2 * Copyright (C) 2015 Cavium, Inc.
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms of version 2 of the GNU General Public License
6 * as published by the Free Software Foundation.
10 #include <linux/netdevice.h>
12 #include <linux/etherdevice.h>
19 #include "nicvf_queues.h"
21 static void nicvf_get_page(struct nicvf *nic)
23 if (!nic->rb_pageref || !nic->rb_page)
26 page_ref_add(nic->rb_page, nic->rb_pageref);
30 /* Poll a register for a specific value */
31 static int nicvf_poll_reg(struct nicvf *nic, int qidx,
32 u64 reg, int bit_pos, int bits, int val)
38 bit_mask = (1ULL << bits) - 1;
39 bit_mask = (bit_mask << bit_pos);
42 reg_val = nicvf_queue_reg_read(nic, reg, qidx);
43 if (((reg_val & bit_mask) >> bit_pos) == val)
45 usleep_range(1000, 2000);
48 netdev_err(nic->netdev, "Poll on reg 0x%llx failed\n", reg);
52 /* Allocate memory for a queue's descriptors */
53 static int nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem,
54 int q_len, int desc_size, int align_bytes)
57 dmem->size = (desc_size * q_len) + align_bytes;
58 /* Save address, need it while freeing */
59 dmem->unalign_base = dma_zalloc_coherent(&nic->pdev->dev, dmem->size,
60 &dmem->dma, GFP_KERNEL);
61 if (!dmem->unalign_base)
64 /* Align memory address for 'align_bytes' */
65 dmem->phys_base = NICVF_ALIGNED_ADDR((u64)dmem->dma, align_bytes);
66 dmem->base = dmem->unalign_base + (dmem->phys_base - dmem->dma);
70 /* Free queue's descriptor memory */
71 static void nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem)
76 dma_free_coherent(&nic->pdev->dev, dmem->size,
77 dmem->unalign_base, dmem->dma);
78 dmem->unalign_base = NULL;
82 /* Allocate buffer for packet reception
83 * HW returns memory address where packet is DMA'ed but not a pointer
84 * into RBDR ring, so save buffer address at the start of fragment and
85 * align the start address to a cache aligned address
87 static inline int nicvf_alloc_rcv_buffer(struct nicvf *nic, gfp_t gfp,
88 u32 buf_len, u64 **rbuf)
90 int order = (PAGE_SIZE <= 4096) ? PAGE_ALLOC_COSTLY_ORDER : 0;
92 /* Check if request can be accomodated in previous allocated page */
94 ((nic->rb_page_offset + buf_len) < (PAGE_SIZE << order))) {
102 /* Allocate a new page */
104 nic->rb_page = alloc_pages(gfp | __GFP_COMP | __GFP_NOWARN,
107 nic->drv_stats.rcv_buffer_alloc_failures++;
110 nic->rb_page_offset = 0;
114 *rbuf = (u64 *)((u64)page_address(nic->rb_page) + nic->rb_page_offset);
115 nic->rb_page_offset += buf_len;
120 /* Build skb around receive buffer */
121 static struct sk_buff *nicvf_rb_ptr_to_skb(struct nicvf *nic,
127 data = phys_to_virt(rb_ptr);
129 /* Now build an skb to give to stack */
130 skb = build_skb(data, RCV_FRAG_LEN);
132 put_page(virt_to_page(data));
140 /* Allocate RBDR ring and populate receive buffers */
141 static int nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr,
142 int ring_len, int buf_size)
146 struct rbdr_entry_t *desc;
149 err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len,
150 sizeof(struct rbdr_entry_t),
151 NICVF_RCV_BUF_ALIGN_BYTES);
155 rbdr->desc = rbdr->dmem.base;
156 /* Buffer size has to be in multiples of 128 bytes */
157 rbdr->dma_size = buf_size;
159 rbdr->thresh = RBDR_THRESH;
162 for (idx = 0; idx < ring_len; idx++) {
163 err = nicvf_alloc_rcv_buffer(nic, GFP_KERNEL, RCV_FRAG_LEN,
168 desc = GET_RBDR_DESC(rbdr, idx);
169 desc->buf_addr = virt_to_phys(rbuf) >> NICVF_RCV_BUF_ALIGN;
177 /* Free RBDR ring and its receive buffers */
178 static void nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr)
182 struct rbdr_entry_t *desc;
187 rbdr->enable = false;
188 if (!rbdr->dmem.base)
195 while (head != tail) {
196 desc = GET_RBDR_DESC(rbdr, head);
197 buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN;
198 put_page(virt_to_page(phys_to_virt(buf_addr)));
200 head &= (rbdr->dmem.q_len - 1);
202 /* Free SKB of tail desc */
203 desc = GET_RBDR_DESC(rbdr, tail);
204 buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN;
205 put_page(virt_to_page(phys_to_virt(buf_addr)));
208 nicvf_free_q_desc_mem(nic, &rbdr->dmem);
211 /* Refill receive buffer descriptors with new buffers.
