4 * Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34 #include <sys/queue.h>
44 #include <rte_interrupts.h>
45 #include <rte_byteorder.h>
46 #include <rte_common.h>
48 #include <rte_debug.h>
50 #include <rte_memory.h>
51 #include <rte_memcpy.h>
52 #include <rte_memzone.h>
53 #include <rte_launch.h>
54 #include <rte_tailq.h>
56 #include <rte_per_lcore.h>
57 #include <rte_lcore.h>
58 #include <rte_atomic.h>
59 #include <rte_branch_prediction.h>
61 #include <rte_mempool.h>
62 #include <rte_malloc.h>
64 #include <rte_ether.h>
65 #include <rte_ethdev.h>
66 #include <rte_prefetch.h>
71 #include <rte_string_fns.h>
73 #include "e1000_logs.h"
74 #include "e1000/e1000_api.h"
75 #include "e1000_ethdev.h"
76 #include "e1000/e1000_osdep.h"
78 #define E1000_TXD_VLAN_SHIFT 16
80 #define E1000_RXDCTL_GRAN 0x01000000 /* RXDCTL Granularity */
82 static inline struct rte_mbuf *
83 rte_rxmbuf_alloc(struct rte_mempool *mp)
87 m = __rte_mbuf_raw_alloc(mp);
88 __rte_mbuf_sanity_check_raw(m, 0);
92 #define RTE_MBUF_DATA_DMA_ADDR(mb) \
93 (uint64_t) ((mb)->buf_physaddr + (mb)->data_off)
95 #define RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mb) \
96 (uint64_t) ((mb)->buf_physaddr + RTE_PKTMBUF_HEADROOM)
99 * Structure associated with each descriptor of the RX ring of a RX queue.
102 struct rte_mbuf *mbuf; /**< mbuf associated with RX descriptor. */
106 * Structure associated with each descriptor of the TX ring of a TX queue.
109 struct rte_mbuf *mbuf; /**< mbuf associated with TX desc, if any. */
110 uint16_t next_id; /**< Index of next descriptor in ring. */
111 uint16_t last_id; /**< Index of last scattered descriptor. */
115 * Structure associated with each RX queue.
118 struct rte_mempool *mb_pool; /**< mbuf pool to populate RX ring. */
119 volatile struct e1000_rx_desc *rx_ring; /**< RX ring virtual address. */
120 uint64_t rx_ring_phys_addr; /**< RX ring DMA address. */
121 volatile uint32_t *rdt_reg_addr; /**< RDT register address. */
122 volatile uint32_t *rdh_reg_addr; /**< RDH register address. */
123 struct em_rx_entry *sw_ring; /**< address of RX software ring. */
124 struct rte_mbuf *pkt_first_seg; /**< First segment of current packet. */
125 struct rte_mbuf *pkt_last_seg; /**< Last segment of current packet. */
126 uint16_t nb_rx_desc; /**< number of RX descriptors. */
127 uint16_t rx_tail; /**< current value of RDT register. */
128 uint16_t nb_rx_hold; /**< number of held free RX desc. */
129 uint16_t rx_free_thresh; /**< max free RX desc to hold. */
130 uint16_t queue_id; /**< RX queue index. */
131 uint8_t port_id; /**< Device port identifier. */
132 uint8_t pthresh; /**< Prefetch threshold register. */
133 uint8_t hthresh; /**< Host threshold register. */
134 uint8_t wthresh; /**< Write-back threshold register. */
135 uint8_t crc_len; /**< 0 if CRC stripped, 4 otherwise. */
139 * Hardware context number
142 EM_CTX_0 = 0, /**< CTX0 */
143 EM_CTX_NUM = 1, /**< CTX NUM */
146 /** Offload features */
147 union em_vlan_macip {
150 uint16_t l3_len:9; /**< L3 (IP) Header Length. */
151 uint16_t l2_len:7; /**< L2 (MAC) Header Length. */
153 /**< VLAN Tag Control Identifier (CPU order). */
158 * Compare mask for vlan_macip_len.data,
159 * should be in sync with em_vlan_macip.f layout.
161 #define TX_VLAN_CMP_MASK 0xFFFF0000 /**< VLAN length - 16-bits. */
162 #define TX_MAC_LEN_CMP_MASK 0x0000FE00 /**< MAC length - 7-bits. */
163 #define TX_IP_LEN_CMP_MASK 0x000001FF /**< IP length - 9-bits. */
164 /** MAC+IP length. */
165 #define TX_MACIP_LEN_CMP_MASK (TX_MAC_LEN_CMP_MASK | TX_IP_LEN_CMP_MASK)
168 * Structure to check if new context need be built
171 uint64_t flags; /**< ol_flags related to context build. */
172 uint32_t cmp_mask; /**< compare mask */
173 union em_vlan_macip hdrlen; /**< L2 and L3 header lenghts */
177 * Structure associated with each TX queue.
180 volatile struct e1000_data_desc *tx_ring; /**< TX ring address */
181 uint64_t tx_ring_phys_addr; /**< TX ring DMA address. */
182 struct em_tx_entry *sw_ring; /**< virtual address of SW ring. */
183 volatile uint32_t *tdt_reg_addr; /**< Address of TDT register. */
184 uint16_t nb_tx_desc; /**< number of TX descriptors. */
185 uint16_t tx_tail; /**< Current value of TDT register. */
186 uint16_t tx_free_thresh;/**< minimum TX before freeing. */
187 /**< Number of TX descriptors to use before RS bit is set. */
188 uint16_t tx_rs_thresh;
189 /** Number of TX descriptors used since RS bit was set. */
191 /** Index to last TX descriptor to have been cleaned. */
192 uint16_t last_desc_cleaned;
193 /** Total number of TX descriptors ready to be allocated. */
195 uint16_t queue_id; /**< TX queue index. */
196 uint8_t port_id; /**< Device port identifier. */
197 uint8_t pthresh; /**< Prefetch threshold register. */
198 uint8_t hthresh; /**< Host threshold register. */
199 uint8_t wthresh; /**< Write-back threshold register. */
200 struct em_ctx_info ctx_cache;
201 /**< Hardware context history.*/
205 #define RTE_PMD_USE_PREFETCH
208 #ifdef RTE_PMD_USE_PREFETCH
209 #define rte_em_prefetch(p) rte_prefetch0(p)
211 #define rte_em_prefetch(p) do {} while(0)
214 #ifdef RTE_PMD_PACKET_PREFETCH
215 #define rte_packet_prefetch(p) rte_prefetch1(p)
217 #define rte_packet_prefetch(p) do {} while(0)
220 #ifndef DEFAULT_TX_FREE_THRESH
221 #define DEFAULT_TX_FREE_THRESH 32
222 #endif /* DEFAULT_TX_FREE_THRESH */
224 #ifndef DEFAULT_TX_RS_THRESH
225 #define DEFAULT_TX_RS_THRESH 32
226 #endif /* DEFAULT_TX_RS_THRESH */
229 /*********************************************************************
233 **********************************************************************/
236 * Populates TX context descriptor.
239 em_set_xmit_ctx(struct em_tx_queue* txq,
240 volatile struct e1000_context_desc *ctx_txd,
242 union em_vlan_macip hdrlen)
244 uint32_t cmp_mask, cmd_len;
245 uint16_t ipcse, l2len;
246 struct e1000_context_desc ctx;
249 cmd_len = E1000_TXD_CMD_DEXT | E1000_TXD_DTYP_C;
251 l2len = hdrlen.f.l2_len;
252 ipcse = (uint16_t)(l2len + hdrlen.f.l3_len);
254 /* setup IPCS* fields */
255 ctx.lower_setup.ip_fields.ipcss = (uint8_t)l2len;
256 ctx.lower_setup.ip_fields.ipcso = (uint8_t)(l2len +
257 offsetof(struct ipv4_hdr, hdr_checksum));
260 * When doing checksum or TCP segmentation with IPv6 headers,
261 * IPCSE field should be set t0 0.
