net/i40e: move RSS to flow API
[dpdk.git] / drivers / net / e1000 / igb_rxtx.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2016 Intel Corporation
3  */
4
5 #include <sys/queue.h>
6
7 #include <stdio.h>
8 #include <stdlib.h>
9 #include <string.h>
10 #include <errno.h>
11 #include <stdint.h>
12 #include <stdarg.h>
13 #include <inttypes.h>
14
15 #include <rte_interrupts.h>
16 #include <rte_byteorder.h>
17 #include <rte_common.h>
18 #include <rte_log.h>
19 #include <rte_debug.h>
20 #include <rte_pci.h>
21 #include <rte_memory.h>
22 #include <rte_memcpy.h>
23 #include <rte_memzone.h>
24 #include <rte_launch.h>
25 #include <rte_eal.h>
26 #include <rte_per_lcore.h>
27 #include <rte_lcore.h>
28 #include <rte_atomic.h>
29 #include <rte_branch_prediction.h>
30 #include <rte_mempool.h>
31 #include <rte_malloc.h>
32 #include <rte_mbuf.h>
33 #include <rte_ether.h>
34 #include <rte_ethdev.h>
35 #include <rte_prefetch.h>
36 #include <rte_udp.h>
37 #include <rte_tcp.h>
38 #include <rte_sctp.h>
39 #include <rte_net.h>
40 #include <rte_string_fns.h>
41
42 #include "e1000_logs.h"
43 #include "base/e1000_api.h"
44 #include "e1000_ethdev.h"
45
46 #ifdef RTE_LIBRTE_IEEE1588
47 #define IGB_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
48 #else
49 #define IGB_TX_IEEE1588_TMST 0
50 #endif
51 /* Bit Mask to indicate what bits required for building TX context */
52 #define IGB_TX_OFFLOAD_MASK (                    \
53                 PKT_TX_VLAN_PKT |                \
54                 PKT_TX_IP_CKSUM |                \
55                 PKT_TX_L4_MASK |                 \
56                 PKT_TX_TCP_SEG |                 \
57                 IGB_TX_IEEE1588_TMST)
58
59 #define IGB_TX_OFFLOAD_NOTSUP_MASK \
60                 (PKT_TX_OFFLOAD_MASK ^ IGB_TX_OFFLOAD_MASK)
61
62 /**
63  * Structure associated with each descriptor of the RX ring of a RX queue.
64  */
65 struct igb_rx_entry {
66         struct rte_mbuf *mbuf; /**< mbuf associated with RX descriptor. */
67 };
68
69 /**
70  * Structure associated with each descriptor of the TX ring of a TX queue.
71  */
72 struct igb_tx_entry {
73         struct rte_mbuf *mbuf; /**< mbuf associated with TX desc, if any. */
74         uint16_t next_id; /**< Index of next descriptor in ring. */
75         uint16_t last_id; /**< Index of last scattered descriptor. */
76 };
77
78 /**
79  * rx queue flags
80  */
81 enum igb_rxq_flags {
82         IGB_RXQ_FLAG_LB_BSWAP_VLAN = 0x01,
83 };
84
85 /**
86  * Structure associated with each RX queue.
87  */
88 struct igb_rx_queue {
89         struct rte_mempool  *mb_pool;   /**< mbuf pool to populate RX ring. */
90         volatile union e1000_adv_rx_desc *rx_ring; /**< RX ring virtual address. */
91         uint64_t            rx_ring_phys_addr; /**< RX ring DMA address. */
92         volatile uint32_t   *rdt_reg_addr; /**< RDT register address. */
93         volatile uint32_t   *rdh_reg_addr; /**< RDH register address. */
94         struct igb_rx_entry *sw_ring;   /**< address of RX software ring. */
95         struct rte_mbuf *pkt_first_seg; /**< First segment of current packet. */
96         struct rte_mbuf *pkt_last_seg;  /**< Last segment of current packet. */
97         uint16_t            nb_rx_desc; /**< number of RX descriptors. */
98         uint16_t            rx_tail;    /**< current value of RDT register. */
99         uint16_t            nb_rx_hold; /**< number of held free RX desc. */
100         uint16_t            rx_free_thresh; /**< max free RX desc to hold. */
101         uint16_t            queue_id;   /**< RX queue index. */
102         uint16_t            reg_idx;    /**< RX queue register index. */
103         uint16_t            port_id;    /**< Device port identifier. */
104         uint8_t             pthresh;    /**< Prefetch threshold register. */
105         uint8_t             hthresh;    /**< Host threshold register. */
106         uint8_t             wthresh;    /**< Write-back threshold register. */
107         uint8_t             crc_len;    /**< 0 if CRC stripped, 4 otherwise. */
108         uint8_t             drop_en;  /**< If not 0, set SRRCTL.Drop_En. */
109         uint32_t            flags;      /**< RX flags. */
110 };
111
112 /**
113  * Hardware context number
114  */
115 enum igb_advctx_num {
116         IGB_CTX_0    = 0, /**< CTX0    */
117         IGB_CTX_1    = 1, /**< CTX1    */
118         IGB_CTX_NUM  = 2, /**< CTX_NUM */
119 };
120
121 /** Offload features */
122 union igb_tx_offload {
123         uint64_t data;
124         struct {
125                 uint64_t l3_len:9; /**< L3 (IP) Header Length. */
126                 uint64_t l2_len:7; /**< L2 (MAC) Header Length. */
127                 uint64_t vlan_tci:16;  /**< VLAN Tag Control Identifier(CPU order). */
128                 uint64_t l4_len:8; /**< L4 (TCP/UDP) Header Length. */
129                 uint64_t tso_segsz:16; /**< TCP TSO segment size. */
130
131                 /* uint64_t unused:8; */
132         };
133 };
134
135 /*
136  * Compare mask for igb_tx_offload.data,
137  * should be in sync with igb_tx_offload layout.
138  * */
139 #define TX_MACIP_LEN_CMP_MASK   0x000000000000FFFFULL /**< L2L3 header mask. */
140 #define TX_VLAN_CMP_MASK                0x00000000FFFF0000ULL /**< Vlan mask. */
141 #define TX_TCP_LEN_CMP_MASK             0x000000FF00000000ULL /**< TCP header mask. */
142 #define TX_TSO_MSS_CMP_MASK             0x00FFFF0000000000ULL /**< TSO segsz mask. */
143 /** Mac + IP + TCP + Mss mask. */
144 #define TX_TSO_CMP_MASK \
145         (TX_MACIP_LEN_CMP_MASK | TX_TCP_LEN_CMP_MASK | TX_TSO_MSS_CMP_MASK)
146
147 /**
148  * Strucutre to check if new context need be built
149  */
150 struct igb_advctx_info {
151         uint64_t flags;           /**< ol_flags related to context build. */
152         /** tx offload: vlan, tso, l2-l3-l4 lengths. */
153         union igb_tx_offload tx_offload;
154         /** compare mask for tx offload. */
155         union igb_tx_offload tx_offload_mask;
156 };
157
158 /**
159  * Structure associated with each TX queue.
160  */
161 struct igb_tx_queue {
162         volatile union e1000_adv_tx_desc *tx_ring; /**< TX ring address */
163         uint64_t               tx_ring_phys_addr; /**< TX ring DMA address. */
164         struct igb_tx_entry    *sw_ring; /**< virtual address of SW ring. */
165         volatile uint32_t      *tdt_reg_addr; /**< Address of TDT register. */
166         uint32_t               txd_type;      /**< Device-specific TXD type */
167         uint16_t               nb_tx_desc;    /**< number of TX descriptors. */
168         uint16_t               tx_tail; /**< Current value of TDT register. */
169         uint16_t               tx_head;
170         /**< Index of first used TX descriptor. */
171         uint16_t               queue_id; /**< TX queue index. */
172         uint16_t               reg_idx;  /**< TX queue register index. */
173         uint16_t               port_id;  /**< Device port identifier. */
174         uint8_t                pthresh;  /**< Prefetch threshold register. */
175         uint8_t                hthresh;  /**< Host threshold register. */
176         uint8_t                wthresh;  /**< Write-back threshold register. */
177         uint32_t               ctx_curr;
178         /**< Current used hardware descriptor. */
179         uint32_t               ctx_start;
180         /**< Start context position for transmit queue. */
181         struct igb_advctx_info ctx_cache[IGB_CTX_NUM];
182         /**< Hardware context history.*/
183 };
184
185 #if 1
186 #define RTE_PMD_USE_PREFETCH
187 #endif
188
189 #ifdef RTE_PMD_USE_PREFETCH
190 #define rte_igb_prefetch(p)     rte_prefetch0(p)
191 #else
192 #define rte_igb_prefetch(p)     do {} while(0)
193 #endif
194
195 #ifdef RTE_PMD_PACKET_PREFETCH
196 #define rte_packet_prefetch(p) rte_prefetch1(p)
197 #else
198 #define rte_packet_prefetch(p)  do {} while(0)
199 #endif
200
201 /*
202  * Macro for VMDq feature for 1 GbE NIC.
203  */
204 #define E1000_VMOLR_SIZE                        (8)
205 #define IGB_TSO_MAX_HDRLEN                      (512)
206 #define IGB_TSO_MAX_MSS                         (9216)
207
208 /*********************************************************************
209  *
210  *  TX function
211  *
212  **********************************************************************/
213
214 /*
215  *There're some limitations in hardware for TCP segmentation offload. We
216  *should check whether the parameters are valid.
217  */
218 static inline uint64_t
219 check_tso_para(uint64_t ol_req, union igb_tx_offload ol_para)
220 {
221         if (!(ol_req & PKT_TX_TCP_SEG))
222                 return ol_req;
223         if ((ol_para.tso_segsz > IGB_TSO_MAX_MSS) || (ol_para.l2_len +
224                         ol_para.l3_len + ol_para.l4_len > IGB_TSO_MAX_HDRLEN)) {
225                 ol_req &= ~PKT_TX_TCP_SEG;
226                 ol_req |= PKT_TX_TCP_CKSUM;
227         }
228         return ol_req;
229 }
230
231 /*
232  * Advanced context descriptor are almost same between igb/ixgbe
233  * This is a separate function, looking for optimization opportunity here
234  * Rework required to go with the pre-defined values.
235  */
236
237 static inline void
238 igbe_set_xmit_ctx(struct igb_tx_queue* txq,
239                 volatile struct e1000_adv_tx_context_desc *ctx_txd,
240                 uint64_t ol_flags, union igb_tx_offload tx_offload)
241 {
242         uint32_t type_tucmd_mlhl;
243         uint32_t mss_l4len_idx;
244         uint32_t ctx_idx, ctx_curr;
245         uint32_t vlan_macip_lens;
246         union igb_tx_offload tx_offload_mask;
247
248         ctx_curr = txq->ctx_curr;
249         ctx_idx = ctx_curr + txq->ctx_start;
250
251         tx_offload_mask.data = 0;
252         type_tucmd_mlhl = 0;
253
254         /* Specify which HW CTX to upload. */
255         mss_l4len_idx = (ctx_idx << E1000_ADVTXD_IDX_SHIFT);
256
257         if (ol_flags & PKT_TX_VLAN_PKT)
258                 tx_offload_mask.data |= TX_VLAN_CMP_MASK;
259
260         /* check if TCP segmentation required for this packet */
261         if (ol_flags & PKT_TX_TCP_SEG) {
262                 /* implies IP cksum in IPv4 */
263                 if (ol_flags & PKT_TX_IP_CKSUM)
264                         type_tucmd_mlhl = E1000_ADVTXD_TUCMD_IPV4 |
265                                 E1000_ADVTXD_TUCMD_L4T_TCP |
266                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
267                 else
268                         type_tucmd_mlhl = E1000_ADVTXD_TUCMD_IPV6 |
269                                 E1000_ADVTXD_TUCMD_L4T_TCP |
270                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
271
272                 tx_offload_mask.data |= TX_TSO_CMP_MASK;
273                 mss_l4len_idx |= tx_offload.tso_segsz << E1000_ADVTXD_MSS_SHIFT;
274                 mss_l4len_idx |= tx_offload.l4_len << E1000_ADVTXD_L4LEN_SHIFT;
275         } else { /* no TSO, check if hardware checksum is needed */
276                 if (ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK))
277                         tx_offload_mask.data |= TX_MACIP_LEN_CMP_MASK;
278
279                 if (ol_flags & PKT_TX_IP_CKSUM)
280                         type_tucmd_mlhl = E1000_ADVTXD_TUCMD_IPV4;
281
282                 switch (ol_flags & PKT_TX_L4_MASK) {
283                 case PKT_TX_UDP_CKSUM:
284                         type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_UDP |
285                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
286                         mss_l4len_idx |= sizeof(struct udp_hdr) << E1000_ADVTXD_L4LEN_SHIFT;
287                         break;
288                 case PKT_TX_TCP_CKSUM:
289                         type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_TCP |
290                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
291                         mss_l4len_idx |= sizeof(struct tcp_hdr) << E1000_ADVTXD_L4LEN_SHIFT;
292                         break;
293                 case PKT_TX_SCTP_CKSUM:
294                         type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_SCTP |
295                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
296                         mss_l4len_idx |= sizeof(struct sctp_hdr) << E1000_ADVTXD_L4LEN_SHIFT;
297                         break;
298                 default:
299                         type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_RSV |
300                                 E1000_ADVTXD_DTYP_CTXT | E1000_ADVTXD_DCMD_DEXT;
301                         break;
302                 }
303         }
304
305         txq->ctx_cache[ctx_curr].flags = ol_flags;
306         txq->ctx_cache[ctx_curr].tx_offload.data =
307                 tx_offload_mask.data & tx_offload.data;
308         txq->ctx_cache[ctx_curr].tx_offload_mask = tx_offload_mask;
309
310         ctx_txd->type_tucmd_mlhl = rte_cpu_to_le_32(type_tucmd_mlhl);
311         vlan_macip_lens = (uint32_t)tx_offload.data;
312         ctx_txd->vlan_macip_lens = rte_cpu_to_le_32(vlan_macip_lens);
313         ctx_txd->mss_l4len_idx = rte_cpu_to_le_32(mss_l4len_idx);
314         ctx_txd->seqnum_seed = 0;
315 }
316
317 /*
318  * Check which hardware context can be used. Use the existing match
319  * or create a new context descriptor.