213 static void nicvf_refill_rbdr(struct nicvf *nic, gfp_t gfp)
215 struct queue_set *qs = nic->qs;
216 int rbdr_idx = qs->rbdr_cnt;
220 struct rbdr_entry_t *desc;
228 rbdr = &qs->rbdr[rbdr_idx];
229 /* Check if it's enabled */
233 /* Get no of desc's to be refilled */
234 qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx);
236 /* Doorbell can be ringed with a max of ring size minus 1 */
237 if (qcount >= (qs->rbdr_len - 1))
240 refill_rb_cnt = qs->rbdr_len - qcount - 1;
242 /* Start filling descs from tail */
243 tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3;
244 while (refill_rb_cnt) {
246 tail &= (rbdr->dmem.q_len - 1);
248 if (nicvf_alloc_rcv_buffer(nic, gfp, RCV_FRAG_LEN, &rbuf))
251 desc = GET_RBDR_DESC(rbdr, tail);
252 desc->buf_addr = virt_to_phys(rbuf) >> NICVF_RCV_BUF_ALIGN;
259 /* make sure all memory stores are done before ringing doorbell */
262 /* Check if buffer allocation failed */
264 nic->rb_alloc_fail = true;
266 nic->rb_alloc_fail = false;
269 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
272 /* Re-enable RBDR interrupts only if buffer allocation is success */
273 if (!nic->rb_alloc_fail && rbdr->enable)
274 nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx);
280 /* Alloc rcv buffers in non-atomic mode for better success */
281 void nicvf_rbdr_work(struct work_struct *work)
283 struct nicvf *nic = container_of(work, struct nicvf, rbdr_work.work);
285 nicvf_refill_rbdr(nic, GFP_KERNEL);
286 if (nic->rb_alloc_fail)
287 schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
289 nic->rb_work_scheduled = false;
292 /* In Softirq context, alloc rcv buffers in atomic mode */
293 void nicvf_rbdr_task(unsigned long data)
295 struct nicvf *nic = (struct nicvf *)data;
297 nicvf_refill_rbdr(nic, GFP_ATOMIC);
298 if (nic->rb_alloc_fail) {
299 nic->rb_work_scheduled = true;
300 schedule_delayed_work(&nic->rbdr_work, msecs_to_jiffies(10));
304 /* Initialize completion queue */
305 static int nicvf_init_cmp_queue(struct nicvf *nic,
306 struct cmp_queue *cq, int q_len)
310 err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE,
311 NICVF_CQ_BASE_ALIGN_BYTES);
315 cq->desc = cq->dmem.base;
316 cq->thresh = pass1_silicon(nic->pdev) ? 0 : CMP_QUEUE_CQE_THRESH;
317 nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1;
322 static void nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq)
329 nicvf_free_q_desc_mem(nic, &cq->dmem);
332 /* Initialize transmit queue */
333 static int nicvf_init_snd_queue(struct nicvf *nic,
334 struct snd_queue *sq, int q_len)
338 err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE,
339 NICVF_SQ_BASE_ALIGN_BYTES);
343 sq->desc = sq->dmem.base;
344 sq->skbuff = kcalloc(q_len, sizeof(u64), GFP_KERNEL);
349 atomic_set(&sq->free_cnt, q_len - 1);
350 sq->thresh = SND_QUEUE_THRESH;
352 /* Preallocate memory for TSO segment's header */
353 sq->tso_hdrs = dma_alloc_coherent(&nic->pdev->dev,
354 q_len * TSO_HEADER_SIZE,
355 &sq->tso_hdrs_phys, GFP_KERNEL);
362 static void nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq)
370 dma_free_coherent(&nic->pdev->dev,
371 sq->dmem.q_len * TSO_HEADER_SIZE,
372 sq->tso_hdrs, sq->tso_hdrs_phys);
375 nicvf_free_q_desc_mem(nic, &sq->dmem);
378 static void nicvf_reclaim_snd_queue(struct nicvf *nic,
379 struct queue_set *qs, int qidx)
381 /* Disable send queue */
382 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0);
383 /* Check if SQ is stopped */
384 if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01))
386 /* Reset send queue */
387 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
390 static void nicvf_reclaim_rcv_queue(struct nicvf *nic,
391 struct queue_set *qs, int qidx)
393 union nic_mbx mbx = {};
395 /* Make sure all packets in the pipeline are written back into mem */
396 mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC;
397 nicvf_send_msg_to_pf(nic, &mbx);
400 static void nicvf_reclaim_cmp_queue(struct nicvf *nic,
401 struct queue_set *qs, int qidx)
403 /* Disable timer threshold (doesn't get reset upon CQ reset */
404 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0);
405 /* Disable completion queue */
406 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0);
407 /* Reset completion queue */
408 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
411 static void nicvf_reclaim_rbdr(struct nicvf *nic,
412 struct rbdr *rbdr, int qidx)
417 /* Save head and tail pointers for feeing up buffers */