263 if (flags & PKT_TX_IP_CKSUM) {
264 ctx.lower_setup.ip_fields.ipcse =
265 (uint16_t)rte_cpu_to_le_16(ipcse - 1);
266 cmd_len |= E1000_TXD_CMD_IP;
267 cmp_mask |= TX_MACIP_LEN_CMP_MASK;
269 ctx.lower_setup.ip_fields.ipcse = 0;
272 /* setup TUCS* fields */
273 ctx.upper_setup.tcp_fields.tucss = (uint8_t)ipcse;
274 ctx.upper_setup.tcp_fields.tucse = 0;
276 switch (flags & PKT_TX_L4_MASK) {
277 case PKT_TX_UDP_CKSUM:
278 ctx.upper_setup.tcp_fields.tucso = (uint8_t)(ipcse +
279 offsetof(struct udp_hdr, dgram_cksum));
280 cmp_mask |= TX_MACIP_LEN_CMP_MASK;
282 case PKT_TX_TCP_CKSUM:
283 ctx.upper_setup.tcp_fields.tucso = (uint8_t)(ipcse +
284 offsetof(struct tcp_hdr, cksum));
285 cmd_len |= E1000_TXD_CMD_TCP;
286 cmp_mask |= TX_MACIP_LEN_CMP_MASK;
289 ctx.upper_setup.tcp_fields.tucso = 0;
292 ctx.cmd_and_length = rte_cpu_to_le_32(cmd_len);
293 ctx.tcp_seg_setup.data = 0;
297 txq->ctx_cache.flags = flags;
298 txq->ctx_cache.cmp_mask = cmp_mask;
299 txq->ctx_cache.hdrlen = hdrlen;
303 * Check which hardware context can be used. Use the existing match
304 * or create a new context descriptor.
306 static inline uint32_t
307 what_ctx_update(struct em_tx_queue *txq, uint64_t flags,
308 union em_vlan_macip hdrlen)
310 /* If match with the current context */
311 if (likely (txq->ctx_cache.flags == flags &&
312 ((txq->ctx_cache.hdrlen.data ^ hdrlen.data) &
313 txq->ctx_cache.cmp_mask) == 0))
320 /* Reset transmit descriptors after they have been used */
322 em_xmit_cleanup(struct em_tx_queue *txq)
324 struct em_tx_entry *sw_ring = txq->sw_ring;
325 volatile struct e1000_data_desc *txr = txq->tx_ring;
326 uint16_t last_desc_cleaned = txq->last_desc_cleaned;
327 uint16_t nb_tx_desc = txq->nb_tx_desc;
328 uint16_t desc_to_clean_to;
329 uint16_t nb_tx_to_clean;
331 /* Determine the last descriptor needing to be cleaned */
332 desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
333 if (desc_to_clean_to >= nb_tx_desc)
334 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
336 /* Check to make sure the last descriptor to clean is done */
337 desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
338 if (! (txr[desc_to_clean_to].upper.fields.status & E1000_TXD_STAT_DD))
340 PMD_TX_FREE_LOG(DEBUG,
341 "TX descriptor %4u is not done"
342 "(port=%d queue=%d)", desc_to_clean_to,
343 txq->port_id, txq->queue_id);
344 /* Failed to clean any descriptors, better luck next time */
348 /* Figure out how many descriptors will be cleaned */
349 if (last_desc_cleaned > desc_to_clean_to)
350 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
353 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
356 PMD_TX_FREE_LOG(DEBUG,
357 "Cleaning %4u TX descriptors: %4u to %4u "
358 "(port=%d queue=%d)", nb_tx_to_clean,
359 last_desc_cleaned, desc_to_clean_to, txq->port_id,
363 * The last descriptor to clean is done, so that means all the
364 * descriptors from the last descriptor that was cleaned
365 * up to the last descriptor with the RS bit set
366 * are done. Only reset the threshold descriptor.
368 txr[desc_to_clean_to].upper.fields.status = 0;
370 /* Update the txq to reflect the last descriptor that was cleaned */
371 txq->last_desc_cleaned = desc_to_clean_to;
372 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
378 static inline uint32_t
379 tx_desc_cksum_flags_to_upper(uint64_t ol_flags)
381 static const uint32_t l4_olinfo[2] = {0, E1000_TXD_POPTS_TXSM << 8};
382 static const uint32_t l3_olinfo[2] = {0, E1000_TXD_POPTS_IXSM << 8};
385 tmp = l4_olinfo[(ol_flags & PKT_TX_L4_MASK) != PKT_TX_L4_NO_CKSUM];
386 tmp |= l3_olinfo[(ol_flags & PKT_TX_IP_CKSUM) != 0];
391 eth_em_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
394 struct em_tx_queue *txq;
395 struct em_tx_entry *sw_ring;
396 struct em_tx_entry *txe, *txn;
397 volatile struct e1000_data_desc *txr;
398 volatile struct e1000_data_desc *txd;
399 struct rte_mbuf *tx_pkt;
400 struct rte_mbuf *m_seg;
401 uint64_t buf_dma_addr;
403 uint32_t cmd_type_len;
413 union em_vlan_macip hdrlen;
416 sw_ring = txq->sw_ring;
418 tx_id = txq->tx_tail;
419 txe = &sw_ring[tx_id];
421 /* Determine if the descriptor ring needs to be cleaned. */
422 if ((txq->nb_tx_desc - txq->nb_tx_free) > txq->tx_free_thresh) {
423 em_xmit_cleanup(txq);
427 for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
431 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
434 * Determine how many (if any) context descriptors
435 * are needed for offload functionality.
437 ol_flags = tx_pkt->ol_flags;
439 /* If hardware offload required */
440 tx_ol_req = (ol_flags & (PKT_TX_IP_CKSUM | PKT_TX_L4_MASK));
442 hdrlen.f.vlan_tci = tx_pkt->vlan_tci;
443 hdrlen.f.l2_len = tx_pkt->l2_len;
444 hdrlen.f.l3_len = tx_pkt->l3_len;
445 /* If new context to be built or reuse the exist ctx. */
446 ctx = what_ctx_update(txq, tx_ol_req, hdrlen);
448 /* Only allocate context descriptor if required*/
449 new_ctx = (ctx == EM_CTX_NUM);
453 * Keep track of how many descriptors are used this loop
454 * This will always be the number of segments + the number of
455 * Context descriptors required to transmit the packet
457 nb_used = (uint16_t)(tx_pkt->nb_segs + new_ctx);
460 * The number of descriptors that must be allocated for a
461 * packet is the number of segments of that packet, plus 1
462 * Context Descriptor for the hardware offload, if any.
463 * Determine the last TX descriptor to allocate in the TX ring
464 * for the packet, starting from the current position (tx_id)
467 tx_last = (uint16_t) (tx_id + nb_used - 1);
470 if (tx_last >= txq->nb_tx_desc)
471 tx_last = (uint16_t) (tx_last - txq->nb_tx_desc);
473 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u pktlen=%u"
474 " tx_first=%u tx_last=%u",
475 (unsigned) txq->port_id,
476 (unsigned) txq->queue_id,
477 (unsigned) tx_pkt->pkt_len,
482 * Make sure there are enough TX descriptors available to
483 * transmit the entire packet.
