring_mask);
274                 ag_buf = &rxr->ag_buf_ring[ag_cons];
275                 ag_mbuf = ag_buf->mbuf;
276                 RTE_ASSERT(ag_mbuf != NULL);
277
278                 ag_mbuf->data_len = rte_le_to_cpu_16(rxcmp->len);
279
280                 mbuf->nb_segs++;
281                 mbuf->pkt_len += ag_mbuf->data_len;
282
283                 last->next = ag_mbuf;
284                 last = ag_mbuf;
285
286                 ag_buf->mbuf = NULL;
287
288                 /*
289                  * As aggregation buffer consumed out of order in TPA module,
290                  * use bitmap to track freed slots to be allocated and notified
291                  * to NIC
292                  */
293                 rte_bitmap_set(rxr->ag_bitmap, ag_cons);
294         }
295         bnxt_prod_ag_mbuf(rxq);
296         return 0;
297 }
298
299 static inline struct rte_mbuf *bnxt_tpa_end(
300                 struct bnxt_rx_queue *rxq,
301                 uint32_t *raw_cp_cons,
302                 struct rx_tpa_end_cmpl *tpa_end,
303                 struct rx_tpa_end_cmpl_hi *tpa_end1 __rte_unused)
304 {
305         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
306         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
307         uint8_t agg_id = (tpa_end->agg_id & RX_TPA_END_CMPL_AGG_ID_MASK)
308                         >> RX_TPA_END_CMPL_AGG_ID_SFT;
309         struct rte_mbuf *mbuf;
310         uint8_t agg_bufs;
311         struct bnxt_tpa_info *tpa_info;
312
313         tpa_info = &rxr->tpa_info[agg_id];
314         mbuf = tpa_info->mbuf;
315         RTE_ASSERT(mbuf != NULL);
316
317         rte_prefetch0(mbuf);
318         agg_bufs = (rte_le_to_cpu_32(tpa_end->agg_bufs_v1) &
319                 RX_TPA_END_CMPL_AGG_BUFS_MASK) >> RX_TPA_END_CMPL_AGG_BUFS_SFT;
320         if (agg_bufs) {
321                 if (!bnxt_agg_bufs_valid(cpr, agg_bufs, *raw_cp_cons))
322                         return NULL;
323                 bnxt_rx_pages(rxq, mbuf, raw_cp_cons, agg_bufs);
324         }
325         mbuf->l4_len = tpa_end->payload_offset;
326
327         struct rte_mbuf *new_data = __bnxt_alloc_rx_data(rxq->mb_pool);
328         RTE_ASSERT(new_data != NULL);
329         if (!new_data) {
330                 rte_atomic64_inc(&rxq->bp->rx_mbuf_alloc_fail);
331                 return NULL;
332         }
333         tpa_info->mbuf = new_data;
334
335         return mbuf;
336 }
337
338 static int bnxt_rx_pkt(struct rte_mbuf **rx_pkt,
339                             struct bnxt_rx_queue *rxq, uint32_t *raw_cons)
340 {
341         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
342         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
343         struct rx_pkt_cmpl *rxcmp;
344         struct rx_pkt_cmpl_hi *rxcmp1;
345         uint32_t tmp_raw_cons = *raw_cons;
346         uint16_t cons, prod, cp_cons =
347             RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
348 #ifdef BNXT_DEBUG
349         uint16_t ag_cons;
350 #endif
351         struct rte_mbuf *mbuf;
352         int rc = 0;
353         uint8_t agg_buf = 0;
354         uint16_t cmp_type;
355
356         rxcmp = (struct rx_pkt_cmpl *)
357             &cpr->cp_desc_ring[cp_cons];
358
359         tmp_raw_cons = NEXT_RAW_CMP(tmp_raw_cons);
360         cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
361         rxcmp1 = (struct rx_pkt_cmpl_hi *)&cpr->cp_desc_ring[cp_cons];
362
363         if (!CMP_VALID(rxcmp1, tmp_raw_cons, cpr->cp_ring_struct))
364                 return -EBUSY;
365
366         cpr->valid = FLIP_VALID(cp_cons,
