xref: /linux/drivers/net/ethernet/meta/fbnic/fbnic_txrx.c (revision 23313771c7b99b3b8dba169bc71dae619d41ab56)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) Meta Platforms, Inc. and affiliates. */
3 
4 #include <linux/bitfield.h>
5 #include <linux/bpf.h>
6 #include <linux/bpf_trace.h>
7 #include <linux/iopoll.h>
8 #include <linux/pci.h>
9 #include <net/netdev_queues.h>
10 #include <net/page_pool/helpers.h>
11 #include <net/tcp.h>
12 #include <net/xdp.h>
13 
14 #include "fbnic.h"
15 #include "fbnic_csr.h"
16 #include "fbnic_netdev.h"
17 #include "fbnic_txrx.h"
18 
19 enum {
20 	FBNIC_XDP_PASS = 0,
21 	FBNIC_XDP_CONSUME,
22 	FBNIC_XDP_TX,
23 	FBNIC_XDP_LEN_ERR,
24 };
25 
26 enum {
27 	FBNIC_XMIT_CB_TS	= 0x01,
28 };
29 
30 struct fbnic_xmit_cb {
31 	u32 bytecount;
32 	u16 gso_segs;
33 	u8 desc_count;
34 	u8 flags;
35 	int hw_head;
36 };
37 
38 #define FBNIC_XMIT_CB(__skb) ((struct fbnic_xmit_cb *)((__skb)->cb))
39 
40 static u32 __iomem *fbnic_ring_csr_base(const struct fbnic_ring *ring)
41 {
42 	unsigned long csr_base = (unsigned long)ring->doorbell;
43 
44 	csr_base &= ~(FBNIC_QUEUE_STRIDE * sizeof(u32) - 1);
45 
46 	return (u32 __iomem *)csr_base;
47 }
48 
49 static u32 fbnic_ring_rd32(struct fbnic_ring *ring, unsigned int csr)
50 {
51 	u32 __iomem *csr_base = fbnic_ring_csr_base(ring);
52 
53 	return readl(csr_base + csr);
54 }
55 
56 static void fbnic_ring_wr32(struct fbnic_ring *ring, unsigned int csr, u32 val)
57 {
58 	u32 __iomem *csr_base = fbnic_ring_csr_base(ring);
59 
60 	writel(val, csr_base + csr);
61 }
62 
63 /**
64  * fbnic_ts40_to_ns() - convert descriptor timestamp to PHC time
65  * @fbn: netdev priv of the FB NIC
66  * @ts40: timestamp read from a descriptor
67  *
68  * Return: u64 value of PHC time in nanoseconds
69  *
70  * Convert truncated 40 bit device timestamp as read from a descriptor
71  * to the full PHC time in nanoseconds.
72  */
73 static __maybe_unused u64 fbnic_ts40_to_ns(struct fbnic_net *fbn, u64 ts40)
74 {
75 	unsigned int s;
76 	u64 time_ns;
77 	s64 offset;
78 	u8 ts_top;
79 	u32 high;
80 
81 	do {
82 		s = u64_stats_fetch_begin(&fbn->time_seq);
83 		offset = READ_ONCE(fbn->time_offset);
84 	} while (u64_stats_fetch_retry(&fbn->time_seq, s));
85 
86 	high = READ_ONCE(fbn->time_high);
87 
88 	/* Bits 63..40 from periodic clock reads, 39..0 from ts40 */
89 	time_ns = (u64)(high >> 8) << 40 | ts40;
90 
91 	/* Compare bits 32-39 between periodic reads and ts40,
92 	 * see if HW clock may have wrapped since last read. We are sure
93 	 * that periodic reads are always at least ~1 minute behind, so
94 	 * this logic works perfectly fine.
95 	 */
96 	ts_top = ts40 >> 32;
97 	if (ts_top < (u8)high && (u8)high - ts_top > U8_MAX / 2)
98 		time_ns += 1ULL << 40;
99 
100 	return time_ns + offset;
101 }
102 
103 static unsigned int fbnic_desc_unused(struct fbnic_ring *ring)
104 {
105 	return (ring->head - ring->tail - 1) & ring->size_mask;
106 }
107 
108 static unsigned int fbnic_desc_used(struct fbnic_ring *ring)
109 {
110 	return (ring->tail - ring->head) & ring->size_mask;
111 }
112 
113 static struct netdev_queue *txring_txq(const struct net_device *dev,
114 				       const struct fbnic_ring *ring)
115 {
116 	return netdev_get_tx_queue(dev, ring->q_idx);
117 }
118 
119 static int fbnic_maybe_stop_tx(const struct net_device *dev,
120 			       struct fbnic_ring *ring,
121 			       const unsigned int size)
122 {
123 	struct netdev_queue *txq = txring_txq(dev, ring);
124 	int res;
125 
126 	res = netif_txq_maybe_stop(txq, fbnic_desc_unused(ring), size,
127 				   FBNIC_TX_DESC_WAKEUP);
128 	if (!res) {
129 		u64_stats_update_begin(&ring->stats.syncp);
130 		ring->stats.twq.stop++;
131 		u64_stats_update_end(&ring->stats.syncp);
132 	}
133 
134 	return !res;
135 }
136 
137 static bool fbnic_tx_sent_queue(struct sk_buff *skb, struct fbnic_ring *ring)
138 {
139 	struct netdev_queue *dev_queue = txring_txq(skb->dev, ring);
140 	unsigned int bytecount = FBNIC_XMIT_CB(skb)->bytecount;
141 	bool xmit_more = netdev_xmit_more();
142 
143 	/* TBD: Request completion more often if xmit_more becomes large */
144 
145 	return __netdev_tx_sent_queue(dev_queue, bytecount, xmit_more);
146 }
147 
148 static void fbnic_unmap_single_twd(struct device *dev, __le64 *twd)
149 {
150 	u64 raw_twd = le64_to_cpu(*twd);
151 	unsigned int len;
152 	dma_addr_t dma;
153 
154 	dma = FIELD_GET(FBNIC_TWD_ADDR_MASK, raw_twd);
155 	len = FIELD_GET(FBNIC_TWD_LEN_MASK, raw_twd);
156 
157 	dma_unmap_single(dev, dma, len, DMA_TO_DEVICE);
158 }
159 
160 static void fbnic_unmap_page_twd(struct device *dev, __le64 *twd)
161 {
162 	u64 raw_twd = le64_to_cpu(*twd);
163 	unsigned int len;
164 	dma_addr_t dma;
165 
166 	dma = FIELD_GET(FBNIC_TWD_ADDR_MASK, raw_twd);
167 	len = FIELD_GET(FBNIC_TWD_LEN_MASK, raw_twd);
168 
169 	dma_unmap_page(dev, dma, len, DMA_TO_DEVICE);
170 }
171 
172 #define FBNIC_TWD_TYPE(_type) \
173 	cpu_to_le64(FIELD_PREP(FBNIC_TWD_TYPE_MASK, FBNIC_TWD_TYPE_##_type))
174 
175 static bool fbnic_tx_tstamp(struct sk_buff *skb)
176 {
177 	struct fbnic_net *fbn;
178 
179 	if (!unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
180 		return false;
181 
182 	fbn = netdev_priv(skb->dev);
183 	if (fbn->hwtstamp_config.tx_type == HWTSTAMP_TX_OFF)
184 		return false;
185 
186 	skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
187 	FBNIC_XMIT_CB(skb)->flags |= FBNIC_XMIT_CB_TS;
188 	FBNIC_XMIT_CB(skb)->hw_head = -1;
189 
190 	return true;
191 }
192 
193 static bool
194 fbnic_tx_lso(struct fbnic_ring *ring, struct sk_buff *skb,
195 	     struct skb_shared_info *shinfo, __le64 *meta,
196 	     unsigned int *l2len, unsigned int *i3len)
197 {
198 	unsigned int l3_type, l4_type, l4len, hdrlen;
199 	unsigned char *l4hdr;
200 	__be16 payload_len;
201 
202 	if (unlikely(skb_cow_head(skb, 0)))
203 		return true;
204 
205 	if (shinfo->gso_type & SKB_GSO_PARTIAL) {
206 		l3_type = FBNIC_TWD_L3_TYPE_OTHER;
207 	} else if (!skb->encapsulation) {
208 		if (ip_hdr(skb)->version == 4)
209 			l3_type = FBNIC_TWD_L3_TYPE_IPV4;
210 		else
211 			l3_type = FBNIC_TWD_L3_TYPE_IPV6;
212 	} else {
213 		unsigned int o3len;
214 
215 		o3len = skb_inner_network_header(skb) - skb_network_header(skb);
216 		*i3len -= o3len;
217 		*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_L3_OHLEN_MASK,
218 						o3len / 2));
219 		l3_type = FBNIC_TWD_L3_TYPE_V6V6;
220 	}
221 
222 	l4hdr = skb_checksum_start(skb);