418 rbdr->head = nicvf_queue_reg_read(nic,
419 NIC_QSET_RBDR_0_1_HEAD,
421 rbdr->tail = nicvf_queue_reg_read(nic,
422 NIC_QSET_RBDR_0_1_TAIL,
425 /* If RBDR FIFO is in 'FAIL' state then do a reset first
428 fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx);
429 if (((fifo_state >> 62) & 0x03) == 0x3)
430 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
431 qidx, NICVF_RBDR_RESET);
434 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0);
435 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
438 tmp = nicvf_queue_reg_read(nic,
439 NIC_QSET_RBDR_0_1_PREFETCH_STATUS,
441 if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF))
443 usleep_range(1000, 2000);
446 netdev_err(nic->netdev,
447 "Failed polling on prefetch status\n");
451 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
452 qidx, NICVF_RBDR_RESET);
454 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02))
456 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00);
457 if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
461 void nicvf_config_vlan_stripping(struct nicvf *nic, netdev_features_t features)
466 rq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_RQ_GEN_CFG, 0);
468 /* Enable first VLAN stripping */
469 if (features & NETIF_F_HW_VLAN_CTAG_RX)
470 rq_cfg |= (1ULL << 25);
472 rq_cfg &= ~(1ULL << 25);
473 nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
475 /* Configure Secondary Qsets, if any */
476 for (sqs = 0; sqs < nic->sqs_count; sqs++)
477 if (nic->snicvf[sqs])
478 nicvf_queue_reg_write(nic->snicvf[sqs],
479 NIC_QSET_RQ_GEN_CFG, 0, rq_cfg);
482 /* Configures receive queue */
483 static void nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs,
484 int qidx, bool enable)
486 union nic_mbx mbx = {};
487 struct rcv_queue *rq;
488 struct rq_cfg rq_cfg;
493 /* Disable receive queue */
494 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0);
497 nicvf_reclaim_rcv_queue(nic, qs, qidx);
501 rq->cq_qs = qs->vnic_id;
503 rq->start_rbdr_qs = qs->vnic_id;
504 rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1;
505 rq->cont_rbdr_qs = qs->vnic_id;
506 rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1;
507 /* all writes of RBDR data to be loaded into L2 Cache as well*/
510 /* Send a mailbox msg to PF to config RQ */
511 mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG;
512 mbx.rq.qs_num = qs->vnic_id;
513 mbx.rq.rq_num = qidx;
514 mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) |
515 (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) |
516 (rq->cont_qs_rbdr_idx << 8) |
517 (rq->start_rbdr_qs << 1) | (rq->start_qs_rbdr_idx);
518 nicvf_send_msg_to_pf(nic, &mbx);
520 mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG;
521 mbx.rq.cfg = (1ULL << 63) | (1ULL << 62) | (qs->vnic_id << 0);
522 nicvf_send_msg_to_pf(nic, &mbx);
525 * Enable CQ drop to reserve sufficient CQEs for all tx packets
527 mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG;
528 mbx.rq.cfg = (1ULL << 62) | (RQ_CQ_DROP << 8);
529 nicvf_send_msg_to_pf(nic, &mbx);
531 nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, 0x00);
533 nicvf_config_vlan_stripping(nic, nic->netdev->features);
535 /* Enable Receive queue */
536 memset(&rq_cfg, 0, sizeof(struct rq_cfg));
539 nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, *(u64 *)&rq_cfg);
542 /* Configures completion queue */
543 void nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs,
544 int qidx, bool enable)
546 struct cmp_queue *cq;
547 struct cq_cfg cq_cfg;
553 nicvf_reclaim_cmp_queue(nic, qs, qidx);
557 /* Reset completion queue */
558 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
563 spin_lock_init(&cq->lock);
564 /* Set completion queue base address */
565 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE,
566 qidx, (u64)(cq->dmem.phys_base));
568 /* Enable Completion queue */
569 memset(&cq_cfg, 0, sizeof(struct cq_cfg));
573 cq_cfg.qsize = CMP_QSIZE;
575 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(u64 *)&cq_cfg);
577 /* Set threshold value for interrupt generation */
578 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh);
579 nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2,
580 qidx, CMP_QUEUE_TIMER_THRESH);
583 /* Configures transmit queue */
584 static void nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs,
585 int qidx, bool enable)
587 union nic_mbx mbx = {};
588 struct snd_queue *sq;
589 struct sq_cfg sq_cfg;