367                                 cpr->cp_ring_struct->ring_mask,
368                                 cpr->valid);
369
370         cmp_type = CMP_TYPE(rxcmp);
371         if (cmp_type == RX_PKT_CMPL_TYPE_RX_L2_TPA_START) {
372                 bnxt_tpa_start(rxq, (struct rx_tpa_start_cmpl *)rxcmp,
373                                (struct rx_tpa_start_cmpl_hi *)rxcmp1);
374                 rc = -EINVAL; /* Continue w/o new mbuf */
375                 goto next_rx;
376         } else if (cmp_type == RX_PKT_CMPL_TYPE_RX_L2_TPA_END) {
377                 mbuf = bnxt_tpa_end(rxq, &tmp_raw_cons,
378                                    (struct rx_tpa_end_cmpl *)rxcmp,
379                                    (struct rx_tpa_end_cmpl_hi *)rxcmp1);
380                 if (unlikely(!mbuf))
381                         return -EBUSY;
382                 *rx_pkt = mbuf;
383                 goto next_rx;
384         } else if (cmp_type != 0x11) {
385                 rc = -EINVAL;
386                 goto next_rx;
387         }
388
389         agg_buf = (rxcmp->agg_bufs_v1 & RX_PKT_CMPL_AGG_BUFS_MASK)
390                         >> RX_PKT_CMPL_AGG_BUFS_SFT;
391         if (agg_buf && !bnxt_agg_bufs_valid(cpr, agg_buf, tmp_raw_cons))
392                 return -EBUSY;
393
394         prod = rxr->rx_prod;
395
396         cons = rxcmp->opaque;
397         mbuf = bnxt_consume_rx_buf(rxr, cons);
398         rte_prefetch0(mbuf);
399
400         if (mbuf == NULL)
401                 return -EBUSY;
402
403         mbuf->nb_segs = 1;
404         mbuf->next = NULL;
405         mbuf->pkt_len = rxcmp->len;
406         mbuf->data_len = mbuf->pkt_len;
407         mbuf->port = rxq->port_id;
408         mbuf->ol_flags = 0;
409         if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID) {
410                 mbuf->hash.rss = rxcmp->rss_hash;
411                 mbuf->ol_flags |= PKT_RX_RSS_HASH;
412         } else {
413                 mbuf->hash.fdir.id = rxcmp1->cfa_code;
414                 mbuf->ol_flags |= PKT_RX_FDIR | PKT_RX_FDIR_ID;
415         }
416
417         if (agg_buf)
418                 bnxt_rx_pages(rxq, mbuf, &tmp_raw_cons, agg_buf);
419
420         if (rxcmp1->flags2 & RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
421                 mbuf->vlan_tci = rxcmp1->metadata &
422                         (RX_PKT_CMPL_METADATA_VID_MASK |
423                         RX_PKT_CMPL_METADATA_DE |
424                         RX_PKT_CMPL_METADATA_PRI_MASK);
425                 mbuf->ol_flags |= PKT_RX_VLAN_PKT;
426         }
427
428         if (likely(RX_CMP_IP_CS_OK(rxcmp1)))
429                 mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
430         else
431                 mbuf->ol_flags |= PKT_RX_IP_CKSUM_NONE;
432
433         if (likely(RX_CMP_L4_CS_OK(rxcmp1)))
434                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
435         else
436                 mbuf->ol_flags |= PKT_RX_L4_CKSUM_NONE;
437
438
439 #ifdef BNXT_DEBUG
440         if (rxcmp1->errors_v2 & RX_CMP_L2_ERRORS) {
441                 /* Re-install the mbuf back to the rx ring */
442                 bnxt_reuse_rx_mbuf(rxr, cons, mbuf);
443                 if (agg_buf)
444                         bnxt_reuse_ag_mbuf(rxr, ag_cons, mbuf);
445
446                 rc = -EIO;
447                 goto next_rx;
448         }
449 #endif
450         /*
451          * TODO: Redesign this....
452          * If the allocation fails, the packet does not get received.
453          * Simply returning this will result in slowly falling behind
454          * on the producer ring buffers.