223 	payload_len = cpu_to_be16(skb->len - (l4hdr - skb->data));
224 
225 	if (shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)) {
226 		struct tcphdr *tcph = (struct tcphdr *)l4hdr;
227 
228 		l4_type = FBNIC_TWD_L4_TYPE_TCP;
229 		l4len = __tcp_hdrlen((struct tcphdr *)l4hdr);
230 		csum_replace_by_diff(&tcph->check, (__force __wsum)payload_len);
231 	} else {
232 		struct udphdr *udph = (struct udphdr *)l4hdr;
233 
234 		l4_type = FBNIC_TWD_L4_TYPE_UDP;
235 		l4len = sizeof(struct udphdr);
236 		csum_replace_by_diff(&udph->check, (__force __wsum)payload_len);
237 	}
238 
239 	hdrlen = (l4hdr - skb->data) + l4len;
240 	*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_L3_TYPE_MASK, l3_type) |
241 			     FIELD_PREP(FBNIC_TWD_L4_TYPE_MASK, l4_type) |
242 			     FIELD_PREP(FBNIC_TWD_L4_HLEN_MASK, l4len / 4) |
243 			     FIELD_PREP(FBNIC_TWD_MSS_MASK, shinfo->gso_size) |
244 			     FBNIC_TWD_FLAG_REQ_LSO);
245 
246 	FBNIC_XMIT_CB(skb)->bytecount += (shinfo->gso_segs - 1) * hdrlen;
247 	FBNIC_XMIT_CB(skb)->gso_segs = shinfo->gso_segs;
248 
249 	u64_stats_update_begin(&ring->stats.syncp);
250 	ring->stats.twq.lso += shinfo->gso_segs;
251 	u64_stats_update_end(&ring->stats.syncp);
252 
253 	return false;
254 }
255 
256 static bool
257 fbnic_tx_offloads(struct fbnic_ring *ring, struct sk_buff *skb, __le64 *meta)
258 {
259 	struct skb_shared_info *shinfo = skb_shinfo(skb);
260 	unsigned int l2len, i3len;
261 
262 	if (fbnic_tx_tstamp(skb))
263 		*meta |= cpu_to_le64(FBNIC_TWD_FLAG_REQ_TS);
264 
265 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL))
266 		return false;
267 
268 	l2len = skb_mac_header_len(skb);
269 	i3len = skb_checksum_start(skb) - skb_network_header(skb);
270 
271 	*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_CSUM_OFFSET_MASK,
272 					skb->csum_offset / 2));
273 
274 	if (shinfo->gso_size) {
275 		if (fbnic_tx_lso(ring, skb, shinfo, meta, &l2len, &i3len))
276 			return true;
277 	} else {
278 		*meta |= cpu_to_le64(FBNIC_TWD_FLAG_REQ_CSO);
279 		u64_stats_update_begin(&ring->stats.syncp);
280 		ring->stats.twq.csum_partial++;
281 		u64_stats_update_end(&ring->stats.syncp);
282 	}
283 
284 	*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_L2_HLEN_MASK, l2len / 2) |
285 			     FIELD_PREP(FBNIC_TWD_L3_IHLEN_MASK, i3len / 2));
286 	return false;
287 }
288 
289 static void
290 fbnic_rx_csum(u64 rcd, struct sk_buff *skb, struct fbnic_ring *rcq,
291 	      u64 *csum_cmpl, u64 *csum_none)
292 {
293 	skb_checksum_none_assert(skb);
294 
295 	if (unlikely(!(skb->dev->features & NETIF_F_RXCSUM))) {
296 		(*csum_none)++;
297 		return;
298 	}
299 
300 	if (FIELD_GET(FBNIC_RCD_META_L4_CSUM_UNNECESSARY, rcd)) {
301 		skb->ip_summed = CHECKSUM_UNNECESSARY;
302 	} else {
303 		u16 csum = FIELD_GET(FBNIC_RCD_META_L2_CSUM_MASK, rcd);
304 
305 		skb->ip_summed = CHECKSUM_COMPLETE;
306 		skb->csum = (__force __wsum)csum;
307 		(*csum_cmpl)++;
308 	}
309 }
310 
311 static bool
312 fbnic_tx_map(struct fbnic_ring *ring, struct sk_buff *skb, __le64 *meta)
313 {
314 	struct device *dev = skb->dev->dev.parent;
315 	unsigned int tail = ring->tail, first;
316 	unsigned int size, data_len;
317 	skb_frag_t *frag;
318 	dma_addr_t dma;
319 	__le64 *twd;
320 
321 	ring->tx_buf[tail] = skb;
322 
323 	tail++;
324 	tail &= ring->size_mask;
325 	first = tail;
326 
327 	size = skb_headlen(skb);
328 	data_len = skb->data_len;
329 
330 	if (size > FIELD_MAX(FBNIC_TWD_LEN_MASK))
331 		goto dma_error;
332 
333 	dma = dma_map_single(dev, skb->data, size, DMA_TO_DEVICE);
334 
335 	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
336 		twd = &ring->desc[tail];
337 
338 		if (dma_mapping_error(dev, dma))
339 			goto dma_error;
340 
341 		*twd = cpu_to_le64(FIELD_PREP(FBNIC_TWD_ADDR_MASK, dma) |
342 				   FIELD_PREP(FBNIC_TWD_LEN_MASK, size) |
343 				   FIELD_PREP(FBNIC_TWD_TYPE_MASK,
344 					      FBNIC_TWD_TYPE_AL));
345 
346 		tail++;
347 		tail &= ring->size_mask;
348 
349 		if (!data_len)
350 			break;
351 
352 		size = skb_frag_size(frag);
353 		data_len -= size;
354 
355 		if (size > FIELD_MAX(FBNIC_TWD_LEN_MASK))
356 			goto dma_error;
357 
358 		dma = skb_frag_dma_map(dev, frag, 0, size, DMA_TO_DEVICE);
359 	}
360 
361 	*twd |= FBNIC_TWD_TYPE(LAST_AL);
362 
363 	FBNIC_XMIT_CB(skb)->desc_count = ((twd - meta) + 1) & ring->size_mask;
364 
365 	ring->tail = tail;
366 
367 	/* Record SW timestamp */
368 	skb_tx_timestamp(skb);
369 
370 	/* Verify there is room for another packet */
371 	fbnic_maybe_stop_tx(skb->dev, ring, FBNIC_MAX_SKB_DESC);
372 
373 	if (fbnic_tx_sent_queue(skb, ring)) {
374 		*meta |= cpu_to_le64(FBNIC_TWD_FLAG_REQ_COMPLETION);
375 
376 		/* Force DMA writes to flush before writing to tail */
377 		dma_wmb();
378 
379 		writel(tail, ring->doorbell);
380 	}
381 
382 	return false;
383 dma_error:
384 	if (net_ratelimit())
385 		netdev_err(skb->dev, "TX DMA map failed\n");
386 
387 	while (tail != first) {
388 		tail--;
389 		tail &= ring->size_mask;
390 		twd = &ring->desc[tail];
391 		if (tail == first)
392 			fbnic_unmap_single_twd(dev, twd);
393 		else
394 			fbnic_unmap_page_twd(dev, twd);
395 	}
396 
397 	return true;
398 }
399 
400 #define FBNIC_MIN_FRAME_LEN	60
401 
402 static netdev_tx_t
403 fbnic_xmit_frame_ring(struct sk_buff *skb, struct fbnic_ring *ring)
404 {
405 	__le64 *meta = &ring->desc[ring->tail];
406 	u16 desc_needed;
407 
408 	if (skb_put_padto(skb, FBNIC_MIN_FRAME_LEN))
409 		goto err_count;
410 
411 	/* Need: 1 descriptor per page,
412 	 *       + 1 desc for skb_head,
413 	 *       + 2 desc for metadata and timestamp metadata
414 	 *       + 7 desc gap to keep tail from touching head
415 	 * otherwise try next time
416 	 */
417 	desc_needed = skb_shinfo(skb)->nr_frags + 10;
418 	if (fbnic_maybe_stop_tx(skb->dev, ring, desc_needed))
419 		return NETDEV_TX_BUSY;
420 
421 	*meta = cpu_to_le64(FBNIC_TWD_FLAG_DEST_MAC);
422 
423 	/* Write all members within DWORD to condense this into 2 4B writes */
424 	FBNIC_XMIT_CB(skb)->bytecount = skb->len;
425 	FBNIC_XMIT_CB(skb)->gso_segs = 1;
426 	FBNIC_XMIT_CB(skb)->desc_count = 0;
427 	FBNIC_XMIT_CB(skb)->flags = 0;
428 
429 	if (fbnic_tx_offloads(ring, skb, meta))
430 		goto err_free;
431 
432 	if (fbnic_tx_map(ring, skb, meta))
433 		goto err_free;
434 
435 	return NETDEV_TX_OK;
436 
437 err_free:
438 	dev_kfree_skb_any(skb);
439 err_count:
440 	u64_stats_update_begin(&ring->stats.syncp);