595 nicvf_reclaim_snd_queue(nic, qs, qidx);
599 /* Reset send queue */
600 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
602 sq->cq_qs = qs->vnic_id;
605 /* Send a mailbox msg to PF to config SQ */
606 mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG;
607 mbx.sq.qs_num = qs->vnic_id;
608 mbx.sq.sq_num = qidx;
609 mbx.sq.sqs_mode = nic->sqs_mode;
610 mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx;
611 nicvf_send_msg_to_pf(nic, &mbx);
613 /* Set queue base address */
614 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE,
615 qidx, (u64)(sq->dmem.phys_base));
617 /* Enable send queue & set queue size */
618 memset(&sq_cfg, 0, sizeof(struct sq_cfg));
622 sq_cfg.qsize = SND_QSIZE;
623 sq_cfg.tstmp_bgx_intf = 0;
624 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(u64 *)&sq_cfg);
626 /* Set threshold value for interrupt generation */
627 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh);
629 /* Set queue:cpu affinity for better load distribution */
630 if (cpu_online(qidx)) {
631 cpumask_set_cpu(qidx, &sq->affinity_mask);
632 netif_set_xps_queue(nic->netdev,
633 &sq->affinity_mask, qidx);
637 /* Configures receive buffer descriptor ring */
638 static void nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs,
639 int qidx, bool enable)
642 struct rbdr_cfg rbdr_cfg;
644 rbdr = &qs->rbdr[qidx];
645 nicvf_reclaim_rbdr(nic, rbdr, qidx);
649 /* Set descriptor base address */
650 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE,
651 qidx, (u64)(rbdr->dmem.phys_base));
653 /* Enable RBDR & set queue size */
654 /* Buffer size should be in multiples of 128 bytes */
655 memset(&rbdr_cfg, 0, sizeof(struct rbdr_cfg));
659 rbdr_cfg.qsize = RBDR_SIZE;
660 rbdr_cfg.avg_con = 0;
661 rbdr_cfg.lines = rbdr->dma_size / 128;
662 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
663 qidx, *(u64 *)&rbdr_cfg);
666 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
667 qidx, qs->rbdr_len - 1);
669 /* Set threshold value for interrupt generation */
670 nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH,
671 qidx, rbdr->thresh - 1);
674 /* Requests PF to assign and enable Qset */
675 void nicvf_qset_config(struct nicvf *nic, bool enable)
677 union nic_mbx mbx = {};
678 struct queue_set *qs = nic->qs;
679 struct qs_cfg *qs_cfg;
682 netdev_warn(nic->netdev,
683 "Qset is still not allocated, don't init queues\n");
688 qs->vnic_id = nic->vf_id;
690 /* Send a mailbox msg to PF to config Qset */
691 mbx.qs.msg = NIC_MBOX_MSG_QS_CFG;
692 mbx.qs.num = qs->vnic_id;
693 mbx.qs.sqs_count = nic->sqs_count;
696 qs_cfg = (struct qs_cfg *)&mbx.qs.cfg;
702 qs_cfg->vnic = qs->vnic_id;
704 nicvf_send_msg_to_pf(nic, &mbx);
707 static void nicvf_free_resources(struct nicvf *nic)
710 struct queue_set *qs = nic->qs;
712 /* Free receive buffer descriptor ring */
713 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
714 nicvf_free_rbdr(nic, &qs->rbdr[qidx]);
716 /* Free completion queue */
717 for (qidx = 0; qidx < qs->cq_cnt; qidx++)
718 nicvf_free_cmp_queue(nic, &qs->cq[qidx]);
720 /* Free send queue */
721 for (qidx = 0; qidx < qs->sq_cnt; qidx++)
722 nicvf_free_snd_queue(nic, &qs->sq[qidx]);
725 static int nicvf_alloc_resources(struct nicvf *nic)
728 struct queue_set *qs = nic->qs;
730 /* Alloc receive buffer descriptor ring */
731 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
732 if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len,
737 /* Alloc send queue */
738 for (qidx = 0; qidx < qs->sq_cnt; qidx++) {
739 if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len))
743 /* Alloc completion queue */
744 for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
745 if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len))
751 nicvf_free_resources(nic);
755 int nicvf_set_qset_resources(struct nicvf *nic)
757 struct queue_set *qs;
759 qs = devm_kzalloc(&nic->pdev->dev, sizeof(*qs), GFP_KERNEL);
764 /* Set count of each queue */
765 qs->rbdr_cnt = RBDR_CNT;
766 qs->rq_cnt = RCV_QUEUE_CNT;
767 qs->sq_cnt = SND_QUEUE_CNT;
768 qs->cq_cnt = CMP_QUEUE_CNT;
770 /* Set queue lengths */
771 qs->rbdr_len = RCV_BUF_COUNT;
772 qs->sq_len = SND_QUEUE_LEN;
773 qs->cq_len = CMP_QUEUE_LEN;
775 nic->rx_queues = qs->rq_cnt;
776 nic->tx_queues = qs->sq_cnt;
781 int nicvf_config_data_transfer(struct nicvf *nic, bool enable)
783 bool disable = false;
784 struct queue_set *qs = nic->qs;
791 if (nicvf_alloc_resources(nic))