455          * Instead, "filling up" the producer just before ringing the
456          * doorbell could be a better solution since it will let the
457          * producer ring starve until memory is available again pushing
458          * the drops into hardware and getting them out of the driver
459          * allowing recovery to a full producer ring.
460          *
461          * This could also help with cache usage by preventing per-packet
462          * calls in favour of a tight loop with the same function being called
463          * in it.
464          */
465         prod = RING_NEXT(rxr->rx_ring_struct, prod);
466         if (bnxt_alloc_rx_data(rxq, rxr, prod)) {
467                 RTE_LOG(ERR, PMD, "mbuf alloc failed with prod=0x%x\n", prod);
468                 rc = -ENOMEM;
469                 goto rx;
470         }
471         rxr->rx_prod = prod;
472         /*
473          * All MBUFs are allocated with the same size under DPDK,
474          * no optimization for rx_copy_thresh
475          */
476 rx:
477         *rx_pkt = mbuf;
478
479 next_rx:
480
481         *raw_cons = tmp_raw_cons;
482
483         return rc;
484 }
485
486 uint16_t bnxt_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
487                                uint16_t nb_pkts)
488 {
489         struct bnxt_rx_queue *rxq = rx_queue;
490         struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
491         struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
492         uint32_t raw_cons = cpr->cp_raw_cons;
493         uint32_t cons;
494         int nb_rx_pkts = 0;
495         struct rx_pkt_cmpl *rxcmp;
496         uint16_t prod = rxr->rx_prod;
497         uint16_t ag_prod = rxr->ag_prod;
498         int rc = 0;
499
500         /* Handle RX burst request */
501         while (1) {
502                 cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
503                 rte_prefetch0(&cpr->cp_desc_ring[cons]);
504                 rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
505
506                 if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
507                         break;
508                 cpr->valid = FLIP_VALID(cons,
509                                         cpr->cp_ring_struct->ring_mask,
510                                         cpr->valid);
511
512                 /* TODO: Avoid magic numbers... */
513                 if ((CMP_TYPE(rxcmp) & 0x30) == 0x10) {
514                         rc = bnxt_rx_pkt(&rx_pkts[nb_rx_pkts], rxq, &raw_cons);
515                         if (likely(!rc) || rc == -ENOMEM)
516                                 nb_rx_pkts++;
517                         if (rc == -EBUSY)       /* partial completion */
518                                 break;
519                 }
520                 raw_cons = NEXT_RAW_CMP(raw_cons);
521                 if (nb_rx_pkts == nb_pkts)
522                         break;
523         }
524
525         cpr->cp_raw_cons = raw_cons;
526         if (prod == rxr->rx_prod && ag_prod == rxr->ag_prod) {
527                 /*
528                  * For PMD, there is no need to keep on pushing to REARM
529                  * the doorbell if there are no new completions
530                  */
531                 return nb_rx_pkts;
532         }
533
534         B_CP_DIS_DB(cpr, cpr->cp_raw_cons);
535         B_RX_DB(rxr->rx_doorbell, rxr->rx_prod);
536         /* Ring the AGG ring DB */
537         B_RX_DB(rxr->ag_doorbell, rxr->ag_prod);
538
539         /* Attempt to alloc Rx buf in case of a previous allocation failure. */
540         if (rc == -ENOMEM) {
541                 int i;
542
543                 for (i = prod; i <= nb_rx_pkts;
544                         i = RING_NEXT(rxr->rx_ring_struct, i)) {
545                         struct bnxt_sw_rx_bd *rx_buf = &rxr->rx_buf_ring[i];
546
547                         /* Buffer already allocated for this index. */