441 	ring->stats.dropped++;
442 	u64_stats_update_end(&ring->stats.syncp);
443 	return NETDEV_TX_OK;
444 }
445 
446 netdev_tx_t fbnic_xmit_frame(struct sk_buff *skb, struct net_device *dev)
447 {
448 	struct fbnic_net *fbn = netdev_priv(dev);
449 	unsigned int q_map = skb->queue_mapping;
450 
451 	return fbnic_xmit_frame_ring(skb, fbn->tx[q_map]);
452 }
453 
454 static netdev_features_t
455 fbnic_features_check_encap_gso(struct sk_buff *skb, struct net_device *dev,
456 			       netdev_features_t features, unsigned int l3len)
457 {
458 	netdev_features_t skb_gso_features;
459 	struct ipv6hdr *ip6_hdr;
460 	unsigned char l4_hdr;
461 	unsigned int start;
462 	__be16 frag_off;
463 
464 	/* Require MANGLEID for GSO_PARTIAL of IPv4.
465 	 * In theory we could support TSO with single, innermost v4 header
466 	 * by pretending everything before it is L2, but that needs to be
467 	 * parsed case by case.. so leaving it for when the need arises.
468 	 */
469 	if (!(features & NETIF_F_TSO_MANGLEID))
470 		features &= ~NETIF_F_TSO;
471 
472 	skb_gso_features = skb_shinfo(skb)->gso_type;
473 	skb_gso_features <<= NETIF_F_GSO_SHIFT;
474 
475 	/* We'd only clear the native GSO features, so don't bother validating
476 	 * if the match can only be on those supported thru GSO_PARTIAL.
477 	 */
478 	if (!(skb_gso_features & FBNIC_TUN_GSO_FEATURES))
479 		return features;
480 
481 	/* We can only do IPv6-in-IPv6, not v4-in-v6. It'd be nice
482 	 * to fall back to partial for this, or any failure below.
483 	 * This is just an optimization, UDPv4 will be caught later on.
484 	 */
485 	if (skb_gso_features & NETIF_F_TSO)
486 		return features & ~FBNIC_TUN_GSO_FEATURES;
487 
488 	/* Inner headers multiple of 2 */
489 	if ((skb_inner_network_header(skb) - skb_network_header(skb)) % 2)
490 		return features & ~FBNIC_TUN_GSO_FEATURES;
491 
492 	/* Encapsulated GSO packet, make 100% sure it's IPv6-in-IPv6. */
493 	ip6_hdr = ipv6_hdr(skb);
494 	if (ip6_hdr->version != 6)
495 		return features & ~FBNIC_TUN_GSO_FEATURES;
496 
497 	l4_hdr = ip6_hdr->nexthdr;
498 	start = (unsigned char *)ip6_hdr - skb->data + sizeof(struct ipv6hdr);
499 	start = ipv6_skip_exthdr(skb, start, &l4_hdr, &frag_off);
500 	if (frag_off || l4_hdr != IPPROTO_IPV6 ||
501 	    skb->data + start != skb_inner_network_header(skb))
502 		return features & ~FBNIC_TUN_GSO_FEATURES;
503 
504 	return features;
505 }
506 
507 netdev_features_t
508 fbnic_features_check(struct sk_buff *skb, struct net_device *dev,
509 		     netdev_features_t features)
510 {
511 	unsigned int l2len, l3len;
512 
513 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL))
514 		return features;
515 
516 	l2len = skb_mac_header_len(skb);
517 	l3len = skb_checksum_start(skb) - skb_network_header(skb);
518 
519 	/* Check header lengths are multiple of 2.
520 	 * In case of 6in6 we support longer headers (IHLEN + OHLEN)
521 	 * but keep things simple for now, 512B is plenty.
522 	 */
523 	if ((l2len | l3len | skb->csum_offset) % 2 ||
524 	    !FIELD_FIT(FBNIC_TWD_L2_HLEN_MASK, l2len / 2) ||
525 	    !FIELD_FIT(FBNIC_TWD_L3_IHLEN_MASK, l3len / 2) ||
526 	    !FIELD_FIT(FBNIC_TWD_CSUM_OFFSET_MASK, skb->csum_offset / 2))
527 		return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
528 
529 	if (likely(!skb->encapsulation) || !skb_is_gso(skb))
530 		return features;
531 
532 	return fbnic_features_check_encap_gso(skb, dev, features, l3len);
533 }
534 
535 static void fbnic_clean_twq0(struct fbnic_napi_vector *nv, int napi_budget,
536 			     struct fbnic_ring *ring, bool discard,
537 			     unsigned int hw_head)
538 {
539 	u64 total_bytes = 0, total_packets = 0, ts_lost = 0;
540 	unsigned int head = ring->head;
541 	struct netdev_queue *txq;
542 	unsigned int clean_desc;
543 
544 	clean_desc = (hw_head - head) & ring->size_mask;
545 
546 	while (clean_desc) {
547 		struct sk_buff *skb = ring->tx_buf[head];
548 		unsigned int desc_cnt;
549 
550 		desc_cnt = FBNIC_XMIT_CB(skb)->desc_count;
551 		if (desc_cnt > clean_desc)
552 			break;
553 
554 		if (unlikely(FBNIC_XMIT_CB(skb)->flags & FBNIC_XMIT_CB_TS)) {
555 			FBNIC_XMIT_CB(skb)->hw_head = hw_head;
556 			if (likely(!discard))
557 				break;
558 			ts_lost++;
559 		}
560 
561 		ring->tx_buf[head] = NULL;
562 
563 		clean_desc -= desc_cnt;
564 
565 		while (!(ring->desc[head] & FBNIC_TWD_TYPE(AL))) {
566 			head++;
567 			head &= ring->size_mask;
568 			desc_cnt--;
569 		}
570 
571 		fbnic_unmap_single_twd(nv->dev, &ring->desc[head]);
572 		head++;
573 		head &= ring->size_mask;
574 		desc_cnt--;
575 
576 		while (desc_cnt--) {
577 			fbnic_unmap_page_twd(nv->dev, &ring->desc[head]);
578 			head++;
579 			head &= ring->size_mask;
580 		}
581 
582 		total_bytes += FBNIC_XMIT_CB(skb)->bytecount;
583 		total_packets += FBNIC_XMIT_CB(skb)->gso_segs;
584 
585 		napi_consume_skb(skb, napi_budget);
586 	}
587 
588 	if (!total_bytes)
589 		return;
590 
591 	ring->head = head;
592 
593 	txq = txring_txq(nv->napi.dev, ring);
594 
595 	if (unlikely(discard)) {
596 		u64_stats_update_begin(&ring->stats.syncp);
597 		ring->stats.dropped += total_packets;
598 		ring->stats.twq.ts_lost += ts_lost;
599 		u64_stats_update_end(&ring->stats.syncp);
600 
601 		netdev_tx_completed_queue(txq, total_packets, total_bytes);
602 		return;
603 	}
604 
605 	u64_stats_update_begin(&ring->stats.syncp);
606 	ring->stats.bytes += total_bytes;
607 	ring->stats.packets += total_packets;
608 	u64_stats_update_end(&ring->stats.syncp);
609 
610 	if (!netif_txq_completed_wake(txq, total_packets, total_bytes,
611 				      fbnic_desc_unused(ring),
612 				      FBNIC_TX_DESC_WAKEUP)) {
613 		u64_stats_update_begin(&ring->stats.syncp);
614 		ring->stats.twq.wake++;
615 		u64_stats_update_end(&ring->stats.syncp);
616 	}
617 }
618 
619 static void fbnic_clean_twq1(struct fbnic_napi_vector *nv, bool pp_allow_direct,
620 			     struct fbnic_ring *ring, bool discard,
621 			     unsigned int hw_head)
622 {
623 	u64 total_bytes = 0, total_packets = 0;
624 	unsigned int head = ring->head;
625 
626 	while (hw_head != head) {
627 		struct page *page;
628 		u64 twd;
629 
630 		if (unlikely(!(ring->desc[head] & FBNIC_TWD_TYPE(AL))))
631 			goto next_desc;
632 
633 		twd = le64_to_cpu(ring->desc[head]);
634 		page = ring->tx_buf[head];
635 
636 		/* TYPE_AL is 2, TYPE_LAST_AL is 3. So this trick gives
637 		 * us one increment per packet, with no branches.
