xref: /freebsd/sys/dev/vnic/nicvf_queues.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*
2  * Copyright (C) 2015 Cavium Inc.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  *
26  * $FreeBSD$
27  *
28  */
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/bitset.h>
35 #include <sys/bitstring.h>
36 #include <sys/buf_ring.h>
37 #include <sys/bus.h>
38 #include <sys/endian.h>
39 #include <sys/kernel.h>
40 #include <sys/malloc.h>
41 #include <sys/module.h>
42 #include <sys/rman.h>
43 #include <sys/pciio.h>
44 #include <sys/pcpu.h>
45 #include <sys/proc.h>
46 #include <sys/sockio.h>
47 #include <sys/socket.h>
48 #include <sys/stdatomic.h>
49 #include <sys/cpuset.h>
50 #include <sys/lock.h>
51 #include <sys/mutex.h>
52 #include <sys/smp.h>
53 #include <sys/taskqueue.h>
54 
55 #include <vm/vm.h>
56 #include <vm/pmap.h>
57 
58 #include <machine/bus.h>
59 #include <machine/vmparam.h>
60 
61 #include <net/ethernet.h>
62 #include <net/if.h>
63 #include <net/if_var.h>
64 #include <net/if_media.h>
65 #include <net/ifq.h>
66 
67 #include <dev/pci/pcireg.h>
68 #include <dev/pci/pcivar.h>
69 
70 #include "thunder_bgx.h"
71 #include "nic_reg.h"
72 #include "nic.h"
73 #include "q_struct.h"
74 #include "nicvf_queues.h"
75 
76 #define	DEBUG
77 #undef DEBUG
78 
79 #ifdef DEBUG
80 #define	dprintf(dev, fmt, ...)	device_printf(dev, fmt, ##__VA_ARGS__)
81 #else
82 #define	dprintf(dev, fmt, ...)
83 #endif
84 
85 MALLOC_DECLARE(M_NICVF);
86 
87 static void nicvf_free_snd_queue(struct nicvf *, struct snd_queue *);
88 static int nicvf_tx_mbuf_locked(struct snd_queue *, struct mbuf *);
89 static struct mbuf * nicvf_get_rcv_mbuf(struct nicvf *, struct cqe_rx_t *);
90 static void nicvf_sq_disable(struct nicvf *, int);
91 static void nicvf_sq_enable(struct nicvf *, struct snd_queue *, int);
92 static void nicvf_put_sq_desc(struct snd_queue *, int);
93 static void nicvf_cmp_queue_config(struct nicvf *, struct queue_set *, int,
94     boolean_t);
95 static void nicvf_sq_free_used_descs(struct nicvf *, struct snd_queue *, int);
96 
97 static void nicvf_rbdr_task(void *, int);
98 static void nicvf_rbdr_task_nowait(void *, int);
99 
100 struct rbuf_info {
101 	bus_dma_tag_t	dmat;
102 	bus_dmamap_t	dmap;
103 	struct mbuf *	mbuf;
104 };
105 
106 #define GET_RBUF_INFO(x) ((struct rbuf_info *)((x) - NICVF_RCV_BUF_ALIGN_BYTES))
107 
108 /* Poll a register for a specific value */
109 static int nicvf_poll_reg(struct nicvf *nic, int qidx,
110 			  uint64_t reg, int bit_pos, int bits, int val)
111 {
112 	uint64_t bit_mask;
113 	uint64_t reg_val;
114 	int timeout = 10;
115 
116 	bit_mask = (1UL << bits) - 1;
117 	bit_mask = (bit_mask << bit_pos);
118 
119 	while (timeout) {
120 		reg_val = nicvf_queue_reg_read(nic, reg, qidx);
121 		if (((reg_val & bit_mask) >> bit_pos) == val)
122 			return (0);
123 
124 		DELAY(1000);
125 		timeout--;
126 	}
127 	device_printf(nic->dev, "Poll on reg 0x%lx failed\n", reg);
128 	return (ETIMEDOUT);
129 }
130 
131 /* Callback for bus_dmamap_load() */
132 static void
133 nicvf_dmamap_q_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
134 {
135 	bus_addr_t *paddr;
136 
137 	KASSERT(nseg == 1, ("wrong number of segments, should be 1"));
138 	paddr = arg;
139 	*paddr = segs->ds_addr;
140 }
141 
142 /* Allocate memory for a queue's descriptors */
143 static int
144 nicvf_alloc_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem,
145     int q_len, int desc_size, int align_bytes)
146 {
147 	int err, err_dmat;
148 
149 	/* Create DMA tag first */
150 	err = bus_dma_tag_create(
151 	    bus_get_dma_tag(nic->dev),		/* parent tag */
152 	    align_bytes,			/* alignment */
153 	    0,					/* boundary */
154 	    BUS_SPACE_MAXADDR,			/* lowaddr */
155 	    BUS_SPACE_MAXADDR,			/* highaddr */
156 	    NULL, NULL,				/* filtfunc, filtfuncarg */
157 	    (q_len * desc_size),		/* maxsize */
158 	    1,					/* nsegments */
159 	    (q_len * desc_size),		/* maxsegsize */
160 	    0,					/* flags */
161 	    NULL, NULL,				/* lockfunc, lockfuncarg */
162 	    &dmem->dmat);			/* dmat */
163 
164 	if (err != 0) {
165 		device_printf(nic->dev,
166 		    "Failed to create busdma tag for descriptors ring\n");
167 		return (err);
168 	}
169 
170 	/* Allocate segment of continuous DMA safe memory */
171 	err = bus_dmamem_alloc(
172 	    dmem->dmat,				/* DMA tag */
173 	    &dmem->base,			/* virtual address */
174 	    (BUS_DMA_NOWAIT | BUS_DMA_ZERO),	/* flags */
175 	    &dmem->dmap);			/* DMA map */
176 	if (err != 0) {
177 		device_printf(nic->dev, "Failed to allocate DMA safe memory for"
178 		    "descriptors ring\n");
179 		goto dmamem_fail;
180 	}
181 
182 	err = bus_dmamap_load(
183 	    dmem->dmat,
184 	    dmem->dmap,
185 	    dmem->base,
186 	    (q_len * desc_size),		/* allocation size */
187 	    nicvf_dmamap_q_cb,			/* map to DMA address cb. */
188 	    &dmem->phys_base,			/* physical address */
189 	    BUS_DMA_NOWAIT);
190 	if (err != 0) {
191 		device_printf(nic->dev,
192 		    "Cannot load DMA map of descriptors ring\n");
193 		goto dmamap_fail;
194 	}
195 
196 	dmem->q_len = q_len;
197 	dmem->size = (desc_size * q_len);
198 
199 	return (0);
200 
201 dmamap_fail:
202 	bus_dmamem_free(dmem->dmat, dmem->base, dmem->dmap);
203 	dmem->phys_base = 0;
204 dmamem_fail:
205 	err_dmat = bus_dma_tag_destroy(dmem->dmat);
206 	dmem->base = NULL;
207 	KASSERT(err_dmat == 0,
208 	    ("%s: Trying to destroy BUSY DMA tag", __func__));
209 
210 	return (err);
211 }
212 
213 /* Free queue's descriptor memory */
214 static void
215 nicvf_free_q_desc_mem(struct nicvf *nic, struct q_desc_mem *dmem)
216 {
217 	int err;
218 
219 	if ((dmem == NULL) || (dmem->base == NULL))
220 		return;
221 
222 	/* Unload a map */
223 	bus_dmamap_sync(dmem->dmat, dmem->dmap, BUS_DMASYNC_POSTREAD);
224 	bus_dmamap_unload(dmem->dmat, dmem->dmap);
225 	/* Free DMA memory */
226 	bus_dmamem_free(dmem->dmat, dmem->base, dmem->dmap);
227 	/* Destroy DMA tag */
228 	err = bus_dma_tag_destroy(dmem->dmat);
229 
230 	KASSERT(err == 0,
231 	    ("%s: Trying to destroy BUSY DMA tag", __func__));
232 
233 	dmem->phys_base = 0;
234 	dmem->base = NULL;
235 }
236 
237 /*
238  * Allocate buffer for packet reception
239  * HW returns memory address where packet is DMA'ed but not a pointer
240  * into RBDR ring, so save buffer address at the start of fragment and
241  * align the start address to a cache aligned address
242  */
243 static __inline int
244 nicvf_alloc_rcv_buffer(struct nicvf *nic, struct rbdr *rbdr,
245     bus_dmamap_t dmap, int mflags, uint32_t buf_len, bus_addr_t *rbuf)
246 {
247 	struct mbuf *mbuf;
248 	struct rbuf_info *rinfo;
249 	bus_dma_segment_t segs[1];
250 	int nsegs;
251 	int err;
252 
253 	mbuf = m_getjcl(mflags, MT_DATA, M_PKTHDR, MCLBYTES);
254 	if (mbuf == NULL)
255 		return (ENOMEM);
256 
257 	/*
258 	 * The length is equal to the actual length + one 128b line
259 	 * used as a room for rbuf_info structure.
260 	 */
261 	mbuf->m_len = mbuf->m_pkthdr.len = buf_len;
262 
263 	err = bus_dmamap_load_mbuf_sg(rbdr->rbdr_buff_dmat, dmap, mbuf, segs,
264 	    &nsegs, BUS_DMA_NOWAIT);
265 	if (err != 0) {
266 		device_printf(nic->dev,
267 		    "Failed to map mbuf into DMA visible memory, err: %d\n",
268 		    err);
269 		m_freem(mbuf);
270 		bus_dmamap_destroy(rbdr->rbdr_buff_dmat, dmap);
271 		return (err);
272 	}
273 	if (nsegs != 1)
274 		panic("Unexpected number of DMA segments for RB: %d", nsegs);
275 	/*
276 	 * Now use the room for rbuf_info structure
277 	 * and adjust mbuf data and length.
278 	 */
279 	rinfo = (struct rbuf_info *)mbuf->m_data;
280 	m_adj(mbuf, NICVF_RCV_BUF_ALIGN_BYTES);
281 
282 	rinfo->dmat = rbdr->rbdr_buff_dmat;
283 	rinfo->dmap = dmap;
284 	rinfo->mbuf = mbuf;
285 
286 	*rbuf = segs[0].ds_addr + NICVF_RCV_BUF_ALIGN_BYTES;
287 
288 	return (0);
289 }
290 
291 /* Retrieve mbuf for received packet */
292 static struct mbuf *
293 nicvf_rb_ptr_to_mbuf(struct nicvf *nic, bus_addr_t rb_ptr)
294 {
295 	struct mbuf *mbuf;
296 	struct rbuf_info *rinfo;
297 
298 	/* Get buffer start address and alignment offset */
299 	rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(rb_ptr));
300 
301 	/* Now retrieve mbuf to give to stack */
302 	mbuf = rinfo->mbuf;
303 	if (__predict_false(mbuf == NULL)) {
304 		panic("%s: Received packet fragment with NULL mbuf",
305 		    device_get_nameunit(nic->dev));
306 	}
307 	/*
308 	 * Clear the mbuf in the descriptor to indicate
309 	 * that this slot is processed and free to use.