320  */
321 static inline uint32_t
322 what_advctx_update(struct igb_tx_queue *txq, uint64_t flags,
323                 union igb_tx_offload tx_offload)
324 {
325         /* If match with the current context */
326         if (likely((txq->ctx_cache[txq->ctx_curr].flags == flags) &&
327                 (txq->ctx_cache[txq->ctx_curr].tx_offload.data ==
328                 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data & tx_offload.data)))) {
329                         return txq->ctx_curr;
330         }
331
332         /* If match with the second context */
333         txq->ctx_curr ^= 1;
334         if (likely((txq->ctx_cache[txq->ctx_curr].flags == flags) &&
335                 (txq->ctx_cache[txq->ctx_curr].tx_offload.data ==
336                 (txq->ctx_cache[txq->ctx_curr].tx_offload_mask.data & tx_offload.data)))) {
337                         return txq->ctx_curr;
338         }
339
340         /* Mismatch, use the previous context */
341         return IGB_CTX_NUM;
342 }
343
344 static inline uint32_t
345 tx_desc_cksum_flags_to_olinfo(uint64_t ol_flags)
346 {
347         static const uint32_t l4_olinfo[2] = {0, E1000_ADVTXD_POPTS_TXSM};
348         static const uint32_t l3_olinfo[2] = {0, E1000_ADVTXD_POPTS_IXSM};
349         uint32_t tmp;
350
351         tmp  = l4_olinfo[(ol_flags & PKT_TX_L4_MASK)  != PKT_TX_L4_NO_CKSUM];
352         tmp |= l3_olinfo[(ol_flags & PKT_TX_IP_CKSUM) != 0];
353         tmp |= l4_olinfo[(ol_flags & PKT_TX_TCP_SEG) != 0];
354         return tmp;
355 }
356
357 static inline uint32_t
358 tx_desc_vlan_flags_to_cmdtype(uint64_t ol_flags)
359 {
360         uint32_t cmdtype;
361         static uint32_t vlan_cmd[2] = {0, E1000_ADVTXD_DCMD_VLE};
362         static uint32_t tso_cmd[2] = {0, E1000_ADVTXD_DCMD_TSE};
363         cmdtype = vlan_cmd[(ol_flags & PKT_TX_VLAN_PKT) != 0];
364         cmdtype |= tso_cmd[(ol_flags & PKT_TX_TCP_SEG) != 0];
365         return cmdtype;
366 }
367
368 uint16_t
369 eth_igb_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
370                uint16_t nb_pkts)
371 {
372         struct igb_tx_queue *txq;
373         struct igb_tx_entry *sw_ring;
374         struct igb_tx_entry *txe, *txn;
375         volatile union e1000_adv_tx_desc *txr;
376         volatile union e1000_adv_tx_desc *txd;
377         struct rte_mbuf     *tx_pkt;
378         struct rte_mbuf     *m_seg;
379         uint64_t buf_dma_addr;
380         uint32_t olinfo_status;
381         uint32_t cmd_type_len;
382         uint32_t pkt_len;
383         uint16_t slen;
384         uint64_t ol_flags;
385         uint16_t tx_end;
386         uint16_t tx_id;
387         uint16_t tx_last;
388         uint16_t nb_tx;
389         uint64_t tx_ol_req;
390         uint32_t new_ctx = 0;
391         uint32_t ctx = 0;
392         union igb_tx_offload tx_offload = {0};
393
394         txq = tx_queue;
395         sw_ring = txq->sw_ring;
396         txr     = txq->tx_ring;
397         tx_id   = txq->tx_tail;
398         txe = &sw_ring[tx_id];
399
400         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
401                 tx_pkt = *tx_pkts++;
402                 pkt_len = tx_pkt->pkt_len;
403
404                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
405
406                 /*
407                  * The number of descriptors that must be allocated for a
408                  * packet is the number of segments of that packet, plus 1
409                  * Context Descriptor for the VLAN Tag Identifier, if any.
410                  * Determine the last TX descriptor to allocate in the TX ring
411                  * for the packet, starting from the current position (tx_id)
412                  * in the ring.
413                  */
414                 tx_last = (uint16_t) (tx_id + tx_pkt->nb_segs - 1);
415
416                 ol_flags = tx_pkt->ol_flags;
417                 tx_ol_req = ol_flags & IGB_TX_OFFLOAD_MASK;
418
419                 /* If a Context Descriptor need be built . */
420                 if (tx_ol_req) {
421                         tx_offload.l2_len = tx_pkt->l2_len;
422                         tx_offload.l3_len = tx_pkt->l3_len;
423                         tx_offload.l4_len = tx_pkt->l4_len;
424                         tx_offload.vlan_tci = tx_pkt->vlan_tci;
425                         tx_offload.tso_segsz = tx_pkt->tso_segsz;
426                         tx_ol_req = check_tso_para(tx_ol_req, tx_offload);
427
428                         ctx = what_advctx_update(txq, tx_ol_req, tx_offload);
429                         /* Only allocate context descriptor if required*/
430                         new_ctx = (ctx == IGB_CTX_NUM);
431                         ctx = txq->ctx_curr + txq->ctx_start;
432                         tx_last = (uint16_t) (tx_last + new_ctx);
433                 }
434                 if (tx_last >= txq->nb_tx_desc)
435                         tx_last = (uint16_t) (tx_last - txq->nb_tx_desc);
436
437                 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u"
438                            " tx_first=%u tx_last=%u",
439                            (unsigned) txq->port_id,
440                            (unsigned) txq->queue_id,
441                            (unsigned) pkt_len,
442                            (unsigned) tx_id,
443                            (unsigned) tx_last);
444
445                 /*
446                  * Check if there are enough free descriptors in the TX ring
447                  * to transmit the next packet.
448                  * This operation is based on the two following rules:
449                  *
450                  *   1- Only check that the last needed TX descriptor can be
451                  *      allocated (by construction, if that descriptor is free,
452                  *      all intermediate ones are also free).
453                  *
454                  *      For this purpose, the index of the last TX descriptor
455                  *      used for a packet (the "last descriptor" of a packet)
456                  *      is recorded in the TX entries (the last one included)
457                  *      that are associated with all TX descriptors allocated
458                  *      for that packet.
459                  *
460                  *   2- Avoid to allocate the last free TX descriptor of the
461                  *      ring, in order to never set the TDT register with the
462                  *      same value stored in parallel by the NIC in the TDH
463                  *      register, which makes the TX engine of the NIC enter
464                  *      in a deadlock situation.
465                  *
466                  *      By extension, avoid to allocate a free descriptor that
467                  *      belongs to the last set of free descriptors allocated
468                  *      to the same packet previously transmitted.
469                  */
470
471                 /*
472                  * The "last descriptor" of the previously sent packet, if any,
473                  * which used the last descriptor to allocate.
474                  */
475                 tx_end = sw_ring[tx_last].last_id;
476
477                 /*
478                  * The next descriptor following that "last descriptor" in the
479                  * ring.
480                  */
481                 tx_end = sw_ring[tx_end].next_id;
482
483                 /*
484                  * The "last descriptor" associated with that next descriptor.
485                  */
486                 tx_end = sw_ring[tx_end].last_id;
487
488                 /*
489                  * Check that this descriptor is free.
490                  */
491                 if (! (txr[tx_end].wb.status & E1000_TXD_STAT_DD)) {
492                         if (nb_tx == 0)
493                                 return 0;
494                         goto end_of_tx;
495                 }
496
497                 /*
498                  * Set common flags of all TX Data Descriptors.
499                  *
500                  * The following bits must be set in all Data Descriptors:
501                  *   - E1000_ADVTXD_DTYP_DATA
502                  *   - E1000_ADVTXD_DCMD_DEXT
503                  *
504                  * The following bits must be set in the first Data Descriptor
505                  * and are ignored in the other ones:
506                  *   - E1000_ADVTXD_DCMD_IFCS
507                  *   - E1000_ADVTXD_MAC_1588
508                  *   - E1000_ADVTXD_DCMD_VLE
509                  *
510                  * The following bits must only be set in the last Data
511                  * Descriptor:
512                  *   - E1000_TXD_CMD_EOP
513                  *
514                  * The following bits can be set in any Data Descriptor, but
515                  * are only set in the last Data Descriptor:
516                  *   - E1000_TXD_CMD_RS
517                  */
518                 cmd_type_len = txq->txd_type |
519                         E1000_ADVTXD_DCMD_IFCS | E1000_ADVTXD_DCMD_DEXT;
520                 if (tx_ol_req & PKT_TX_TCP_SEG)
521                         pkt_len -= (tx_pkt->l2_len + tx_pkt->l3_len + tx_pkt->l4_len);
522                 olinfo_status = (pkt_len << E1000_ADVTXD_PAYLEN_SHIFT);
523 #if defined(RTE_LIBRTE_IEEE1588)
524                 if (ol_flags & PKT_TX_IEEE1588_TMST)
525                         cmd_type_len |= E1000_ADVTXD_MAC_TSTAMP;
526 #endif
527                 if (tx_ol_req) {
528                         /* Setup TX Advanced context descriptor if required */
529                         if (new_ctx) {
530                                 volatile struct e1000_adv_tx_context_desc *
531                                     ctx_txd;
532
533                                 ctx_txd = (volatile struct
534                                     e1000_adv_tx_context_desc *)
535                                     &txr[tx_id];
536
537                                 txn = &sw_ring[txe->next_id];
538                                 RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
539
540                                 if (txe->mbuf != NULL) {
541                                         rte_pktmbuf_free_seg(txe->mbuf);
542                                         txe->mbuf = NULL;
543                                 }
544
545                                 igbe_set_xmit_ctx(txq, ctx_txd, tx_ol_req, tx_offload);
546
547                                 txe->last_id = tx_last;
548                                 tx_id = txe->next_id;
549                                 txe = txn;
550                         }
551
552                         /* Setup the TX Advanced Data Descriptor */
553                         cmd_type_len  |= tx_desc_vlan_flags_to_cmdtype(tx_ol_req);
554                         olinfo_status |= tx_desc_cksum_flags_to_olinfo(tx_ol_req);
555                         olinfo_status |= (ctx << E1000_ADVTXD_IDX_SHIFT);
556                 }
557
558                 m_seg = tx_pkt;
559                 do {
560                         txn = &sw_ring[txe->next_id];
561                         txd = &txr[tx_id];
562
563                         if (txe->mbuf != NULL)
564                                 rte_pktmbuf_free_seg(txe->mbuf);
565                         txe->mbuf = m_seg;
566
567                         /*
568                          * Set up transmit descriptor.
569                          */
570                         slen = (uint16_t) m_seg->data_len;
571                         buf_dma_addr = rte_mbuf_data_iova(m_seg);
572                         txd->read.buffer_addr =
573                                 rte_cpu_to_le_64(buf_dma_addr);
574                         txd->read.cmd_type_len =
575                                 rte_cpu_to_le_32(cmd_type_len | slen);
576                         txd->read.olinfo_status =
577                                 rte_cpu_to_le_32(olinfo_status);
578                         txe->last_id = tx_last;
579                         tx_id = txe->next_id;
580                         txe = txn;
581                         m_seg = m_seg->next;
582                 } while (m_seg != NULL);
583
584                 /*
585                  * The last packet data descriptor needs End Of Packet (EOP)
586                  * and Report Status (RS).
587                  */
588                 txd->read.cmd_type_len |=
589                         rte_cpu_to_le_32(E1000_TXD_CMD_EOP | E1000_TXD_CMD_RS);
590         }
591  end_of_tx:
592         rte_wmb();
593
594         /*
595          * Set the Transmit Descriptor Tail (TDT).