638 		 */
639 		total_packets += FIELD_GET(FBNIC_TWD_TYPE_MASK, twd) -
640 				 FBNIC_TWD_TYPE_AL;
641 		total_bytes += FIELD_GET(FBNIC_TWD_LEN_MASK, twd);
642 
643 		page_pool_put_page(nv->page_pool, page, -1, pp_allow_direct);
644 next_desc:
645 		head++;
646 		head &= ring->size_mask;
647 	}
648 
649 	if (!total_bytes)
650 		return;
651 
652 	ring->head = head;
653 
654 	if (discard) {
655 		u64_stats_update_begin(&ring->stats.syncp);
656 		ring->stats.dropped += total_packets;
657 		u64_stats_update_end(&ring->stats.syncp);
658 		return;
659 	}
660 
661 	u64_stats_update_begin(&ring->stats.syncp);
662 	ring->stats.bytes += total_bytes;
663 	ring->stats.packets += total_packets;
664 	u64_stats_update_end(&ring->stats.syncp);
665 }
666 
667 static void fbnic_clean_tsq(struct fbnic_napi_vector *nv,
668 			    struct fbnic_ring *ring,
669 			    u64 tcd, int *ts_head, int *head0)
670 {
671 	struct skb_shared_hwtstamps hwtstamp;
672 	struct fbnic_net *fbn;
673 	struct sk_buff *skb;
674 	int head;
675 	u64 ns;
676 
677 	head = (*ts_head < 0) ? ring->head : *ts_head;
678 
679 	do {
680 		unsigned int desc_cnt;
681 
682 		if (head == ring->tail) {
683 			if (unlikely(net_ratelimit()))
684 				netdev_err(nv->napi.dev,
685 					   "Tx timestamp without matching packet\n");
686 			return;
687 		}
688 
689 		skb = ring->tx_buf[head];
690 		desc_cnt = FBNIC_XMIT_CB(skb)->desc_count;
691 
692 		head += desc_cnt;
693 		head &= ring->size_mask;
694 	} while (!(FBNIC_XMIT_CB(skb)->flags & FBNIC_XMIT_CB_TS));
695 
696 	fbn = netdev_priv(nv->napi.dev);
697 	ns = fbnic_ts40_to_ns(fbn, FIELD_GET(FBNIC_TCD_TYPE1_TS_MASK, tcd));
698 
699 	memset(&hwtstamp, 0, sizeof(hwtstamp));
700 	hwtstamp.hwtstamp = ns_to_ktime(ns);
701 
702 	*ts_head = head;
703 
704 	FBNIC_XMIT_CB(skb)->flags &= ~FBNIC_XMIT_CB_TS;
705 	if (*head0 < 0) {
706 		head = FBNIC_XMIT_CB(skb)->hw_head;
707 		if (head >= 0)
708 			*head0 = head;
709 	}
710 
711 	skb_tstamp_tx(skb, &hwtstamp);
712 	u64_stats_update_begin(&ring->stats.syncp);
713 	ring->stats.twq.ts_packets++;
714 	u64_stats_update_end(&ring->stats.syncp);
715 }
716 
717 static void fbnic_page_pool_init(struct fbnic_ring *ring, unsigned int idx,
718 				 struct page *page)
719 {
720 	struct fbnic_rx_buf *rx_buf = &ring->rx_buf[idx];
721 
722 	page_pool_fragment_page(page, FBNIC_PAGECNT_BIAS_MAX);
723 	rx_buf->pagecnt_bias = FBNIC_PAGECNT_BIAS_MAX;
724 	rx_buf->page = page;
725 }
726 
727 static struct page *fbnic_page_pool_get(struct fbnic_ring *ring,
728 					unsigned int idx)
729 {
730 	struct fbnic_rx_buf *rx_buf = &ring->rx_buf[idx];
731 
732 	rx_buf->pagecnt_bias--;
733 
734 	return rx_buf->page;
735 }
736 
737 static void fbnic_page_pool_drain(struct fbnic_ring *ring, unsigned int idx,
738 				  struct fbnic_napi_vector *nv, int budget)
739 {
740 	struct fbnic_rx_buf *rx_buf = &ring->rx_buf[idx];
741 	struct page *page = rx_buf->page;
742 
743 	if (!page_pool_unref_page(page, rx_buf->pagecnt_bias))
744 		page_pool_put_unrefed_page(nv->page_pool, page, -1, !!budget);
745 
746 	rx_buf->page = NULL;
747 }
748 
749 static void fbnic_clean_twq(struct fbnic_napi_vector *nv, int napi_budget,
750 			    struct fbnic_q_triad *qt, s32 ts_head, s32 head0,
751 			    s32 head1)
752 {
753 	if (head0 >= 0)
754 		fbnic_clean_twq0(nv, napi_budget, &qt->sub0, false, head0);
755 	else if (ts_head >= 0)
756 		fbnic_clean_twq0(nv, napi_budget, &qt->sub0, false, ts_head);
757 
758 	if (head1 >= 0) {
759 		qt->cmpl.deferred_head = -1;
760 		if (napi_budget)
761 			fbnic_clean_twq1(nv, true, &qt->sub1, false, head1);
762 		else
763 			qt->cmpl.deferred_head = head1;
764 	}
765 }
766 
767 static void
768 fbnic_clean_tcq(struct fbnic_napi_vector *nv, struct fbnic_q_triad *qt,
769 		int napi_budget)
770 {
771 	struct fbnic_ring *cmpl = &qt->cmpl;
772 	s32 head1 = cmpl->deferred_head;
773 	s32 head0 = -1, ts_head = -1;
774 	__le64 *raw_tcd, done;
775 	u32 head = cmpl->head;
776 
777 	done = (head & (cmpl->size_mask + 1)) ? 0 : cpu_to_le64(FBNIC_TCD_DONE);
778 	raw_tcd = &cmpl->desc[head & cmpl->size_mask];
779 
780 	/* Walk the completion queue collecting the heads reported by NIC */
781 	while ((*raw_tcd & cpu_to_le64(FBNIC_TCD_DONE)) == done) {
782 		u64 tcd;
783 
784 		dma_rmb();
785 
786 		tcd = le64_to_cpu(*raw_tcd);
787 
788 		switch (FIELD_GET(FBNIC_TCD_TYPE_MASK, tcd)) {
789 		case FBNIC_TCD_TYPE_0:
790 			if (tcd & FBNIC_TCD_TWQ1)
791 				head1 = FIELD_GET(FBNIC_TCD_TYPE0_HEAD1_MASK,
792 						  tcd);
793 			else
794 				head0 = FIELD_GET(FBNIC_TCD_TYPE0_HEAD0_MASK,
795 						  tcd);
796 			/* Currently all err status bits are related to
797 			 * timestamps and as those have yet to be added
798 			 * they are skipped for now.