310 	 */
311 	rinfo->mbuf = NULL;
312 
313 	bus_dmamap_sync(rinfo->dmat, rinfo->dmap, BUS_DMASYNC_POSTREAD);
314 	bus_dmamap_unload(rinfo->dmat, rinfo->dmap);
315 
316 	return (mbuf);
317 }
318 
319 /* Allocate RBDR ring and populate receive buffers */
320 static int
321 nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr, int ring_len,
322     int buf_size, int qidx)
323 {
324 	bus_dmamap_t dmap;
325 	bus_addr_t rbuf;
326 	struct rbdr_entry_t *desc;
327 	int idx;
328 	int err;
329 
330 	/* Allocate rbdr descriptors ring */
331 	err = nicvf_alloc_q_desc_mem(nic, &rbdr->dmem, ring_len,
332 	    sizeof(struct rbdr_entry_t), NICVF_RCV_BUF_ALIGN_BYTES);
333 	if (err != 0) {
334 		device_printf(nic->dev,
335 		    "Failed to create RBDR descriptors ring\n");
336 		return (err);
337 	}
338 
339 	rbdr->desc = rbdr->dmem.base;
340 	/*
341 	 * Buffer size has to be in multiples of 128 bytes.
342 	 * Make room for metadata of size of one line (128 bytes).
343 	 */
344 	rbdr->dma_size = buf_size - NICVF_RCV_BUF_ALIGN_BYTES;
345 	rbdr->enable = TRUE;
346 	rbdr->thresh = RBDR_THRESH;
347 	rbdr->nic = nic;
348 	rbdr->idx = qidx;
349 
350 	/*
351 	 * Create DMA tag for Rx buffers.
352 	 * Each map created using this tag is intended to store Rx payload for
353 	 * one fragment and one header structure containing rbuf_info (thus
354 	 * additional 128 byte line since RB must be a multiple of 128 byte
355 	 * cache line).
356 	 */
357 	if (buf_size > MCLBYTES) {
358 		device_printf(nic->dev,
359 		    "Buffer size to large for mbuf cluster\n");
360 		return (EINVAL);
361 	}
362 	err = bus_dma_tag_create(
363 	    bus_get_dma_tag(nic->dev),		/* parent tag */
364 	    NICVF_RCV_BUF_ALIGN_BYTES,		/* alignment */
365 	    0,					/* boundary */
366 	    DMAP_MAX_PHYSADDR,			/* lowaddr */
367 	    DMAP_MIN_PHYSADDR,			/* highaddr */
368 	    NULL, NULL,				/* filtfunc, filtfuncarg */
369 	    roundup2(buf_size, MCLBYTES),	/* maxsize */
370 	    1,					/* nsegments */
371 	    roundup2(buf_size, MCLBYTES),	/* maxsegsize */
372 	    0,					/* flags */
373 	    NULL, NULL,				/* lockfunc, lockfuncarg */
374 	    &rbdr->rbdr_buff_dmat);		/* dmat */
375 
376 	if (err != 0) {
377 		device_printf(nic->dev,
378 		    "Failed to create busdma tag for RBDR buffers\n");
379 		return (err);
380 	}
381 
382 	rbdr->rbdr_buff_dmaps = malloc(sizeof(*rbdr->rbdr_buff_dmaps) *
383 	    ring_len, M_NICVF, (M_WAITOK | M_ZERO));
384 
385 	for (idx = 0; idx < ring_len; idx++) {
386 		err = bus_dmamap_create(rbdr->rbdr_buff_dmat, 0, &dmap);
387 		if (err != 0) {
388 			device_printf(nic->dev,
389 			    "Failed to create DMA map for RB\n");
390 			return (err);
391 		}
392 		rbdr->rbdr_buff_dmaps[idx] = dmap;
393 
394 		err = nicvf_alloc_rcv_buffer(nic, rbdr, dmap, M_WAITOK,
395 		    DMA_BUFFER_LEN, &rbuf);
396 		if (err != 0)
397 			return (err);
398 
399 		desc = GET_RBDR_DESC(rbdr, idx);
400 		desc->buf_addr = (rbuf >> NICVF_RCV_BUF_ALIGN);
401 	}
402 
403 	/* Allocate taskqueue */
404 	TASK_INIT(&rbdr->rbdr_task, 0, nicvf_rbdr_task, rbdr);
405 	TASK_INIT(&rbdr->rbdr_task_nowait, 0, nicvf_rbdr_task_nowait, rbdr);
406 	rbdr->rbdr_taskq = taskqueue_create_fast("nicvf_rbdr_taskq", M_WAITOK,
407 	    taskqueue_thread_enqueue, &rbdr->rbdr_taskq);
408 	taskqueue_start_threads(&rbdr->rbdr_taskq, 1, PI_NET, "%s: rbdr_taskq",
409 	    device_get_nameunit(nic->dev));
410 
411 	return (0);
412 }
413 
414 /* Free RBDR ring and its receive buffers */
415 static void
416 nicvf_free_rbdr(struct nicvf *nic, struct rbdr *rbdr)
417 {
418 	struct mbuf *mbuf;
419 	struct queue_set *qs;
420 	struct rbdr_entry_t *desc;
421 	struct rbuf_info *rinfo;
422 	bus_addr_t buf_addr;
423 	int head, tail, idx;
424 	int err;
425 
426 	qs = nic->qs;
427 
428 	if ((qs == NULL) || (rbdr == NULL))
429 		return;
430 
431 	rbdr->enable = FALSE;
432 	if (rbdr->rbdr_taskq != NULL) {
433 		/* Remove tasks */
434 		while (taskqueue_cancel(rbdr->rbdr_taskq,
435 		    &rbdr->rbdr_task_nowait, NULL) != 0) {
436 			/* Finish the nowait task first */
437 			taskqueue_drain(rbdr->rbdr_taskq,
438 			    &rbdr->rbdr_task_nowait);
439 		}
440 		taskqueue_free(rbdr->rbdr_taskq);
441 		rbdr->rbdr_taskq = NULL;
442 
443 		while (taskqueue_cancel(taskqueue_thread,
444 		    &rbdr->rbdr_task, NULL) != 0) {
445 			/* Now finish the sleepable task */
446 			taskqueue_drain(taskqueue_thread, &rbdr->rbdr_task);
447 		}
448 	}
449 
450 	/*
451 	 * Free all of the memory under the RB descriptors.
452 	 * There are assumptions here:
453 	 * 1. Corresponding RBDR is disabled
454 	 *    - it is safe to operate using head and tail indexes
455 	 * 2. All bffers that were received are properly freed by
456 	 *    the receive handler
457 	 *    - there is no need to unload DMA map and free MBUF for other
458 	 *      descriptors than unused ones
459 	 */
460 	if (rbdr->rbdr_buff_dmat != NULL) {
461 		head = rbdr->head;
462 		tail = rbdr->tail;
463 		while (head != tail) {
464 			desc = GET_RBDR_DESC(rbdr, head);
465 			buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN;
466 			rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(buf_addr));
467 			bus_dmamap_unload(rbdr->rbdr_buff_dmat, rinfo->dmap);
468 			mbuf = rinfo->mbuf;
469 			/* This will destroy everything including rinfo! */
470 			m_freem(mbuf);
471 			head++;
472 			head &= (rbdr->dmem.q_len - 1);
473 		}
474 		/* Free tail descriptor */
475 		desc = GET_RBDR_DESC(rbdr, tail);
476 		buf_addr = desc->buf_addr << NICVF_RCV_BUF_ALIGN;
477 		rinfo = GET_RBUF_INFO(PHYS_TO_DMAP(buf_addr));
478 		bus_dmamap_unload(rbdr->rbdr_buff_dmat, rinfo->dmap);
479 		mbuf = rinfo->mbuf;
480 		/* This will destroy everything including rinfo! */
481 		m_freem(mbuf);
482 
483 		/* Destroy DMA maps */
484 		for (idx = 0; idx < qs->rbdr_len; idx++) {
485 			if (rbdr->rbdr_buff_dmaps[idx] == NULL)
486 				continue;
487 			err = bus_dmamap_destroy(rbdr->rbdr_buff_dmat,
488 			    rbdr->rbdr_buff_dmaps[idx]);
489 			KASSERT(err == 0,
490 			    ("%s: Could not destroy DMA map for RB, desc: %d",
491 			    __func__, idx));
492 			rbdr->rbdr_buff_dmaps[idx] = NULL;
493 		}
494 
495 		/* Now destroy the tag */
496 		err = bus_dma_tag_destroy(rbdr->rbdr_buff_dmat);
497 		KASSERT(err == 0,
498 		    ("%s: Trying to destroy BUSY DMA tag", __func__));
499 
500 		rbdr->head = 0;
501 		rbdr->tail = 0;
502 	}
503 
504 	/* Free RBDR ring */
505 	nicvf_free_q_desc_mem(nic, &rbdr->dmem);
506 }
507 
508 /*
509  * Refill receive buffer descriptors with new buffers.
510  */
511 static int
512 nicvf_refill_rbdr(struct rbdr *rbdr, int mflags)
513 {
514 	struct nicvf *nic;
515 	struct queue_set *qs;
516 	int rbdr_idx;
517 	int tail, qcount;
518 	int refill_rb_cnt;
519 	struct rbdr_entry_t *desc;
520 	bus_dmamap_t dmap;
521 	bus_addr_t rbuf;
522 	boolean_t rb_alloc_fail;
523 	int new_rb;
524 
525 	rb_alloc_fail = TRUE;
526 	new_rb = 0;
527 	nic = rbdr->nic;
528 	qs = nic->qs;
529 	rbdr_idx = rbdr->idx;
530 
531 	/* Check if it's enabled */
532 	if (!rbdr->enable)
533 		return (0);
534 
535 	/* Get no of desc's to be refilled */
536 	qcount = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, rbdr_idx);
537 	qcount &= 0x7FFFF;
538 	/* Doorbell can be ringed with a max of ring size minus 1 */
539 	if (qcount >= (qs->rbdr_len - 1)) {
540 		rb_alloc_fail = FALSE;
541 		goto out;
542 	} else
543 		refill_rb_cnt = qs->rbdr_len - qcount - 1;
544 
545 	/* Start filling descs from tail */
546 	tail = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, rbdr_idx) >> 3;
547 	while (refill_rb_cnt) {
548 		tail++;
549 		tail &= (rbdr->dmem.q_len - 1);
550 
551 		dmap = rbdr->rbdr_buff_dmaps[tail];
552 		if (nicvf_alloc_rcv_buffer(nic, rbdr, dmap, mflags,
553 		    DMA_BUFFER_LEN, &rbuf)) {
554 			/* Something went wrong. Resign */
555 			break;
556 		}
557 		desc = GET_RBDR_DESC(rbdr, tail);
558 		desc->buf_addr = (rbuf >> NICVF_RCV_BUF_ALIGN);
559 		refill_rb_cnt--;
560 		new_rb++;
561 	}
562 
563 	/* make sure all memory stores are done before ringing doorbell */
564 	wmb();
565 
566 	/* Check if buffer allocation failed */
567 	if (refill_rb_cnt == 0)
568 		rb_alloc_fail = FALSE;
569 
570 	/* Notify HW */
571 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR,
572 			      rbdr_idx, new_rb);
573 out:
574 	if (!rb_alloc_fail) {
575 		/*
576 		 * Re-enable RBDR interrupts only
577 		 * if buffer allocation is success.