596          */
597         E1000_PCI_REG_WRITE_RELAXED(txq->tdt_reg_addr, tx_id);
598         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
599                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
600                    (unsigned) tx_id, (unsigned) nb_tx);
601         txq->tx_tail = tx_id;
602
603         return nb_tx;
604 }
605
606 /*********************************************************************
607  *
608  *  TX prep functions
609  *
610  **********************************************************************/
611 uint16_t
612 eth_igb_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
613                 uint16_t nb_pkts)
614 {
615         int i, ret;
616         struct rte_mbuf *m;
617
618         for (i = 0; i < nb_pkts; i++) {
619                 m = tx_pkts[i];
620
621                 /* Check some limitations for TSO in hardware */
622                 if (m->ol_flags & PKT_TX_TCP_SEG)
623                         if ((m->tso_segsz > IGB_TSO_MAX_MSS) ||
624                                         (m->l2_len + m->l3_len + m->l4_len >
625                                         IGB_TSO_MAX_HDRLEN)) {
626                                 rte_errno = -EINVAL;
627                                 return i;
628                         }
629
630                 if (m->ol_flags & IGB_TX_OFFLOAD_NOTSUP_MASK) {
631                         rte_errno = -ENOTSUP;
632                         return i;
633                 }
634
635 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
636                 ret = rte_validate_tx_offload(m);
637                 if (ret != 0) {
638                         rte_errno = ret;
639                         return i;
640                 }
641 #endif
642                 ret = rte_net_intel_cksum_prepare(m);
643                 if (ret != 0) {
644                         rte_errno = ret;
645                         return i;
646                 }
647         }
648
649         return i;
650 }
651
652 /*********************************************************************
653  *
654  *  RX functions
655  *
656  **********************************************************************/
657 #define IGB_PACKET_TYPE_IPV4              0X01
658 #define IGB_PACKET_TYPE_IPV4_TCP          0X11
659 #define IGB_PACKET_TYPE_IPV4_UDP          0X21
660 #define IGB_PACKET_TYPE_IPV4_SCTP         0X41
661 #define IGB_PACKET_TYPE_IPV4_EXT          0X03
662 #define IGB_PACKET_TYPE_IPV4_EXT_SCTP     0X43
663 #define IGB_PACKET_TYPE_IPV6              0X04
664 #define IGB_PACKET_TYPE_IPV6_TCP          0X14
665 #define IGB_PACKET_TYPE_IPV6_UDP          0X24
666 #define IGB_PACKET_TYPE_IPV6_EXT          0X0C
667 #define IGB_PACKET_TYPE_IPV6_EXT_TCP      0X1C
668 #define IGB_PACKET_TYPE_IPV6_EXT_UDP      0X2C
669 #define IGB_PACKET_TYPE_IPV4_IPV6         0X05
670 #define IGB_PACKET_TYPE_IPV4_IPV6_TCP     0X15
671 #define IGB_PACKET_TYPE_IPV4_IPV6_UDP     0X25
672 #define IGB_PACKET_TYPE_IPV4_IPV6_EXT     0X0D
673 #define IGB_PACKET_TYPE_IPV4_IPV6_EXT_TCP 0X1D
674 #define IGB_PACKET_TYPE_IPV4_IPV6_EXT_UDP 0X2D
675 #define IGB_PACKET_TYPE_MAX               0X80
676 #define IGB_PACKET_TYPE_MASK              0X7F
677 #define IGB_PACKET_TYPE_SHIFT             0X04
678 static inline uint32_t
679 igb_rxd_pkt_info_to_pkt_type(uint16_t pkt_info)
680 {
681         static const uint32_t
682                 ptype_table[IGB_PACKET_TYPE_MAX] __rte_cache_aligned = {
683                 [IGB_PACKET_TYPE_IPV4] = RTE_PTYPE_L2_ETHER |
684                         RTE_PTYPE_L3_IPV4,
685                 [IGB_PACKET_TYPE_IPV4_EXT] = RTE_PTYPE_L2_ETHER |
686                         RTE_PTYPE_L3_IPV4_EXT,
687                 [IGB_PACKET_TYPE_IPV6] = RTE_PTYPE_L2_ETHER |
688                         RTE_PTYPE_L3_IPV6,
689                 [IGB_PACKET_TYPE_IPV4_IPV6] = RTE_PTYPE_L2_ETHER |
690                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_IP |
691                         RTE_PTYPE_INNER_L3_IPV6,
692                 [IGB_PACKET_TYPE_IPV6_EXT] = RTE_PTYPE_L2_ETHER |
693                         RTE_PTYPE_L3_IPV6_EXT,
694                 [IGB_PACKET_TYPE_IPV4_IPV6_EXT] = RTE_PTYPE_L2_ETHER |
695                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_IP |
696                         RTE_PTYPE_INNER_L3_IPV6_EXT,
697                 [IGB_PACKET_TYPE_IPV4_TCP] = RTE_PTYPE_L2_ETHER |
698                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_TCP,
699                 [IGB_PACKET_TYPE_IPV6_TCP] = RTE_PTYPE_L2_ETHER |
700                         RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_TCP,
701                 [IGB_PACKET_TYPE_IPV4_IPV6_TCP] = RTE_PTYPE_L2_ETHER |
702                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_IP |
703                         RTE_PTYPE_INNER_L3_IPV6 | RTE_PTYPE_INNER_L4_TCP,
704                 [IGB_PACKET_TYPE_IPV6_EXT_TCP] = RTE_PTYPE_L2_ETHER |
705                         RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_L4_TCP,
706                 [IGB_PACKET_TYPE_IPV4_IPV6_EXT_TCP] = RTE_PTYPE_L2_ETHER |
707                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_IP |
708                         RTE_PTYPE_INNER_L3_IPV6_EXT | RTE_PTYPE_INNER_L4_TCP,
709                 [IGB_PACKET_TYPE_IPV4_UDP] = RTE_PTYPE_L2_ETHER |
710                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_UDP,
711                 [IGB_PACKET_TYPE_IPV6_UDP] = RTE_PTYPE_L2_ETHER |
712                         RTE_PTYPE_L3_IPV6 | RTE_PTYPE_L4_UDP,
713                 [IGB_PACKET_TYPE_IPV4_IPV6_UDP] =  RTE_PTYPE_L2_ETHER |
714                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_IP |
715                         RTE_PTYPE_INNER_L3_IPV6 | RTE_PTYPE_INNER_L4_UDP,
716                 [IGB_PACKET_TYPE_IPV6_EXT_UDP] = RTE_PTYPE_L2_ETHER |
717                         RTE_PTYPE_L3_IPV6_EXT | RTE_PTYPE_L4_UDP,
718                 [IGB_PACKET_TYPE_IPV4_IPV6_EXT_UDP] = RTE_PTYPE_L2_ETHER |
719                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_TUNNEL_IP |
720                         RTE_PTYPE_INNER_L3_IPV6_EXT | RTE_PTYPE_INNER_L4_UDP,
721                 [IGB_PACKET_TYPE_IPV4_SCTP] = RTE_PTYPE_L2_ETHER |
722                         RTE_PTYPE_L3_IPV4 | RTE_PTYPE_L4_SCTP,
723                 [IGB_PACKET_TYPE_IPV4_EXT_SCTP] = RTE_PTYPE_L2_ETHER |
724                         RTE_PTYPE_L3_IPV4_EXT | RTE_PTYPE_L4_SCTP,
725         };
726         if (unlikely(pkt_info & E1000_RXDADV_PKTTYPE_ETQF))
727                 return RTE_PTYPE_UNKNOWN;
728
729         pkt_info = (pkt_info >> IGB_PACKET_TYPE_SHIFT) & IGB_PACKET_TYPE_MASK;
730
731         return ptype_table[pkt_info];
732 }
733
734 static inline uint64_t
735 rx_desc_hlen_type_rss_to_pkt_flags(struct igb_rx_queue *rxq, uint32_t hl_tp_rs)
736 {
737         uint64_t pkt_flags = ((hl_tp_rs & 0x0F) == 0) ?  0 : PKT_RX_RSS_HASH;
738
739 #if defined(RTE_LIBRTE_IEEE1588)
740         static uint32_t ip_pkt_etqf_map[8] = {
741                 0, 0, 0, PKT_RX_IEEE1588_PTP,
742                 0, 0, 0, 0,
743         };
744
745         struct rte_eth_dev dev = rte_eth_devices[rxq->port_id];
746         struct e1000_hw *hw = E1000_DEV_PRIVATE_TO_HW(dev.data->dev_private);
747
748         /* EtherType is in bits 8:10 in Packet Type, and not in the default 0:2 */
749         if (hw->mac.type == e1000_i210)
750                 pkt_flags |= ip_pkt_etqf_map[(hl_tp_rs >> 12) & 0x07];
751         else
752                 pkt_flags |= ip_pkt_etqf_map[(hl_tp_rs >> 4) & 0x07];
753 #else
754         RTE_SET_USED(rxq);
755 #endif
756
757         return pkt_flags;
758 }
759
760 static inline uint64_t
761 rx_desc_status_to_pkt_flags(uint32_t rx_status)
762 {
763         uint64_t pkt_flags;
764
765         /* Check if VLAN present */
766         pkt_flags = ((rx_status & E1000_RXD_STAT_VP) ?
767                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED : 0);
768
769 #if defined(RTE_LIBRTE_IEEE1588)
770         if (rx_status & E1000_RXD_STAT_TMST)
771                 pkt_flags = pkt_flags | PKT_RX_IEEE1588_TMST;
772 #endif
773         return pkt_flags;
774 }
775
776 static inline uint64_t
777 rx_desc_error_to_pkt_flags(uint32_t rx_status)
778 {
779         /*
780          * Bit 30: IPE, IPv4 checksum error
781          * Bit 29: L4I, L4I integrity error
782          */
783
784         static uint64_t error_to_pkt_flags_map[4] = {
785                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD,
786                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD,
787                 PKT_RX_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD,
788                 PKT_RX_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD
789         };
790         return error_to_pkt_flags_map[(rx_status >>
791                 E1000_RXD_ERR_CKSUM_BIT) & E1000_RXD_ERR_CKSUM_MSK];
792 }
793
794 uint16_t
795 eth_igb_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
796                uint16_t nb_pkts)
797 {
798         struct igb_rx_queue *rxq;
799         volatile union e1000_adv_rx_desc *rx_ring;
800         volatile union e1000_adv_rx_desc *rxdp;
801         struct igb_rx_entry *sw_ring;
802         struct igb_rx_entry *rxe;
803         struct rte_mbuf *rxm;
804         struct rte_mbuf *nmb;
805         union e1000_adv_rx_desc rxd;
806         uint64_t dma_addr;
807         uint32_t staterr;
808         uint32_t hlen_type_rss;
809         uint16_t pkt_len;
810         uint16_t rx_id;
811         uint16_t nb_rx;
812         uint16_t nb_hold;
813         uint64_t pkt_flags;
814
815         nb_rx = 0;
816         nb_hold = 0;
817         rxq = rx_queue;
818         rx_id = rxq->rx_tail;
819         rx_ring = rxq->rx_ring;
820         sw_ring = rxq->sw_ring;
821         while (nb_rx < nb_pkts) {
822                 /*
823                  * The order of operations here is important as the DD status
824                  * bit must not be read after any other descriptor fields.
825                  * rx_ring and rxdp are pointing to volatile data so the order
826                  * of accesses cannot be reordered by the compiler. If they were
827                  * not volatile, they could be reordered which could lead to
828                  * using invalid descriptor fields when read from rxd.
829                  */
830                 rxdp = &rx_ring[rx_id];
831                 staterr = rxdp->wb.upper.status_error;
832                 if (! (staterr & rte_cpu_to_le_32(E1000_RXD_STAT_DD)))
833                         break;
834                 rxd = *rxdp;
835
836                 /*
837                  * End of packet.
838                  *
839                  * If the E1000_RXD_STAT_EOP flag is not set, the RX packet is
840                  * likely to be invalid and to be dropped by the various
841                  * validation checks performed by the network stack.
842                  *
843                  * Allocate a new mbuf to replenish the RX ring descriptor.
844                  * If the allocation fails:
845                  *    - arrange for that RX descriptor to be the first one
846                  *      being parsed the next time the receive function is
847                  *      invoked [on the same queue].
848                  *
849                  *    - Stop parsing the RX ring and return immediately.
850                  *
851                  * This policy do not drop the packet received in the RX
852                  * descriptor for which the allocation of a new mbuf failed.
853                  * Thus, it allows that packet to be later retrieved if
854                  * mbuf have been freed in the mean time.
855                  * As a side effect, holding RX descriptors instead of
856                  * systematically giving them back to the NIC may lead to
857                  * RX ring exhaustion situations.
858                  * However, the NIC can gracefully prevent such situations
859                  * to happen by sending specific "back-pressure" flow control
860                  * frames to its peer(s).
861                  */
862                 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
863                            "staterr=0x%x pkt_len=%u",
864                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
865                            (unsigned) rx_id, (unsigned) staterr,
866                            (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
867
868                 nmb = rte_mbuf_raw_alloc(rxq->mb_pool);
869                 if (nmb == NULL) {
870                         PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
871                                    "queue_id=%u", (unsigned) rxq->port_id,
872                                    (unsigned) rxq->queue_id);
873                         rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
874                         break;
875                 }
876
877                 nb_hold++;
878                 rxe = &sw_ring[rx_id];
879                 rx_id++;
880                 if (rx_id == rxq->nb_rx_desc)
881                         rx_id = 0;
882
883                 /* Prefetch next mbuf while processing current one. */
884                 rte_igb_prefetch(sw_ring[rx_id].mbuf);
885
886                 /*
887                  * When next RX descriptor is on a cache-line boundary,
888                  * prefetch the next 4 RX descriptors and the next 8 pointers
889                  * to mbufs.
890                  */
891                 if ((rx_id & 0x3) == 0) {
892                         rte_igb_prefetch(&rx_ring[rx_id]);
893                         rte_igb_prefetch(&sw_ring[rx_id]);
894                 }
895
896                 rxm = rxe->mbuf;
897                 rxe->mbuf = nmb;
898                 dma_addr =
899                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
900                 rxdp->read.hdr_addr = 0;
901                 rxdp->read.pkt_addr = dma_addr;
902
903                 /*
904                  * Initialize the returned mbuf.
905                  * 1) setup generic mbuf fields:
906                  *    - number of segments,
907                  *    - next segment,
908                  *    - packet length,
909                  *    - RX port identifier.
910                  * 2) integrate hardware offload data, if any:
911                  *    - RSS flag & hash,
912                  *    - IP checksum flag,
913                  *    - VLAN TCI, if any,
914                  *    - error flags.
915                  */
916                 pkt_len = (uint16_t) (rte_le_to_cpu_16(rxd.wb.upper.length) -
917                                       rxq->crc_len);
918                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
919                 rte_packet_prefetch((char *)rxm->buf_addr + rxm->data_off);
920                 rxm->nb_segs = 1;
921                 rxm->next = NULL;
922                 rxm->pkt_len = pkt_len;
923                 rxm->data_len = pkt_len;
924                 rxm->port = rxq->port_id;
925
926                 rxm->hash.rss = rxd.wb.lower.hi_dword.rss;
927                 hlen_type_rss = rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
928
929                 /*
930                  * The vlan_tci field is only valid when PKT_RX_VLAN is
931                  * set in the pkt_flags field and must be in CPU byte order.
932                  */
933                 if ((staterr & rte_cpu_to_le_32(E1000_RXDEXT_STATERR_LB)) &&
934                                 (rxq->flags & IGB_RXQ_FLAG_LB_BSWAP_VLAN)) {
935                         rxm->vlan_tci = rte_be_to_cpu_16(rxd.wb.upper.vlan);
936                 } else {
937                         rxm->vlan_tci = rte_le_to_cpu_16(rxd.wb.upper.vlan);
938                 }
939                 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(rxq, hlen_type_rss);
940                 pkt_flags = pkt_flags | rx_desc_status_to_pkt_flags(staterr);
941                 pkt_flags = pkt_flags | rx_desc_error_to_pkt_flags(staterr);
942                 rxm->ol_flags = pkt_flags;
943                 rxm->packet_type = igb_rxd_pkt_info_to_pkt_type(rxd.wb.lower.
944                                                 lo_dword.hs_rss.pkt_info);
945
946                 /*
947                  * Store the mbuf address into the next entry of the array
948                  * of returned packets.
949                  */
950                 rx_pkts[nb_rx++] = rxm;
951         }
952         rxq->rx_tail = rx_id;
953
954         /*
955          * If the number of free RX descriptors is greater than the RX free
956          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
957          * register.
958          * Update the RDT with the value of the last processed RX descriptor
959          * minus 1, to guarantee that the RDT register is never equal to the
960          * RDH register, which creates a "full" ring situtation from the
961          * hardware point of view...
962          */
963         nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
964         if (nb_hold > rxq->rx_free_thresh) {
965                 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
966                            "nb_hold=%u nb_rx=%u",
967                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
968                            (unsigned) rx_id, (unsigned) nb_hold,
969                            (unsigned) nb_rx);
970                 rx_id = (uint16_t) ((rx_id == 0) ?
971                                      (rxq->nb_rx_desc - 1) : (rx_id - 1));
972                 E1000_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
973                 nb_hold = 0;
974         }
975         rxq->nb_rx_hold = nb_hold;
976         return nb_rx;
977 }
978
979 uint16_t
980 eth_igb_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
981                          uint16_t nb_pkts)
982 {
983         struct igb_rx_queue *rxq;
984         volatile union e1000_adv_rx_desc *rx_ring;
985         volatile union e1000_adv_rx_desc *rxdp;
986         struct igb_rx_entry *sw_ring;
987         struct igb_rx_entry *rxe;
988         struct rte_mbuf *first_seg;
989         struct rte_mbuf *last_seg;
990         struct rte_mbuf *rxm;
991         struct rte_mbuf *nmb;
992         union e1000_adv_rx_desc rxd;
993         uint64_t dma; /* Physical address of mbuf data buffer */
994         uint32_t staterr;
995         uint32_t hlen_type_rss;
996         uint16_t rx_id;
997         uint16_t nb_rx;
998         uint16_t nb_hold;
999         uint16_t data_len;
1000         uint64_t pkt_flags;
1001
1002         nb_rx = 0;
1003         nb_hold = 0;
1004         rxq = rx_queue;
1005         rx_id = rxq->rx_tail;
1006         rx_ring = rxq->rx_ring;
1007         sw_ring = rxq->sw_ring;
1008
1009         /*
1010          * Retrieve RX context of current packet, if any.
1011          */
1012         first_seg = rxq->pkt_first_seg;
1013         last_seg = rxq->pkt_last_seg;
1014
1015         while (nb_rx < nb_pkts) {
1016         next_desc:
1017                 /*
1018                  * The order of operations here is important as the DD status
1019                  * bit must not be read after any other descriptor fields.
1020                  * rx_ring and rxdp are pointing to volatile data so the order
1021                  * of accesses cannot be reordered by the compiler. If they were
1022                  * not volatile, they could be reordered which could lead to
1023                  * using invalid descriptor fields when read from rxd.
1024                  */
1025                 rxdp = &rx_ring[rx_id];
1026                 staterr = rxdp->wb.upper.status_error;
1027                 if (! (staterr & rte_cpu_to_le_32(E1000_RXD_STAT_DD)))
1028                         break;
1029                 rxd = *rxdp;
1030
1031                 /*
1032                  * Descriptor done.
1033                  *
1034                  * Allocate a new mbuf to replenish the RX ring descriptor.
1035                  * If the allocation fails:
1036                  *    - arrange for that RX descriptor to be the first one
1037                  *      being parsed the next time the receive function is
1038                  *      invoked [on the same queue].