799 			 */
800 			break;
801 		case FBNIC_TCD_TYPE_1:
802 			if (WARN_ON_ONCE(tcd & FBNIC_TCD_TWQ1))
803 				break;
804 
805 			fbnic_clean_tsq(nv, &qt->sub0, tcd, &ts_head, &head0);
806 			break;
807 		default:
808 			break;
809 		}
810 
811 		raw_tcd++;
812 		head++;
813 		if (!(head & cmpl->size_mask)) {
814 			done ^= cpu_to_le64(FBNIC_TCD_DONE);
815 			raw_tcd = &cmpl->desc[0];
816 		}
817 	}
818 
819 	/* Record the current head/tail of the queue */
820 	if (cmpl->head != head) {
821 		cmpl->head = head;
822 		writel(head & cmpl->size_mask, cmpl->doorbell);
823 	}
824 
825 	/* Unmap and free processed buffers */
826 	fbnic_clean_twq(nv, napi_budget, qt, ts_head, head0, head1);
827 }
828 
829 static void fbnic_clean_bdq(struct fbnic_napi_vector *nv, int napi_budget,
830 			    struct fbnic_ring *ring, unsigned int hw_head)
831 {
832 	unsigned int head = ring->head;
833 
834 	if (head == hw_head)
835 		return;
836 
837 	do {
838 		fbnic_page_pool_drain(ring, head, nv, napi_budget);
839 
840 		head++;
841 		head &= ring->size_mask;
842 	} while (head != hw_head);
843 
844 	ring->head = head;
845 }
846 
847 static void fbnic_bd_prep(struct fbnic_ring *bdq, u16 id, struct page *page)
848 {
849 	__le64 *bdq_desc = &bdq->desc[id * FBNIC_BD_FRAG_COUNT];
850 	dma_addr_t dma = page_pool_get_dma_addr(page);
851 	u64 bd, i = FBNIC_BD_FRAG_COUNT;
852 
853 	bd = (FBNIC_BD_PAGE_ADDR_MASK & dma) |
854 	     FIELD_PREP(FBNIC_BD_PAGE_ID_MASK, id);
855 
856 	/* In the case that a page size is larger than 4K we will map a
857 	 * single page to multiple fragments. The fragments will be
858 	 * FBNIC_BD_FRAG_COUNT in size and the lower n bits will be use
859 	 * to indicate the individual fragment IDs.
860 	 */
861 	do {
862 		*bdq_desc = cpu_to_le64(bd);
863 		bd += FIELD_PREP(FBNIC_BD_DESC_ADDR_MASK, 1) |
864 		      FIELD_PREP(FBNIC_BD_DESC_ID_MASK, 1);
865 	} while (--i);
866 }
867 
868 static void fbnic_fill_bdq(struct fbnic_napi_vector *nv, struct fbnic_ring *bdq)
869 {
870 	unsigned int count = fbnic_desc_unused(bdq);
871 	unsigned int i = bdq->tail;
872 
873 	if (!count)
874 		return;
875 
876 	do {
877 		struct page *page;
878 
879 		page = page_pool_dev_alloc_pages(nv->page_pool);
880 		if (!page) {
881 			u64_stats_update_begin(&bdq->stats.syncp);
882 			bdq->stats.rx.alloc_failed++;
883 			u64_stats_update_end(&bdq->stats.syncp);
884 
885 			break;
886 		}
887 
888 		fbnic_page_pool_init(bdq, i, page);
889 		fbnic_bd_prep(bdq, i, page);
890 
891 		i++;
892 		i &= bdq->size_mask;
893 
894 		count--;
895 	} while (count);
896 
897 	if (bdq->tail != i) {
898 		bdq->tail = i;
899 
900 		/* Force DMA writes to flush before writing to tail */
901 		dma_wmb();
902 
903 		writel(i, bdq->doorbell);
904 	}
905 }
906 
907 static unsigned int fbnic_hdr_pg_start(unsigned int pg_off)
908 {
909 	/* The headroom of the first header may be larger than FBNIC_RX_HROOM
910 	 * due to alignment. So account for that by just making the page
911 	 * offset 0 if we are starting at the first header.
912 	 */
913 	if (ALIGN(FBNIC_RX_HROOM, 128) > FBNIC_RX_HROOM &&
914 	    pg_off == ALIGN(FBNIC_RX_HROOM, 128))
915 		return 0;
916 
917 	return pg_off - FBNIC_RX_HROOM;
918 }
919 
920 static unsigned int fbnic_hdr_pg_end(unsigned int pg_off, unsigned int len)
921 {
922 	/* Determine the end of the buffer by finding the start of the next
923 	 * and then subtracting the headroom from that frame.
924 	 */
925 	pg_off += len + FBNIC_RX_TROOM + FBNIC_RX_HROOM;
926 
927 	return ALIGN(pg_off, 128) - FBNIC_RX_HROOM;
928 }
929 
930 static void fbnic_pkt_prepare(struct fbnic_napi_vector *nv, u64 rcd,
931 			      struct fbnic_pkt_buff *pkt,
932 			      struct fbnic_q_triad *qt)
933 {
934 	unsigned int hdr_pg_idx = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
935 	unsigned int hdr_pg_off = FIELD_GET(FBNIC_RCD_AL_BUFF_OFF_MASK, rcd);
936 	struct page *page = fbnic_page_pool_get(&qt->sub0, hdr_pg_idx);
937 	unsigned int len = FIELD_GET(FBNIC_RCD_AL_BUFF_LEN_MASK, rcd);
938 	unsigned int frame_sz, hdr_pg_start, hdr_pg_end, headroom;
939 	unsigned char *hdr_start;
940 
941 	/* data_hard_start should always be NULL when this is called */
942 	WARN_ON_ONCE(pkt->buff.data_hard_start);
943 
944 	/* Short-cut the end calculation if we know page is fully consumed */
945 	hdr_pg_end = FIELD_GET(FBNIC_RCD_AL_PAGE_FIN, rcd) ?
946 		     FBNIC_BD_FRAG_SIZE : fbnic_hdr_pg_end(hdr_pg_off, len);
947 	hdr_pg_start = fbnic_hdr_pg_start(hdr_pg_off);
948 
949 	headroom = hdr_pg_off - hdr_pg_start + FBNIC_RX_PAD;
950 	frame_sz = hdr_pg_end - hdr_pg_start;
951 	xdp_init_buff(&pkt->buff, frame_sz, &qt->xdp_rxq);
952 	hdr_pg_start += (FBNIC_RCD_AL_BUFF_FRAG_MASK & rcd) *
953 			FBNIC_BD_FRAG_SIZE;
954 
955 	/* Sync DMA buffer */
956 	dma_sync_single_range_for_cpu(nv->dev, page_pool_get_dma_addr(page),
957 				      hdr_pg_start, frame_sz,
958 				      DMA_BIDIRECTIONAL);
959 
960 	/* Build frame around buffer */
961 	hdr_start = page_address(page) + hdr_pg_start;
962 	net_prefetch(pkt->buff.data);
963 	xdp_prepare_buff(&pkt->buff, hdr_start, headroom,
964 			 len - FBNIC_RX_PAD, true);
965 
966 	pkt->hwtstamp = 0;
967 	pkt->add_frag_failed = false;
968 }
969 
970 static void fbnic_add_rx_frag(struct fbnic_napi_vector *nv, u64 rcd,
971 			      struct fbnic_pkt_buff *pkt,
972 			      struct fbnic_q_triad *qt)
973 {
974 	unsigned int pg_idx = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
975 	unsigned int pg_off = FIELD_GET(FBNIC_RCD_AL_BUFF_OFF_MASK, rcd);
976 	unsigned int len = FIELD_GET(FBNIC_RCD_AL_BUFF_LEN_MASK, rcd);
977 	struct page *page = fbnic_page_pool_get(&qt->sub1, pg_idx);
978 	unsigned int truesize;
979 	bool added;
980 
981 	truesize = FIELD_GET(FBNIC_RCD_AL_PAGE_FIN, rcd) ?