578 		 */
579 		nicvf_enable_intr(nic, NICVF_INTR_RBDR, rbdr_idx);
580 
581 		return (0);
582 	}
583 
584 	return (ENOMEM);
585 }
586 
587 /* Refill RBs even if sleep is needed to reclaim memory */
588 static void
589 nicvf_rbdr_task(void *arg, int pending)
590 {
591 	struct rbdr *rbdr;
592 	int err;
593 
594 	rbdr = (struct rbdr *)arg;
595 
596 	err = nicvf_refill_rbdr(rbdr, M_WAITOK);
597 	if (__predict_false(err != 0)) {
598 		panic("%s: Failed to refill RBs even when sleep enabled",
599 		    __func__);
600 	}
601 }
602 
603 /* Refill RBs as soon as possible without waiting */
604 static void
605 nicvf_rbdr_task_nowait(void *arg, int pending)
606 {
607 	struct rbdr *rbdr;
608 	int err;
609 
610 	rbdr = (struct rbdr *)arg;
611 
612 	err = nicvf_refill_rbdr(rbdr, M_NOWAIT);
613 	if (err != 0) {
614 		/*
615 		 * Schedule another, sleepable kernel thread
616 		 * that will for sure refill the buffers.
617 		 */
618 		taskqueue_enqueue(taskqueue_thread, &rbdr->rbdr_task);
619 	}
620 }
621 
622 static int
623 nicvf_rcv_pkt_handler(struct nicvf *nic, struct cmp_queue *cq,
624     struct cqe_rx_t *cqe_rx, int cqe_type)
625 {
626 	struct mbuf *mbuf;
627 	int rq_idx;
628 	int err = 0;
629 
630 	rq_idx = cqe_rx->rq_idx;
631 
632 	/* Check for errors */
633 	err = nicvf_check_cqe_rx_errs(nic, cq, cqe_rx);
634 	if (err && !cqe_rx->rb_cnt)
635 		return (0);
636 
637 	mbuf = nicvf_get_rcv_mbuf(nic, cqe_rx);
638 	if (mbuf == NULL) {
639 		dprintf(nic->dev, "Packet not received\n");
640 		return (0);
641 	}
642 
643 	/* If error packet */
644 	if (err != 0) {
645 		m_freem(mbuf);
646 		return (0);
647 	}
648 
649 	/*
650 	 * Push this packet to the stack later to avoid
651 	 * unlocking completion task in the middle of work.
652 	 */
653 	err = buf_ring_enqueue(cq->rx_br, mbuf);
654 	if (err != 0) {
655 		/*
656 		 * Failed to enqueue this mbuf.
657 		 * We don't drop it, just schedule another task.
658 		 */
659 		return (err);
660 	}
661 
662 	return (0);
663 }
664 
665 static int
666 nicvf_snd_pkt_handler(struct nicvf *nic, struct cmp_queue *cq,
667     struct cqe_send_t *cqe_tx, int cqe_type)
668 {
669 	bus_dmamap_t dmap;
670 	struct mbuf *mbuf;
671 	struct snd_queue *sq;
672 	struct sq_hdr_subdesc *hdr;
673 
674 	mbuf = NULL;
675 	sq = &nic->qs->sq[cqe_tx->sq_idx];
676 	/* Avoid blocking here since we hold a non-sleepable NICVF_CMP_LOCK */
677 	if (NICVF_TX_TRYLOCK(sq) == 0)
678 		return (EAGAIN);
679 
680 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, cqe_tx->sqe_ptr);
681 	if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
682 		NICVF_TX_UNLOCK(sq);
683 		return (0);
684 	}
685 
686 	dprintf(nic->dev,
687 	    "%s Qset #%d SQ #%d SQ ptr #%d subdesc count %d\n",
688 	    __func__, cqe_tx->sq_qs, cqe_tx->sq_idx,
689 	    cqe_tx->sqe_ptr, hdr->subdesc_cnt);
690 
691 	dmap = (bus_dmamap_t)sq->snd_buff[cqe_tx->sqe_ptr].dmap;
692 	bus_dmamap_unload(sq->snd_buff_dmat, dmap);
693 
694 	mbuf = (struct mbuf *)sq->snd_buff[cqe_tx->sqe_ptr].mbuf;
695 	if (mbuf != NULL) {
696 		m_freem(mbuf);
697 		sq->snd_buff[cqe_tx->sqe_ptr].mbuf = NULL;
698 	}
699 
700 	nicvf_check_cqe_tx_errs(nic, cq, cqe_tx);
701 	nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
702 
703 	NICVF_TX_UNLOCK(sq);
704 	return (0);
705 }
706 
707 static int
708 nicvf_cq_intr_handler(struct nicvf *nic, uint8_t cq_idx)
709 {
710 	struct mbuf *mbuf;
711 	struct ifnet *ifp;
712 	int processed_cqe, work_done = 0, tx_done = 0;
713 	int cqe_count, cqe_head;
714 	struct queue_set *qs = nic->qs;
715 	struct cmp_queue *cq = &qs->cq[cq_idx];
716 	struct cqe_rx_t *cq_desc;
717 	int cmp_err;
718 
719 	NICVF_CMP_LOCK(cq);
720 	cmp_err = 0;
721 	processed_cqe = 0;
722 	/* Get no of valid CQ entries to process */
723 	cqe_count = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_STATUS, cq_idx);
724 	cqe_count &= CQ_CQE_COUNT;
725 	if (cqe_count == 0)
726 		goto out;
727 
728 	/* Get head of the valid CQ entries */
729 	cqe_head = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_HEAD, cq_idx) >> 9;
730 	cqe_head &= 0xFFFF;
731 
732 	dprintf(nic->dev, "%s CQ%d cqe_count %d cqe_head %d\n",
733 	    __func__, cq_idx, cqe_count, cqe_head);
734 	while (processed_cqe < cqe_count) {
735 		/* Get the CQ descriptor */
736 		cq_desc = (struct cqe_rx_t *)GET_CQ_DESC(cq, cqe_head);
737 		cqe_head++;
738 		cqe_head &= (cq->dmem.q_len - 1);
739 
740 		dprintf(nic->dev, "CQ%d cq_desc->cqe_type %d\n", cq_idx,
741 		    cq_desc->cqe_type);
742 		switch (cq_desc->cqe_type) {
743 		case CQE_TYPE_RX:
744 			cmp_err = nicvf_rcv_pkt_handler(nic, cq, cq_desc,
745 			    CQE_TYPE_RX);
746 			if (__predict_false(cmp_err != 0)) {
747 				/*
748 				 * Ups. Cannot finish now.
749 				 * Let's try again later.
750 				 */
751 				goto done;
752 			}
753 			work_done++;
754 			break;
755 		case CQE_TYPE_SEND:
756 			cmp_err = nicvf_snd_pkt_handler(nic, cq,
757 			    (void *)cq_desc, CQE_TYPE_SEND);
758 			if (__predict_false(cmp_err != 0)) {
759 				/*
760 				 * Ups. Cannot finish now.
761 				 * Let's try again later.
762 				 */
763 				goto done;
764 			}
765 
766 			tx_done++;
767 			break;
768 		case CQE_TYPE_INVALID:
769 		case CQE_TYPE_RX_SPLIT:
770 		case CQE_TYPE_RX_TCP:
771 		case CQE_TYPE_SEND_PTP:
772 			/* Ignore for now */
773 			break;
774 		}
775 		processed_cqe++;
776 	}
777 done:
778 	dprintf(nic->dev,
779 	    "%s CQ%d processed_cqe %d work_done %d\n",
780 	    __func__, cq_idx, processed_cqe, work_done);
781 
782 	/* Ring doorbell to inform H/W to reuse processed CQEs */
783 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_DOOR, cq_idx, processed_cqe);
784 
785 	if ((tx_done > 0) &&
786 	    ((if_getdrvflags(nic->ifp) & IFF_DRV_RUNNING) != 0)) {
787 		/* Reenable TXQ if its stopped earlier due to SQ full */
788 		if_setdrvflagbits(nic->ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
789 	}
790 out:
791 	NICVF_CMP_UNLOCK(cq);
792 
793 	ifp = nic->ifp;
794 	/* Push received MBUFs to the stack */
795 	while (!buf_ring_empty(cq->rx_br)) {
796 		mbuf = buf_ring_dequeue_mc(cq->rx_br);
797 		if (__predict_true(mbuf != NULL))
798 			(*ifp->if_input)(ifp, mbuf);
799 	}
800 
801 	return (cmp_err);
802 }
803 
804 /*
805  * Qset error interrupt handler
806  *
807  * As of now only CQ errors are handled
808  */
809 static void
810 nicvf_qs_err_task(void *arg, int pending)
811 {
812 	struct nicvf *nic;
813 	struct queue_set *qs;
814 	int qidx;
815 	uint64_t status;
816 	boolean_t enable = TRUE;
817 
818 	nic = (struct nicvf *)arg;
819 	qs = nic->qs;
820 
821 	/* Deactivate network interface */
822 	if_setdrvflagbits(nic->ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
823 
824 	/* Check if it is CQ err */
825 	for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
826 		status = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_STATUS,
827 		    qidx);
828 		if ((status & CQ_ERR_MASK) == 0)
829 			continue;
830 		/* Process already queued CQEs and reconfig CQ */
831 		nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx);
832 		nicvf_sq_disable(nic, qidx);
833 		(void)nicvf_cq_intr_handler(nic, qidx);
834 		nicvf_cmp_queue_config(nic, qs, qidx, enable);
835 		nicvf_sq_free_used_descs(nic, &qs->sq[qidx], qidx);
836 		nicvf_sq_enable(nic, &qs->sq[qidx], qidx);
837 		nicvf_enable_intr(nic, NICVF_INTR_CQ, qidx);
838 	}
839 
840 	if_setdrvflagbits(nic->ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
841 	/* Re-enable Qset error interrupt */
842 	nicvf_enable_intr(nic, NICVF_INTR_QS_ERR, 0);
843 }
844 
845 static void
846 nicvf_cmp_task(void *arg, int pending)
847 {
848 	uint64_t cq_head;
849 	struct cmp_queue *cq;
850 	struct nicvf *nic;
851 	int cmp_err;
852 
853 	cq = (struct cmp_queue *)arg;
854 	nic = cq->nic;
855 
856 	/* Handle CQ descriptors */
857 	cmp_err = nicvf_cq_intr_handler(nic, cq->idx);
858 	/* Re-enable interrupts */
859 	cq_head = nicvf_queue_reg_read(nic, NIC_QSET_CQ_0_7_HEAD, cq->idx);
860 	nicvf_clear_intr(nic, NICVF_INTR_CQ, cq->idx);
861 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_HEAD, cq->idx, cq_head);
862 
863 	if (__predict_false(cmp_err != 0)) {
864 		/*
865 		 * Schedule another thread here since we did not
866 		 * process the entire CQ due to Tx or Rx CQ parse error.