1039                  *
1040                  *    - Stop parsing the RX ring and return immediately.
1041                  *
1042                  * This policy does not drop the packet received in the RX
1043                  * descriptor for which the allocation of a new mbuf failed.
1044                  * Thus, it allows that packet to be later retrieved if
1045                  * mbuf have been freed in the mean time.
1046                  * As a side effect, holding RX descriptors instead of
1047                  * systematically giving them back to the NIC may lead to
1048                  * RX ring exhaustion situations.
1049                  * However, the NIC can gracefully prevent such situations
1050                  * to happen by sending specific "back-pressure" flow control
1051                  * frames to its peer(s).
1052                  */
1053                 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
1054                            "staterr=0x%x data_len=%u",
1055                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1056                            (unsigned) rx_id, (unsigned) staterr,
1057                            (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
1058
1059                 nmb = rte_mbuf_raw_alloc(rxq->mb_pool);
1060                 if (nmb == NULL) {
1061                         PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
1062                                    "queue_id=%u", (unsigned) rxq->port_id,
1063                                    (unsigned) rxq->queue_id);
1064                         rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
1065                         break;
1066                 }
1067
1068                 nb_hold++;
1069                 rxe = &sw_ring[rx_id];
1070                 rx_id++;
1071                 if (rx_id == rxq->nb_rx_desc)
1072                         rx_id = 0;
1073
1074                 /* Prefetch next mbuf while processing current one. */
1075                 rte_igb_prefetch(sw_ring[rx_id].mbuf);
1076
1077                 /*
1078                  * When next RX descriptor is on a cache-line boundary,
1079                  * prefetch the next 4 RX descriptors and the next 8 pointers
1080                  * to mbufs.
1081                  */
1082                 if ((rx_id & 0x3) == 0) {
1083                         rte_igb_prefetch(&rx_ring[rx_id]);
1084                         rte_igb_prefetch(&sw_ring[rx_id]);
1085                 }
1086
1087                 /*
1088                  * Update RX descriptor with the physical address of the new
1089                  * data buffer of the new allocated mbuf.
1090                  */
1091                 rxm = rxe->mbuf;
1092                 rxe->mbuf = nmb;
1093                 dma = rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
1094                 rxdp->read.pkt_addr = dma;
1095                 rxdp->read.hdr_addr = 0;
1096
1097                 /*
1098                  * Set data length & data buffer address of mbuf.
1099                  */
1100                 data_len = rte_le_to_cpu_16(rxd.wb.upper.length);
1101                 rxm->data_len = data_len;
1102                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
1103
1104                 /*
1105                  * If this is the first buffer of the received packet,
1106                  * set the pointer to the first mbuf of the packet and
1107                  * initialize its context.
1108                  * Otherwise, update the total length and the number of segments
1109                  * of the current scattered packet, and update the pointer to
1110                  * the last mbuf of the current packet.
1111                  */
1112                 if (first_seg == NULL) {
1113                         first_seg = rxm;
1114                         first_seg->pkt_len = data_len;
1115                         first_seg->nb_segs = 1;
1116                 } else {
1117                         first_seg->pkt_len += data_len;
1118                         first_seg->nb_segs++;
1119                         last_seg->next = rxm;
1120                 }
1121
1122                 /*
1123                  * If this is not the last buffer of the received packet,
1124                  * update the pointer to the last mbuf of the current scattered
1125                  * packet and continue to parse the RX ring.
1126                  */
1127                 if (! (staterr & E1000_RXD_STAT_EOP)) {
1128                         last_seg = rxm;
1129                         goto next_desc;
1130                 }
1131
1132                 /*
1133                  * This is the last buffer of the received packet.
1134                  * If the CRC is not stripped by the hardware:
1135                  *   - Subtract the CRC length from the total packet length.
1136                  *   - If the last buffer only contains the whole CRC or a part
1137                  *     of it, free the mbuf associated to the last buffer.
1138                  *     If part of the CRC is also contained in the previous
1139                  *     mbuf, subtract the length of that CRC part from the
1140                  *     data length of the previous mbuf.
1141                  */
1142                 rxm->next = NULL;
1143                 if (unlikely(rxq->crc_len > 0)) {
1144                         first_seg->pkt_len -= ETHER_CRC_LEN;
1145                         if (data_len <= ETHER_CRC_LEN) {
1146                                 rte_pktmbuf_free_seg(rxm);
1147                                 first_seg->nb_segs--;
1148                                 last_seg->data_len = (uint16_t)
1149                                         (last_seg->data_len -
1150                                          (ETHER_CRC_LEN - data_len));
1151                                 last_seg->next = NULL;
1152                         } else
1153                                 rxm->data_len =
1154                                         (uint16_t) (data_len - ETHER_CRC_LEN);
1155                 }
1156
1157                 /*
1158                  * Initialize the first mbuf of the returned packet:
1159                  *    - RX port identifier,
1160                  *    - hardware offload data, if any:
1161                  *      - RSS flag & hash,
1162                  *      - IP checksum flag,
1163                  *      - VLAN TCI, if any,
1164                  *      - error flags.
1165                  */
1166                 first_seg->port = rxq->port_id;
1167                 first_seg->hash.rss = rxd.wb.lower.hi_dword.rss;
1168
1169                 /*
1170                  * The vlan_tci field is only valid when PKT_RX_VLAN is
1171                  * set in the pkt_flags field and must be in CPU byte order.
1172                  */
1173                 if ((staterr & rte_cpu_to_le_32(E1000_RXDEXT_STATERR_LB)) &&
1174                                 (rxq->flags & IGB_RXQ_FLAG_LB_BSWAP_VLAN)) {
1175                         first_seg->vlan_tci =
1176                                 rte_be_to_cpu_16(rxd.wb.upper.vlan);
1177                 } else {
1178                         first_seg->vlan_tci =
1179                                 rte_le_to_cpu_16(rxd.wb.upper.vlan);
1180                 }
1181                 hlen_type_rss = rte_le_to_cpu_32(rxd.wb.lower.lo_dword.data);
1182                 pkt_flags = rx_desc_hlen_type_rss_to_pkt_flags(rxq, hlen_type_rss);
1183                 pkt_flags = pkt_flags | rx_desc_status_to_pkt_flags(staterr);
1184                 pkt_flags = pkt_flags | rx_desc_error_to_pkt_flags(staterr);
1185                 first_seg->ol_flags = pkt_flags;
1186                 first_seg->packet_type = igb_rxd_pkt_info_to_pkt_type(rxd.wb.
1187                                         lower.lo_dword.hs_rss.pkt_info);
1188
1189                 /* Prefetch data of first segment, if configured to do so. */
1190                 rte_packet_prefetch((char *)first_seg->buf_addr +
1191                         first_seg->data_off);
1192
1193                 /*
1194                  * Store the mbuf address into the next entry of the array
1195                  * of returned packets.
1196                  */
1197                 rx_pkts[nb_rx++] = first_seg;
1198
1199                 /*
1200                  * Setup receipt context for a new packet.
1201                  */
1202                 first_seg = NULL;
1203         }
1204
1205         /*
1206          * Record index of the next RX descriptor to probe.
1207          */
1208         rxq->rx_tail = rx_id;
1209
1210         /*
1211          * Save receive context.
1212          */
1213         rxq->pkt_first_seg = first_seg;
1214         rxq->pkt_last_seg = last_seg;
1215
1216         /*
1217          * If the number of free RX descriptors is greater than the RX free
1218          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1219          * register.
1220          * Update the RDT with the value of the last processed RX descriptor
1221          * minus 1, to guarantee that the RDT register is never equal to the
1222          * RDH register, which creates a "full" ring situtation from the
1223          * hardware point of view...
1224          */
1225         nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
1226         if (nb_hold > rxq->rx_free_thresh) {
1227                 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1228                            "nb_hold=%u nb_rx=%u",
1229                            (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1230                            (unsigned) rx_id, (unsigned) nb_hold,
1231                            (unsigned) nb_rx);
1232                 rx_id = (uint16_t) ((rx_id == 0) ?
1233                                      (rxq->nb_rx_desc - 1) : (rx_id - 1));
1234                 E1000_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
1235                 nb_hold = 0;
1236         }
1237         rxq->nb_rx_hold = nb_hold;
1238         return nb_rx;
1239 }
1240
1241 /*
1242  * Maximum number of Ring Descriptors.
1243  *
1244  * Since RDLEN/TDLEN should be multiple of 128bytes, the number of ring
1245  * desscriptors should meet the following condition:
1246  *      (num_ring_desc * sizeof(struct e1000_rx/tx_desc)) % 128 == 0
1247  */
1248
1249 static void
1250 igb_tx_queue_release_mbufs(struct igb_tx_queue *txq)
1251 {
1252         unsigned i;
1253
1254         if (txq->sw_ring != NULL) {
1255                 for (i = 0; i < txq->nb_tx_desc; i++) {
1256                         if (txq->sw_ring[i].mbuf != NULL) {
1257                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
1258                                 txq->sw_ring[i].mbuf = NULL;
1259                         }
1260                 }
1261         }
1262 }
1263
1264 static void
1265 igb_tx_queue_release(struct igb_tx_queue *txq)
1266 {
1267         if (txq != NULL) {
1268                 igb_tx_queue_release_mbufs(txq);
1269                 rte_free(txq->sw_ring);
1270                 rte_free(txq);
1271         }
1272 }
1273
1274 void
1275 eth_igb_tx_queue_release(void *txq)
1276 {
1277         igb_tx_queue_release(txq);
1278 }
1279
1280 static int
1281 igb_tx_done_cleanup(struct igb_tx_queue *txq, uint32_t free_cnt)
1282 {
1283         struct igb_tx_entry *sw_ring;
1284         volatile union e1000_adv_tx_desc *txr;
1285         uint16_t tx_first; /* First segment analyzed. */
1286         uint16_t tx_id;    /* Current segment being processed. */
1287         uint16_t tx_last;  /* Last segment in the current packet. */
1288         uint16_t tx_next;  /* First segment of the next packet. */
1289         int count;
1290
1291         if (txq != NULL) {
1292                 count = 0;
1293                 sw_ring = txq->sw_ring;
1294                 txr = txq->tx_ring;
1295
1296                 /*
1297                  * tx_tail is the last sent packet on the sw_ring. Goto the end
1298                  * of that packet (the last segment in the packet chain) and
1299                  * then the next segment will be the start of the oldest segment
1300                  * in the sw_ring. This is the first packet that will be
1301                  * attempted to be freed.
1302                  */
1303
1304                 /* Get last segment in most recently added packet. */
1305                 tx_first = sw_ring[txq->tx_tail].last_id;
1306
1307                 /* Get the next segment, which is the oldest segment in ring. */
1308                 tx_first = sw_ring[tx_first].next_id;
1309
1310                 /* Set the current index to the first. */
1311                 tx_id = tx_first;
1312
1313                 /*
1314                  * Loop through each packet. For each packet, verify that an
1315                  * mbuf exists and that the last segment is free. If so, free
1316                  * it and move on.
1317                  */
1318                 while (1) {
1319                         tx_last = sw_ring[tx_id].last_id;
1320
1321                         if (sw_ring[tx_last].mbuf) {
1322                                 if (txr[tx_last].wb.status &
1323                                                 E1000_TXD_STAT_DD) {
1324                                         /*
1325                                          * Increment the number of packets
1326                                          * freed.
1327                                          */
1328                                         count++;
1329
1330                                         /* Get the start of the next packet. */
1331                                         tx_next = sw_ring[tx_last].next_id;
1332
1333                                         /*
1334                                          * Loop through all segments in a
1335                                          * packet.
1336                                          */
1337                                         do {
1338                                                 rte_pktmbuf_free_seg(sw_ring[tx_id].mbuf);
1339                                                 sw_ring[tx_id].mbuf = NULL;
1340                                                 sw_ring[tx_id].last_id = tx_id;
1341
1342                                                 /* Move to next segemnt. */
1343                                                 tx_id = sw_ring[tx_id].next_id;
1344
1345                                         } while (tx_id != tx_next);
1346
1347                                         if (unlikely(count == (int)free_cnt))
1348                                                 break;
1349                                 } else
1350                                         /*
1351                                          * mbuf still in use, nothing left to
1352                                          * free.
1353                                          */
1354                                         break;
1355                         } else {
1356                                 /*
1357                                  * There are multiple reasons to be here:
1358                                  * 1) All the packets on the ring have been
1359                                  *    freed - tx_id is equal to tx_first
1360                                  *    and some packets have been freed.
1361                                  *    - Done, exit
1362                                  * 2) Interfaces has not sent a rings worth of
1363                                  *    packets yet, so the segment after tail is
1364                                  *    still empty. Or a previous call to this
1365                                  *    function freed some of the segments but
1366                                  *    not all so there is a hole in the list.
1367                                  *    Hopefully this is a rare case.
1368                                  *    - Walk the list and find the next mbuf. If
1369                                  *      there isn't one, then done.
1370                                  */
1371                                 if (likely((tx_id == tx_first) && (count != 0)))
1372                                         break;
1373
1374                                 /*
1375                                  * Walk the list and find the next mbuf, if any.
1376                                  */
1377                                 do {
1378                                         /* Move to next segemnt. */
1379                                         tx_id = sw_ring[tx_id].next_id;
1380
1381                                         if (sw_ring[tx_id].mbuf)
1382                                                 break;
1383
1384                                 } while (tx_id != tx_first);
1385
1386                                 /*
1387                                  * Determine why previous loop bailed. If there
1388                                  * is not an mbuf, done.
1389                                  */
1390                                 if (sw_ring[tx_id].mbuf == NULL)
1391                                         break;
1392                         }
1393                 }
1394         } else
1395                 count = -ENODEV;
1396
1397         return count;
1398 }
1399
1400 int
1401 eth_igb_tx_done_cleanup(void *txq, uint32_t free_cnt)
1402 {
1403         return igb_tx_done_cleanup(txq, free_cnt);
1404 }
1405
1406 static void
1407 igb_reset_tx_queue_stat(struct igb_tx_queue *txq)
1408 {
1409         txq->tx_head = 0;
1410         txq->tx_tail = 0;
1411         txq->ctx_curr = 0;
1412         memset((void*)&txq->ctx_cache, 0,
1413                 IGB_CTX_NUM * sizeof(struct igb_advctx_info));
1414 }
1415
1416 static void
1417 igb_reset_tx_queue(struct igb_tx_queue *txq, struct rte_eth_dev *dev)
1418 {
1419         static const union e1000_adv_tx_desc zeroed_desc = {{0}};
1420         struct igb_tx_entry *txe = txq->sw_ring;
1421         uint16_t i, prev;
1422         struct e1000_hw *hw;
1423
1424         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1425         /* Zero out HW ring memory */
1426         for (i = 0; i < txq->nb_tx_desc; i++) {
1427                 txq->tx_ring[i] = zeroed_desc;
1428         }
1429
1430         /* Initialize ring entries */
1431         prev = (uint16_t)(txq->nb_tx_desc - 1);
1432         for (i = 0; i < txq->nb_tx_desc; i++) {
1433                 volatile union e1000_adv_tx_desc *txd = &(txq->tx_ring[i]);
1434
1435                 txd->wb.status = E1000_TXD_STAT_DD;
1436                 txe[i].mbuf = NULL;
1437                 txe[i].last_id = i;
1438                 txe[prev].next_id = i;
1439                 prev = i;
1440         }
1441
1442         txq->txd_type = E1000_ADVTXD_DTYP_DATA;
1443         /* 82575 specific, each tx queue will use 2 hw contexts */
1444         if (hw->mac.type == e1000_82575)
1445                 txq->ctx_start = txq->queue_id * IGB_CTX_NUM;
1446
1447         igb_reset_tx_queue_stat(txq);
1448 }
1449
1450 int
1451 eth_igb_tx_queue_setup(struct rte_eth_dev *dev,
1452                          uint16_t queue_idx,
1453                          uint16_t nb_desc,
1454                          unsigned int socket_id,
1455                          const struct rte_eth_txconf *tx_conf)
1456 {
1457         const struct rte_memzone *tz;
1458         struct igb_tx_queue *txq;
1459         struct e1000_hw     *hw;
1460         uint32_t size;
1461
1462         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1463
1464         /*
1465          * Validate number of transmit descriptors.