982 		   FBNIC_BD_FRAG_SIZE - pg_off : ALIGN(len, 128);
983 
984 	pg_off += (FBNIC_RCD_AL_BUFF_FRAG_MASK & rcd) *
985 		  FBNIC_BD_FRAG_SIZE;
986 
987 	/* Sync DMA buffer */
988 	dma_sync_single_range_for_cpu(nv->dev, page_pool_get_dma_addr(page),
989 				      pg_off, truesize, DMA_BIDIRECTIONAL);
990 
991 	added = xdp_buff_add_frag(&pkt->buff, page_to_netmem(page), pg_off, len,
992 				  truesize);
993 	if (unlikely(!added)) {
994 		pkt->add_frag_failed = true;
995 		netdev_err_once(nv->napi.dev,
996 				"Failed to add fragment to xdp_buff\n");
997 	}
998 }
999 
1000 static void fbnic_put_pkt_buff(struct fbnic_napi_vector *nv,
1001 			       struct fbnic_pkt_buff *pkt, int budget)
1002 {
1003 	struct page *page;
1004 
1005 	if (!pkt->buff.data_hard_start)
1006 		return;
1007 
1008 	if (xdp_buff_has_frags(&pkt->buff)) {
1009 		struct skb_shared_info *shinfo;
1010 		int nr_frags;
1011 
1012 		shinfo = xdp_get_shared_info_from_buff(&pkt->buff);
1013 		nr_frags = shinfo->nr_frags;
1014 
1015 		while (nr_frags--) {
1016 			page = skb_frag_page(&shinfo->frags[nr_frags]);
1017 			page_pool_put_full_page(nv->page_pool, page, !!budget);
1018 		}
1019 	}
1020 
1021 	page = virt_to_page(pkt->buff.data_hard_start);
1022 	page_pool_put_full_page(nv->page_pool, page, !!budget);
1023 }
1024 
1025 static struct sk_buff *fbnic_build_skb(struct fbnic_napi_vector *nv,
1026 				       struct fbnic_pkt_buff *pkt)
1027 {
1028 	struct sk_buff *skb;
1029 
1030 	skb = xdp_build_skb_from_buff(&pkt->buff);
1031 	if (!skb)
1032 		return NULL;
1033 
1034 	/* Add timestamp if present */
1035 	if (pkt->hwtstamp)
1036 		skb_hwtstamps(skb)->hwtstamp = pkt->hwtstamp;
1037 
1038 	return skb;
1039 }
1040 
1041 static long fbnic_pkt_tx(struct fbnic_napi_vector *nv,
1042 			 struct fbnic_pkt_buff *pkt)
1043 {
1044 	struct fbnic_ring *ring = &nv->qt[0].sub1;
1045 	int size, offset, nsegs = 1, data_len = 0;
1046 	unsigned int tail = ring->tail;
1047 	struct skb_shared_info *shinfo;
1048 	skb_frag_t *frag = NULL;
1049 	struct page *page;
1050 	dma_addr_t dma;
1051 	__le64 *twd;
1052 
1053 	if (unlikely(xdp_buff_has_frags(&pkt->buff))) {
1054 		shinfo = xdp_get_shared_info_from_buff(&pkt->buff);
1055 		nsegs += shinfo->nr_frags;
1056 		data_len = shinfo->xdp_frags_size;
1057 		frag = &shinfo->frags[0];
1058 	}
1059 
1060 	if (fbnic_desc_unused(ring) < nsegs) {
1061 		u64_stats_update_begin(&ring->stats.syncp);
1062 		ring->stats.dropped++;
1063 		u64_stats_update_end(&ring->stats.syncp);
1064 		return -FBNIC_XDP_CONSUME;
1065 	}
1066 
1067 	page = virt_to_page(pkt->buff.data_hard_start);
1068 	offset = offset_in_page(pkt->buff.data);
1069 	dma = page_pool_get_dma_addr(page);
1070 
1071 	size = pkt->buff.data_end - pkt->buff.data;
1072 
1073 	while (nsegs--) {
1074 		dma_sync_single_range_for_device(nv->dev, dma, offset, size,
1075 						 DMA_BIDIRECTIONAL);
1076 		dma += offset;
1077 
1078 		ring->tx_buf[tail] = page;
1079 
1080 		twd = &ring->desc[tail];
1081 		*twd = cpu_to_le64(FIELD_PREP(FBNIC_TWD_ADDR_MASK, dma) |
1082 				   FIELD_PREP(FBNIC_TWD_LEN_MASK, size) |
1083 				   FIELD_PREP(FBNIC_TWD_TYPE_MASK,
1084 					      FBNIC_TWD_TYPE_AL));
1085 
1086 		tail++;
1087 		tail &= ring->size_mask;
1088 
1089 		if (!data_len)
1090 			break;
1091 
1092 		offset = skb_frag_off(frag);
1093 		page = skb_frag_page(frag);
1094 		dma = page_pool_get_dma_addr(page);
1095 
1096 		size = skb_frag_size(frag);
1097 		data_len -= size;
1098 		frag++;
1099 	}
1100 
1101 	*twd |= FBNIC_TWD_TYPE(LAST_AL);
1102 
1103 	ring->tail = tail;
1104 
1105 	return -FBNIC_XDP_TX;
1106 }
1107 
1108 static void fbnic_pkt_commit_tail(struct fbnic_napi_vector *nv,
1109 				  unsigned int pkt_tail)
1110 {
1111 	struct fbnic_ring *ring = &nv->qt[0].sub1;
1112 
1113 	/* Force DMA writes to flush before writing to tail */
1114 	dma_wmb();
1115 
1116 	writel(pkt_tail, ring->doorbell);
1117 }
1118 
1119 static struct sk_buff *fbnic_run_xdp(struct fbnic_napi_vector *nv,
1120 				     struct fbnic_pkt_buff *pkt)
1121 {
1122 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
1123 	struct bpf_prog *xdp_prog;
1124 	int act;
1125 
1126 	xdp_prog = READ_ONCE(fbn->xdp_prog);
1127 	if (!xdp_prog)
1128 		goto xdp_pass;
1129 
1130 	/* Should never happen, config paths enforce HDS threshold > MTU */
1131 	if (xdp_buff_has_frags(&pkt->buff) && !xdp_prog->aux->xdp_has_frags)
1132 		return ERR_PTR(-FBNIC_XDP_LEN_ERR);
1133 
1134 	act = bpf_prog_run_xdp(xdp_prog, &pkt->buff);
1135 	switch (act) {
1136 	case XDP_PASS:
1137 xdp_pass:
1138 		return fbnic_build_skb(nv, pkt);
1139 	case XDP_TX:
1140 		return ERR_PTR(fbnic_pkt_tx(nv, pkt));
1141 	default:
1142 		bpf_warn_invalid_xdp_action(nv->napi.dev, xdp_prog, act);
1143 		fallthrough;
1144 	case XDP_ABORTED:
1145 		trace_xdp_exception(nv->napi.dev, xdp_prog, act);
1146 		fallthrough;
1147 	case XDP_DROP:
1148 		break;
1149 	}
1150 
1151 	return ERR_PTR(-FBNIC_XDP_CONSUME);
1152 }
1153 
1154 static enum pkt_hash_types fbnic_skb_hash_type(u64 rcd)
1155 {
1156 	return (FBNIC_RCD_META_L4_TYPE_MASK & rcd) ? PKT_HASH_TYPE_L4 :
1157 	       (FBNIC_RCD_META_L3_TYPE_MASK & rcd) ? PKT_HASH_TYPE_L3 :
1158 						     PKT_HASH_TYPE_L2;
1159 }
1160 
1161 static void fbnic_rx_tstamp(struct fbnic_napi_vector *nv, u64 rcd,