867 		 */
868 		taskqueue_enqueue(cq->cmp_taskq, &cq->cmp_task);
869 
870 	}
871 
872 	/* Reenable interrupt (previously disabled in nicvf_intr_handler() */
873 	nicvf_enable_intr(nic, NICVF_INTR_CQ, cq->idx);
874 
875 }
876 
877 /* Initialize completion queue */
878 static int
879 nicvf_init_cmp_queue(struct nicvf *nic, struct cmp_queue *cq, int q_len,
880     int qidx)
881 {
882 	int err;
883 
884 	/* Initizalize lock */
885 	snprintf(cq->mtx_name, sizeof(cq->mtx_name), "%s: CQ(%d) lock",
886 	    device_get_nameunit(nic->dev), qidx);
887 	mtx_init(&cq->mtx, cq->mtx_name, NULL, MTX_DEF);
888 
889 	err = nicvf_alloc_q_desc_mem(nic, &cq->dmem, q_len, CMP_QUEUE_DESC_SIZE,
890 				     NICVF_CQ_BASE_ALIGN_BYTES);
891 
892 	if (err != 0) {
893 		device_printf(nic->dev,
894 		    "Could not allocate DMA memory for CQ\n");
895 		return (err);
896 	}
897 
898 	cq->desc = cq->dmem.base;
899 	cq->thresh = CMP_QUEUE_CQE_THRESH;
900 	cq->nic = nic;
901 	cq->idx = qidx;
902 	nic->cq_coalesce_usecs = (CMP_QUEUE_TIMER_THRESH * 0.05) - 1;
903 
904 	cq->rx_br = buf_ring_alloc(CMP_QUEUE_LEN * 8, M_DEVBUF, M_WAITOK,
905 	    &cq->mtx);
906 
907 	/* Allocate taskqueue */
908 	TASK_INIT(&cq->cmp_task, 0, nicvf_cmp_task, cq);
909 	cq->cmp_taskq = taskqueue_create_fast("nicvf_cmp_taskq", M_WAITOK,
910 	    taskqueue_thread_enqueue, &cq->cmp_taskq);
911 	taskqueue_start_threads(&cq->cmp_taskq, 1, PI_NET, "%s: cmp_taskq(%d)",
912 	    device_get_nameunit(nic->dev), qidx);
913 
914 	return (0);
915 }
916 
917 static void
918 nicvf_free_cmp_queue(struct nicvf *nic, struct cmp_queue *cq)
919 {
920 
921 	if (cq == NULL)
922 		return;
923 	/*
924 	 * The completion queue itself should be disabled by now
925 	 * (ref. nicvf_snd_queue_config()).
926 	 * Ensure that it is safe to disable it or panic.
927 	 */
928 	if (cq->enable)
929 		panic("%s: Trying to free working CQ(%d)", __func__, cq->idx);
930 
931 	if (cq->cmp_taskq != NULL) {
932 		/* Remove task */
933 		while (taskqueue_cancel(cq->cmp_taskq, &cq->cmp_task, NULL) != 0)
934 			taskqueue_drain(cq->cmp_taskq, &cq->cmp_task);
935 
936 		taskqueue_free(cq->cmp_taskq);
937 		cq->cmp_taskq = NULL;
938 	}
939 	/*
940 	 * Completion interrupt will possibly enable interrupts again
941 	 * so disable interrupting now after we finished processing
942 	 * completion task. It is safe to do so since the corresponding CQ
943 	 * was already disabled.
944 	 */
945 	nicvf_disable_intr(nic, NICVF_INTR_CQ, cq->idx);
946 	nicvf_clear_intr(nic, NICVF_INTR_CQ, cq->idx);
947 
948 	NICVF_CMP_LOCK(cq);
949 	nicvf_free_q_desc_mem(nic, &cq->dmem);
950 	drbr_free(cq->rx_br, M_DEVBUF);
951 	NICVF_CMP_UNLOCK(cq);
952 	mtx_destroy(&cq->mtx);
953 	memset(cq->mtx_name, 0, sizeof(cq->mtx_name));
954 }
955 
956 static void
957 nicvf_snd_task(void *arg, int pending)
958 {
959 	struct snd_queue *sq = (struct snd_queue *)arg;
960 	struct mbuf *mbuf;
961 
962 	NICVF_TX_LOCK(sq);
963 	while (1) {
964 		mbuf = drbr_dequeue(NULL, sq->br);
965 		if (mbuf == NULL)
966 			break;
967 
968 		if (nicvf_tx_mbuf_locked(sq, mbuf) != 0) {
969 			/* XXX ARM64TODO: Increase Tx drop counter */
970 			m_freem(mbuf);
971 			break;
972 		}
973 	}
974 	NICVF_TX_UNLOCK(sq);
975 }
976 
977 /* Initialize transmit queue */
978 static int
979 nicvf_init_snd_queue(struct nicvf *nic, struct snd_queue *sq, int q_len,
980     int qidx)
981 {
982 	size_t i;
983 	int err;
984 
985 	/* Initizalize TX lock for this queue */
986 	snprintf(sq->mtx_name, sizeof(sq->mtx_name), "%s: SQ(%d) lock",
987 	    device_get_nameunit(nic->dev), qidx);
988 	mtx_init(&sq->mtx, sq->mtx_name, NULL, MTX_DEF);
989 
990 	NICVF_TX_LOCK(sq);
991 	/* Allocate buffer ring */
992 	sq->br = buf_ring_alloc(q_len / MIN_SQ_DESC_PER_PKT_XMIT, M_DEVBUF,
993 	    M_NOWAIT, &sq->mtx);
994 	if (sq->br == NULL) {
995 		device_printf(nic->dev,
996 		    "ERROR: Could not set up buf ring for SQ(%d)\n", qidx);
997 		err = ENOMEM;
998 		goto error;
999 	}
1000 
1001 	/* Allocate DMA memory for Tx descriptors */
1002 	err = nicvf_alloc_q_desc_mem(nic, &sq->dmem, q_len, SND_QUEUE_DESC_SIZE,
1003 				     NICVF_SQ_BASE_ALIGN_BYTES);
1004 	if (err != 0) {
1005 		device_printf(nic->dev,
1006 		    "Could not allocate DMA memory for SQ\n");
1007 		goto error;
1008 	}
1009 
1010 	sq->desc = sq->dmem.base;
1011 	sq->head = sq->tail = 0;
1012 	atomic_store_rel_int(&sq->free_cnt, q_len - 1);
1013 	sq->thresh = SND_QUEUE_THRESH;
1014 	sq->idx = qidx;
1015 	sq->nic = nic;
1016 
1017 	/*
1018 	 * Allocate DMA maps for Tx buffers
1019 	 */
1020 
1021 	/* Create DMA tag first */
1022 	err = bus_dma_tag_create(
1023 	    bus_get_dma_tag(nic->dev),		/* parent tag */
1024 	    1,					/* alignment */
1025 	    0,					/* boundary */
1026 	    BUS_SPACE_MAXADDR,			/* lowaddr */
1027 	    BUS_SPACE_MAXADDR,			/* highaddr */
1028 	    NULL, NULL,				/* filtfunc, filtfuncarg */
1029 	    NICVF_TXBUF_MAXSIZE,		/* maxsize */
1030 	    NICVF_TXBUF_NSEGS,			/* nsegments */
1031 	    MCLBYTES,				/* maxsegsize */
1032 	    0,					/* flags */
1033 	    NULL, NULL,				/* lockfunc, lockfuncarg */
1034 	    &sq->snd_buff_dmat);		/* dmat */
1035 
1036 	if (err != 0) {
1037 		device_printf(nic->dev,
1038 		    "Failed to create busdma tag for Tx buffers\n");
1039 		goto error;
1040 	}
1041 
1042 	/* Allocate send buffers array */
1043 	sq->snd_buff = malloc(sizeof(*sq->snd_buff) * q_len, M_NICVF,
1044 	    (M_NOWAIT | M_ZERO));
1045 	if (sq->snd_buff == NULL) {
1046 		device_printf(nic->dev,
1047 		    "Could not allocate memory for Tx buffers array\n");
1048 		err = ENOMEM;
1049 		goto error;
1050 	}
1051 
1052 	/* Now populate maps */
1053 	for (i = 0; i < q_len; i++) {
1054 		err = bus_dmamap_create(sq->snd_buff_dmat, 0,
1055 		    &sq->snd_buff[i].dmap);
1056 		if (err != 0) {
1057 			device_printf(nic->dev,
1058 			    "Failed to create DMA maps for Tx buffers\n");
1059 			goto error;
1060 		}
1061 	}
1062 	NICVF_TX_UNLOCK(sq);
1063 
1064 	/* Allocate taskqueue */
1065 	TASK_INIT(&sq->snd_task, 0, nicvf_snd_task, sq);
1066 	sq->snd_taskq = taskqueue_create_fast("nicvf_snd_taskq", M_WAITOK,
1067 	    taskqueue_thread_enqueue, &sq->snd_taskq);
1068 	taskqueue_start_threads(&sq->snd_taskq, 1, PI_NET, "%s: snd_taskq(%d)",
1069 	    device_get_nameunit(nic->dev), qidx);
1070 
1071 	return (0);
1072 error:
1073 	NICVF_TX_UNLOCK(sq);
1074 	return (err);
1075 }
1076 
1077 static void
1078 nicvf_free_snd_queue(struct nicvf *nic, struct snd_queue *sq)
1079 {
1080 	struct queue_set *qs = nic->qs;
1081 	size_t i;
1082 	int err;
1083 
1084 	if (sq == NULL)
1085 		return;
1086 
1087 	if (sq->snd_taskq != NULL) {
1088 		/* Remove task */
1089 		while (taskqueue_cancel(sq->snd_taskq, &sq->snd_task, NULL) != 0)
1090 			taskqueue_drain(sq->snd_taskq, &sq->snd_task);
1091 
1092 		taskqueue_free(sq->snd_taskq);
1093 		sq->snd_taskq = NULL;
1094 	}
1095 
1096 	NICVF_TX_LOCK(sq);
1097 	if (sq->snd_buff_dmat != NULL) {
1098 		if (sq->snd_buff != NULL) {
1099 			for (i = 0; i < qs->sq_len; i++) {
1100 				m_freem(sq->snd_buff[i].mbuf);
1101 				sq->snd_buff[i].mbuf = NULL;
1102 
1103 				bus_dmamap_unload(sq->snd_buff_dmat,
1104 				    sq->snd_buff[i].dmap);
1105 				err = bus_dmamap_destroy(sq->snd_buff_dmat,
1106 				    sq->snd_buff[i].dmap);
1107 				/*
1108 				 * If bus_dmamap_destroy fails it can cause
1109 				 * random panic later if the tag is also
1110 				 * destroyed in the process.