1466          * It must not exceed hardware maximum, and must be multiple
1467          * of E1000_ALIGN.
1468          */
1469         if (nb_desc % IGB_TXD_ALIGN != 0 ||
1470                         (nb_desc > E1000_MAX_RING_DESC) ||
1471                         (nb_desc < E1000_MIN_RING_DESC)) {
1472                 return -EINVAL;
1473         }
1474
1475         /*
1476          * The tx_free_thresh and tx_rs_thresh values are not used in the 1G
1477          * driver.
1478          */
1479         if (tx_conf->tx_free_thresh != 0)
1480                 PMD_INIT_LOG(INFO, "The tx_free_thresh parameter is not "
1481                              "used for the 1G driver.");
1482         if (tx_conf->tx_rs_thresh != 0)
1483                 PMD_INIT_LOG(INFO, "The tx_rs_thresh parameter is not "
1484                              "used for the 1G driver.");
1485         if (tx_conf->tx_thresh.wthresh == 0 && hw->mac.type != e1000_82576)
1486                 PMD_INIT_LOG(INFO, "To improve 1G driver performance, "
1487                              "consider setting the TX WTHRESH value to 4, 8, "
1488                              "or 16.");
1489
1490         /* Free memory prior to re-allocation if needed */
1491         if (dev->data->tx_queues[queue_idx] != NULL) {
1492                 igb_tx_queue_release(dev->data->tx_queues[queue_idx]);
1493                 dev->data->tx_queues[queue_idx] = NULL;
1494         }
1495
1496         /* First allocate the tx queue data structure */
1497         txq = rte_zmalloc("ethdev TX queue", sizeof(struct igb_tx_queue),
1498                                                         RTE_CACHE_LINE_SIZE);
1499         if (txq == NULL)
1500                 return -ENOMEM;
1501
1502         /*
1503          * Allocate TX ring hardware descriptors. A memzone large enough to
1504          * handle the maximum ring size is allocated in order to allow for
1505          * resizing in later calls to the queue setup function.
1506          */
1507         size = sizeof(union e1000_adv_tx_desc) * E1000_MAX_RING_DESC;
1508         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx, size,
1509                                       E1000_ALIGN, socket_id);
1510         if (tz == NULL) {
1511                 igb_tx_queue_release(txq);
1512                 return -ENOMEM;
1513         }
1514
1515         txq->nb_tx_desc = nb_desc;
1516         txq->pthresh = tx_conf->tx_thresh.pthresh;
1517         txq->hthresh = tx_conf->tx_thresh.hthresh;
1518         txq->wthresh = tx_conf->tx_thresh.wthresh;
1519         if (txq->wthresh > 0 && hw->mac.type == e1000_82576)
1520                 txq->wthresh = 1;
1521         txq->queue_id = queue_idx;
1522         txq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
1523                 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
1524         txq->port_id = dev->data->port_id;
1525
1526         txq->tdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_TDT(txq->reg_idx));
1527         txq->tx_ring_phys_addr = tz->iova;
1528
1529         txq->tx_ring = (union e1000_adv_tx_desc *) tz->addr;
1530         /* Allocate software ring */
1531         txq->sw_ring = rte_zmalloc("txq->sw_ring",
1532                                    sizeof(struct igb_tx_entry) * nb_desc,
1533                                    RTE_CACHE_LINE_SIZE);
1534         if (txq->sw_ring == NULL) {
1535                 igb_tx_queue_release(txq);
1536                 return -ENOMEM;
1537         }
1538         PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64,
1539                      txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
1540
1541         igb_reset_tx_queue(txq, dev);
1542         dev->tx_pkt_burst = eth_igb_xmit_pkts;
1543         dev->tx_pkt_prepare = &eth_igb_prep_pkts;
1544         dev->data->tx_queues[queue_idx] = txq;
1545
1546         return 0;
1547 }
1548
1549 static void
1550 igb_rx_queue_release_mbufs(struct igb_rx_queue *rxq)
1551 {
1552         unsigned i;
1553
1554         if (rxq->sw_ring != NULL) {
1555                 for (i = 0; i < rxq->nb_rx_desc; i++) {
1556                         if (rxq->sw_ring[i].mbuf != NULL) {
1557                                 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
1558                                 rxq->sw_ring[i].mbuf = NULL;
1559                         }
1560                 }
1561         }
1562 }
1563
1564 static void
1565 igb_rx_queue_release(struct igb_rx_queue *rxq)
1566 {
1567         if (rxq != NULL) {
1568                 igb_rx_queue_release_mbufs(rxq);
1569                 rte_free(rxq->sw_ring);
1570                 rte_free(rxq);
1571         }
1572 }
1573
1574 void
1575 eth_igb_rx_queue_release(void *rxq)
1576 {
1577         igb_rx_queue_release(rxq);
1578 }
1579
1580 static void
1581 igb_reset_rx_queue(struct igb_rx_queue *rxq)
1582 {
1583         static const union e1000_adv_rx_desc zeroed_desc = {{0}};
1584         unsigned i;
1585
1586         /* Zero out HW ring memory */
1587         for (i = 0; i < rxq->nb_rx_desc; i++) {
1588                 rxq->rx_ring[i] = zeroed_desc;
1589         }
1590
1591         rxq->rx_tail = 0;
1592         rxq->pkt_first_seg = NULL;
1593         rxq->pkt_last_seg = NULL;
1594 }
1595
1596 int
1597 eth_igb_rx_queue_setup(struct rte_eth_dev *dev,
1598                          uint16_t queue_idx,
1599                          uint16_t nb_desc,
1600                          unsigned int socket_id,
1601                          const struct rte_eth_rxconf *rx_conf,
1602                          struct rte_mempool *mp)
1603 {
1604         const struct rte_memzone *rz;
1605         struct igb_rx_queue *rxq;
1606         struct e1000_hw     *hw;
1607         unsigned int size;
1608
1609         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1610
1611         /*
1612          * Validate number of receive descriptors.
1613          * It must not exceed hardware maximum, and must be multiple
1614          * of E1000_ALIGN.
1615          */
1616         if (nb_desc % IGB_RXD_ALIGN != 0 ||
1617                         (nb_desc > E1000_MAX_RING_DESC) ||
1618                         (nb_desc < E1000_MIN_RING_DESC)) {
1619                 return -EINVAL;
1620         }
1621
1622         /* Free memory prior to re-allocation if needed */
1623         if (dev->data->rx_queues[queue_idx] != NULL) {
1624                 igb_rx_queue_release(dev->data->rx_queues[queue_idx]);
1625                 dev->data->rx_queues[queue_idx] = NULL;
1626         }
1627
1628         /* First allocate the RX queue data structure. */
1629         rxq = rte_zmalloc("ethdev RX queue", sizeof(struct igb_rx_queue),
1630                           RTE_CACHE_LINE_SIZE);
1631         if (rxq == NULL)
1632                 return -ENOMEM;
1633         rxq->mb_pool = mp;
1634         rxq->nb_rx_desc = nb_desc;
1635         rxq->pthresh = rx_conf->rx_thresh.pthresh;
1636         rxq->hthresh = rx_conf->rx_thresh.hthresh;
1637         rxq->wthresh = rx_conf->rx_thresh.wthresh;
1638         if (rxq->wthresh > 0 &&
1639             (hw->mac.type == e1000_82576 || hw->mac.type == e1000_vfadapt_i350))
1640                 rxq->wthresh = 1;
1641         rxq->drop_en = rx_conf->rx_drop_en;
1642         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1643         rxq->queue_id = queue_idx;
1644         rxq->reg_idx = (uint16_t)((RTE_ETH_DEV_SRIOV(dev).active == 0) ?
1645                 queue_idx : RTE_ETH_DEV_SRIOV(dev).def_pool_q_idx + queue_idx);
1646         rxq->port_id = dev->data->port_id;
1647         rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ? 0 :
1648                                   ETHER_CRC_LEN);
1649
1650         /*
1651          *  Allocate RX ring hardware descriptors. A memzone large enough to
1652          *  handle the maximum ring size is allocated in order to allow for
1653          *  resizing in later calls to the queue setup function.
1654          */
1655         size = sizeof(union e1000_adv_rx_desc) * E1000_MAX_RING_DESC;
1656         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx, size,
1657                                       E1000_ALIGN, socket_id);
1658         if (rz == NULL) {
1659                 igb_rx_queue_release(rxq);
1660                 return -ENOMEM;
1661         }
1662         rxq->rdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_RDT(rxq->reg_idx));
1663         rxq->rdh_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_RDH(rxq->reg_idx));
1664         rxq->rx_ring_phys_addr = rz->iova;
1665         rxq->rx_ring = (union e1000_adv_rx_desc *) rz->addr;
1666
1667         /* Allocate software ring. */
1668         rxq->sw_ring = rte_zmalloc("rxq->sw_ring",
1669                                    sizeof(struct igb_rx_entry) * nb_desc,
1670                                    RTE_CACHE_LINE_SIZE);
1671         if (rxq->sw_ring == NULL) {
1672                 igb_rx_queue_release(rxq);
1673                 return -ENOMEM;
1674         }
1675         PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64,
1676                      rxq->sw_ring, rxq->rx_ring, rxq->rx_ring_phys_addr);
1677
1678         dev->data->rx_queues[queue_idx] = rxq;
1679         igb_reset_rx_queue(rxq);
1680
1681         return 0;
1682 }
1683
1684 uint32_t
1685 eth_igb_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1686 {
1687 #define IGB_RXQ_SCAN_INTERVAL 4
1688         volatile union e1000_adv_rx_desc *rxdp;
1689         struct igb_rx_queue *rxq;
1690         uint32_t desc = 0;
1691
1692         rxq = dev->data->rx_queues[rx_queue_id];
1693         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1694
1695         while ((desc < rxq->nb_rx_desc) &&
1696                 (rxdp->wb.upper.status_error & E1000_RXD_STAT_DD)) {
1697                 desc += IGB_RXQ_SCAN_INTERVAL;
1698                 rxdp += IGB_RXQ_SCAN_INTERVAL;
1699                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
1700                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
1701                                 desc - rxq->nb_rx_desc]);
1702         }
1703
1704         return desc;
1705 }
1706
1707 int
1708 eth_igb_rx_descriptor_done(void *rx_queue, uint16_t offset)
1709 {
1710         volatile union e1000_adv_rx_desc *rxdp;
1711         struct igb_rx_queue *rxq = rx_queue;
1712         uint32_t desc;
1713
1714         if (unlikely(offset >= rxq->nb_rx_desc))
1715                 return 0;
1716         desc = rxq->rx_tail + offset;
1717         if (desc >= rxq->nb_rx_desc)
1718                 desc -= rxq->nb_rx_desc;
1719
1720         rxdp = &rxq->rx_ring[desc];
1721         return !!(rxdp->wb.upper.status_error & E1000_RXD_STAT_DD);
1722 }
1723
1724 int
1725 eth_igb_rx_descriptor_status(void *rx_queue, uint16_t offset)
1726 {
1727         struct igb_rx_queue *rxq = rx_queue;
1728         volatile uint32_t *status;
1729         uint32_t desc;
1730
1731         if (unlikely(offset >= rxq->nb_rx_desc))
1732                 return -EINVAL;
1733
1734         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
1735                 return RTE_ETH_RX_DESC_UNAVAIL;
1736
1737         desc = rxq->rx_tail + offset;
1738         if (desc >= rxq->nb_rx_desc)
1739                 desc -= rxq->nb_rx_desc;
1740
1741         status = &rxq->rx_ring[desc].wb.upper.status_error;
1742         if (*status & rte_cpu_to_le_32(E1000_RXD_STAT_DD))
1743                 return RTE_ETH_RX_DESC_DONE;
1744
1745         return RTE_ETH_RX_DESC_AVAIL;
1746 }
1747
1748 int
1749 eth_igb_tx_descriptor_status(void *tx_queue, uint16_t offset)
1750 {
1751         struct igb_tx_queue *txq = tx_queue;
1752         volatile uint32_t *status;
1753         uint32_t desc;
1754
1755         if (unlikely(offset >= txq->nb_tx_desc))
1756                 return -EINVAL;
1757
1758         desc = txq->tx_tail + offset;
1759         if (desc >= txq->nb_tx_desc)
1760                 desc -= txq->nb_tx_desc;
1761
1762         status = &txq->tx_ring[desc].wb.status;
1763         if (*status & rte_cpu_to_le_32(E1000_TXD_STAT_DD))
1764                 return RTE_ETH_TX_DESC_DONE;
1765
1766         return RTE_ETH_TX_DESC_FULL;
1767 }
1768
1769 void
1770 igb_dev_clear_queues(struct rte_eth_dev *dev)
1771 {
1772         uint16_t i;
1773         struct igb_tx_queue *txq;
1774         struct igb_rx_queue *rxq;
1775
1776         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1777                 txq = dev->data->tx_queues[i];
1778                 if (txq != NULL) {
1779                         igb_tx_queue_release_mbufs(txq);
1780                         igb_reset_tx_queue(txq, dev);
1781                 }
1782         }
1783
1784         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1785                 rxq = dev->data->rx_queues[i];
1786                 if (rxq != NULL) {
1787                         igb_rx_queue_release_mbufs(rxq);
1788                         igb_reset_rx_queue(rxq);
1789                 }
1790         }
1791 }
1792
1793 void
1794 igb_dev_free_queues(struct rte_eth_dev *dev)
1795 {
1796         uint16_t i;
1797
1798         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1799                 eth_igb_rx_queue_release(dev->data->rx_queues[i]);
1800                 dev->data->rx_queues[i] = NULL;
1801         }
1802         dev->data->nb_rx_queues = 0;
1803
1804         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1805                 eth_igb_tx_queue_release(dev->data->tx_queues[i]);
1806                 dev->data->tx_queues[i] = NULL;
1807         }
1808         dev->data->nb_tx_queues = 0;
1809 }
1810
1811 /**
1812  * Receive Side Scaling (RSS).
1813  * See section 7.1.1.7 in the following document:
1814  *     "Intel 82576 GbE Controller Datasheet" - Revision 2.45 October 2009
1815  *
1816  * Principles:
1817  * The source and destination IP addresses of the IP header and the source and
1818  * destination ports of TCP/UDP headers, if any, of received packets are hashed
1819  * against a configurable random key to compute a 32-bit RSS hash result.