1162 			    struct fbnic_pkt_buff *pkt)
1163 {
1164 	struct fbnic_net *fbn;
1165 	u64 ns, ts;
1166 
1167 	if (!FIELD_GET(FBNIC_RCD_OPT_META_TS, rcd))
1168 		return;
1169 
1170 	fbn = netdev_priv(nv->napi.dev);
1171 	ts = FIELD_GET(FBNIC_RCD_OPT_META_TS_MASK, rcd);
1172 	ns = fbnic_ts40_to_ns(fbn, ts);
1173 
1174 	/* Add timestamp to shared info */
1175 	pkt->hwtstamp = ns_to_ktime(ns);
1176 }
1177 
1178 static void fbnic_populate_skb_fields(struct fbnic_napi_vector *nv,
1179 				      u64 rcd, struct sk_buff *skb,
1180 				      struct fbnic_q_triad *qt,
1181 				      u64 *csum_cmpl, u64 *csum_none)
1182 {
1183 	struct net_device *netdev = nv->napi.dev;
1184 	struct fbnic_ring *rcq = &qt->cmpl;
1185 
1186 	fbnic_rx_csum(rcd, skb, rcq, csum_cmpl, csum_none);
1187 
1188 	if (netdev->features & NETIF_F_RXHASH)
1189 		skb_set_hash(skb,
1190 			     FIELD_GET(FBNIC_RCD_META_RSS_HASH_MASK, rcd),
1191 			     fbnic_skb_hash_type(rcd));
1192 
1193 	skb_record_rx_queue(skb, rcq->q_idx);
1194 }
1195 
1196 static bool fbnic_rcd_metadata_err(u64 rcd)
1197 {
1198 	return !!(FBNIC_RCD_META_UNCORRECTABLE_ERR_MASK & rcd);
1199 }
1200 
1201 static int fbnic_clean_rcq(struct fbnic_napi_vector *nv,
1202 			   struct fbnic_q_triad *qt, int budget)
1203 {
1204 	unsigned int packets = 0, bytes = 0, dropped = 0, alloc_failed = 0;
1205 	u64 csum_complete = 0, csum_none = 0, length_errors = 0;
1206 	s32 head0 = -1, head1 = -1, pkt_tail = -1;
1207 	struct fbnic_ring *rcq = &qt->cmpl;
1208 	struct fbnic_pkt_buff *pkt;
1209 	__le64 *raw_rcd, done;
1210 	u32 head = rcq->head;
1211 
1212 	done = (head & (rcq->size_mask + 1)) ? cpu_to_le64(FBNIC_RCD_DONE) : 0;
1213 	raw_rcd = &rcq->desc[head & rcq->size_mask];
1214 	pkt = rcq->pkt;
1215 
1216 	/* Walk the completion queue collecting the heads reported by NIC */
1217 	while (likely(packets < budget)) {
1218 		struct sk_buff *skb = ERR_PTR(-EINVAL);
1219 		u64 rcd;
1220 
1221 		if ((*raw_rcd & cpu_to_le64(FBNIC_RCD_DONE)) == done)
1222 			break;
1223 
1224 		dma_rmb();
1225 
1226 		rcd = le64_to_cpu(*raw_rcd);
1227 
1228 		switch (FIELD_GET(FBNIC_RCD_TYPE_MASK, rcd)) {
1229 		case FBNIC_RCD_TYPE_HDR_AL:
1230 			head0 = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
1231 			fbnic_pkt_prepare(nv, rcd, pkt, qt);
1232 
1233 			break;
1234 		case FBNIC_RCD_TYPE_PAY_AL:
1235 			head1 = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
1236 			fbnic_add_rx_frag(nv, rcd, pkt, qt);
1237 
1238 			break;
1239 		case FBNIC_RCD_TYPE_OPT_META:
1240 			/* Only type 0 is currently supported */
1241 			if (FIELD_GET(FBNIC_RCD_OPT_META_TYPE_MASK, rcd))
1242 				break;
1243 
1244 			fbnic_rx_tstamp(nv, rcd, pkt);
1245 
1246 			/* We currently ignore the action table index */
1247 			break;
1248 		case FBNIC_RCD_TYPE_META:
1249 			if (unlikely(pkt->add_frag_failed))
1250 				skb = NULL;
1251 			else if (likely(!fbnic_rcd_metadata_err(rcd)))
1252 				skb = fbnic_run_xdp(nv, pkt);
1253 
1254 			/* Populate skb and invalidate XDP */
1255 			if (!IS_ERR_OR_NULL(skb)) {
1256 				fbnic_populate_skb_fields(nv, rcd, skb, qt,
1257 							  &csum_complete,
1258 							  &csum_none);
1259 
1260 				packets++;
1261 				bytes += skb->len;
1262 
1263 				napi_gro_receive(&nv->napi, skb);
1264 			} else if (skb == ERR_PTR(-FBNIC_XDP_TX)) {
1265 				pkt_tail = nv->qt[0].sub1.tail;
1266 				bytes += xdp_get_buff_len(&pkt->buff);
1267 			} else {
1268 				if (!skb) {
1269 					alloc_failed++;
1270 					dropped++;
1271 				} else if (skb == ERR_PTR(-FBNIC_XDP_LEN_ERR)) {
1272 					length_errors++;
1273 				} else {
1274 					dropped++;
1275 				}
1276 
1277 				fbnic_put_pkt_buff(nv, pkt, 1);
1278 			}
1279 
1280 			pkt->buff.data_hard_start = NULL;
1281 
1282 			break;
1283 		}
1284 
1285 		raw_rcd++;
1286 		head++;
1287 		if (!(head & rcq->size_mask)) {
1288 			done ^= cpu_to_le64(FBNIC_RCD_DONE);
1289 			raw_rcd = &rcq->desc[0];
1290 		}
1291 	}
1292 
1293 	u64_stats_update_begin(&rcq->stats.syncp);
1294 	rcq->stats.packets += packets;
1295 	rcq->stats.bytes += bytes;
1296 	/* Re-add ethernet header length (removed in fbnic_build_skb) */
1297 	rcq->stats.bytes += ETH_HLEN * packets;
1298 	rcq->stats.dropped += dropped;
1299 	rcq->stats.rx.alloc_failed += alloc_failed;
1300 	rcq->stats.rx.csum_complete += csum_complete;
1301 	rcq->stats.rx.csum_none += csum_none;
1302 	rcq->stats.rx.length_errors += length_errors;
1303 	u64_stats_update_end(&rcq->stats.syncp);
1304 
1305 	if (pkt_tail >= 0)
1306 		fbnic_pkt_commit_tail(nv, pkt_tail);
1307 
1308 	/* Unmap and free processed buffers */
1309 	if (head0 >= 0)
1310 		fbnic_clean_bdq(nv, budget, &qt->sub0, head0);
1311 	fbnic_fill_bdq(nv, &qt->sub0);
1312 
1313 	if (head1 >= 0)
1314 		fbnic_clean_bdq(nv, budget, &qt->sub1, head1);
1315 	fbnic_fill_bdq(nv, &qt->sub1);
1316 
1317 	/* Record the current head/tail of the queue */
1318 	if (rcq->head != head) {
1319 		rcq->head = head;
1320 		writel(head & rcq->size_mask, rcq->doorbell);
1321 	}
1322 
1323 	return packets;
1324 }
1325 
1326 static void fbnic_nv_irq_disable(struct fbnic_napi_vector *nv)
1327 {
1328 	struct fbnic_dev *fbd = nv->fbd;
1329 	u32 v_idx = nv->v_idx;
1330 
1331 	fbnic_wr32(fbd, FBNIC_INTR_MASK_SET(v_idx / 32), 1 << (v_idx % 32));
1332 }
1333 