1111 				 */
1112 				KASSERT(err == 0,
1113 				    ("%s: Could not destroy DMA map for SQ",
1114 				    __func__));
1115 			}
1116 		}
1117 
1118 		free(sq->snd_buff, M_NICVF);
1119 
1120 		err = bus_dma_tag_destroy(sq->snd_buff_dmat);
1121 		KASSERT(err == 0,
1122 		    ("%s: Trying to destroy BUSY DMA tag", __func__));
1123 	}
1124 
1125 	/* Free private driver ring for this send queue */
1126 	if (sq->br != NULL)
1127 		drbr_free(sq->br, M_DEVBUF);
1128 
1129 	if (sq->dmem.base != NULL)
1130 		nicvf_free_q_desc_mem(nic, &sq->dmem);
1131 
1132 	NICVF_TX_UNLOCK(sq);
1133 	/* Destroy Tx lock */
1134 	mtx_destroy(&sq->mtx);
1135 	memset(sq->mtx_name, 0, sizeof(sq->mtx_name));
1136 }
1137 
1138 static void
1139 nicvf_reclaim_snd_queue(struct nicvf *nic, struct queue_set *qs, int qidx)
1140 {
1141 
1142 	/* Disable send queue */
1143 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, 0);
1144 	/* Check if SQ is stopped */
1145 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_SQ_0_7_STATUS, 21, 1, 0x01))
1146 		return;
1147 	/* Reset send queue */
1148 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
1149 }
1150 
1151 static void
1152 nicvf_reclaim_rcv_queue(struct nicvf *nic, struct queue_set *qs, int qidx)
1153 {
1154 	union nic_mbx mbx = {};
1155 
1156 	/* Make sure all packets in the pipeline are written back into mem */
1157 	mbx.msg.msg = NIC_MBOX_MSG_RQ_SW_SYNC;
1158 	nicvf_send_msg_to_pf(nic, &mbx);
1159 }
1160 
1161 static void
1162 nicvf_reclaim_cmp_queue(struct nicvf *nic, struct queue_set *qs, int qidx)
1163 {
1164 
1165 	/* Disable timer threshold (doesn't get reset upon CQ reset */
1166 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx, 0);
1167 	/* Disable completion queue */
1168 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, 0);
1169 	/* Reset completion queue */
1170 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
1171 }
1172 
1173 static void
1174 nicvf_reclaim_rbdr(struct nicvf *nic, struct rbdr *rbdr, int qidx)
1175 {
1176 	uint64_t tmp, fifo_state;
1177 	int timeout = 10;
1178 
1179 	/* Save head and tail pointers for feeing up buffers */
1180 	rbdr->head =
1181 	    nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_HEAD, qidx) >> 3;
1182 	rbdr->tail =
1183 	    nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_TAIL, qidx) >> 3;
1184 
1185 	/*
1186 	 * If RBDR FIFO is in 'FAIL' state then do a reset first
1187 	 * before relaiming.
1188 	 */
1189 	fifo_state = nicvf_queue_reg_read(nic, NIC_QSET_RBDR_0_1_STATUS0, qidx);
1190 	if (((fifo_state >> 62) & 0x03) == 0x3) {
1191 		nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG,
1192 		    qidx, NICVF_RBDR_RESET);
1193 	}
1194 
1195 	/* Disable RBDR */
1196 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0);
1197 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
1198 		return;
1199 	while (1) {
1200 		tmp = nicvf_queue_reg_read(nic,
1201 		    NIC_QSET_RBDR_0_1_PREFETCH_STATUS, qidx);
1202 		if ((tmp & 0xFFFFFFFF) == ((tmp >> 32) & 0xFFFFFFFF))
1203 			break;
1204 
1205 		DELAY(1000);
1206 		timeout--;
1207 		if (!timeout) {
1208 			device_printf(nic->dev,
1209 			    "Failed polling on prefetch status\n");
1210 			return;
1211 		}
1212 	}
1213 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx,
1214 	    NICVF_RBDR_RESET);
1215 
1216 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x02))
1217 		return;
1218 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx, 0x00);
1219 	if (nicvf_poll_reg(nic, qidx, NIC_QSET_RBDR_0_1_STATUS0, 62, 2, 0x00))
1220 		return;
1221 }
1222 
1223 /* Configures receive queue */
1224 static void
1225 nicvf_rcv_queue_config(struct nicvf *nic, struct queue_set *qs,
1226     int qidx, bool enable)
1227 {
1228 	union nic_mbx mbx = {};
1229 	struct rcv_queue *rq;
1230 	struct rq_cfg rq_cfg;
1231 
1232 	rq = &qs->rq[qidx];
1233 	rq->enable = enable;
1234 
1235 	/* Disable receive queue */
1236 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx, 0);
1237 
1238 	if (!rq->enable) {
1239 		nicvf_reclaim_rcv_queue(nic, qs, qidx);
1240 		return;
1241 	}
1242 
1243 	rq->cq_qs = qs->vnic_id;
1244 	rq->cq_idx = qidx;
1245 	rq->start_rbdr_qs = qs->vnic_id;
1246 	rq->start_qs_rbdr_idx = qs->rbdr_cnt - 1;
1247 	rq->cont_rbdr_qs = qs->vnic_id;
1248 	rq->cont_qs_rbdr_idx = qs->rbdr_cnt - 1;
1249 	/* all writes of RBDR data to be loaded into L2 Cache as well*/
1250 	rq->caching = 1;
1251 
1252 	/* Send a mailbox msg to PF to config RQ */
1253 	mbx.rq.msg = NIC_MBOX_MSG_RQ_CFG;
1254 	mbx.rq.qs_num = qs->vnic_id;
1255 	mbx.rq.rq_num = qidx;
1256 	mbx.rq.cfg = (rq->caching << 26) | (rq->cq_qs << 19) |
1257 	    (rq->cq_idx << 16) | (rq->cont_rbdr_qs << 9) |
1258 	    (rq->cont_qs_rbdr_idx << 8) | (rq->start_rbdr_qs << 1) |
1259 	    (rq->start_qs_rbdr_idx);
1260 	nicvf_send_msg_to_pf(nic, &mbx);
1261 
1262 	mbx.rq.msg = NIC_MBOX_MSG_RQ_BP_CFG;
1263 	mbx.rq.cfg = (1UL << 63) | (1UL << 62) | (qs->vnic_id << 0);
1264 	nicvf_send_msg_to_pf(nic, &mbx);
1265 
1266 	/*
1267 	 * RQ drop config
1268 	 * Enable CQ drop to reserve sufficient CQEs for all tx packets
1269 	 */
1270 	mbx.rq.msg = NIC_MBOX_MSG_RQ_DROP_CFG;
1271 	mbx.rq.cfg = (1UL << 62) | (RQ_CQ_DROP << 8);
1272 	nicvf_send_msg_to_pf(nic, &mbx);
1273 
1274 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_GEN_CFG, 0, 0x00);
1275 
1276 	/* Enable Receive queue */
1277 	rq_cfg.ena = 1;
1278 	rq_cfg.tcp_ena = 0;
1279 	nicvf_queue_reg_write(nic, NIC_QSET_RQ_0_7_CFG, qidx,
1280 	    *(uint64_t *)&rq_cfg);
1281 }
1282 
1283 /* Configures completion queue */
1284 static void
1285 nicvf_cmp_queue_config(struct nicvf *nic, struct queue_set *qs,
1286     int qidx, boolean_t enable)
1287 {
1288 	struct cmp_queue *cq;
1289 	struct cq_cfg cq_cfg;
1290 
1291 	cq = &qs->cq[qidx];
1292 	cq->enable = enable;
1293 
1294 	if (!cq->enable) {
1295 		nicvf_reclaim_cmp_queue(nic, qs, qidx);
1296 		return;
1297 	}
1298 
1299 	/* Reset completion queue */
1300 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, NICVF_CQ_RESET);
1301 
1302 	/* Set completion queue base address */
1303 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_BASE, qidx,
1304 	    (uint64_t)(cq->dmem.phys_base));
1305 
1306 	/* Enable Completion queue */
1307 	cq_cfg.ena = 1;
1308 	cq_cfg.reset = 0;
1309 	cq_cfg.caching = 0;
1310 	cq_cfg.qsize = CMP_QSIZE;
1311 	cq_cfg.avg_con = 0;
1312 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG, qidx, *(uint64_t *)&cq_cfg);
1313 
1314 	/* Set threshold value for interrupt generation */
1315 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_THRESH, qidx, cq->thresh);
1316 	nicvf_queue_reg_write(nic, NIC_QSET_CQ_0_7_CFG2, qidx,
1317 	    nic->cq_coalesce_usecs);
1318 }
1319 
1320 /* Configures transmit queue */
1321 static void
1322 nicvf_snd_queue_config(struct nicvf *nic, struct queue_set *qs, int qidx,
1323     boolean_t enable)
1324 {
1325 	union nic_mbx mbx = {};
1326 	struct snd_queue *sq;
1327 	struct sq_cfg sq_cfg;
1328 
1329 	sq = &qs->sq[qidx];
1330 	sq->enable = enable;
1331 
1332 	if (!sq->enable) {
1333 		nicvf_reclaim_snd_queue(nic, qs, qidx);
1334 		return;
1335 	}
1336 
1337 	/* Reset send queue */
1338 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, NICVF_SQ_RESET);
1339 
1340 	sq->cq_qs = qs->vnic_id;
1341 	sq->cq_idx = qidx;
1342 
1343 	/* Send a mailbox msg to PF to config SQ */
1344 	mbx.sq.msg = NIC_MBOX_MSG_SQ_CFG;
1345 	mbx.sq.qs_num = qs->vnic_id;
1346 	mbx.sq.sq_num = qidx;
1347 	mbx.sq.sqs_mode = nic->sqs_mode;
1348 	mbx.sq.cfg = (sq->cq_qs << 3) | sq->cq_idx;
1349 	nicvf_send_msg_to_pf(nic, &mbx);
1350 
1351 	/* Set queue base address */
1352 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_BASE, qidx,
1353 	    (uint64_t)(sq->dmem.phys_base));
1354 
1355 	/* Enable send queue  & set queue size */
1356 	sq_cfg.ena = 1;
1357 	sq_cfg.reset = 0;
1358 	sq_cfg.ldwb = 0;
1359 	sq_cfg.qsize = SND_QSIZE;
1360 	sq_cfg.tstmp_bgx_intf = 0;
1361 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, *(uint64_t *)&sq_cfg);
1362 
1363 	/* Set threshold value for interrupt generation */
1364 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_THRESH, qidx, sq->thresh);
1365 }
1366 
1367 /* Configures receive buffer descriptor ring */
1368 static void
1369 nicvf_rbdr_config(struct nicvf *nic, struct queue_set *qs, int qidx,
1370     boolean_t enable)
1371 {
1372 	struct rbdr *rbdr;
1373 	struct rbdr_cfg rbdr_cfg;
1374 
1375 	rbdr = &qs->rbdr[qidx];
1376 	nicvf_reclaim_rbdr(nic, rbdr, qidx);