548                         if (rx_buf->mbuf != NULL)
549                                 continue;
550
551                         /* This slot is empty. Alloc buffer for Rx */
552                         if (!bnxt_alloc_rx_data(rxq, rxr, i)) {
553                                 rxr->rx_prod = i;
554                                 B_RX_DB(rxr->rx_doorbell, rxr->rx_prod);
555                         } else {
556                                 RTE_LOG(ERR, PMD, "Alloc  mbuf failed\n");
557                                 break;
558                         }
559                 }
560         }
561
562         return nb_rx_pkts;
563 }
564
565 void bnxt_free_rx_rings(struct bnxt *bp)
566 {
567         int i;
568
569         for (i = 0; i < (int)bp->rx_nr_rings; i++) {
570                 struct bnxt_rx_queue *rxq = bp->rx_queues[i];
571
572                 if (!rxq)
573                         continue;
574
575                 bnxt_free_ring(rxq->rx_ring->rx_ring_struct);
576                 rte_free(rxq->rx_ring->rx_ring_struct);
577
578                 /* Free the Aggregator ring */
579                 bnxt_free_ring(rxq->rx_ring->ag_ring_struct);
580                 rte_free(rxq->rx_ring->ag_ring_struct);
581                 rxq->rx_ring->ag_ring_struct = NULL;
582
583                 rte_free(rxq->rx_ring);
584
585                 bnxt_free_ring(rxq->cp_ring->cp_ring_struct);
586                 rte_free(rxq->cp_ring->cp_ring_struct);
587                 rte_free(rxq->cp_ring);
588
589                 rte_free(rxq);
590                 bp->rx_queues[i] = NULL;
591         }
592 }
593
594 int bnxt_init_rx_ring_struct(struct bnxt_rx_queue *rxq, unsigned int socket_id)
595 {
596         struct bnxt_cp_ring_info *cpr;
597         struct bnxt_rx_ring_info *rxr;
598         struct bnxt_ring *ring;
599
600         rxq->rx_buf_use_size = BNXT_MAX_MTU + ETHER_HDR_LEN + ETHER_CRC_LEN +
601                                (2 * VLAN_TAG_SIZE);
602         rxq->rx_buf_size = rxq->rx_buf_use_size + sizeof(struct rte_mbuf);
603
604         rxr = rte_zmalloc_socket("bnxt_rx_ring",
605                                  sizeof(struct bnxt_rx_ring_info),
606                                  RTE_CACHE_LINE_SIZE, socket_id);
607         if (rxr == NULL)
608                 return -ENOMEM;
609         rxq->rx_ring = rxr;
610
611         ring = rte_zmalloc_socket("bnxt_rx_ring_struct",
612                                    sizeof(struct bnxt_ring),
613                                    RTE_CACHE_LINE_SIZE, socket_id);
614         if (ring == NULL)
615                 return -ENOMEM;
616         rxr->rx_ring_struct = ring;
617         ring->ring_size = rte_align32pow2(rxq->nb_rx_desc);
618         ring->ring_mask = ring->ring_size - 1;
619         ring->bd = (void *)rxr->rx_desc_ring;
620         ring->bd_dma = rxr->rx_desc_mapping;
621         ring->vmem_size = ring->ring_size * sizeof(struct bnxt_sw_rx_bd);
622         ring->vmem = (void **)&rxr->rx_buf_ring;
623
624         cpr = rte_zmalloc_socket("bnxt_rx_ring",
625                                  sizeof(struct bnxt_cp_ring_info),
626                                  RTE_CACHE_LINE_SIZE, socket_id);
627         if (cpr == NULL)
628                 return -ENOMEM;
629         rxq->cp_ring = cpr;
630
631         ring = rte_zmalloc_socket("bnxt_rx_ring_struct",
632                                    sizeof(struct bnxt_ring),
633                                    RTE_CACHE_LINE_SIZE, socket_id);
634         if (ring == NULL)
635                 return -ENOMEM;
636         cpr->cp_ring_struct = ring;
637         ring->ring_size = rte_align32pow2(rxr->rx_ring_struct->ring_size *
638                                           (2 + AGG_RING_SIZE_FACTOR));
639         ring->ring_mask = ring->ring_size - 1;
640         ring->bd = (void *)cpr->cp_desc_ring;
641         ring->bd_dma = cpr->cp_desc_mapping;