1820  * The seven (7) LSBs of the 32-bit hash result are used as an index into a
1821  * 128-entry redirection table (RETA).  Each entry of the RETA provides a 3-bit
1822  * RSS output index which is used as the RX queue index where to store the
1823  * received packets.
1824  * The following output is supplied in the RX write-back descriptor:
1825  *     - 32-bit result of the Microsoft RSS hash function,
1826  *     - 4-bit RSS type field.
1827  */
1828
1829 /*
1830  * RSS random key supplied in section 7.1.1.7.3 of the Intel 82576 datasheet.
1831  * Used as the default key.
1832  */
1833 static uint8_t rss_intel_key[40] = {
1834         0x6D, 0x5A, 0x56, 0xDA, 0x25, 0x5B, 0x0E, 0xC2,
1835         0x41, 0x67, 0x25, 0x3D, 0x43, 0xA3, 0x8F, 0xB0,
1836         0xD0, 0xCA, 0x2B, 0xCB, 0xAE, 0x7B, 0x30, 0xB4,
1837         0x77, 0xCB, 0x2D, 0xA3, 0x80, 0x30, 0xF2, 0x0C,
1838         0x6A, 0x42, 0xB7, 0x3B, 0xBE, 0xAC, 0x01, 0xFA,
1839 };
1840
1841 static void
1842 igb_rss_disable(struct rte_eth_dev *dev)
1843 {
1844         struct e1000_hw *hw;
1845         uint32_t mrqc;
1846
1847         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1848         mrqc = E1000_READ_REG(hw, E1000_MRQC);
1849         mrqc &= ~E1000_MRQC_ENABLE_MASK;
1850         E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
1851 }
1852
1853 static void
1854 igb_hw_rss_hash_set(struct e1000_hw *hw, struct rte_eth_rss_conf *rss_conf)
1855 {
1856         uint8_t  *hash_key;
1857         uint32_t rss_key;
1858         uint32_t mrqc;
1859         uint64_t rss_hf;
1860         uint16_t i;
1861
1862         hash_key = rss_conf->rss_key;
1863         if (hash_key != NULL) {
1864                 /* Fill in RSS hash key */
1865                 for (i = 0; i < 10; i++) {
1866                         rss_key  = hash_key[(i * 4)];
1867                         rss_key |= hash_key[(i * 4) + 1] << 8;
1868                         rss_key |= hash_key[(i * 4) + 2] << 16;
1869                         rss_key |= hash_key[(i * 4) + 3] << 24;
1870                         E1000_WRITE_REG_ARRAY(hw, E1000_RSSRK(0), i, rss_key);
1871                 }
1872         }
1873
1874         /* Set configured hashing protocols in MRQC register */
1875         rss_hf = rss_conf->rss_hf;
1876         mrqc = E1000_MRQC_ENABLE_RSS_4Q; /* RSS enabled. */
1877         if (rss_hf & ETH_RSS_IPV4)
1878                 mrqc |= E1000_MRQC_RSS_FIELD_IPV4;
1879         if (rss_hf & ETH_RSS_NONFRAG_IPV4_TCP)
1880                 mrqc |= E1000_MRQC_RSS_FIELD_IPV4_TCP;
1881         if (rss_hf & ETH_RSS_IPV6)
1882                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6;
1883         if (rss_hf & ETH_RSS_IPV6_EX)
1884                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_EX;
1885         if (rss_hf & ETH_RSS_NONFRAG_IPV6_TCP)
1886                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_TCP;
1887         if (rss_hf & ETH_RSS_IPV6_TCP_EX)
1888                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_TCP_EX;
1889         if (rss_hf & ETH_RSS_NONFRAG_IPV4_UDP)
1890                 mrqc |= E1000_MRQC_RSS_FIELD_IPV4_UDP;
1891         if (rss_hf & ETH_RSS_NONFRAG_IPV6_UDP)
1892                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_UDP;
1893         if (rss_hf & ETH_RSS_IPV6_UDP_EX)
1894                 mrqc |= E1000_MRQC_RSS_FIELD_IPV6_UDP_EX;
1895         E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
1896 }
1897
1898 int
1899 eth_igb_rss_hash_update(struct rte_eth_dev *dev,
1900                         struct rte_eth_rss_conf *rss_conf)
1901 {
1902         struct e1000_hw *hw;
1903         uint32_t mrqc;
1904         uint64_t rss_hf;
1905
1906         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1907
1908         /*
1909          * Before changing anything, first check that the update RSS operation
1910          * does not attempt to disable RSS, if RSS was enabled at
1911          * initialization time, or does not attempt to enable RSS, if RSS was
1912          * disabled at initialization time.
1913          */
1914         rss_hf = rss_conf->rss_hf & IGB_RSS_OFFLOAD_ALL;
1915         mrqc = E1000_READ_REG(hw, E1000_MRQC);
1916         if (!(mrqc & E1000_MRQC_ENABLE_MASK)) { /* RSS disabled */
1917                 if (rss_hf != 0) /* Enable RSS */
1918                         return -(EINVAL);
1919                 return 0; /* Nothing to do */
1920         }
1921         /* RSS enabled */
1922         if (rss_hf == 0) /* Disable RSS */
1923                 return -(EINVAL);
1924         igb_hw_rss_hash_set(hw, rss_conf);
1925         return 0;
1926 }
1927
1928 int eth_igb_rss_hash_conf_get(struct rte_eth_dev *dev,
1929                               struct rte_eth_rss_conf *rss_conf)
1930 {
1931         struct e1000_hw *hw;
1932         uint8_t *hash_key;
1933         uint32_t rss_key;
1934         uint32_t mrqc;
1935         uint64_t rss_hf;
1936         uint16_t i;
1937
1938         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1939         hash_key = rss_conf->rss_key;
1940         if (hash_key != NULL) {
1941                 /* Return RSS hash key */
1942                 for (i = 0; i < 10; i++) {
1943                         rss_key = E1000_READ_REG_ARRAY(hw, E1000_RSSRK(0), i);
1944                         hash_key[(i * 4)] = rss_key & 0x000000FF;
1945                         hash_key[(i * 4) + 1] = (rss_key >> 8) & 0x000000FF;
1946                         hash_key[(i * 4) + 2] = (rss_key >> 16) & 0x000000FF;
1947                         hash_key[(i * 4) + 3] = (rss_key >> 24) & 0x000000FF;
1948                 }
1949         }
1950
1951         /* Get RSS functions configured in MRQC register */
1952         mrqc = E1000_READ_REG(hw, E1000_MRQC);
1953         if ((mrqc & E1000_MRQC_ENABLE_RSS_4Q) == 0) { /* RSS is disabled */
1954                 rss_conf->rss_hf = 0;
1955                 return 0;
1956         }
1957         rss_hf = 0;
1958         if (mrqc & E1000_MRQC_RSS_FIELD_IPV4)
1959                 rss_hf |= ETH_RSS_IPV4;
1960         if (mrqc & E1000_MRQC_RSS_FIELD_IPV4_TCP)
1961                 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
1962         if (mrqc & E1000_MRQC_RSS_FIELD_IPV6)
1963                 rss_hf |= ETH_RSS_IPV6;
1964         if (mrqc & E1000_MRQC_RSS_FIELD_IPV6_EX)
1965                 rss_hf |= ETH_RSS_IPV6_EX;
1966         if (mrqc & E1000_MRQC_RSS_FIELD_IPV6_TCP)
1967                 rss_hf |= ETH_RSS_NONFRAG_IPV6_TCP;
1968         if (mrqc & E1000_MRQC_RSS_FIELD_IPV6_TCP_EX)
1969                 rss_hf |= ETH_RSS_IPV6_TCP_EX;
1970         if (mrqc & E1000_MRQC_RSS_FIELD_IPV4_UDP)
1971                 rss_hf |= ETH_RSS_NONFRAG_IPV4_UDP;
1972         if (mrqc & E1000_MRQC_RSS_FIELD_IPV6_UDP)
1973                 rss_hf |= ETH_RSS_NONFRAG_IPV6_UDP;
1974         if (mrqc & E1000_MRQC_RSS_FIELD_IPV6_UDP_EX)
1975                 rss_hf |= ETH_RSS_IPV6_UDP_EX;
1976         rss_conf->rss_hf = rss_hf;
1977         return 0;
1978 }
1979
1980 static void
1981 igb_rss_configure(struct rte_eth_dev *dev)
1982 {
1983         struct rte_eth_rss_conf rss_conf;
1984         struct e1000_hw *hw;
1985         uint32_t shift;
1986         uint16_t i;
1987
1988         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1989
1990         /* Fill in redirection table. */
1991         shift = (hw->mac.type == e1000_82575) ? 6 : 0;
1992         for (i = 0; i < 128; i++) {
1993                 union e1000_reta {
1994                         uint32_t dword;
1995                         uint8_t  bytes[4];
1996                 } reta;
1997                 uint8_t q_idx;
1998
1999                 q_idx = (uint8_t) ((dev->data->nb_rx_queues > 1) ?
2000                                    i % dev->data->nb_rx_queues : 0);
2001                 reta.bytes[i & 3] = (uint8_t) (q_idx << shift);
2002                 if ((i & 3) == 3)
2003                         E1000_WRITE_REG(hw, E1000_RETA(i >> 2), reta.dword);
2004         }
2005
2006         /*
2007          * Configure the RSS key and the RSS protocols used to compute
2008          * the RSS hash of input packets.
2009          */
2010         rss_conf = dev->data->dev_conf.rx_adv_conf.rss_conf;
2011         if ((rss_conf.rss_hf & IGB_RSS_OFFLOAD_ALL) == 0) {
2012                 igb_rss_disable(dev);
2013                 return;
2014         }
2015         if (rss_conf.rss_key == NULL)
2016                 rss_conf.rss_key = rss_intel_key; /* Default hash key */
2017         igb_hw_rss_hash_set(hw, &rss_conf);
2018 }
2019
2020 /*
2021  * Check if the mac type support VMDq or not.
2022  * Return 1 if it supports, otherwise, return 0.
2023  */
2024 static int
2025 igb_is_vmdq_supported(const struct rte_eth_dev *dev)
2026 {
2027         const struct e1000_hw *hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2028
2029         switch (hw->mac.type) {
2030         case e1000_82576:
2031         case e1000_82580:
2032         case e1000_i350:
2033                 return 1;
2034         case e1000_82540:
2035         case e1000_82541:
2036         case e1000_82542:
2037         case e1000_82543:
2038         case e1000_82544:
2039         case e1000_82545:
2040         case e1000_82546:
2041         case e1000_82547:
2042         case e1000_82571:
2043         case e1000_82572:
2044         case e1000_82573:
2045         case e1000_82574:
2046         case e1000_82583:
2047         case e1000_i210:
2048         case e1000_i211:
2049         default:
2050                 PMD_INIT_LOG(ERR, "Cannot support VMDq feature");
2051                 return 0;
2052         }
2053 }
2054
2055 static int
2056 igb_vmdq_rx_hw_configure(struct rte_eth_dev *dev)
2057 {
2058         struct rte_eth_vmdq_rx_conf *cfg;
2059         struct e1000_hw *hw;
2060         uint32_t mrqc, vt_ctl, vmolr, rctl;
2061         int i;
2062
2063         PMD_INIT_FUNC_TRACE();
2064
2065         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2066         cfg = &dev->data->dev_conf.rx_adv_conf.vmdq_rx_conf;
2067
2068         /* Check if mac type can support VMDq, return value of 0 means NOT support */
2069         if (igb_is_vmdq_supported(dev) == 0)
2070                 return -1;
2071
2072         igb_rss_disable(dev);
2073
2074         /* RCTL: eanble VLAN filter */
2075         rctl = E1000_READ_REG(hw, E1000_RCTL);
2076         rctl |= E1000_RCTL_VFE;
2077         E1000_WRITE_REG(hw, E1000_RCTL, rctl);
2078
2079         /* MRQC: enable vmdq */
2080         mrqc = E1000_READ_REG(hw, E1000_MRQC);
2081         mrqc |= E1000_MRQC_ENABLE_VMDQ;
2082         E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
2083
2084         /* VTCTL:  pool selection according to VLAN tag */
2085         vt_ctl = E1000_READ_REG(hw, E1000_VT_CTL);
2086         if (cfg->enable_default_pool)
2087                 vt_ctl |= (cfg->default_pool << E1000_VT_CTL_DEFAULT_POOL_SHIFT);
2088         vt_ctl |= E1000_VT_CTL_IGNORE_MAC;
2089         E1000_WRITE_REG(hw, E1000_VT_CTL, vt_ctl);
2090
2091         for (i = 0; i < E1000_VMOLR_SIZE; i++) {
2092                 vmolr = E1000_READ_REG(hw, E1000_VMOLR(i));
2093                 vmolr &= ~(E1000_VMOLR_AUPE | E1000_VMOLR_ROMPE |
2094                         E1000_VMOLR_ROPE | E1000_VMOLR_BAM |
2095                         E1000_VMOLR_MPME);
2096
2097                 if (cfg->rx_mode & ETH_VMDQ_ACCEPT_UNTAG)
2098                         vmolr |= E1000_VMOLR_AUPE;
2099                 if (cfg->rx_mode & ETH_VMDQ_ACCEPT_HASH_MC)
2100                         vmolr |= E1000_VMOLR_ROMPE;
2101                 if (cfg->rx_mode & ETH_VMDQ_ACCEPT_HASH_UC)
2102                         vmolr |= E1000_VMOLR_ROPE;
2103                 if (cfg->rx_mode & ETH_VMDQ_ACCEPT_BROADCAST)
2104                         vmolr |= E1000_VMOLR_BAM;
2105                 if (cfg->rx_mode & ETH_VMDQ_ACCEPT_MULTICAST)
2106                         vmolr |= E1000_VMOLR_MPME;
2107
2108                 E1000_WRITE_REG(hw, E1000_VMOLR(i), vmolr);
2109         }
2110
2111         /*
2112          * VMOLR: set STRVLAN as 1 if IGMAC in VTCTL is set as 1
2113          * Both 82576 and 82580 support it
2114          */
2115         if (hw->mac.type != e1000_i350) {
2116                 for (i = 0; i < E1000_VMOLR_SIZE; i++) {
2117                         vmolr = E1000_READ_REG(hw, E1000_VMOLR(i));
2118                         vmolr |= E1000_VMOLR_STRVLAN;
2119                         E1000_WRITE_REG(hw, E1000_VMOLR(i), vmolr);
2120                 }
2121         }
2122
2123         /* VFTA - enable all vlan filters */
2124         for (i = 0; i < IGB_VFTA_SIZE; i++)
2125                 E1000_WRITE_REG(hw, (E1000_VFTA+(i*4)), UINT32_MAX);
2126
2127         /* VFRE: 8 pools enabling for rx, both 82576 and i350 support it */
2128         if (hw->mac.type != e1000_82580)
2129                 E1000_WRITE_REG(hw, E1000_VFRE, E1000_MBVFICR_VFREQ_MASK);
2130
2131         /*
2132          * RAH/RAL - allow pools to read specific mac addresses
2133          * In this case, all pools should be able to read from mac addr 0
2134          */
2135         E1000_WRITE_REG(hw, E1000_RAH(0), (E1000_RAH_AV | UINT16_MAX));
2136         E1000_WRITE_REG(hw, E1000_RAL(0), UINT32_MAX);
2137
2138         /* VLVF: set up filters for vlan tags as configured */
2139         for (i = 0; i < cfg->nb_pool_maps; i++) {
2140                 /* set vlan id in VF register and set the valid bit */
2141                 E1000_WRITE_REG(hw, E1000_VLVF(i), (E1000_VLVF_VLANID_ENABLE | \
2142                         (cfg->pool_map[i].vlan_id & ETH_VLAN_ID_MAX) | \
2143                         ((cfg->pool_map[i].pools << E1000_VLVF_POOLSEL_SHIFT ) & \
2144                         E1000_VLVF_POOLSEL_MASK)));
2145         }
2146
2147         E1000_WRITE_FLUSH(hw);
2148
2149         return 0;
2150 }
2151
2152
2153 /*********************************************************************
2154  *
2155  *  Enable receive unit.