1377 	if (!enable)
1378 		return;
1379 
1380 	/* Set descriptor base address */
1381 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_BASE, qidx,
1382 	    (uint64_t)(rbdr->dmem.phys_base));
1383 
1384 	/* Enable RBDR  & set queue size */
1385 	/* Buffer size should be in multiples of 128 bytes */
1386 	rbdr_cfg.ena = 1;
1387 	rbdr_cfg.reset = 0;
1388 	rbdr_cfg.ldwb = 0;
1389 	rbdr_cfg.qsize = RBDR_SIZE;
1390 	rbdr_cfg.avg_con = 0;
1391 	rbdr_cfg.lines = rbdr->dma_size / 128;
1392 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_CFG, qidx,
1393 	    *(uint64_t *)&rbdr_cfg);
1394 
1395 	/* Notify HW */
1396 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_DOOR, qidx,
1397 	    qs->rbdr_len - 1);
1398 
1399 	/* Set threshold value for interrupt generation */
1400 	nicvf_queue_reg_write(nic, NIC_QSET_RBDR_0_1_THRESH, qidx,
1401 	    rbdr->thresh - 1);
1402 }
1403 
1404 /* Requests PF to assign and enable Qset */
1405 void
1406 nicvf_qset_config(struct nicvf *nic, boolean_t enable)
1407 {
1408 	union nic_mbx mbx = {};
1409 	struct queue_set *qs;
1410 	struct qs_cfg *qs_cfg;
1411 
1412 	qs = nic->qs;
1413 	if (qs == NULL) {
1414 		device_printf(nic->dev,
1415 		    "Qset is still not allocated, don't init queues\n");
1416 		return;
1417 	}
1418 
1419 	qs->enable = enable;
1420 	qs->vnic_id = nic->vf_id;
1421 
1422 	/* Send a mailbox msg to PF to config Qset */
1423 	mbx.qs.msg = NIC_MBOX_MSG_QS_CFG;
1424 	mbx.qs.num = qs->vnic_id;
1425 
1426 	mbx.qs.cfg = 0;
1427 	qs_cfg = (struct qs_cfg *)&mbx.qs.cfg;
1428 	if (qs->enable) {
1429 		qs_cfg->ena = 1;
1430 		qs_cfg->vnic = qs->vnic_id;
1431 	}
1432 	nicvf_send_msg_to_pf(nic, &mbx);
1433 }
1434 
1435 static void
1436 nicvf_free_resources(struct nicvf *nic)
1437 {
1438 	int qidx;
1439 	struct queue_set *qs;
1440 
1441 	qs = nic->qs;
1442 	/*
1443 	 * Remove QS error task first since it has to be dead
1444 	 * to safely free completion queue tasks.
1445 	 */
1446 	if (qs->qs_err_taskq != NULL) {
1447 		/* Shut down QS error tasks */
1448 		while (taskqueue_cancel(qs->qs_err_taskq,
1449 		    &qs->qs_err_task,  NULL) != 0) {
1450 			taskqueue_drain(qs->qs_err_taskq, &qs->qs_err_task);
1451 
1452 		}
1453 		taskqueue_free(qs->qs_err_taskq);
1454 		qs->qs_err_taskq = NULL;
1455 	}
1456 	/* Free receive buffer descriptor ring */
1457 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1458 		nicvf_free_rbdr(nic, &qs->rbdr[qidx]);
1459 
1460 	/* Free completion queue */
1461 	for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1462 		nicvf_free_cmp_queue(nic, &qs->cq[qidx]);
1463 
1464 	/* Free send queue */
1465 	for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1466 		nicvf_free_snd_queue(nic, &qs->sq[qidx]);
1467 }
1468 
1469 static int
1470 nicvf_alloc_resources(struct nicvf *nic)
1471 {
1472 	struct queue_set *qs = nic->qs;
1473 	int qidx;
1474 
1475 	/* Alloc receive buffer descriptor ring */
1476 	for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
1477 		if (nicvf_init_rbdr(nic, &qs->rbdr[qidx], qs->rbdr_len,
1478 				    DMA_BUFFER_LEN, qidx))
1479 			goto alloc_fail;
1480 	}
1481 
1482 	/* Alloc send queue */
1483 	for (qidx = 0; qidx < qs->sq_cnt; qidx++) {
1484 		if (nicvf_init_snd_queue(nic, &qs->sq[qidx], qs->sq_len, qidx))
1485 			goto alloc_fail;
1486 	}
1487 
1488 	/* Alloc completion queue */
1489 	for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
1490 		if (nicvf_init_cmp_queue(nic, &qs->cq[qidx], qs->cq_len, qidx))
1491 			goto alloc_fail;
1492 	}
1493 
1494 	/* Allocate QS error taskqueue */
1495 	TASK_INIT(&qs->qs_err_task, 0, nicvf_qs_err_task, nic);
1496 	qs->qs_err_taskq = taskqueue_create_fast("nicvf_qs_err_taskq", M_WAITOK,
1497 	    taskqueue_thread_enqueue, &qs->qs_err_taskq);
1498 	taskqueue_start_threads(&qs->qs_err_taskq, 1, PI_NET, "%s: qs_taskq",
1499 	    device_get_nameunit(nic->dev));
1500 
1501 	return (0);
1502 alloc_fail:
1503 	nicvf_free_resources(nic);
1504 	return (ENOMEM);
1505 }
1506 
1507 int
1508 nicvf_set_qset_resources(struct nicvf *nic)
1509 {
1510 	struct queue_set *qs;
1511 
1512 	qs = malloc(sizeof(*qs), M_NICVF, (M_ZERO | M_WAITOK));
1513 	nic->qs = qs;
1514 
1515 	/* Set count of each queue */
1516 	qs->rbdr_cnt = RBDR_CNT;
1517 	/* With no RSS we stay with single RQ */
1518 	qs->rq_cnt = 1;
1519 
1520 	qs->sq_cnt = SND_QUEUE_CNT;
1521 	qs->cq_cnt = CMP_QUEUE_CNT;
1522 
1523 	/* Set queue lengths */
1524 	qs->rbdr_len = RCV_BUF_COUNT;
1525 	qs->sq_len = SND_QUEUE_LEN;
1526 	qs->cq_len = CMP_QUEUE_LEN;
1527 
1528 	nic->rx_queues = qs->rq_cnt;
1529 	nic->tx_queues = qs->sq_cnt;
1530 
1531 	return (0);
1532 }
1533 
1534 int
1535 nicvf_config_data_transfer(struct nicvf *nic, boolean_t enable)
1536 {
1537 	boolean_t disable = FALSE;
1538 	struct queue_set *qs;
1539 	int qidx;
1540 
1541 	qs = nic->qs;
1542 	if (qs == NULL)
1543 		return (0);
1544 
1545 	if (enable) {
1546 		if (nicvf_alloc_resources(nic) != 0)
1547 			return (ENOMEM);
1548 
1549 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1550 			nicvf_snd_queue_config(nic, qs, qidx, enable);
1551 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1552 			nicvf_cmp_queue_config(nic, qs, qidx, enable);
1553 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1554 			nicvf_rbdr_config(nic, qs, qidx, enable);
1555 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1556 			nicvf_rcv_queue_config(nic, qs, qidx, enable);
1557 	} else {
1558 		for (qidx = 0; qidx < qs->rq_cnt; qidx++)
1559 			nicvf_rcv_queue_config(nic, qs, qidx, disable);
1560 		for (qidx = 0; qidx < qs->rbdr_cnt; qidx++)
1561 			nicvf_rbdr_config(nic, qs, qidx, disable);
1562 		for (qidx = 0; qidx < qs->sq_cnt; qidx++)
1563 			nicvf_snd_queue_config(nic, qs, qidx, disable);
1564 		for (qidx = 0; qidx < qs->cq_cnt; qidx++)
1565 			nicvf_cmp_queue_config(nic, qs, qidx, disable);
1566 
1567 		nicvf_free_resources(nic);
1568 	}
1569 
1570 	return (0);
1571 }
1572 
1573 /*
1574  * Get a free desc from SQ
1575  * returns descriptor ponter & descriptor number
1576  */
1577 static __inline int
1578 nicvf_get_sq_desc(struct snd_queue *sq, int desc_cnt)
1579 {
1580 	int qentry;
1581 
1582 	qentry = sq->tail;
1583 	atomic_subtract_int(&sq->free_cnt, desc_cnt);
1584 	sq->tail += desc_cnt;
1585 	sq->tail &= (sq->dmem.q_len - 1);
1586 
1587 	return (qentry);
1588 }
1589 
1590 /* Free descriptor back to SQ for future use */
1591 static void
1592 nicvf_put_sq_desc(struct snd_queue *sq, int desc_cnt)
1593 {
1594 
1595 	atomic_add_int(&sq->free_cnt, desc_cnt);
1596 	sq->head += desc_cnt;
1597 	sq->head &= (sq->dmem.q_len - 1);
1598 }
1599 
1600 static __inline int
1601 nicvf_get_nxt_sqentry(struct snd_queue *sq, int qentry)
1602 {
1603 	qentry++;
1604 	qentry &= (sq->dmem.q_len - 1);
1605 	return (qentry);
1606 }
1607 
1608 static void
1609 nicvf_sq_enable(struct nicvf *nic, struct snd_queue *sq, int qidx)
1610 {
1611 	uint64_t sq_cfg;
1612 
1613 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1614 	sq_cfg |= NICVF_SQ_EN;
1615 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1616 	/* Ring doorbell so that H/W restarts processing SQEs */
1617 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_DOOR, qidx, 0);
1618 }
1619 
1620 static void
1621 nicvf_sq_disable(struct nicvf *nic, int qidx)
1622 {
1623 	uint64_t sq_cfg;
1624 
1625 	sq_cfg = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_CFG, qidx);
1626 	sq_cfg &= ~NICVF_SQ_EN;
1627 	nicvf_queue_reg_write(nic, NIC_QSET_SQ_0_7_CFG, qidx, sq_cfg);
1628 }
1629 
1630 static void
1631 nicvf_sq_free_used_descs(struct nicvf *nic, struct snd_queue *sq, int qidx)
1632 {
1633 	uint64_t head, tail;
1634 	struct snd_buff *snd_buff;
1635 	struct sq_hdr_subdesc *hdr;
1636 
1637 	NICVF_TX_LOCK(sq);
1638 	head = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_HEAD, qidx) >> 4;
1639 	tail = nicvf_queue_reg_read(nic, NIC_QSET_SQ_0_7_TAIL, qidx) >> 4;
1640 	while (sq->head != head) {
1641 		hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, sq->head);
1642 		if (hdr->subdesc_type != SQ_DESC_TYPE_HEADER) {
1643 			nicvf_put_sq_desc(sq, 1);
1644 			continue;
1645 		}
1646 		snd_buff = &sq->snd_buff[sq->head];
1647 		if (snd_buff->mbuf != NULL) {
1648 			bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap);
1649 			m_freem(snd_buff->mbuf);
1650 			sq->snd_buff[sq->head].mbuf = NULL;
1651 		}
1652 		nicvf_put_sq_desc(sq, hdr->subdesc_cnt + 1);
1653 	}
1654 	NICVF_TX_UNLOCK(sq);
1655 }
1656 
1657 /*
1658  * Add SQ HEADER subdescriptor.