642         ring->vmem_size = 0;
643         ring->vmem = NULL;
644
645         /* Allocate Aggregator rings */
646         ring = rte_zmalloc_socket("bnxt_rx_ring_struct",
647                                    sizeof(struct bnxt_ring),
648                                    RTE_CACHE_LINE_SIZE, socket_id);
649         if (ring == NULL)
650                 return -ENOMEM;
651         rxr->ag_ring_struct = ring;
652         ring->ring_size = rte_align32pow2(rxq->nb_rx_desc *
653                                           AGG_RING_SIZE_FACTOR);
654         ring->ring_mask = ring->ring_size - 1;
655         ring->bd = (void *)rxr->ag_desc_ring;
656         ring->bd_dma = rxr->ag_desc_mapping;
657         ring->vmem_size = ring->ring_size * sizeof(struct bnxt_sw_rx_bd);
658         ring->vmem = (void **)&rxr->ag_buf_ring;
659
660         return 0;
661 }
662
663 static void bnxt_init_rxbds(struct bnxt_ring *ring, uint32_t type,
664                             uint16_t len)
665 {
666         uint32_t j;
667         struct rx_prod_pkt_bd *rx_bd_ring = (struct rx_prod_pkt_bd *)ring->bd;
668
669         if (!rx_bd_ring)
670                 return;
671         for (j = 0; j < ring->ring_size; j++) {
672                 rx_bd_ring[j].flags_type = rte_cpu_to_le_16(type);
673                 rx_bd_ring[j].len = rte_cpu_to_le_16(len);
674                 rx_bd_ring[j].opaque = j;
675         }
676 }
677
678 int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq)
679 {
680         struct bnxt_rx_ring_info *rxr;
681         struct bnxt_ring *ring;
682         uint32_t prod, type;
683         unsigned int i;
684         uint16_t size;
685
686         size = rte_pktmbuf_data_room_size(rxq->mb_pool) - RTE_PKTMBUF_HEADROOM;
687         if (rxq->rx_buf_use_size <= size)
688                 size = rxq->rx_buf_use_size;
689
690         type = RX_PROD_PKT_BD_TYPE_RX_PROD_PKT;
691
692         rxr = rxq->rx_ring;
693         ring = rxr->rx_ring_struct;
694         bnxt_init_rxbds(ring, type, size);
695
696         prod = rxr->rx_prod;
697         for (i = 0; i < ring->ring_size; i++) {
698                 if (bnxt_alloc_rx_data(rxq, rxr, prod) != 0) {
699                         RTE_LOG(WARNING, PMD,
700                                 "init'ed rx ring %d with %d/%d mbufs only\n",
701                                 rxq->queue_id, i, ring->ring_size);
702                         break;
703                 }
704                 rxr->rx_prod = prod;
705                 prod = RING_NEXT(rxr->rx_ring_struct, prod);
706         }
707         RTE_LOG(DEBUG, PMD, "%s\n", __func__);
708
709         ring = rxr->ag_ring_struct;
710         type = RX_PROD_AGG_BD_TYPE_RX_PROD_AGG;
711         bnxt_init_rxbds(ring, type, size);
712         prod = rxr->ag_prod;
713
714         for (i = 0; i < ring->ring_size; i++) {
715                 if (bnxt_alloc_ag_data(rxq, rxr, prod) != 0) {
716                         RTE_LOG(WARNING, PMD,
717                         "init'ed AG ring %d with %d/%d mbufs only\n",
718                         rxq->queue_id, i, ring->ring_size);
719                         break;
720                 }
721                 rxr->ag_prod = prod;
722                 prod = RING_NEXT(rxr->ag_ring_struct, prod);
723         }
724         RTE_LOG(DEBUG, PMD, "%s AGG Done!\n", __func__);
725
726         if (rxr->tpa_info) {
727                 for (i = 0; i < BNXT_TPA_MAX; i++) {
728                         rxr->tpa_info[i].mbuf =
729                                 __bnxt_alloc_rx_data(rxq->mb_pool);
730                         if (!rxr->tpa_info[i].mbuf) {
731                                 rte_atomic64_inc(&rxq->bp->rx_mbuf_alloc_fail);
732                                 return -ENOMEM;
733                         }
734                 }
735         }
736         RTE_LOG(DEBUG, PMD, "%s TPA alloc Done!\n", __func__);
737
738         return 0;
739 }