794 for (qidx = 0; qidx < qs->sq_cnt; qidx++)
795 nicvf_snd_queue_config(nic, qs, qidx, enable);
796 for (qidx = 0; qidx < qs->cq_cnt; qidx++)
797 nicvf_cmp_queue_config(nic, qs, qidx, enable);
798 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
799 nicvf_rbdr_config(nic, qs, qidx, enable);
800 for (qidx = 0; qidx < qs->rq_cnt; qidx++)
801 nicvf_rcv_queue_config(nic, qs, qidx, enable);
803 for (qidx = 0; qidx < qs->rq_cnt; qidx++)
804 nicvf_rcv_queue_config(nic, qs, qidx, disable);
805 for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
806 nicvf_rbdr_config(nic, qs, qidx, disable);
807 for (qidx = 0; qidx < qs->sq_cnt; qidx++)
808 nicvf_snd_queue_config(nic, qs, qidx, disable);
809 for (qidx = 0; qidx < qs->cq_cnt; qidx++)
810 nicvf_cmp_queue_config(nic, qs, qidx, disable);
812 nicvf_free_resources(nic);
818 /* Get a free desc from SQ
819 * returns descriptor ponter & descriptor number
821 static inline int nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt)
826 atomic_sub(desc_cnt, &sq->free_cnt);
827 sq->tail += desc_cnt;
828 sq->tail &= (sq->dmem.q_len - 1);
833 /* Free descriptor back to SQ for future use */
834 void nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt)
836 atomic_add(desc_cnt, &sq->free_cnt);
837 sq->head += desc_cnt;
838 sq->head &= (sq->dmem.q_len - 1);
841 static inline int nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry)
844 qentry &= (sq->dmem.q_len - 1);
848 void nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx)
852 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
853 sq_cfg |= NICVF_SQ_EN;
854 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
855 /* Ring doorbell so that H/W restarts processing SQEs */
856 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0);
859 void nicvf_sq_disable(struct nicvf *nic, int qidx)
863 sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
864 sq_cfg &= ~NICVF_SQ_EN;
865 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
868 void nicvf_sq_free_used_descs(struct net_device *netdev, struct snd_queue *sq,
873 struct nicvf *nic = netdev_priv(netdev);
874 struct sq_hdr_subdesc *hdr;
876 head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4;
877 tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4;
878 while (sq->head != head) {
879 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
880 if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
881 nicvf_put_sq_desc(sq, 1);
884 skb = (struct sk_buff *)sq->skbuff[sq->head];
886 dev_kfree_skb_any(skb);
887 atomic64_add(1, (atomic64_t *)&netdev->stats.tx_packets);
888 atomic64_add(hdr->tot_len,
889 (atomic64_t *)&netdev->stats.tx_bytes);
890 nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
894 /* Calculate no of SQ subdescriptors needed to transmit all
895 * segments of this TSO packet.
896 * Taken from 'Tilera network driver' with a minor modification.
898 static int nicvf_tso_count_subdescs(struct sk_buff *skb)
900 struct skb_shared_info *sh = skb_shinfo(skb);
901 unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
902 unsigned int data_len = skb->len - sh_len;
903 unsigned int p_len = sh->gso_size;
904 long f_id = -1; /* id of the current fragment */
905 long f_size = skb_headlen(skb) - sh_len; /* current fragment size */
906 long f_used = 0; /* bytes used from the current fragment */
907 long n; /* size of the current piece of payload */
911 for (segment = 0; segment < sh->gso_segs; segment++) {
912 unsigned int p_used = 0;
914 /* One edesc for header and for each piece of the payload. */
915 for (num_edescs++; p_used < p_len; num_edescs++) {
916 /* Advance as needed. */
917 while (f_used >= f_size) {
919 f_size = skb_frag_size(&sh->frags[f_id]);
923 /* Use bytes from the current fragment. */
925 if (n > f_size - f_used)
931 /* The last segment may be less than gso_size. */
933 if (data_len < p_len)
937 /* '+ gso_segs' for SQ_HDR_SUDESCs for each segment */
938 return num_edescs + sh->gso_segs;
941 /* Get the number of SQ descriptors needed to xmit this skb */
942 static int nicvf_sq_subdesc_required(struct nicvf *nic, struct sk_buff *skb)
944 int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT;
946 if (skb_shinfo(skb)->gso_size && !nic->hw_tso) {
947 subdesc_cnt = nicvf_tso_count_subdescs(skb);
951 if (skb_shinfo(skb)->nr_frags)
952 subdesc_cnt += skb_shinfo(skb)->nr_frags;
957 /* Add SQ HEADER subdescriptor.
958 * First subdescriptor for every send descriptor.