484 * nb_used better be less than or equal to txq->tx_rs_thresh
486 while (unlikely (nb_used > txq->nb_tx_free)) {
487 PMD_TX_FREE_LOG(DEBUG, "Not enough free TX descriptors "
488 "nb_used=%4u nb_free=%4u "
489 "(port=%d queue=%d)",
490 nb_used, txq->nb_tx_free,
491 txq->port_id, txq->queue_id);
493 if (em_xmit_cleanup(txq) != 0) {
494 /* Could not clean any descriptors */
502 * By now there are enough free TX descriptors to transmit
507 * Set common flags of all TX Data Descriptors.
509 * The following bits must be set in all Data Descriptors:
510 * - E1000_TXD_DTYP_DATA
511 * - E1000_TXD_DTYP_DEXT
513 * The following bits must be set in the first Data Descriptor
514 * and are ignored in the other ones:
515 * - E1000_TXD_POPTS_IXSM
516 * - E1000_TXD_POPTS_TXSM
518 * The following bits must be set in the last Data Descriptor
519 * and are ignored in the other ones:
520 * - E1000_TXD_CMD_VLE
521 * - E1000_TXD_CMD_IFCS
523 * The following bits must only be set in the last Data
525 * - E1000_TXD_CMD_EOP
527 * The following bits can be set in any Data Descriptor, but
528 * are only set in the last Data Descriptor:
531 cmd_type_len = E1000_TXD_CMD_DEXT | E1000_TXD_DTYP_D |
535 /* Set VLAN Tag offload fields. */
536 if (ol_flags & PKT_TX_VLAN_PKT) {
537 cmd_type_len |= E1000_TXD_CMD_VLE;
538 popts_spec = tx_pkt->vlan_tci << E1000_TXD_VLAN_SHIFT;
543 * Setup the TX Context Descriptor if required
546 volatile struct e1000_context_desc *ctx_txd;
548 ctx_txd = (volatile struct e1000_context_desc *)
551 txn = &sw_ring[txe->next_id];
552 RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
554 if (txe->mbuf != NULL) {
555 rte_pktmbuf_free_seg(txe->mbuf);
559 em_set_xmit_ctx(txq, ctx_txd, tx_ol_req,
562 txe->last_id = tx_last;
563 tx_id = txe->next_id;
568 * Setup the TX Data Descriptor,
569 * This path will go through
570 * whatever new/reuse the context descriptor
572 popts_spec |= tx_desc_cksum_flags_to_upper(ol_flags);
578 txn = &sw_ring[txe->next_id];
580 if (txe->mbuf != NULL)
581 rte_pktmbuf_free_seg(txe->mbuf);
585 * Set up Transmit Data Descriptor.
587 slen = m_seg->data_len;
588 buf_dma_addr = RTE_MBUF_DATA_DMA_ADDR(m_seg);
590 txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
591 txd->lower.data = rte_cpu_to_le_32(cmd_type_len | slen);
592 txd->upper.data = rte_cpu_to_le_32(popts_spec);
594 txe->last_id = tx_last;
595 tx_id = txe->next_id;
598 } while (m_seg != NULL);
601 * The last packet data descriptor needs End Of Packet (EOP)
603 cmd_type_len |= E1000_TXD_CMD_EOP;
604 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
605 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
607 /* Set RS bit only on threshold packets' last descriptor */
608 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
609 PMD_TX_FREE_LOG(DEBUG,
610 "Setting RS bit on TXD id=%4u "
611 "(port=%d queue=%d)",
612 tx_last, txq->port_id, txq->queue_id);
614 cmd_type_len |= E1000_TXD_CMD_RS;
616 /* Update txq RS bit counters */
619 txd->lower.data |= rte_cpu_to_le_32(cmd_type_len);
625 * Set the Transmit Descriptor Tail (TDT)
627 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
628 (unsigned) txq->port_id, (unsigned) txq->queue_id,
629 (unsigned) tx_id, (unsigned) nb_tx);
630 E1000_PCI_REG_WRITE(txq->tdt_reg_addr, tx_id);
631 txq->tx_tail = tx_id;
636 /*********************************************************************
640 **********************************************************************/
642 static inline uint64_t
643 rx_desc_status_to_pkt_flags(uint32_t rx_status)
647 /* Check if VLAN present */
648 pkt_flags = ((rx_status & E1000_RXD_STAT_VP) ? PKT_RX_VLAN_PKT : 0);
653 static inline uint64_t
654 rx_desc_error_to_pkt_flags(uint32_t rx_error)
656 uint64_t pkt_flags = 0;
658 if (rx_error & E1000_RXD_ERR_IPE)
659 pkt_flags |= PKT_RX_IP_CKSUM_BAD;
660 if (rx_error & E1000_RXD_ERR_TCPE)
661 pkt_flags |= PKT_RX_L4_CKSUM_BAD;
666 eth_em_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
669 volatile struct e1000_rx_desc *rx_ring;
670 volatile struct e1000_rx_desc *rxdp;
671 struct em_rx_queue *rxq;
672 struct em_rx_entry *sw_ring;
673 struct em_rx_entry *rxe;
674 struct rte_mbuf *rxm;
675 struct rte_mbuf *nmb;
676 struct e1000_rx_desc rxd;
688 rx_id = rxq->rx_tail;
689 rx_ring = rxq->rx_ring;
690 sw_ring = rxq->sw_ring;
691 while (nb_rx < nb_pkts) {
693 * The order of operations here is important as the DD status
694 * bit must not be read after any other descriptor fields.
695 * rx_ring and rxdp are pointing to volatile data so the order
696 * of accesses cannot be reordered by the compiler. If they were
697 * not volatile, they could be reordered which could lead to
698 * using invalid descriptor fields when read from rxd.
700 rxdp = &rx_ring[rx_id];
701 status = rxdp->status;
702 if (! (status & E1000_RXD_STAT_DD))
709 * If the E1000_RXD_STAT_EOP flag is not set, the RX packet is
710 * likely to be invalid and to be dropped by the various
711 * validation checks performed by the network stack.
713 * Allocate a new mbuf to replenish the RX ring descriptor.
714 * If the allocation fails:
715 * - arrange for that RX descriptor to be the first one
716 * being parsed the next time the receive function is
717 * invoked [on the same queue].
719 * - Stop parsing the RX ring and return immediately.
721 * This policy do not drop the packet received in the RX
722 * descriptor for which the allocation of a new mbuf failed.
723 * Thus, it allows that packet to be later retrieved if
724 * mbuf have been freed in the mean time.
725 * As a side effect, holding RX descriptors instead of
726 * systematically giving them back to the NIC may lead to
727 * RX ring exhaustion situations.
728 * However, the NIC can gracefully prevent such situations
729 * to happen by sending specific "back-pressure" flow control
730 * frames to its peer(s).
732 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
733 "status=0x%x pkt_len=%u",
734 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
735 (unsigned) rx_id, (unsigned) status,
736 (unsigned) rte_le_to_cpu_16(rxd.length));
738 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
740 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
742 (unsigned) rxq->port_id,
743 (unsigned) rxq->queue_id);
744 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
749 rxe = &sw_ring[rx_id];
751 if (rx_id == rxq->nb_rx_desc)
754 /* Prefetch next mbuf while processing current one. */
755 rte_em_prefetch(sw_ring[rx_id].mbuf);
758 * When next RX descriptor is on a cache-line boundary,
759 * prefetch the next 4 RX descriptors and the next 8 pointers
762 if ((rx_id & 0x3) == 0) {
763 rte_em_prefetch(&rx_ring[rx_id]);
764 rte_em_prefetch(&sw_ring[rx_id]);
767 /* Rearm RXD: attach new mbuf and reset status to zero. */
772 rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
773 rxdp->buffer_addr = dma_addr;
777 * Initialize the returned mbuf.