1659  * First subdescriptor for every send descriptor.
1660  */
1661 static __inline void
1662 nicvf_sq_add_hdr_subdesc(struct snd_queue *sq, int qentry,
1663 			 int subdesc_cnt, struct mbuf *mbuf, int len)
1664 {
1665 	struct sq_hdr_subdesc *hdr;
1666 
1667 	hdr = (struct sq_hdr_subdesc *)GET_SQ_DESC(sq, qentry);
1668 	sq->snd_buff[qentry].mbuf = mbuf;
1669 
1670 	memset(hdr, 0, SND_QUEUE_DESC_SIZE);
1671 	hdr->subdesc_type = SQ_DESC_TYPE_HEADER;
1672 	/* Enable notification via CQE after processing SQE */
1673 	hdr->post_cqe = 1;
1674 	/* No of subdescriptors following this */
1675 	hdr->subdesc_cnt = subdesc_cnt;
1676 	hdr->tot_len = len;
1677 
1678 	/* ARM64TODO: Implement HW checksums calculation */
1679 }
1680 
1681 /*
1682  * SQ GATHER subdescriptor
1683  * Must follow HDR descriptor
1684  */
1685 static inline void nicvf_sq_add_gather_subdesc(struct snd_queue *sq, int qentry,
1686 					       int size, uint64_t data)
1687 {
1688 	struct sq_gather_subdesc *gather;
1689 
1690 	qentry &= (sq->dmem.q_len - 1);
1691 	gather = (struct sq_gather_subdesc *)GET_SQ_DESC(sq, qentry);
1692 
1693 	memset(gather, 0, SND_QUEUE_DESC_SIZE);
1694 	gather->subdesc_type = SQ_DESC_TYPE_GATHER;
1695 	gather->ld_type = NIC_SEND_LD_TYPE_E_LDD;
1696 	gather->size = size;
1697 	gather->addr = data;
1698 }
1699 
1700 /* Put an mbuf to a SQ for packet transfer. */
1701 static int
1702 nicvf_tx_mbuf_locked(struct snd_queue *sq, struct mbuf *mbuf)
1703 {
1704 	bus_dma_segment_t segs[256];
1705 	struct snd_buff *snd_buff;
1706 	size_t seg;
1707 	int nsegs, qentry;
1708 	int subdesc_cnt = MIN_SQ_DESC_PER_PKT_XMIT - 1;
1709 	int err;
1710 
1711 	NICVF_TX_LOCK_ASSERT(sq);
1712 
1713 	if (sq->free_cnt == 0)
1714 		return (ENOBUFS);
1715 
1716 	snd_buff = &sq->snd_buff[sq->tail];
1717 
1718 	err = bus_dmamap_load_mbuf_sg(sq->snd_buff_dmat, snd_buff->dmap,
1719 	    mbuf, segs, &nsegs, BUS_DMA_NOWAIT);
1720 	if (err != 0) {
1721 		/* ARM64TODO: Add mbuf defragmenting if we lack maps */
1722 		return (err);
1723 	}
1724 
1725 	/* Set how many subdescriptors is required */
1726 	subdesc_cnt += nsegs;
1727 
1728 	if (subdesc_cnt > sq->free_cnt) {
1729 		/* ARM64TODO: Add mbuf defragmentation if we lack descriptors */
1730 		bus_dmamap_unload(sq->snd_buff_dmat, snd_buff->dmap);
1731 		return (ENOBUFS);
1732 	}
1733 
1734 	qentry = nicvf_get_sq_desc(sq, subdesc_cnt);
1735 
1736 	/* Add SQ header subdesc */
1737 	nicvf_sq_add_hdr_subdesc(sq, qentry, subdesc_cnt - 1, mbuf,
1738 	    mbuf->m_pkthdr.len);
1739 
1740 	/* Add SQ gather subdescs */
1741 	for (seg = 0; seg < nsegs; seg++) {
1742 		qentry = nicvf_get_nxt_sqentry(sq, qentry);
1743 		nicvf_sq_add_gather_subdesc(sq, qentry, segs[seg].ds_len,
1744 		    segs[seg].ds_addr);
1745 	}
1746 
1747 	/* make sure all memory stores are done before ringing doorbell */
1748 	bus_dmamap_sync(sq->dmem.dmat, sq->dmem.dmap, BUS_DMASYNC_PREWRITE);
1749 
1750 	dprintf(sq->nic->dev, "%s: sq->idx: %d, subdesc_cnt: %d\n",
1751 	    __func__, sq->idx, subdesc_cnt);
1752 	/* Inform HW to xmit new packet */
1753 	nicvf_queue_reg_write(sq->nic, NIC_QSET_SQ_0_7_DOOR,
1754 	    sq->idx, subdesc_cnt);
1755 	return (0);
1756 }
1757 
1758 static __inline u_int
1759 frag_num(u_int i)
1760 {
1761 #if BYTE_ORDER == BIG_ENDIAN
1762 	return ((i & ~3) + 3 - (i & 3));
1763 #else
1764 	return (i);
1765 #endif
1766 }
1767 
1768 /* Returns MBUF for a received packet */
1769 struct mbuf *
1770 nicvf_get_rcv_mbuf(struct nicvf *nic, struct cqe_rx_t *cqe_rx)
1771 {
1772 	int frag;
1773 	int payload_len = 0;
1774 	struct mbuf *mbuf;
1775 	struct mbuf *mbuf_frag;
1776 	uint16_t *rb_lens = NULL;
1777 	uint64_t *rb_ptrs = NULL;
1778 
1779 	mbuf = NULL;
1780 	rb_lens = (uint16_t *)((uint8_t *)cqe_rx + (3 * sizeof(uint64_t)));
1781 	rb_ptrs = (uint64_t *)((uint8_t *)cqe_rx + (6 * sizeof(uint64_t)));
1782 
1783 	dprintf(nic->dev, "%s rb_cnt %d rb0_ptr %lx rb0_sz %d\n",
1784 	    __func__, cqe_rx->rb_cnt, cqe_rx->rb0_ptr, cqe_rx->rb0_sz);
1785 
1786 	for (frag = 0; frag < cqe_rx->rb_cnt; frag++) {
1787 		payload_len = rb_lens[frag_num(frag)];
1788 		if (frag == 0) {
1789 			/* First fragment */
1790 			mbuf = nicvf_rb_ptr_to_mbuf(nic,
1791 			    (*rb_ptrs - cqe_rx->align_pad));
1792 			mbuf->m_len = payload_len;
1793 			mbuf->m_data += cqe_rx->align_pad;
1794 			if_setrcvif(mbuf, nic->ifp);
1795 		} else {
1796 			/* Add fragments */
1797 			mbuf_frag = nicvf_rb_ptr_to_mbuf(nic, *rb_ptrs);
1798 			m_append(mbuf, payload_len, mbuf_frag->m_data);
1799 			m_freem(mbuf_frag);
1800 		}
1801 		/* Next buffer pointer */
1802 		rb_ptrs++;
1803 	}
1804 
1805 	if (__predict_true(mbuf != NULL)) {
1806 		m_fixhdr(mbuf);
1807 		mbuf->m_pkthdr.flowid = cqe_rx->rq_idx;
1808 		M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE);
1809 	}
1810 
1811 	return (mbuf);
1812 }
1813 
1814 /* Enable interrupt */
1815 void
1816 nicvf_enable_intr(struct nicvf *nic, int int_type, int q_idx)
1817 {
1818 	uint64_t reg_val;
1819 
1820 	reg_val = nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1821 
1822 	switch (int_type) {
1823 	case NICVF_INTR_CQ:
1824 		reg_val |= ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT);
1825 		break;
1826 	case NICVF_INTR_SQ:
1827 		reg_val |= ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT);
1828 		break;
1829 	case NICVF_INTR_RBDR:
1830 		reg_val |= ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1831 		break;
1832 	case NICVF_INTR_PKT_DROP:
1833 		reg_val |= (1UL << NICVF_INTR_PKT_DROP_SHIFT);
1834 		break;
1835 	case NICVF_INTR_TCP_TIMER:
1836 		reg_val |= (1UL << NICVF_INTR_TCP_TIMER_SHIFT);
1837 		break;
1838 	case NICVF_INTR_MBOX:
1839 		reg_val |= (1UL << NICVF_INTR_MBOX_SHIFT);
1840 		break;
1841 	case NICVF_INTR_QS_ERR:
1842 		reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT);
1843 		break;
1844 	default:
1845 		device_printf(nic->dev,
1846 			   "Failed to enable interrupt: unknown type\n");
1847 		break;
1848 	}
1849 
1850 	nicvf_reg_write(nic, NIC_VF_ENA_W1S, reg_val);
1851 }
1852 
1853 /* Disable interrupt */
1854 void
1855 nicvf_disable_intr(struct nicvf *nic, int int_type, int q_idx)
1856 {
1857 	uint64_t reg_val = 0;
1858 
1859 	switch (int_type) {
1860 	case NICVF_INTR_CQ:
1861 		reg_val |= ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT);
1862 		break;
1863 	case NICVF_INTR_SQ:
1864 		reg_val |= ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT);
1865 		break;
1866 	case NICVF_INTR_RBDR:
1867 		reg_val |= ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1868 		break;
1869 	case NICVF_INTR_PKT_DROP:
1870 		reg_val |= (1UL << NICVF_INTR_PKT_DROP_SHIFT);
1871 		break;
1872 	case NICVF_INTR_TCP_TIMER:
1873 		reg_val |= (1UL << NICVF_INTR_TCP_TIMER_SHIFT);
1874 		break;
1875 	case NICVF_INTR_MBOX:
1876 		reg_val |= (1UL << NICVF_INTR_MBOX_SHIFT);
1877 		break;
1878 	case NICVF_INTR_QS_ERR:
1879 		reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT);
1880 		break;
1881 	default:
1882 		device_printf(nic->dev,
1883 			   "Failed to disable interrupt: unknown type\n");