961 nicvf_sq_add_hdr_subdesc(struct nicvf *nic, struct snd_queue *sq, int qentry,
962 int subdesc_cnt, struct sk_buff *skb, int len)
965 struct sq_hdr_subdesc *hdr;
967 hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
968 sq->skbuff[qentry] = (u64)skb;
970 memset(hdr, 0, SND_QUEUE_DESC_SIZE);
971 hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
972 /* Enable notification via CQE after processing SQE */
974 /* No of subdescriptors following this */
975 hdr->subdesc_cnt = subdesc_cnt;
978 /* Offload checksum calculation to HW */
979 if (skb->ip_summed == CHECKSUM_PARTIAL) {
980 hdr->csum_l3 = 1; /* Enable IP csum calculation */
981 hdr->l3_offset = skb_network_offset(skb);
982 hdr->l4_offset = skb_transport_offset(skb);
984 proto = ip_hdr(skb)->protocol;
987 hdr->csum_l4 = SEND_L4_CSUM_TCP;
990 hdr->csum_l4 = SEND_L4_CSUM_UDP;
993 hdr->csum_l4 = SEND_L4_CSUM_SCTP;
998 if (nic->hw_tso && skb_shinfo(skb)->gso_size) {
1000 hdr->tso_start = skb_transport_offset(skb) + tcp_hdrlen(skb);
1001 hdr->tso_max_paysize = skb_shinfo(skb)->gso_size;
1002 /* For non-tunneled pkts, point this to L2 ethertype */
1003 hdr->inner_l3_offset = skb_network_offset(skb) - 2;
1004 nic->drv_stats.tx_tso++;
1008 /* SQ GATHER subdescriptor
1009 * Must follow HDR descriptor
1011 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
1014 struct sq_gather_subdesc *gather;
1016 qentry &= (sq->dmem.q_len - 1);
1017 gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry);
1019 memset(gather, 0, SND_QUEUE_DESC_SIZE);
1020 gather->subdesc_type = SQ_DESC_TYPE_GATHER;
1021 gather->ld_type = NIC_SEND_LD_TYPE_E_LDD;
1022 gather->size = size;
1023 gather->addr = data;
1026 /* Segment a TSO packet into 'gso_size' segments and append
1027 * them to SQ for transfer
1029 static int nicvf_sq_append_tso(struct nicvf *nic, struct snd_queue *sq,
1030 int sq_num, int qentry, struct sk_buff *skb)
1033 int seg_subdescs = 0, desc_cnt = 0;
1034 int seg_len, total_len, data_left;
1035 int hdr_qentry = qentry;
1036 int hdr_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
1038 tso_start(skb, &tso);
1039 total_len = skb->len - hdr_len;
1040 while (total_len > 0) {
1043 /* Save Qentry for adding HDR_SUBDESC at the end */
1044 hdr_qentry = qentry;
1046 data_left = min_t(int, skb_shinfo(skb)->gso_size, total_len);
1047 total_len -= data_left;
1049 /* Add segment's header */
1050 qentry = nicvf_get_nxt_sqentry(sq, qentry);
1051 hdr = sq->tso_hdrs + qentry * TSO_HEADER_SIZE;
1052 tso_build_hdr(skb, hdr, &tso, data_left, total_len == 0);
1053 nicvf_sq_add_gather_subdesc(sq, qentry, hdr_len,
1055 qentry * TSO_HEADER_SIZE);
1056 /* HDR_SUDESC + GATHER */
1060 /* Add segment's payload fragments */
1061 while (data_left > 0) {
1064 size = min_t(int, tso.size, data_left);
1066 qentry = nicvf_get_nxt_sqentry(sq, qentry);
1067 nicvf_sq_add_gather_subdesc(sq, qentry, size,
1068 virt_to_phys(tso.data));
1073 tso_build_data(skb, &tso, size);
1075 nicvf_sq_add_hdr_subdesc(nic, sq, hdr_qentry,
1076 seg_subdescs - 1, skb, seg_len);
1077 sq->skbuff[hdr_qentry] = (u64)NULL;
1078 qentry = nicvf_get_nxt_sqentry(sq, qentry);
1080 desc_cnt += seg_subdescs;
1082 /* Save SKB in the last segment for freeing */
1083 sq->skbuff[hdr_qentry] = (u64)skb;
1085 /* make sure all memory stores are done before ringing doorbell */
1088 /* Inform HW to xmit all TSO segments */
1089 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1091 nic->drv_stats.tx_tso++;
1095 /* Append an skb to a SQ for packet transfer. */
1096 int nicvf_sq_append_skb(struct nicvf *nic, struct sk_buff *skb)
1101 struct queue_set *qs;
1102 struct snd_queue *sq;
1104 sq_num = skb_get_queue_mapping(skb);
1105 if (sq_num >= MAX_SND_QUEUES_PER_QS) {
1106 /* Get secondary Qset's SQ structure */
1107 i = sq_num / MAX_SND_QUEUES_PER_QS;
1108 if (!nic->snicvf[i - 1]) {
1109 netdev_warn(nic->netdev,
1110 "Secondary Qset#%d's ptr not initialized\n",
1114 nic = (struct nicvf *)nic->snicvf[i - 1];
1115 sq_num = sq_num % MAX_SND_QUEUES_PER_QS;
1119 sq = &qs->sq[sq_num];
1121 subdesc_cnt = nicvf_sq_subdesc_required(nic, skb);
1122 if (subdesc_cnt > atomic_read(&sq->free_cnt))