778 * 1) setup generic mbuf fields:
779 * - number of segments,
782 * - RX port identifier.
783 * 2) integrate hardware offload data, if any:
785 * - IP checksum flag,
786 * - VLAN TCI, if any,
789 pkt_len = (uint16_t) (rte_le_to_cpu_16(rxd.length) -
791 rxm->data_off = RTE_PKTMBUF_HEADROOM;
792 rte_packet_prefetch((char *)rxm->buf_addr + rxm->data_off);
795 rxm->pkt_len = pkt_len;
796 rxm->data_len = pkt_len;
797 rxm->port = rxq->port_id;
799 rxm->ol_flags = rx_desc_status_to_pkt_flags(status);
800 rxm->ol_flags = rxm->ol_flags |
801 rx_desc_error_to_pkt_flags(rxd.errors);
803 /* Only valid if PKT_RX_VLAN_PKT set in pkt_flags */
804 rxm->vlan_tci = rte_le_to_cpu_16(rxd.special);
807 * Store the mbuf address into the next entry of the array
808 * of returned packets.
810 rx_pkts[nb_rx++] = rxm;
812 rxq->rx_tail = rx_id;
815 * If the number of free RX descriptors is greater than the RX free
816 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
818 * Update the RDT with the value of the last processed RX descriptor
819 * minus 1, to guarantee that the RDT register is never equal to the
820 * RDH register, which creates a "full" ring situtation from the
821 * hardware point of view...
823 nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
824 if (nb_hold > rxq->rx_free_thresh) {
825 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
826 "nb_hold=%u nb_rx=%u",
827 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
828 (unsigned) rx_id, (unsigned) nb_hold,
830 rx_id = (uint16_t) ((rx_id == 0) ?
831 (rxq->nb_rx_desc - 1) : (rx_id - 1));
832 E1000_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
835 rxq->nb_rx_hold = nb_hold;
840 eth_em_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
843 struct em_rx_queue *rxq;
844 volatile struct e1000_rx_desc *rx_ring;
845 volatile struct e1000_rx_desc *rxdp;
846 struct em_rx_entry *sw_ring;
847 struct em_rx_entry *rxe;
848 struct rte_mbuf *first_seg;
849 struct rte_mbuf *last_seg;
850 struct rte_mbuf *rxm;
851 struct rte_mbuf *nmb;
852 struct e1000_rx_desc rxd;
853 uint64_t dma; /* Physical address of mbuf data buffer */
864 rx_id = rxq->rx_tail;
865 rx_ring = rxq->rx_ring;
866 sw_ring = rxq->sw_ring;
869 * Retrieve RX context of current packet, if any.
871 first_seg = rxq->pkt_first_seg;
872 last_seg = rxq->pkt_last_seg;
874 while (nb_rx < nb_pkts) {
877 * The order of operations here is important as the DD status
878 * bit must not be read after any other descriptor fields.
879 * rx_ring and rxdp are pointing to volatile data so the order
880 * of accesses cannot be reordered by the compiler. If they were
881 * not volatile, they could be reordered which could lead to
882 * using invalid descriptor fields when read from rxd.
884 rxdp = &rx_ring[rx_id];
885 status = rxdp->status;
886 if (! (status & E1000_RXD_STAT_DD))
893 * Allocate a new mbuf to replenish the RX ring descriptor.
894 * If the allocation fails:
895 * - arrange for that RX descriptor to be the first one
896 * being parsed the next time the receive function is
897 * invoked [on the same queue].
899 * - Stop parsing the RX ring and return immediately.
901 * This policy does not drop the packet received in the RX
902 * descriptor for which the allocation of a new mbuf failed.
903 * Thus, it allows that packet to be later retrieved if
904 * mbuf have been freed in the mean time.
905 * As a side effect, holding RX descriptors instead of
906 * systematically giving them back to the NIC may lead to
907 * RX ring exhaustion situations.
908 * However, the NIC can gracefully prevent such situations
909 * to happen by sending specific "back-pressure" flow control
910 * frames to its peer(s).
912 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
913 "status=0x%x data_len=%u",
914 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
915 (unsigned) rx_id, (unsigned) status,
916 (unsigned) rte_le_to_cpu_16(rxd.length));
918 nmb = rte_rxmbuf_alloc(rxq->mb_pool);
920 PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
921 "queue_id=%u", (unsigned) rxq->port_id,
922 (unsigned) rxq->queue_id);
923 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
928 rxe = &sw_ring[rx_id];
930 if (rx_id == rxq->nb_rx_desc)
933 /* Prefetch next mbuf while processing current one. */
934 rte_em_prefetch(sw_ring[rx_id].mbuf);
937 * When next RX descriptor is on a cache-line boundary,
938 * prefetch the next 4 RX descriptors and the next 8 pointers
941 if ((rx_id & 0x3) == 0) {
942 rte_em_prefetch(&rx_ring[rx_id]);
943 rte_em_prefetch(&sw_ring[rx_id]);
947 * Update RX descriptor with the physical address of the new
948 * data buffer of the new allocated mbuf.
952 dma = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
953 rxdp->buffer_addr = dma;
957 * Set data length & data buffer address of mbuf.
959 data_len = rte_le_to_cpu_16(rxd.length);
960 rxm->data_len = data_len;
961 rxm->data_off = RTE_PKTMBUF_HEADROOM;
964 * If this is the first buffer of the received packet,
965 * set the pointer to the first mbuf of the packet and
966 * initialize its context.
967 * Otherwise, update the total length and the number of segments
968 * of the current scattered packet, and update the pointer to
969 * the last mbuf of the current packet.
971 if (first_seg == NULL) {
973 first_seg->pkt_len = data_len;
974 first_seg->nb_segs = 1;
976 first_seg->pkt_len += data_len;
977 first_seg->nb_segs++;
978 last_seg->next = rxm;
982 * If this is not the last buffer of the received packet,
983 * update the pointer to the last mbuf of the current scattered
984 * packet and continue to parse the RX ring.
986 if (! (status & E1000_RXD_STAT_EOP)) {
992 * This is the last buffer of the received packet.
993 * If the CRC is not stripped by the hardware:
994 * - Subtract the CRC length from the total packet length.
995 * - If the last buffer only contains the whole CRC or a part
996 * of it, free the mbuf associated to the last buffer.
997 * If part of the CRC is also contained in the previous
998 * mbuf, subtract the length of that CRC part from the
999 * data length of the previous mbuf.
1002 if (unlikely(rxq->crc_len > 0)) {
1003 first_seg->pkt_len -= ETHER_CRC_LEN;
1004 if (data_len <= ETHER_CRC_LEN) {
1005 rte_pktmbuf_free_seg(rxm);
1006 first_seg->nb_segs--;
1007 last_seg->data_len = (uint16_t)
1008 (last_seg->data_len -
1009 (ETHER_CRC_LEN - data_len));
1010 last_seg->next = NULL;
1013 (uint16_t) (data_len - ETHER_CRC_LEN);
1017 * Initialize the first mbuf of the returned packet:
1018 * - RX port identifier,
1019 * - hardware offload data, if any:
1020 * - IP checksum flag,
1023 first_seg->port = rxq->port_id;
1025 first_seg->ol_flags = rx_desc_status_to_pkt_flags(status);
1026 first_seg->ol_flags = first_seg->ol_flags |
1027 rx_desc_error_to_pkt_flags(rxd.errors);
1029 /* Only valid if PKT_RX_VLAN_PKT set in pkt_flags */
1030 rxm->vlan_tci = rte_le_to_cpu_16(rxd.special);
1032 /* Prefetch data of first segment, if configured to do so. */
1033 rte_packet_prefetch((char *)first_seg->buf_addr +
1034 first_seg->data_off);
1037 * Store the mbuf address into the next entry of the array
1038 * of returned packets.