1884 		break;
1885 	}
1886 
1887 	nicvf_reg_write(nic, NIC_VF_ENA_W1C, reg_val);
1888 }
1889 
1890 /* Clear interrupt */
1891 void
1892 nicvf_clear_intr(struct nicvf *nic, int int_type, int q_idx)
1893 {
1894 	uint64_t reg_val = 0;
1895 
1896 	switch (int_type) {
1897 	case NICVF_INTR_CQ:
1898 		reg_val = ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT);
1899 		break;
1900 	case NICVF_INTR_SQ:
1901 		reg_val = ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT);
1902 		break;
1903 	case NICVF_INTR_RBDR:
1904 		reg_val = ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1905 		break;
1906 	case NICVF_INTR_PKT_DROP:
1907 		reg_val = (1UL << NICVF_INTR_PKT_DROP_SHIFT);
1908 		break;
1909 	case NICVF_INTR_TCP_TIMER:
1910 		reg_val = (1UL << NICVF_INTR_TCP_TIMER_SHIFT);
1911 		break;
1912 	case NICVF_INTR_MBOX:
1913 		reg_val = (1UL << NICVF_INTR_MBOX_SHIFT);
1914 		break;
1915 	case NICVF_INTR_QS_ERR:
1916 		reg_val |= (1UL << NICVF_INTR_QS_ERR_SHIFT);
1917 		break;
1918 	default:
1919 		device_printf(nic->dev,
1920 			   "Failed to clear interrupt: unknown type\n");
1921 		break;
1922 	}
1923 
1924 	nicvf_reg_write(nic, NIC_VF_INT, reg_val);
1925 }
1926 
1927 /* Check if interrupt is enabled */
1928 int
1929 nicvf_is_intr_enabled(struct nicvf *nic, int int_type, int q_idx)
1930 {
1931 	uint64_t reg_val;
1932 	uint64_t mask = 0xff;
1933 
1934 	reg_val = nicvf_reg_read(nic, NIC_VF_ENA_W1S);
1935 
1936 	switch (int_type) {
1937 	case NICVF_INTR_CQ:
1938 		mask = ((1UL << q_idx) << NICVF_INTR_CQ_SHIFT);
1939 		break;
1940 	case NICVF_INTR_SQ:
1941 		mask = ((1UL << q_idx) << NICVF_INTR_SQ_SHIFT);
1942 		break;
1943 	case NICVF_INTR_RBDR:
1944 		mask = ((1UL << q_idx) << NICVF_INTR_RBDR_SHIFT);
1945 		break;
1946 	case NICVF_INTR_PKT_DROP:
1947 		mask = NICVF_INTR_PKT_DROP_MASK;
1948 		break;
1949 	case NICVF_INTR_TCP_TIMER:
1950 		mask = NICVF_INTR_TCP_TIMER_MASK;
1951 		break;
1952 	case NICVF_INTR_MBOX:
1953 		mask = NICVF_INTR_MBOX_MASK;
1954 		break;
1955 	case NICVF_INTR_QS_ERR:
1956 		mask = NICVF_INTR_QS_ERR_MASK;
1957 		break;
1958 	default:
1959 		device_printf(nic->dev,
1960 			   "Failed to check interrupt enable: unknown type\n");
1961 		break;
1962 	}
1963 
1964 	return (reg_val & mask);
1965 }
1966 
1967 void
1968 nicvf_update_rq_stats(struct nicvf *nic, int rq_idx)
1969 {
1970 	struct rcv_queue *rq;
1971 
1972 #define GET_RQ_STATS(reg) \
1973 	nicvf_reg_read(nic, NIC_QSET_RQ_0_7_STAT_0_1 |\
1974 			    (rq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1975 
1976 	rq = &nic->qs->rq[rq_idx];
1977 	rq->stats.bytes = GET_RQ_STATS(RQ_SQ_STATS_OCTS);
1978 	rq->stats.pkts = GET_RQ_STATS(RQ_SQ_STATS_PKTS);
1979 }
1980 
1981 void
1982 nicvf_update_sq_stats(struct nicvf *nic, int sq_idx)
1983 {
1984 	struct snd_queue *sq;
1985 
1986 #define GET_SQ_STATS(reg) \
1987 	nicvf_reg_read(nic, NIC_QSET_SQ_0_7_STAT_0_1 |\
1988 			    (sq_idx << NIC_Q_NUM_SHIFT) | (reg << 3))
1989 
1990 	sq = &nic->qs->sq[sq_idx];
1991 	sq->stats.bytes = GET_SQ_STATS(RQ_SQ_STATS_OCTS);
1992 	sq->stats.pkts = GET_SQ_STATS(RQ_SQ_STATS_PKTS);
1993 }
1994 
1995 /* Check for errors in the receive cmp.queue entry */
1996 int
1997 nicvf_check_cqe_rx_errs(struct nicvf *nic, struct cmp_queue *cq,
1998     struct cqe_rx_t *cqe_rx)
1999 {
2000 	struct nicvf_hw_stats *stats = &nic->hw_stats;
2001 	struct nicvf_drv_stats *drv_stats = &nic->drv_stats;
2002 
2003 	if (!cqe_rx->err_level && !cqe_rx->err_opcode) {
2004 		drv_stats->rx_frames_ok++;
2005 		return (0);
2006 	}
2007 
2008 	switch (cqe_rx->err_opcode) {
2009 	case CQ_RX_ERROP_RE_PARTIAL:
2010 		stats->rx_bgx_truncated_pkts++;
2011 		break;
2012 	case CQ_RX_ERROP_RE_JABBER:
2013 		stats->rx_jabber_errs++;
2014 		break;
2015 	case CQ_RX_ERROP_RE_FCS:
2016 		stats->rx_fcs_errs++;
2017 		break;
2018 	case CQ_RX_ERROP_RE_RX_CTL:
2019 		stats->rx_bgx_errs++;
2020 		break;
2021 	case CQ_RX_ERROP_PREL2_ERR:
2022 		stats->rx_prel2_errs++;
2023 		break;
2024 	case CQ_RX_ERROP_L2_MAL:
2025 		stats->rx_l2_hdr_malformed++;
2026 		break;
2027 	case CQ_RX_ERROP_L2_OVERSIZE:
2028 		stats->rx_oversize++;
2029 		break;
2030 	case CQ_RX_ERROP_L2_UNDERSIZE:
2031 		stats->rx_undersize++;
2032 		break;
2033 	case CQ_RX_ERROP_L2_LENMISM:
2034 		stats->rx_l2_len_mismatch++;
2035 		break;
2036 	case CQ_RX_ERROP_L2_PCLP:
2037 		stats->rx_l2_pclp++;
2038 		break;
2039 	case CQ_RX_ERROP_IP_NOT:
2040 		stats->rx_ip_ver_errs++;
2041 		break;
2042 	case CQ_RX_ERROP_IP_CSUM_ERR:
2043 		stats->rx_ip_csum_errs++;
2044 		break;
2045 	case CQ_RX_ERROP_IP_MAL:
2046 		stats->rx_ip_hdr_malformed++;
2047 		break;
2048 	case CQ_RX_ERROP_IP_MALD:
2049 		stats->rx_ip_payload_malformed++;
2050 		break;
2051 	case CQ_RX_ERROP_IP_HOP:
2052 		stats->rx_ip_ttl_errs++;
2053 		break;
2054 	case CQ_RX_ERROP_L3_PCLP:
2055 		stats->rx_l3_pclp++;
2056 		break;
2057 	case CQ_RX_ERROP_L4_MAL:
2058 		stats->rx_l4_malformed++;
2059 		break;
2060 	case CQ_RX_ERROP_L4_CHK:
2061 		stats->rx_l4_csum_errs++;
2062 		break;
2063 	case CQ_RX_ERROP_UDP_LEN:
2064 		stats->rx_udp_len_errs++;
2065 		break;
2066 	case CQ_RX_ERROP_L4_PORT:
2067 		stats->rx_l4_port_errs++;
2068 		break;
2069 	case CQ_RX_ERROP_TCP_FLAG:
2070 		stats->rx_tcp_flag_errs++;
2071 		break;
2072 	case CQ_RX_ERROP_TCP_OFFSET:
2073 		stats->rx_tcp_offset_errs++;
2074 		break;
2075 	case CQ_RX_ERROP_L4_PCLP:
2076 		stats->rx_l4_pclp++;
2077 		break;
2078 	case CQ_RX_ERROP_RBDR_TRUNC:
2079 		stats->rx_truncated_pkts++;
2080 		break;
2081 	}
2082 
2083 	return (1);
2084 }
2085 
2086 /* Check for errors in the send cmp.queue entry */
2087 int
2088 nicvf_check_cqe_tx_errs(struct nicvf *nic, struct cmp_queue *cq,
2089     struct cqe_send_t *cqe_tx)
2090 {
2091 	struct cmp_queue_stats *stats = &cq->stats;
2092 
2093 	switch (cqe_tx->send_status) {
2094 	case CQ_TX_ERROP_GOOD:
2095 		stats->tx.good++;
2096 		return (0);
2097 	case CQ_TX_ERROP_DESC_FAULT:
2098 		stats->tx.desc_fault++;
2099 		break;
2100 	case CQ_TX_ERROP_HDR_CONS_ERR:
2101 		stats->tx.hdr_cons_err++;
2102 		break;
2103 	case CQ_TX_ERROP_SUBDC_ERR:
2104 		stats->tx.subdesc_err++;
2105 		break;
2106 	case CQ_TX_ERROP_IMM_SIZE_OFLOW:
2107 		stats->tx.imm_size_oflow++;
2108 		break;
2109 	case CQ_TX_ERROP_DATA_SEQUENCE_ERR:
2110 		stats->tx.data_seq_err++;
2111 		break;
2112 	case CQ_TX_ERROP_MEM_SEQUENCE_ERR:
2113 		stats->tx.mem_seq_err++;
2114 		break;
2115 	case CQ_TX_ERROP_LOCK_VIOL:
2116 		stats->tx.lock_viol++;
2117 		break;
2118 	case CQ_TX_ERROP_DATA_FAULT:
2119 		stats->tx.data_fault++;
2120 		break;
2121 	case CQ_TX_ERROP_TSTMP_CONFLICT:
2122 		stats->tx.tstmp_conflict++;
2123 		break;
2124 	case CQ_TX_ERROP_TSTMP_TIMEOUT:
2125 		stats->tx.tstmp_timeout++;
2126 		break;
2127 	case CQ_TX_ERROP_MEM_FAULT:
2128 		stats->tx.mem_fault++;
2129 		break;
2130 	case CQ_TX_ERROP_CK_OVERLAP:
2131 		stats->tx.csum_overlap++;
2132 		break;
2133 	case CQ_TX_ERROP_CK_OFLOW:
2134 		stats->tx.csum_overflow++;
2135 		break;
2136 	}
2137 
2138 	return (1);
2139 }
2140