1125 qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1127 /* Check if its a TSO packet */
1128 if (skb_shinfo(skb)->gso_size && !nic->hw_tso)
1129 return nicvf_sq_append_tso(nic, sq, sq_num, qentry, skb);
1131 /* Add SQ header subdesc */
1132 nicvf_sq_add_hdr_subdesc(nic, sq, qentry, subdesc_cnt - 1,
1135 /* Add SQ gather subdescs */
1136 qentry = nicvf_get_nxt_sqentry(sq, qentry);
1137 size = skb_is_nonlinear(skb) ? skb_headlen(skb) : skb->len;
1138 nicvf_sq_add_gather_subdesc(sq, qentry, size, virt_to_phys(skb->data));
1140 /* Check for scattered buffer */
1141 if (!skb_is_nonlinear(skb))
1144 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1145 const struct skb_frag_struct *frag;
1147 frag = &skb_shinfo(skb)->frags[i];
1149 qentry = nicvf_get_nxt_sqentry(sq, qentry);
1150 size = skb_frag_size(frag);
1151 nicvf_sq_add_gather_subdesc(sq, qentry, size,
1153 skb_frag_address(frag)));
1157 /* make sure all memory stores are done before ringing doorbell */
1160 /* Inform HW to xmit new packet */
1161 nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR,
1162 sq_num, subdesc_cnt);
1166 /* Use original PCI dev for debug log */
1168 netdev_dbg(nic->netdev, "Not enough SQ descriptors to xmit pkt\n");
1172 static inline unsigned frag_num(unsigned i)
1175 return (i & ~3) + 3 - (i & 3);
1181 /* Returns SKB for a received packet */
1182 struct sk_buff *nicvf_get_rcv_skb(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1185 int payload_len = 0;
1186 struct sk_buff *skb = NULL;
1187 struct sk_buff *skb_frag = NULL;
1188 struct sk_buff *prev_frag = NULL;
1189 u16 *rb_lens = NULL;
1190 u64 *rb_ptrs = NULL;
1192 rb_lens = (void *)cqe_rx + (3 * sizeof(u64));
1193 rb_ptrs = (void *)cqe_rx + (6 * sizeof(u64));
1195 netdev_dbg(nic->netdev, "%s rb_cnt %d rb0_ptr %llx rb0_sz %d\n",
1196 __func__, cqe_rx->rb_cnt, cqe_rx->rb0_ptr, cqe_rx->rb0_sz);
1198 for (frag = 0; frag < cqe_rx->rb_cnt; frag++) {
1199 payload_len = rb_lens[frag_num(frag)];
1201 /* First fragment */
1202 skb = nicvf_rb_ptr_to_skb(nic,
1203 *rb_ptrs - cqe_rx->align_pad,
1207 skb_reserve(skb, cqe_rx->align_pad);
1208 skb_put(skb, payload_len);
1211 skb_frag = nicvf_rb_ptr_to_skb(nic, *rb_ptrs,
1218 if (!skb_shinfo(skb)->frag_list)
1219 skb_shinfo(skb)->frag_list = skb_frag;
1221 prev_frag->next = skb_frag;
1223 prev_frag = skb_frag;
1224 skb->len += payload_len;
1225 skb->data_len += payload_len;
1226 skb_frag->len = payload_len;
1228 /* Next buffer pointer */
1234 static u64 nicvf_int_type_to_mask(int int_type, int q_idx)
1240 reg_val = ((1ULL << q_idx) << NICVF_INTR_CQ_SHIFT);
1243 reg_val = ((1ULL << q_idx) << NICVF_INTR_SQ_SHIFT);
1245 case NICVF_INTR_RBDR:
1246 reg_val = ((1ULL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1248 case NICVF_INTR_PKT_DROP:
1249 reg_val = (1ULL << NICVF_INTR_PKT_DROP_SHIFT);
1251 case NICVF_INTR_TCP_TIMER:
1252 reg_val = (1ULL << NICVF_INTR_TCP_TIMER_SHIFT);
1254 case NICVF_INTR_MBOX:
1255 reg_val = (1ULL << NICVF_INTR_MBOX_SHIFT);
1257 case NICVF_INTR_QS_ERR:
1258 reg_val = (1ULL << NICVF_INTR_QS_ERR_SHIFT);
1267 /* Enable interrupt */
1268 void nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx)
1270 u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1273 netdev_dbg(nic->netdev,
1274 "Failed to enable interrupt: unknown type\n");
1277 nicvf_reg_write(nic, NIC_VF_ENA_W1S,
1278 nicvf_reg_read(nic, NIC_VF_ENA_W1S) | mask);
1281 /* Disable interrupt */
1282 void nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx)
1284 u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1287 netdev_dbg(nic->netdev,
1288 "Failed to disable interrupt: unknown type\n");
1292 nicvf_reg_write(nic, NIC_VF_ENA_W1C, mask);
1295 /* Clear interrupt */
1296 void nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx)
1298 u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1301 netdev_dbg(nic->netdev,
1302 "Failed to clear interrupt: unknown type\n");
1306 nicvf_reg_write(nic, NIC_VF_INT, mask);
1309 /* Check if interrupt is enabled */
1310 int nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx)
1312 u64 mask = nicvf_int_type_to_mask(int_type, q_idx);