1040 rx_pkts[nb_rx++] = first_seg;
1043 * Setup receipt context for a new packet.
1049 * Record index of the next RX descriptor to probe.
1051 rxq->rx_tail = rx_id;
1054 * Save receive context.
1056 rxq->pkt_first_seg = first_seg;
1057 rxq->pkt_last_seg = last_seg;
1060 * If the number of free RX descriptors is greater than the RX free
1061 * threshold of the queue, advance the Receive Descriptor Tail (RDT)
1063 * Update the RDT with the value of the last processed RX descriptor
1064 * minus 1, to guarantee that the RDT register is never equal to the
1065 * RDH register, which creates a "full" ring situtation from the
1066 * hardware point of view...
1068 nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
1069 if (nb_hold > rxq->rx_free_thresh) {
1070 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
1071 "nb_hold=%u nb_rx=%u",
1072 (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
1073 (unsigned) rx_id, (unsigned) nb_hold,
1075 rx_id = (uint16_t) ((rx_id == 0) ?
1076 (rxq->nb_rx_desc - 1) : (rx_id - 1));
1077 E1000_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
1080 rxq->nb_rx_hold = nb_hold;
1085 * Rings setup and release.
1087 * TDBA/RDBA should be aligned on 16 byte boundary. But TDLEN/RDLEN should be
1088 * multiple of 128 bytes. So we align TDBA/RDBA on 128 byte boundary.
1089 * This will also optimize cache line size effect.
1090 * H/W supports up to cache line size 128.
1092 #define EM_ALIGN 128
1095 * Maximum number of Ring Descriptors.
1097 * Since RDLEN/TDLEN should be multiple of 128 bytes, the number of ring
1098 * desscriptors should meet the following condition:
1099 * (num_ring_desc * sizeof(struct e1000_rx/tx_desc)) % 128 == 0
1101 #define EM_MIN_RING_DESC 32
1102 #define EM_MAX_RING_DESC 4096
1104 #define EM_MAX_BUF_SIZE 16384
1105 #define EM_RCTL_FLXBUF_STEP 1024
1107 static const struct rte_memzone *
1108 ring_dma_zone_reserve(struct rte_eth_dev *dev, const char *ring_name,
1109 uint16_t queue_id, uint32_t ring_size, int socket_id)
1111 const struct rte_memzone *mz;
1112 char z_name[RTE_MEMZONE_NAMESIZE];
1114 snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
1115 dev->driver->pci_drv.name, ring_name, dev->data->port_id,
1118 if ((mz = rte_memzone_lookup(z_name)) != 0)
1121 #ifdef RTE_LIBRTE_XEN_DOM0
1122 return rte_memzone_reserve_bounded(z_name, ring_size,
1123 socket_id, 0, RTE_CACHE_LINE_SIZE, RTE_PGSIZE_2M);
1125 return rte_memzone_reserve(z_name, ring_size, socket_id, 0);
1130 em_tx_queue_release_mbufs(struct em_tx_queue *txq)
1134 if (txq->sw_ring != NULL) {
1135 for (i = 0; i != txq->nb_tx_desc; i++) {
1136 if (txq->sw_ring[i].mbuf != NULL) {
1137 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
1138 txq->sw_ring[i].mbuf = NULL;
1145 em_tx_queue_release(struct em_tx_queue *txq)
1148 em_tx_queue_release_mbufs(txq);
1149 rte_free(txq->sw_ring);
1155 eth_em_tx_queue_release(void *txq)
1157 em_tx_queue_release(txq);
1160 /* (Re)set dynamic em_tx_queue fields to defaults */
1162 em_reset_tx_queue(struct em_tx_queue *txq)
1164 uint16_t i, nb_desc, prev;
1165 static const struct e1000_data_desc txd_init = {
1166 .upper.fields = {.status = E1000_TXD_STAT_DD},
1169 nb_desc = txq->nb_tx_desc;
1171 /* Initialize ring entries */
1173 prev = (uint16_t) (nb_desc - 1);
1175 for (i = 0; i < nb_desc; i++) {
1176 txq->tx_ring[i] = txd_init;
1177 txq->sw_ring[i].mbuf = NULL;
1178 txq->sw_ring[i].last_id = i;
1179 txq->sw_ring[prev].next_id = i;
1184 * Always allow 1 descriptor to be un-allocated to avoid
1185 * a H/W race condition
1187 txq->nb_tx_free = (uint16_t)(nb_desc - 1);
1188 txq->last_desc_cleaned = (uint16_t)(nb_desc - 1);
1189 txq->nb_tx_used = 0;
1192 memset((void*)&txq->ctx_cache, 0, sizeof (txq->ctx_cache));
1196 eth_em_tx_queue_setup(struct rte_eth_dev *dev,
1199 unsigned int socket_id,
1200 const struct rte_eth_txconf *tx_conf)
1202 const struct rte_memzone *tz;
1203 struct em_tx_queue *txq;
1204 struct e1000_hw *hw;
1206 uint16_t tx_rs_thresh, tx_free_thresh;
1208 hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1211 * Validate number of transmit descriptors.
1212 * It must not exceed hardware maximum, and must be multiple
1215 if (((nb_desc * sizeof(*txq->tx_ring)) % EM_ALIGN) != 0 ||
1216 (nb_desc > EM_MAX_RING_DESC) ||
1217 (nb_desc < EM_MIN_RING_DESC)) {
1221 tx_free_thresh = tx_conf->tx_free_thresh;
1222 if (tx_free_thresh == 0)
1223 tx_free_thresh = (uint16_t)RTE_MIN(nb_desc / 4,
1224 DEFAULT_TX_FREE_THRESH);
1226 tx_rs_thresh = tx_conf->tx_rs_thresh;
1227 if (tx_rs_thresh == 0)
1228 tx_rs_thresh = (uint16_t)RTE_MIN(tx_free_thresh,
1229 DEFAULT_TX_RS_THRESH);
1231 if (tx_free_thresh >= (nb_desc - 3)) {
1232 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
1233 "number of TX descriptors minus 3. "
1234 "(tx_free_thresh=%u port=%d queue=%d)",
1235 (unsigned int)tx_free_thresh,
1236 (int)dev->data->port_id, (int)queue_idx);
1239 if (tx_rs_thresh > tx_free_thresh) {
1240 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or equal to "
1241 "tx_free_thresh. (tx_free_thresh=%u "
1242 "tx_rs_thresh=%u port=%d queue=%d)",
1243 (unsigned int)tx_free_thresh,
1244 (unsigned int)tx_rs_thresh,
1245 (int)dev->data->port_id,
1251 * If rs_bit_thresh is greater than 1, then TX WTHRESH should be
1252 * set to 0. If WTHRESH is greater than zero, the RS bit is ignored
1253 * by the NIC and all descriptors are written back after the NIC
1254 * accumulates WTHRESH descriptors.