2156  *
2157  **********************************************************************/
2158
2159 static int
2160 igb_alloc_rx_queue_mbufs(struct igb_rx_queue *rxq)
2161 {
2162         struct igb_rx_entry *rxe = rxq->sw_ring;
2163         uint64_t dma_addr;
2164         unsigned i;
2165
2166         /* Initialize software ring entries. */
2167         for (i = 0; i < rxq->nb_rx_desc; i++) {
2168                 volatile union e1000_adv_rx_desc *rxd;
2169                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
2170
2171                 if (mbuf == NULL) {
2172                         PMD_INIT_LOG(ERR, "RX mbuf alloc failed "
2173                                      "queue_id=%hu", rxq->queue_id);
2174                         return -ENOMEM;
2175                 }
2176                 dma_addr =
2177                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2178                 rxd = &rxq->rx_ring[i];
2179                 rxd->read.hdr_addr = 0;
2180                 rxd->read.pkt_addr = dma_addr;
2181                 rxe[i].mbuf = mbuf;
2182         }
2183
2184         return 0;
2185 }
2186
2187 #define E1000_MRQC_DEF_Q_SHIFT               (3)
2188 static int
2189 igb_dev_mq_rx_configure(struct rte_eth_dev *dev)
2190 {
2191         struct e1000_hw *hw =
2192                 E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2193         uint32_t mrqc;
2194
2195         if (RTE_ETH_DEV_SRIOV(dev).active == ETH_8_POOLS) {
2196                 /*
2197                  * SRIOV active scheme
2198                  * FIXME if support RSS together with VMDq & SRIOV
2199                  */
2200                 mrqc = E1000_MRQC_ENABLE_VMDQ;
2201                 /* 011b Def_Q ignore, according to VT_CTL.DEF_PL */
2202                 mrqc |= 0x3 << E1000_MRQC_DEF_Q_SHIFT;
2203                 E1000_WRITE_REG(hw, E1000_MRQC, mrqc);
2204         } else if(RTE_ETH_DEV_SRIOV(dev).active == 0) {
2205                 /*
2206                  * SRIOV inactive scheme
2207                  */
2208                 switch (dev->data->dev_conf.rxmode.mq_mode) {
2209                         case ETH_MQ_RX_RSS:
2210                                 igb_rss_configure(dev);
2211                                 break;
2212                         case ETH_MQ_RX_VMDQ_ONLY:
2213                                 /*Configure general VMDQ only RX parameters*/
2214                                 igb_vmdq_rx_hw_configure(dev);
2215                                 break;
2216                         case ETH_MQ_RX_NONE:
2217                                 /* if mq_mode is none, disable rss mode.*/
2218                         default:
2219                                 igb_rss_disable(dev);
2220                                 break;
2221                 }
2222         }
2223
2224         return 0;
2225 }
2226
2227 int
2228 eth_igb_rx_init(struct rte_eth_dev *dev)
2229 {
2230         struct e1000_hw     *hw;
2231         struct igb_rx_queue *rxq;
2232         uint32_t rctl;
2233         uint32_t rxcsum;
2234         uint32_t srrctl;
2235         uint16_t buf_size;
2236         uint16_t rctl_bsize;
2237         uint16_t i;
2238         int ret;
2239
2240         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2241         srrctl = 0;
2242
2243         /*
2244          * Make sure receives are disabled while setting
2245          * up the descriptor ring.
2246          */
2247         rctl = E1000_READ_REG(hw, E1000_RCTL);
2248         E1000_WRITE_REG(hw, E1000_RCTL, rctl & ~E1000_RCTL_EN);
2249
2250         /*
2251          * Configure support of jumbo frames, if any.
2252          */
2253         if (dev->data->dev_conf.rxmode.jumbo_frame == 1) {
2254                 rctl |= E1000_RCTL_LPE;
2255
2256                 /*
2257                  * Set maximum packet length by default, and might be updated
2258                  * together with enabling/disabling dual VLAN.
2259                  */
2260                 E1000_WRITE_REG(hw, E1000_RLPML,
2261                         dev->data->dev_conf.rxmode.max_rx_pkt_len +
2262                                                 VLAN_TAG_SIZE);
2263         } else
2264                 rctl &= ~E1000_RCTL_LPE;
2265
2266         /* Configure and enable each RX queue. */
2267         rctl_bsize = 0;
2268         dev->rx_pkt_burst = eth_igb_recv_pkts;
2269         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2270                 uint64_t bus_addr;
2271                 uint32_t rxdctl;
2272
2273                 rxq = dev->data->rx_queues[i];
2274
2275                 rxq->flags = 0;
2276                 /*
2277                  * i350 and i354 vlan packets have vlan tags byte swapped.
2278                  */
2279                 if (hw->mac.type == e1000_i350 || hw->mac.type == e1000_i354) {
2280                         rxq->flags |= IGB_RXQ_FLAG_LB_BSWAP_VLAN;
2281                         PMD_INIT_LOG(DEBUG, "IGB rx vlan bswap required");
2282                 } else {
2283                         PMD_INIT_LOG(DEBUG, "IGB rx vlan bswap not required");
2284                 }
2285
2286                 /* Allocate buffers for descriptor rings and set up queue */
2287                 ret = igb_alloc_rx_queue_mbufs(rxq);
2288                 if (ret)
2289                         return ret;
2290
2291                 /*
2292                  * Reset crc_len in case it was changed after queue setup by a
2293                  *  call to configure
2294                  */
2295                 rxq->crc_len =
2296                         (uint8_t)(dev->data->dev_conf.rxmode.hw_strip_crc ?
2297                                                         0 : ETHER_CRC_LEN);
2298
2299                 bus_addr = rxq->rx_ring_phys_addr;
2300                 E1000_WRITE_REG(hw, E1000_RDLEN(rxq->reg_idx),
2301                                 rxq->nb_rx_desc *
2302                                 sizeof(union e1000_adv_rx_desc));
2303                 E1000_WRITE_REG(hw, E1000_RDBAH(rxq->reg_idx),
2304                                 (uint32_t)(bus_addr >> 32));
2305                 E1000_WRITE_REG(hw, E1000_RDBAL(rxq->reg_idx), (uint32_t)bus_addr);
2306
2307                 srrctl = E1000_SRRCTL_DESCTYPE_ADV_ONEBUF;
2308
2309                 /*
2310                  * Configure RX buffer size.
2311                  */
2312                 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
2313                         RTE_PKTMBUF_HEADROOM);
2314                 if (buf_size >= 1024) {
2315                         /*
2316                          * Configure the BSIZEPACKET field of the SRRCTL
2317                          * register of the queue.
2318                          * Value is in 1 KB resolution, from 1 KB to 127 KB.
2319                          * If this field is equal to 0b, then RCTL.BSIZE
2320                          * determines the RX packet buffer size.
2321                          */
2322                         srrctl |= ((buf_size >> E1000_SRRCTL_BSIZEPKT_SHIFT) &
2323                                    E1000_SRRCTL_BSIZEPKT_MASK);
2324                         buf_size = (uint16_t) ((srrctl &
2325                                                 E1000_SRRCTL_BSIZEPKT_MASK) <<
2326                                                E1000_SRRCTL_BSIZEPKT_SHIFT);
2327
2328                         /* It adds dual VLAN length for supporting dual VLAN */
2329                         if ((dev->data->dev_conf.rxmode.max_rx_pkt_len +
2330                                                 2 * VLAN_TAG_SIZE) > buf_size){
2331                                 if (!dev->data->scattered_rx)
2332                                         PMD_INIT_LOG(DEBUG,
2333                                                      "forcing scatter mode");
2334                                 dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
2335                                 dev->data->scattered_rx = 1;
2336                         }
2337                 } else {
2338                         /*
2339                          * Use BSIZE field of the device RCTL register.
2340                          */
2341                         if ((rctl_bsize == 0) || (rctl_bsize > buf_size))
2342                                 rctl_bsize = buf_size;
2343                         if (!dev->data->scattered_rx)
2344                                 PMD_INIT_LOG(DEBUG, "forcing scatter mode");
2345                         dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
2346                         dev->data->scattered_rx = 1;
2347                 }
2348
2349                 /* Set if packets are dropped when no descriptors available */
2350                 if (rxq->drop_en)
2351                         srrctl |= E1000_SRRCTL_DROP_EN;
2352
2353                 E1000_WRITE_REG(hw, E1000_SRRCTL(rxq->reg_idx), srrctl);
2354
2355                 /* Enable this RX queue. */
2356                 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(rxq->reg_idx));
2357                 rxdctl |= E1000_RXDCTL_QUEUE_ENABLE;
2358                 rxdctl &= 0xFFF00000;
2359                 rxdctl |= (rxq->pthresh & 0x1F);
2360                 rxdctl |= ((rxq->hthresh & 0x1F) << 8);
2361                 rxdctl |= ((rxq->wthresh & 0x1F) << 16);
2362                 E1000_WRITE_REG(hw, E1000_RXDCTL(rxq->reg_idx), rxdctl);
2363         }
2364
2365         if (dev->data->dev_conf.rxmode.enable_scatter) {
2366                 if (!dev->data->scattered_rx)
2367                         PMD_INIT_LOG(DEBUG, "forcing scatter mode");
2368                 dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
2369                 dev->data->scattered_rx = 1;
2370         }
2371
2372         /*
2373          * Setup BSIZE field of RCTL register, if needed.
2374          * Buffer sizes >= 1024 are not [supposed to be] setup in the RCTL
2375          * register, since the code above configures the SRRCTL register of
2376          * the RX queue in such a case.
2377          * All configurable sizes are:
2378          * 16384: rctl |= (E1000_RCTL_SZ_16384 | E1000_RCTL_BSEX);
2379          *  8192: rctl |= (E1000_RCTL_SZ_8192  | E1000_RCTL_BSEX);
2380          *  4096: rctl |= (E1000_RCTL_SZ_4096  | E1000_RCTL_BSEX);
2381          *  2048: rctl |= E1000_RCTL_SZ_2048;
2382          *  1024: rctl |= E1000_RCTL_SZ_1024;
2383          *   512: rctl |= E1000_RCTL_SZ_512;
2384          *   256: rctl |= E1000_RCTL_SZ_256;
2385          */
2386         if (rctl_bsize > 0) {
2387                 if (rctl_bsize >= 512) /* 512 <= buf_size < 1024 - use 512 */
2388                         rctl |= E1000_RCTL_SZ_512;
2389                 else /* 256 <= buf_size < 512 - use 256 */
2390                         rctl |= E1000_RCTL_SZ_256;
2391         }
2392
2393         /*
2394          * Configure RSS if device configured with multiple RX queues.
2395          */
2396         igb_dev_mq_rx_configure(dev);
2397
2398         /* Update the rctl since igb_dev_mq_rx_configure may change its value */
2399         rctl |= E1000_READ_REG(hw, E1000_RCTL);
2400
2401         /*
2402          * Setup the Checksum Register.
2403          * Receive Full-Packet Checksum Offload is mutually exclusive with RSS.
2404          */
2405         rxcsum = E1000_READ_REG(hw, E1000_RXCSUM);
2406         rxcsum |= E1000_RXCSUM_PCSD;
2407
2408         /* Enable both L3/L4 rx checksum offload */
2409         if (dev->data->dev_conf.rxmode.hw_ip_checksum)
2410                 rxcsum |= (E1000_RXCSUM_IPOFL | E1000_RXCSUM_TUOFL |
2411                                 E1000_RXCSUM_CRCOFL);
2412         else
2413                 rxcsum &= ~(E1000_RXCSUM_IPOFL | E1000_RXCSUM_TUOFL |
2414                                 E1000_RXCSUM_CRCOFL);
2415         E1000_WRITE_REG(hw, E1000_RXCSUM, rxcsum);
2416
2417         /* Setup the Receive Control Register. */
2418         if (dev->data->dev_conf.rxmode.hw_strip_crc) {
2419                 rctl |= E1000_RCTL_SECRC; /* Strip Ethernet CRC. */
2420
2421                 /* set STRCRC bit in all queues */
2422                 if (hw->mac.type == e1000_i350 ||
2423                     hw->mac.type == e1000_i210 ||
2424                     hw->mac.type == e1000_i211 ||
2425                     hw->mac.type == e1000_i354) {
2426                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2427                                 rxq = dev->data->rx_queues[i];
2428                                 uint32_t dvmolr = E1000_READ_REG(hw,
2429                                         E1000_DVMOLR(rxq->reg_idx));
2430                                 dvmolr |= E1000_DVMOLR_STRCRC;
2431                                 E1000_WRITE_REG(hw, E1000_DVMOLR(rxq->reg_idx), dvmolr);
2432                         }
2433                 }
2434         } else {
2435                 rctl &= ~E1000_RCTL_SECRC; /* Do not Strip Ethernet CRC. */
2436
2437                 /* clear STRCRC bit in all queues */
2438                 if (hw->mac.type == e1000_i350 ||
2439                     hw->mac.type == e1000_i210 ||
2440                     hw->mac.type == e1000_i211 ||
2441                     hw->mac.type == e1000_i354) {
2442                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2443                                 rxq = dev->data->rx_queues[i];
2444                                 uint32_t dvmolr = E1000_READ_REG(hw,
2445                                         E1000_DVMOLR(rxq->reg_idx));
2446                                 dvmolr &= ~E1000_DVMOLR_STRCRC;
2447                                 E1000_WRITE_REG(hw, E1000_DVMOLR(rxq->reg_idx), dvmolr);
2448                         }
2449                 }
2450         }
2451
2452         rctl &= ~(3 << E1000_RCTL_MO_SHIFT);
2453         rctl |= E1000_RCTL_EN | E1000_RCTL_BAM | E1000_RCTL_LBM_NO |
2454                 E1000_RCTL_RDMTS_HALF |
2455                 (hw->mac.mc_filter_type << E1000_RCTL_MO_SHIFT);
2456
2457         /* Make sure VLAN Filters are off. */
2458         if (dev->data->dev_conf.rxmode.mq_mode != ETH_MQ_RX_VMDQ_ONLY)
2459                 rctl &= ~E1000_RCTL_VFE;
2460         /* Don't store bad packets. */
2461         rctl &= ~E1000_RCTL_SBP;
2462
2463         /* Enable Receives. */
2464         E1000_WRITE_REG(hw, E1000_RCTL, rctl);
2465
2466         /*
2467          * Setup the HW Rx Head and Tail Descriptor Pointers.