1313 /* If interrupt type is unknown, we treat it disabled. */
1315 netdev_dbg(nic->netdev,
1316 "Failed to check interrupt enable: unknown type\n");
1320 return mask & nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1323 void nicvf_update_rq_stats(struct nicvf *nic, int rq_idx)
1325 struct rcv_queue *rq;
1327 #define GET_RQ_STATS(reg) \
1328 nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\
1329 (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1331 rq = &nic->qs->rq[rq_idx];
1332 rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS);
1333 rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS);
1336 void nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
1338 struct snd_queue *sq;
1340 #define GET_SQ_STATS(reg) \
1341 nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\
1342 (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1344 sq = &nic->qs->sq[sq_idx];
1345 sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS);
1346 sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS);
1349 /* Check for errors in the receive cmp.queue entry */
1350 int nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1352 struct nicvf_hw_stats *stats = &nic->hw_stats;
1354 if (!cqe_rx->err_level && !cqe_rx->err_opcode)
1357 if (netif_msg_rx_err(nic))
1358 netdev_err(nic->netdev,
1359 "%s: RX error CQE err_level 0x%x err_opcode 0x%x\n",
1361 cqe_rx->err_level, cqe_rx->err_opcode);
1363 switch (cqe_rx->err_opcode) {
1364 case CQ_RX_ERROP_RE_PARTIAL:
1365 stats->rx_bgx_truncated_pkts++;
1367 case CQ_RX_ERROP_RE_JABBER:
1368 stats->rx_jabber_errs++;
1370 case CQ_RX_ERROP_RE_FCS:
1371 stats->rx_fcs_errs++;
1373 case CQ_RX_ERROP_RE_RX_CTL:
1374 stats->rx_bgx_errs++;
1376 case CQ_RX_ERROP_PREL2_ERR:
1377 stats->rx_prel2_errs++;
1379 case CQ_RX_ERROP_L2_MAL:
1380 stats->rx_l2_hdr_malformed++;
1382 case CQ_RX_ERROP_L2_OVERSIZE:
1383 stats->rx_oversize++;
1385 case CQ_RX_ERROP_L2_UNDERSIZE:
1386 stats->rx_undersize++;
1388 case CQ_RX_ERROP_L2_LENMISM:
1389 stats->rx_l2_len_mismatch++;
1391 case CQ_RX_ERROP_L2_PCLP:
1392 stats->rx_l2_pclp++;
1394 case CQ_RX_ERROP_IP_NOT:
1395 stats->rx_ip_ver_errs++;
1397 case CQ_RX_ERROP_IP_CSUM_ERR:
1398 stats->rx_ip_csum_errs++;
1400 case CQ_RX_ERROP_IP_MAL:
1401 stats->rx_ip_hdr_malformed++;
1403 case CQ_RX_ERROP_IP_MALD:
1404 stats->rx_ip_payload_malformed++;
1406 case CQ_RX_ERROP_IP_HOP:
1407 stats->rx_ip_ttl_errs++;
1409 case CQ_RX_ERROP_L3_PCLP:
1410 stats->rx_l3_pclp++;
1412 case CQ_RX_ERROP_L4_MAL:
1413 stats->rx_l4_malformed++;
1415 case CQ_RX_ERROP_L4_CHK:
1416 stats->rx_l4_csum_errs++;
1418 case CQ_RX_ERROP_UDP_LEN:
1419 stats->rx_udp_len_errs++;
1421 case CQ_RX_ERROP_L4_PORT:
1422 stats->rx_l4_port_errs++;
1424 case CQ_RX_ERROP_TCP_FLAG:
1425 stats->rx_tcp_flag_errs++;
1427 case CQ_RX_ERROP_TCP_OFFSET:
1428 stats->rx_tcp_offset_errs++;
1430 case CQ_RX_ERROP_L4_PCLP:
1431 stats->rx_l4_pclp++;
1433 case CQ_RX_ERROP_RBDR_TRUNC:
1434 stats->rx_truncated_pkts++;
1441 /* Check for errors in the send cmp.queue entry */
1442 int nicvf_check_cqe_tx_errs(struct nicvf *nic,
1443 struct cmp_queue *cq, struct cqe_send_t *cqe_tx)
1445 struct cmp_queue_stats *stats = &cq->stats;
1447 switch (cqe_tx->send_status) {
1448 case CQ_TX_ERROP_GOOD:
1451 case CQ_TX_ERROP_DESC_FAULT:
1452 stats->tx.desc_fault++;
1454 case CQ_TX_ERROP_HDR_CONS_ERR:
1455 stats->tx.hdr_cons_err++;
1457 case CQ_TX_ERROP_SUBDC_ERR:
1458 stats->tx.subdesc_err++;
1460 case CQ_TX_ERROP_IMM_SIZE_OFLOW:
1461 stats->tx.imm_size_oflow++;
1463 case CQ_TX_ERROP_DATA_SEQUENCE_ERR:
1464 stats->tx.data_seq_err++;
1466 case CQ_TX_ERROP_MEM_SEQUENCE_ERR:
1467 stats->tx.mem_seq_err++;
1469 case CQ_TX_ERROP_LOCK_VIOL:
1470 stats->tx.lock_viol++;
1472 case CQ_TX_ERROP_DATA_FAULT:
1473 stats->tx.data_fault++;
1475 case CQ_TX_ERROP_TSTMP_CONFLICT:
1476 stats->tx.tstmp_conflict++;
1478 case CQ_TX_ERROP_TSTMP_TIMEOUT:
1479 stats->tx.tstmp_timeout++;
1481 case CQ_TX_ERROP_MEM_FAULT:
1482 stats->tx.mem_fault++;
1484 case CQ_TX_ERROP_CK_OVERLAP:
1485 stats->tx.csum_overlap++;
1487 case CQ_TX_ERROP_CK_OFLOW:
1488 stats->tx.csum_overflow++;