1256 if (tx_conf->tx_thresh.wthresh != 0 && tx_rs_thresh != 1) {
1257 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
1258 "tx_rs_thresh is greater than 1. (tx_rs_thresh=%u "
1259 "port=%d queue=%d)", (unsigned int)tx_rs_thresh,
1260 (int)dev->data->port_id, (int)queue_idx);
1264 /* Free memory prior to re-allocation if needed... */
1265 if (dev->data->tx_queues[queue_idx] != NULL) {
1266 em_tx_queue_release(dev->data->tx_queues[queue_idx]);
1267 dev->data->tx_queues[queue_idx] = NULL;
1271 * Allocate TX ring hardware descriptors. A memzone large enough to
1272 * handle the maximum ring size is allocated in order to allow for
1273 * resizing in later calls to the queue setup function.
1275 tsize = sizeof (txq->tx_ring[0]) * EM_MAX_RING_DESC;
1276 if ((tz = ring_dma_zone_reserve(dev, "tx_ring", queue_idx, tsize,
1277 socket_id)) == NULL)
1280 /* Allocate the tx queue data structure. */
1281 if ((txq = rte_zmalloc("ethdev TX queue", sizeof(*txq),
1282 RTE_CACHE_LINE_SIZE)) == NULL)
1285 /* Allocate software ring */
1286 if ((txq->sw_ring = rte_zmalloc("txq->sw_ring",
1287 sizeof(txq->sw_ring[0]) * nb_desc,
1288 RTE_CACHE_LINE_SIZE)) == NULL) {
1289 em_tx_queue_release(txq);
1293 txq->nb_tx_desc = nb_desc;
1294 txq->tx_free_thresh = tx_free_thresh;
1295 txq->tx_rs_thresh = tx_rs_thresh;
1296 txq->pthresh = tx_conf->tx_thresh.pthresh;
1297 txq->hthresh = tx_conf->tx_thresh.hthresh;
1298 txq->wthresh = tx_conf->tx_thresh.wthresh;
1299 txq->queue_id = queue_idx;
1300 txq->port_id = dev->data->port_id;
1302 txq->tdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_TDT(queue_idx));
1303 #ifndef RTE_LIBRTE_XEN_DOM0
1304 txq->tx_ring_phys_addr = (uint64_t) tz->phys_addr;
1306 txq->tx_ring_phys_addr = rte_mem_phy2mch(tz->memseg_id, tz->phys_addr);
1308 txq->tx_ring = (struct e1000_data_desc *) tz->addr;
1310 PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64,
1311 txq->sw_ring, txq->tx_ring, txq->tx_ring_phys_addr);
1313 em_reset_tx_queue(txq);
1315 dev->data->tx_queues[queue_idx] = txq;
1320 em_rx_queue_release_mbufs(struct em_rx_queue *rxq)
1324 if (rxq->sw_ring != NULL) {
1325 for (i = 0; i != rxq->nb_rx_desc; i++) {
1326 if (rxq->sw_ring[i].mbuf != NULL) {
1327 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
1328 rxq->sw_ring[i].mbuf = NULL;
1335 em_rx_queue_release(struct em_rx_queue *rxq)
1338 em_rx_queue_release_mbufs(rxq);
1339 rte_free(rxq->sw_ring);
1345 eth_em_rx_queue_release(void *rxq)
1347 em_rx_queue_release(rxq);
1350 /* Reset dynamic em_rx_queue fields back to defaults */
1352 em_reset_rx_queue(struct em_rx_queue *rxq)
1355 rxq->nb_rx_hold = 0;
1356 rxq->pkt_first_seg = NULL;
1357 rxq->pkt_last_seg = NULL;
1361 eth_em_rx_queue_setup(struct rte_eth_dev *dev,
1364 unsigned int socket_id,
1365 const struct rte_eth_rxconf *rx_conf,
1366 struct rte_mempool *mp)
1368 const struct rte_memzone *rz;
1369 struct em_rx_queue *rxq;
1370 struct e1000_hw *hw;
1373 hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1376 * Validate number of receive descriptors.
1377 * It must not exceed hardware maximum, and must be multiple
1380 if (((nb_desc * sizeof(rxq->rx_ring[0])) % EM_ALIGN) != 0 ||
1381 (nb_desc > EM_MAX_RING_DESC) ||
1382 (nb_desc < EM_MIN_RING_DESC)) {
1387 * EM devices don't support drop_en functionality
1389 if (rx_conf->rx_drop_en) {
1390 PMD_INIT_LOG(ERR, "drop_en functionality not supported by "
1395 /* Free memory prior to re-allocation if needed. */
1396 if (dev->data->rx_queues[queue_idx] != NULL) {
1397 em_rx_queue_release(dev->data->rx_queues[queue_idx]);
1398 dev->data->rx_queues[queue_idx] = NULL;
1401 /* Allocate RX ring for max possible mumber of hardware descriptors. */
1402 rsize = sizeof (rxq->rx_ring[0]) * EM_MAX_RING_DESC;
1403 if ((rz = ring_dma_zone_reserve(dev, "rx_ring", queue_idx, rsize,
1404 socket_id)) == NULL)
1407 /* Allocate the RX queue data structure. */
1408 if ((rxq = rte_zmalloc("ethdev RX queue", sizeof(*rxq),
1409 RTE_CACHE_LINE_SIZE)) == NULL)
1412 /* Allocate software ring. */
1413 if ((rxq->sw_ring = rte_zmalloc("rxq->sw_ring",
1414 sizeof (rxq->sw_ring[0]) * nb_desc,
1415 RTE_CACHE_LINE_SIZE)) == NULL) {
1416 em_rx_queue_release(rxq);
1421 rxq->nb_rx_desc = nb_desc;
1422 rxq->pthresh = rx_conf->rx_thresh.pthresh;
1423 rxq->hthresh = rx_conf->rx_thresh.hthresh;
1424 rxq->wthresh = rx_conf->rx_thresh.wthresh;
1425 rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1426 rxq->queue_id = queue_idx;
1427 rxq->port_id = dev->data->port_id;
1428 rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ?
1431 rxq->rdt_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_RDT(queue_idx));
1432 rxq->rdh_reg_addr = E1000_PCI_REG_ADDR(hw, E1000_RDH(queue_idx));
1433 #ifndef RTE_LIBRTE_XEN_DOM0
1434 rxq->rx_ring_phys_addr = (uint64_t) rz->phys_addr;
1436 rxq->rx_ring_phys_addr = rte_mem_phy2mch(rz->memseg_id, rz->phys_addr);
1438 rxq->rx_ring = (struct e1000_rx_desc *) rz->addr;
1440 PMD_INIT_LOG(DEBUG, "sw_ring=%p hw_ring=%p dma_addr=0x%"PRIx64,
1441 rxq->sw_ring, rxq->rx_ring, rxq->rx_ring_phys_addr);
1443 dev->data->rx_queues[queue_idx] = rxq;
1444 em_reset_rx_queue(rxq);
1450 eth_em_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1452 #define EM_RXQ_SCAN_INTERVAL 4
1453 volatile struct e1000_rx_desc *rxdp;
1454 struct em_rx_queue *rxq;
1457 if (rx_queue_id >= dev->data->nb_rx_queues) {
1458 PMD_RX_LOG(DEBUG, "Invalid RX queue_id=%d", rx_queue_id);
1462 rxq = dev->data->rx_queues[rx_queue_id];
1463 rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1465 while ((desc < rxq->nb_rx_desc) &&
1466 (rxdp->status & E1000_RXD_STAT_DD)) {
1467 desc += EM_RXQ_SCAN_INTERVAL;
1468 rxdp += EM_RXQ_SCAN_INTERVAL;
1469 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
1470 rxdp = &(rxq->rx_ring[rxq->rx_tail +
1471 desc - rxq->nb_rx_desc]);
1478 eth_em_rx_descriptor_done(void *rx_queue, uint16_t offset)
1480 volatile struct e1000_rx_desc *rxdp;
1481 struct em_rx_queue *rxq = rx_queue;
1484 if (unlikely(offset >= rxq->nb_rx_desc))
1486 desc = rxq->rx_tail + offset;
1487 if (desc >= rxq->nb_rx_desc)
1488 desc -= rxq->nb_rx_desc;
1490 rxdp = &rxq->rx_ring[desc];
1491 return !!(rxdp->status & E1000_RXD_STAT_DD);
1495 em_dev_clear_queues(struct rte_eth_dev *dev)
1498 struct em_tx_queue *txq;
1499 struct em_rx_queue *rxq;
1501 for (i = 0; i < dev->data->nb_tx_queues; i++) {
1502 txq = dev->data->tx_queues[i];
1504 em_tx_queue_release_mbufs(txq);
1505 em_reset_tx_queue(txq);
1509 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1510 rxq = dev->data->rx_queues[i];
1512 em_rx_queue_release_mbufs(rxq);
1513 em_reset_rx_queue(rxq);
1519 * Takes as input/output parameter RX buffer size.