2468          * This needs to be done after enable.
2469          */
2470         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2471                 rxq = dev->data->rx_queues[i];
2472                 E1000_WRITE_REG(hw, E1000_RDH(rxq->reg_idx), 0);
2473                 E1000_WRITE_REG(hw, E1000_RDT(rxq->reg_idx), rxq->nb_rx_desc - 1);
2474         }
2475
2476         return 0;
2477 }
2478
2479 /*********************************************************************
2480  *
2481  *  Enable transmit unit.
2482  *
2483  **********************************************************************/
2484 void
2485 eth_igb_tx_init(struct rte_eth_dev *dev)
2486 {
2487         struct e1000_hw     *hw;
2488         struct igb_tx_queue *txq;
2489         uint32_t tctl;
2490         uint32_t txdctl;
2491         uint16_t i;
2492
2493         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2494
2495         /* Setup the Base and Length of the Tx Descriptor Rings. */
2496         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2497                 uint64_t bus_addr;
2498                 txq = dev->data->tx_queues[i];
2499                 bus_addr = txq->tx_ring_phys_addr;
2500
2501                 E1000_WRITE_REG(hw, E1000_TDLEN(txq->reg_idx),
2502                                 txq->nb_tx_desc *
2503                                 sizeof(union e1000_adv_tx_desc));
2504                 E1000_WRITE_REG(hw, E1000_TDBAH(txq->reg_idx),
2505                                 (uint32_t)(bus_addr >> 32));
2506                 E1000_WRITE_REG(hw, E1000_TDBAL(txq->reg_idx), (uint32_t)bus_addr);
2507
2508                 /* Setup the HW Tx Head and Tail descriptor pointers. */
2509                 E1000_WRITE_REG(hw, E1000_TDT(txq->reg_idx), 0);
2510                 E1000_WRITE_REG(hw, E1000_TDH(txq->reg_idx), 0);
2511
2512                 /* Setup Transmit threshold registers. */
2513                 txdctl = E1000_READ_REG(hw, E1000_TXDCTL(txq->reg_idx));
2514                 txdctl |= txq->pthresh & 0x1F;
2515                 txdctl |= ((txq->hthresh & 0x1F) << 8);
2516                 txdctl |= ((txq->wthresh & 0x1F) << 16);
2517                 txdctl |= E1000_TXDCTL_QUEUE_ENABLE;
2518                 E1000_WRITE_REG(hw, E1000_TXDCTL(txq->reg_idx), txdctl);
2519         }
2520
2521         /* Program the Transmit Control Register. */
2522         tctl = E1000_READ_REG(hw, E1000_TCTL);
2523         tctl &= ~E1000_TCTL_CT;
2524         tctl |= (E1000_TCTL_PSP | E1000_TCTL_RTLC | E1000_TCTL_EN |
2525                  (E1000_COLLISION_THRESHOLD << E1000_CT_SHIFT));
2526
2527         e1000_config_collision_dist(hw);
2528
2529         /* This write will effectively turn on the transmit unit. */
2530         E1000_WRITE_REG(hw, E1000_TCTL, tctl);
2531 }
2532
2533 /*********************************************************************
2534  *
2535  *  Enable VF receive unit.
2536  *
2537  **********************************************************************/
2538 int
2539 eth_igbvf_rx_init(struct rte_eth_dev *dev)
2540 {
2541         struct e1000_hw     *hw;
2542         struct igb_rx_queue *rxq;
2543         uint32_t srrctl;
2544         uint16_t buf_size;
2545         uint16_t rctl_bsize;
2546         uint16_t i;
2547         int ret;
2548
2549         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2550
2551         /* setup MTU */
2552         e1000_rlpml_set_vf(hw,
2553                 (uint16_t)(dev->data->dev_conf.rxmode.max_rx_pkt_len +
2554                 VLAN_TAG_SIZE));
2555
2556         /* Configure and enable each RX queue. */
2557         rctl_bsize = 0;
2558         dev->rx_pkt_burst = eth_igb_recv_pkts;
2559         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2560                 uint64_t bus_addr;
2561                 uint32_t rxdctl;
2562
2563                 rxq = dev->data->rx_queues[i];
2564
2565                 rxq->flags = 0;
2566                 /*
2567                  * i350VF LB vlan packets have vlan tags byte swapped.
2568                  */
2569                 if (hw->mac.type == e1000_vfadapt_i350) {
2570                         rxq->flags |= IGB_RXQ_FLAG_LB_BSWAP_VLAN;
2571                         PMD_INIT_LOG(DEBUG, "IGB rx vlan bswap required");
2572                 } else {
2573                         PMD_INIT_LOG(DEBUG, "IGB rx vlan bswap not required");
2574                 }
2575
2576                 /* Allocate buffers for descriptor rings and set up queue */
2577                 ret = igb_alloc_rx_queue_mbufs(rxq);
2578                 if (ret)
2579                         return ret;
2580
2581                 bus_addr = rxq->rx_ring_phys_addr;
2582                 E1000_WRITE_REG(hw, E1000_RDLEN(i),
2583                                 rxq->nb_rx_desc *
2584                                 sizeof(union e1000_adv_rx_desc));
2585                 E1000_WRITE_REG(hw, E1000_RDBAH(i),
2586                                 (uint32_t)(bus_addr >> 32));
2587                 E1000_WRITE_REG(hw, E1000_RDBAL(i), (uint32_t)bus_addr);
2588
2589                 srrctl = E1000_SRRCTL_DESCTYPE_ADV_ONEBUF;
2590
2591                 /*
2592                  * Configure RX buffer size.
2593                  */
2594                 buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mb_pool) -
2595                         RTE_PKTMBUF_HEADROOM);
2596                 if (buf_size >= 1024) {
2597                         /*
2598                          * Configure the BSIZEPACKET field of the SRRCTL
2599                          * register of the queue.
2600                          * Value is in 1 KB resolution, from 1 KB to 127 KB.
2601                          * If this field is equal to 0b, then RCTL.BSIZE
2602                          * determines the RX packet buffer size.
2603                          */
2604                         srrctl |= ((buf_size >> E1000_SRRCTL_BSIZEPKT_SHIFT) &
2605                                    E1000_SRRCTL_BSIZEPKT_MASK);
2606                         buf_size = (uint16_t) ((srrctl &
2607                                                 E1000_SRRCTL_BSIZEPKT_MASK) <<
2608                                                E1000_SRRCTL_BSIZEPKT_SHIFT);
2609
2610                         /* It adds dual VLAN length for supporting dual VLAN */
2611                         if ((dev->data->dev_conf.rxmode.max_rx_pkt_len +
2612                                                 2 * VLAN_TAG_SIZE) > buf_size){
2613                                 if (!dev->data->scattered_rx)
2614                                         PMD_INIT_LOG(DEBUG,
2615                                                      "forcing scatter mode");
2616                                 dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
2617                                 dev->data->scattered_rx = 1;
2618                         }
2619                 } else {
2620                         /*
2621                          * Use BSIZE field of the device RCTL register.
2622                          */
2623                         if ((rctl_bsize == 0) || (rctl_bsize > buf_size))
2624                                 rctl_bsize = buf_size;
2625                         if (!dev->data->scattered_rx)
2626                                 PMD_INIT_LOG(DEBUG, "forcing scatter mode");
2627                         dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
2628                         dev->data->scattered_rx = 1;
2629                 }
2630
2631                 /* Set if packets are dropped when no descriptors available */
2632                 if (rxq->drop_en)
2633                         srrctl |= E1000_SRRCTL_DROP_EN;
2634
2635                 E1000_WRITE_REG(hw, E1000_SRRCTL(i), srrctl);
2636
2637                 /* Enable this RX queue. */
2638                 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(i));
2639                 rxdctl |= E1000_RXDCTL_QUEUE_ENABLE;
2640                 rxdctl &= 0xFFF00000;
2641                 rxdctl |= (rxq->pthresh & 0x1F);
2642                 rxdctl |= ((rxq->hthresh & 0x1F) << 8);
2643                 if (hw->mac.type == e1000_vfadapt) {
2644                         /*
2645                          * Workaround of 82576 VF Erratum
2646                          * force set WTHRESH to 1
2647                          * to avoid Write-Back not triggered sometimes
2648                          */
2649                         rxdctl |= 0x10000;
2650                         PMD_INIT_LOG(DEBUG, "Force set RX WTHRESH to 1 !");
2651                 }
2652                 else
2653                         rxdctl |= ((rxq->wthresh & 0x1F) << 16);
2654                 E1000_WRITE_REG(hw, E1000_RXDCTL(i), rxdctl);
2655         }
2656
2657         if (dev->data->dev_conf.rxmode.enable_scatter) {
2658                 if (!dev->data->scattered_rx)
2659                         PMD_INIT_LOG(DEBUG, "forcing scatter mode");
2660                 dev->rx_pkt_burst = eth_igb_recv_scattered_pkts;
2661                 dev->data->scattered_rx = 1;
2662         }
2663
2664         /*
2665          * Setup the HW Rx Head and Tail Descriptor Pointers.
2666          * This needs to be done after enable.
2667          */
2668         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2669                 rxq = dev->data->rx_queues[i];
2670                 E1000_WRITE_REG(hw, E1000_RDH(i), 0);
2671                 E1000_WRITE_REG(hw, E1000_RDT(i), rxq->nb_rx_desc - 1);
2672         }
2673
2674         return 0;
2675 }
2676
2677 /*********************************************************************
2678  *
2679  *  Enable VF transmit unit.
2680  *
2681  **********************************************************************/
2682 void
2683 eth_igbvf_tx_init(struct rte_eth_dev *dev)
2684 {
2685         struct e1000_hw     *hw;
2686         struct igb_tx_queue *txq;
2687         uint32_t txdctl;
2688         uint16_t i;
2689
2690         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2691
2692         /* Setup the Base and Length of the Tx Descriptor Rings. */
2693         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2694                 uint64_t bus_addr;
2695
2696                 txq = dev->data->tx_queues[i];
2697                 bus_addr = txq->tx_ring_phys_addr;
2698                 E1000_WRITE_REG(hw, E1000_TDLEN(i),
2699                                 txq->nb_tx_desc *
2700                                 sizeof(union e1000_adv_tx_desc));
2701                 E1000_WRITE_REG(hw, E1000_TDBAH(i),
2702                                 (uint32_t)(bus_addr >> 32));
2703                 E1000_WRITE_REG(hw, E1000_TDBAL(i), (uint32_t)bus_addr);
2704
2705                 /* Setup the HW Tx Head and Tail descriptor pointers. */
2706                 E1000_WRITE_REG(hw, E1000_TDT(i), 0);
2707                 E1000_WRITE_REG(hw, E1000_TDH(i), 0);
2708
2709                 /* Setup Transmit threshold registers. */
2710                 txdctl = E1000_READ_REG(hw, E1000_TXDCTL(i));
2711                 txdctl |= txq->pthresh & 0x1F;
2712                 txdctl |= ((txq->hthresh & 0x1F) << 8);
2713                 if (hw->mac.type == e1000_82576) {
2714                         /*
2715                          * Workaround of 82576 VF Erratum
2716                          * force set WTHRESH to 1
2717                          * to avoid Write-Back not triggered sometimes
2718                          */
2719                         txdctl |= 0x10000;
2720                         PMD_INIT_LOG(DEBUG, "Force set TX WTHRESH to 1 !");
2721                 }
2722                 else
2723                         txdctl |= ((txq->wthresh & 0x1F) << 16);
2724                 txdctl |= E1000_TXDCTL_QUEUE_ENABLE;
2725                 E1000_WRITE_REG(hw, E1000_TXDCTL(i), txdctl);
2726         }
2727
2728 }
2729
2730 void
2731 igb_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
2732         struct rte_eth_rxq_info *qinfo)
2733 {
2734         struct igb_rx_queue *rxq;
2735
2736         rxq = dev->data->rx_queues[queue_id];
2737
2738         qinfo->mp = rxq->mb_pool;
2739         qinfo->scattered_rx = dev->data->scattered_rx;
2740         qinfo->nb_desc = rxq->nb_rx_desc;
2741
2742         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
2743         qinfo->conf.rx_drop_en = rxq->drop_en;
2744 }
2745
2746 void
2747 igb_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
2748         struct rte_eth_txq_info *qinfo)
2749 {
2750         struct igb_tx_queue *txq;
2751
2752         txq = dev->data->tx_queues[queue_id];
2753
2754         qinfo->nb_desc = txq->nb_tx_desc;
2755
2756         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
2757         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
2758         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
2759 }
2760
2761 int
2762 igb_config_rss_filter(struct rte_eth_dev *dev,
2763                 struct igb_rte_flow_rss_conf *conf, bool add)
2764 {
2765         uint32_t shift;
2766         uint16_t i, j;
2767         struct rte_eth_rss_conf rss_conf = conf->rss_conf;
2768         struct e1000_filter_info *filter_info =
2769                 E1000_DEV_PRIVATE_TO_FILTER_INFO(dev->data->dev_private);
2770         struct e1000_hw *hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2771
2772         hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2773
2774         if (!add) {
2775                 if (memcmp(conf, &filter_info->rss_info,
2776                         sizeof(struct igb_rte_flow_rss_conf)) == 0) {
2777                         igb_rss_disable(dev);
2778                         memset(&filter_info->rss_info, 0,
2779                                 sizeof(struct igb_rte_flow_rss_conf));
2780                         return 0;
2781                 }
2782                 return -EINVAL;
2783         }
2784
2785         if (filter_info->rss_info.num)
2786                 return -EINVAL;
2787
2788         /* Fill in redirection table. */
2789         shift = (hw->mac.type == e1000_82575) ? 6 : 0;
2790         for (i = 0, j = 0; i < 128; i++, j++) {
2791                 union e1000_reta {
2792                         uint32_t dword;
2793                         uint8_t  bytes[4];
2794                 } reta;
2795                 uint8_t q_idx;
2796
2797                 q_idx = conf->queue[j];
2798                 if (j == conf->num)
2799                         j = 0;
2800                 reta.bytes[i & 3] = (uint8_t)(q_idx << shift);
2801                 if ((i & 3) == 3)
2802                         E1000_WRITE_REG(hw, E1000_RETA(i >> 2), reta.dword);
2803         }
2804
2805         /* Configure the RSS key and the RSS protocols used to compute
2806          * the RSS hash of input packets.
2807          */
2808         if ((rss_conf.rss_hf & IGB_RSS_OFFLOAD_ALL) == 0) {
2809                 igb_rss_disable(dev);
2810                 return 0;
2811         }
2812         if (rss_conf.rss_key == NULL)
2813                 rss_conf.rss_key = rss_intel_key; /* Default hash key */
2814         igb_hw_rss_hash_set(hw, &rss_conf);
2815
2816         rte_memcpy(&filter_info->rss_info,
2817                 conf, sizeof(struct igb_rte_flow_rss_conf));
2818
2819         return 0;
2820 }