1520 * Returns (BSIZE | BSEX | FLXBUF) fields of RCTL register.
1523 em_rctl_bsize(__rte_unused enum e1000_mac_type hwtyp, uint32_t *bufsz)
1526 * For BSIZE & BSEX all configurable sizes are:
1527 * 16384: rctl |= (E1000_RCTL_SZ_16384 | E1000_RCTL_BSEX);
1528 * 8192: rctl |= (E1000_RCTL_SZ_8192 | E1000_RCTL_BSEX);
1529 * 4096: rctl |= (E1000_RCTL_SZ_4096 | E1000_RCTL_BSEX);
1530 * 2048: rctl |= E1000_RCTL_SZ_2048;
1531 * 1024: rctl |= E1000_RCTL_SZ_1024;
1532 * 512: rctl |= E1000_RCTL_SZ_512;
1533 * 256: rctl |= E1000_RCTL_SZ_256;
1535 static const struct {
1538 } bufsz_to_rctl[] = {
1539 {16384, (E1000_RCTL_SZ_16384 | E1000_RCTL_BSEX)},
1540 {8192, (E1000_RCTL_SZ_8192 | E1000_RCTL_BSEX)},
1541 {4096, (E1000_RCTL_SZ_4096 | E1000_RCTL_BSEX)},
1542 {2048, E1000_RCTL_SZ_2048},
1543 {1024, E1000_RCTL_SZ_1024},
1544 {512, E1000_RCTL_SZ_512},
1545 {256, E1000_RCTL_SZ_256},
1549 uint32_t rctl_bsize;
1551 rctl_bsize = *bufsz;
1554 * Starting from 82571 it is possible to specify RX buffer size
1555 * by RCTL.FLXBUF. When this field is different from zero, the
1556 * RX buffer size = RCTL.FLXBUF * 1K
1557 * (e.g. t is possible to specify RX buffer size 1,2,...,15KB).
1558 * It is working ok on real HW, but by some reason doesn't work
1559 * on VMware emulated 82574L.
1560 * So for now, always use BSIZE/BSEX to setup RX buffer size.
1561 * If you don't plan to use it on VMware emulated 82574L and
1562 * would like to specify RX buffer size in 1K granularity,
1563 * uncomment the following lines:
1564 * ***************************************************************
1565 * if (hwtyp >= e1000_82571 && hwtyp <= e1000_82574 &&
1566 * rctl_bsize >= EM_RCTL_FLXBUF_STEP) {
1567 * rctl_bsize /= EM_RCTL_FLXBUF_STEP;
1568 * *bufsz = rctl_bsize;
1569 * return (rctl_bsize << E1000_RCTL_FLXBUF_SHIFT &
1570 * E1000_RCTL_FLXBUF_MASK);
1572 * ***************************************************************
1575 for (i = 0; i != sizeof(bufsz_to_rctl) / sizeof(bufsz_to_rctl[0]);
1577 if (rctl_bsize >= bufsz_to_rctl[i].bufsz) {
1578 *bufsz = bufsz_to_rctl[i].bufsz;
1579 return (bufsz_to_rctl[i].rctl);
1583 /* Should never happen. */
1588 em_alloc_rx_queue_mbufs(struct em_rx_queue *rxq)
1590 struct em_rx_entry *rxe = rxq->sw_ring;
1593 static const struct e1000_rx_desc rxd_init = {
1597 /* Initialize software ring entries */
1598 for (i = 0; i < rxq->nb_rx_desc; i++) {
1599 volatile struct e1000_rx_desc *rxd;
1600 struct rte_mbuf *mbuf = rte_rxmbuf_alloc(rxq->mb_pool);
1603 PMD_INIT_LOG(ERR, "RX mbuf alloc failed "
1604 "queue_id=%hu", rxq->queue_id);
1608 dma_addr = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(mbuf));
1610 /* Clear HW ring memory */
1611 rxq->rx_ring[i] = rxd_init;
1613 rxd = &rxq->rx_ring[i];
1614 rxd->buffer_addr = dma_addr;
1621 /*********************************************************************
1623 * Enable receive unit.
1625 **********************************************************************/
1627 eth_em_rx_init(struct rte_eth_dev *dev)
1629 struct e1000_hw *hw;
1630 struct em_rx_queue *rxq;
1634 uint32_t rctl_bsize;
1638 hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1641 * Make sure receives are disabled while setting
1642 * up the descriptor ring.
1644 rctl = E1000_READ_REG(hw, E1000_RCTL);
1645 E1000_WRITE_REG(hw, E1000_RCTL, rctl & ~E1000_RCTL_EN);
1647 rfctl = E1000_READ_REG(hw, E1000_RFCTL);
1649 /* Disable extended descriptor type. */
1650 rfctl &= ~E1000_RFCTL_EXTEN;
1651 /* Disable accelerated acknowledge */
1652 if (hw->mac.type == e1000_82574)
1653 rfctl |= E1000_RFCTL_ACK_DIS;
1655 E1000_WRITE_REG(hw, E1000_RFCTL, rfctl);
1658 * XXX TEMPORARY WORKAROUND: on some systems with 82573
1659 * long latencies are observed, like Lenovo X60. This
1660 * change eliminates the problem, but since having positive
1661 * values in RDTR is a known source of problems on other
1662 * platforms another solution is being sought.
1664 if (hw->mac.type == e1000_82573)
1665 E1000_WRITE_REG(hw, E1000_RDTR, 0x20);
1667 dev->rx_pkt_burst = (eth_rx_burst_t)eth_em_recv_pkts;
1669 /* Determine RX bufsize. */
1670 rctl_bsize = EM_MAX_BUF_SIZE;
1671 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1672 struct rte_pktmbuf_pool_private *mbp_priv;
1675 rxq = dev->data->rx_queues[i];
1676 mbp_priv = rte_mempool_get_priv(rxq->mb_pool);
1677 buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM;
1678 rctl_bsize = RTE_MIN(rctl_bsize, buf_size);
1681 rctl |= em_rctl_bsize(hw->mac.type, &rctl_bsize);
1683 /* Configure and enable each RX queue. */
1684 for (i = 0; i < dev->data->nb_rx_queues; i++) {
1688 rxq = dev->data->rx_queues[i];
1690 /* Allocate buffers for descriptor rings and setup queue */
1691 ret = em_alloc_rx_queue_mbufs(rxq);
1696 * Reset crc_len in case it was changed after queue setup by a
1700 (uint8_t)(dev->data->dev_conf.rxmode.hw_strip_crc ?
1703 bus_addr = rxq->rx_ring_phys_addr;
1704 E1000_WRITE_REG(hw, E1000_RDLEN(i),
1706 sizeof(*rxq->rx_ring));
1707 E1000_WRITE_REG(hw, E1000_RDBAH(i),
1708 (uint32_t)(bus_addr >> 32));
1709 E1000_WRITE_REG(hw, E1000_RDBAL(i), (uint32_t)bus_addr);
1711 E1000_WRITE_REG(hw, E1000_RDH(i), 0);
1712 E1000_WRITE_REG(hw, E1000_RDT(i), rxq->nb_rx_desc - 1);
1714 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(0));
1715 rxdctl &= 0xFE000000;
1716 rxdctl |= rxq->pthresh & 0x3F;
1717 rxdctl |= (rxq->hthresh & 0x3F) << 8;
1718 rxdctl |= (rxq->wthresh & 0x3F) << 16;
1719 rxdctl |= E1000_RXDCTL_GRAN;
1720 E1000_WRITE_REG(hw, E1000_RXDCTL(i), rxdctl);
1723 * Due to EM devices not having any sort of hardware
1724 * limit for packet length, jumbo frame of any size
1725 * can be accepted, thus we have to enable scattered
1726 * rx if jumbo frames are enabled (or if buffer size
1727 * is too small to accommodate non-jumbo packets)
1728 * to avoid splitting packets that don't fit into
1731 if (dev->data->dev_conf.rxmode.jumbo_frame ||
1732 rctl_bsize < ETHER_MAX_LEN) {
1733 if (!dev->data->scattered_rx)
1734 PMD_INIT_LOG(DEBUG, "forcing scatter mode");
1736 (eth_rx_burst_t)eth_em_recv_scattered_pkts;
1737 dev->data->scattered_rx = 1;
1741 if (dev->data->dev_conf.rxmode.enable_scatter) {
1742 if (!dev->data->scattered_rx)
1743 PMD_INIT_LOG(DEBUG, "forcing scatter mode");
1744 dev->rx_pkt_burst = eth_em_recv_scattered_pkts;
1745 dev->data->scattered_rx = 1;
1749 * Setup the Checksum Register.
1750 * Receive Full-Packet Checksum Offload is mutually exclusive with RSS.
1752 rxcsum = E1000_READ_REG(hw, E1000_RXCSUM);
1754 if (dev->data->dev_conf.rxmode.hw_ip_checksum)
1755 rxcsum |= E1000_RXCSUM_IPOFL;
1757 rxcsum &= ~E1000_RXCSUM_IPOFL;
1758 E1000_WRITE_REG(hw, E1000_RXCSUM, rxcsum);
1760 /* No MRQ or RSS support for now */
1762 /* Set early receive threshold on appropriate hw */
1763 if ((hw->mac.type == e1000_ich9lan ||
1764 hw->mac.type == e1000_pch2lan ||
1765 hw->mac.type == e1000_ich10lan) &&
1766 dev->data->dev_conf.rxmode.jumbo_frame == 1) {
1767 u32 rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(0));
1768 E1000_WRITE_REG(hw, E1000_RXDCTL(0), rxdctl | 3);
1769 E1000_WRITE_REG(hw, E1000_ERT, 0x100 | (1 << 13));
1772 if (hw->mac.type == e1000_pch2lan) {
1773 if (dev->data->dev_conf.rxmode.jumbo_frame == 1)
1774 e1000_lv_jumbo_workaround_ich8lan(hw, TRUE);
1776 e1000_lv_jumbo_workaround_ich8lan(hw, FALSE);
1779 /* Setup the Receive Control Register. */
1780 if (dev->data->dev_conf.rxmode.hw_strip_crc)
1781 rctl |= E1000_RCTL_SECRC; /* Strip Ethernet CRC. */
1783 rctl &= ~E1000_RCTL_SECRC; /* Do not Strip Ethernet CRC. */
1785 rctl &= ~(3 << E1000_RCTL_MO_SHIFT);
1786 rctl |= E1000_RCTL_EN | E1000_RCTL_BAM | E1000_RCTL_LBM_NO |
1787 E1000_RCTL_RDMTS_HALF |
1788 (hw->mac.mc_filter_type << E1000_RCTL_MO_SHIFT);
1790 /* Make sure VLAN Filters are off. */
1791 rctl &= ~E1000_RCTL_VFE;
1792 /* Don't store bad packets. */
1793 rctl &= ~E1000_RCTL_SBP;
1794 /* Legacy descriptor type. */
1795 rctl &= ~E1000_RCTL_DTYP_MASK;
1798 * Configure support of jumbo frames, if any.
1800 if (dev->data->dev_conf.rxmode.jumbo_frame == 1)
1801 rctl |= E1000_RCTL_LPE;
1803 rctl &= ~E1000_RCTL_LPE;
1805 /* Enable Receives. */
1806 E1000_WRITE_REG(hw, E1000_RCTL, rctl);
1811 /*********************************************************************
1813 * Enable transmit unit.
1815 **********************************************************************/
1817 eth_em_tx_init(struct rte_eth_dev *dev)
1819 struct e1000_hw *hw;
1820 struct em_tx_queue *txq;
1825 hw = E1000_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1827 /* Setup the Base and Length of the Tx Descriptor Rings. */
1828 for (i = 0; i < dev->data->nb_tx_queues; i++) {
1831 txq = dev->data->tx_queues[i];
1832 bus_addr = txq->tx_ring_phys_addr;
1833 E1000_WRITE_REG(hw, E1000_TDLEN(i),
1835 sizeof(*txq->tx_ring));
1836 E1000_WRITE_REG(hw, E1000_TDBAH(i),
1837 (uint32_t)(bus_addr >> 32));
1838 E1000_WRITE_REG(hw, E1000_TDBAL(i), (uint32_t)bus_addr);
1840 /* Setup the HW Tx Head and Tail descriptor pointers. */
1841 E1000_WRITE_REG(hw, E1000_TDT(i), 0);
1842 E1000_WRITE_REG(hw, E1000_TDH(i), 0);
1844 /* Setup Transmit threshold registers. */
1845 txdctl = E1000_READ_REG(hw, E1000_TXDCTL(i));
1847 * bit 22 is reserved, on some models should always be 0,
1848 * on others - always 1.
1850 txdctl &= E1000_TXDCTL_COUNT_DESC;
1851 txdctl |= txq->pthresh & 0x3F;
1852 txdctl |= (txq->hthresh & 0x3F) << 8;
1853 txdctl |= (txq->wthresh & 0x3F) << 16;
1854 txdctl |= E1000_TXDCTL_GRAN;
1855 E1000_WRITE_REG(hw, E1000_TXDCTL(i), txdctl);
1858 /* Program the Transmit Control Register. */
1859 tctl = E1000_READ_REG(hw, E1000_TCTL);
1860 tctl &= ~E1000_TCTL_CT;
1861 tctl |= (E1000_TCTL_PSP | E1000_TCTL_RTLC | E1000_TCTL_EN |
1862 (E1000_COLLISION_THRESHOLD << E1000_CT_SHIFT));
1864 /* This write will effectively turn on the transmit unit. */
1865 E1000_WRITE_REG(hw, E1000_TCTL, tctl);