1334 static void fbnic_nv_irq_rearm(struct fbnic_napi_vector *nv)
1335 {
1336 	struct fbnic_dev *fbd = nv->fbd;
1337 	u32 v_idx = nv->v_idx;
1338 
1339 	fbnic_wr32(fbd, FBNIC_INTR_CQ_REARM(v_idx),
1340 		   FBNIC_INTR_CQ_REARM_INTR_UNMASK);
1341 }
1342 
1343 static int fbnic_poll(struct napi_struct *napi, int budget)
1344 {
1345 	struct fbnic_napi_vector *nv = container_of(napi,
1346 						    struct fbnic_napi_vector,
1347 						    napi);
1348 	int i, j, work_done = 0;
1349 
1350 	for (i = 0; i < nv->txt_count; i++)
1351 		fbnic_clean_tcq(nv, &nv->qt[i], budget);
1352 
1353 	for (j = 0; j < nv->rxt_count; j++, i++)
1354 		work_done += fbnic_clean_rcq(nv, &nv->qt[i], budget);
1355 
1356 	if (work_done >= budget)
1357 		return budget;
1358 
1359 	if (likely(napi_complete_done(napi, work_done)))
1360 		fbnic_nv_irq_rearm(nv);
1361 
1362 	return work_done;
1363 }
1364 
1365 irqreturn_t fbnic_msix_clean_rings(int __always_unused irq, void *data)
1366 {
1367 	struct fbnic_napi_vector *nv = *(void **)data;
1368 
1369 	napi_schedule_irqoff(&nv->napi);
1370 
1371 	return IRQ_HANDLED;
1372 }
1373 
1374 void fbnic_aggregate_ring_rx_counters(struct fbnic_net *fbn,
1375 				      struct fbnic_ring *rxr)
1376 {
1377 	struct fbnic_queue_stats *stats = &rxr->stats;
1378 
1379 	/* Capture stats from queues before dissasociating them */
1380 	fbn->rx_stats.bytes += stats->bytes;
1381 	fbn->rx_stats.packets += stats->packets;
1382 	fbn->rx_stats.dropped += stats->dropped;
1383 	fbn->rx_stats.rx.alloc_failed += stats->rx.alloc_failed;
1384 	fbn->rx_stats.rx.csum_complete += stats->rx.csum_complete;
1385 	fbn->rx_stats.rx.csum_none += stats->rx.csum_none;
1386 	fbn->rx_stats.rx.length_errors += stats->rx.length_errors;
1387 	/* Remember to add new stats here */
1388 	BUILD_BUG_ON(sizeof(fbn->rx_stats.rx) / 8 != 4);
1389 }
1390 
1391 void fbnic_aggregate_ring_tx_counters(struct fbnic_net *fbn,
1392 				      struct fbnic_ring *txr)
1393 {
1394 	struct fbnic_queue_stats *stats = &txr->stats;
1395 
1396 	/* Capture stats from queues before dissasociating them */
1397 	fbn->tx_stats.bytes += stats->bytes;
1398 	fbn->tx_stats.packets += stats->packets;
1399 	fbn->tx_stats.dropped += stats->dropped;
1400 	fbn->tx_stats.twq.csum_partial += stats->twq.csum_partial;
1401 	fbn->tx_stats.twq.lso += stats->twq.lso;
1402 	fbn->tx_stats.twq.ts_lost += stats->twq.ts_lost;
1403 	fbn->tx_stats.twq.ts_packets += stats->twq.ts_packets;
1404 	fbn->tx_stats.twq.stop += stats->twq.stop;
1405 	fbn->tx_stats.twq.wake += stats->twq.wake;
1406 	/* Remember to add new stats here */
1407 	BUILD_BUG_ON(sizeof(fbn->tx_stats.twq) / 8 != 6);
1408 }
1409 
1410 static void fbnic_aggregate_ring_xdp_counters(struct fbnic_net *fbn,
1411 					      struct fbnic_ring *xdpr)
1412 {
1413 	struct fbnic_queue_stats *stats = &xdpr->stats;
1414 
1415 	if (!(xdpr->flags & FBNIC_RING_F_STATS))
1416 		return;
1417 
1418 	/* Capture stats from queues before dissasociating them */
1419 	fbn->rx_stats.bytes += stats->bytes;
1420 	fbn->rx_stats.packets += stats->packets;
1421 	fbn->rx_stats.dropped += stats->dropped;
1422 	fbn->tx_stats.bytes += stats->bytes;
1423 	fbn->tx_stats.packets += stats->packets;
1424 }
1425 
1426 static void fbnic_remove_tx_ring(struct fbnic_net *fbn,
1427 				 struct fbnic_ring *txr)
1428 {
1429 	if (!(txr->flags & FBNIC_RING_F_STATS))
1430 		return;
1431 
1432 	fbnic_aggregate_ring_tx_counters(fbn, txr);
1433 
1434 	/* Remove pointer to the Tx ring */
1435 	WARN_ON(fbn->tx[txr->q_idx] && fbn->tx[txr->q_idx] != txr);
1436 	fbn->tx[txr->q_idx] = NULL;
1437 }
1438 
1439 static void fbnic_remove_xdp_ring(struct fbnic_net *fbn,
1440 				  struct fbnic_ring *xdpr)
1441 {
1442 	if (!(xdpr->flags & FBNIC_RING_F_STATS))
1443 		return;
1444 
1445 	fbnic_aggregate_ring_xdp_counters(fbn, xdpr);
1446 
1447 	/* Remove pointer to the Tx ring */
1448 	WARN_ON(fbn->tx[xdpr->q_idx] && fbn->tx[xdpr->q_idx] != xdpr);
1449 	fbn->tx[xdpr->q_idx] = NULL;
1450 }
1451 
1452 static void fbnic_remove_rx_ring(struct fbnic_net *fbn,
1453 				 struct fbnic_ring *rxr)
1454 {
1455 	if (!(rxr->flags & FBNIC_RING_F_STATS))
1456 		return;
1457 
1458 	fbnic_aggregate_ring_rx_counters(fbn, rxr);
1459 
1460 	/* Remove pointer to the Rx ring */
1461 	WARN_ON(fbn->rx[rxr->q_idx] && fbn->rx[rxr->q_idx] != rxr);
1462 	fbn->rx[rxr->q_idx] = NULL;
1463 }
1464 
1465 static void fbnic_free_napi_vector(struct fbnic_net *fbn,
1466 				   struct fbnic_napi_vector *nv)
1467 {
1468 	struct fbnic_dev *fbd = nv->fbd;
1469 	int i, j;
1470 
1471 	for (i = 0; i < nv->txt_count; i++) {
1472 		fbnic_remove_tx_ring(fbn, &nv->qt[i].sub0);
1473 		fbnic_remove_xdp_ring(fbn, &nv->qt[i].sub1);
1474 		fbnic_remove_tx_ring(fbn, &nv->qt[i].cmpl);
1475 	}
1476 
1477 	for (j = 0; j < nv->rxt_count; j++, i++) {
1478 		xdp_rxq_info_unreg(&nv->qt[i].xdp_rxq);
1479 		fbnic_remove_rx_ring(fbn, &nv->qt[i].sub0);
1480 		fbnic_remove_rx_ring(fbn, &nv->qt[i].sub1);
1481 		fbnic_remove_rx_ring(fbn, &nv->qt[i].cmpl);
1482 	}
1483 
1484 	fbnic_napi_free_irq(fbd, nv);
1485 	page_pool_destroy(nv->page_pool);
1486 	netif_napi_del(&nv->napi);
1487 	fbn->napi[fbnic_napi_idx(nv)] = NULL;
1488 	kfree(nv);
1489 }
1490 
1491 void fbnic_free_napi_vectors(struct fbnic_net *fbn)
1492 {
1493 	int i;
1494 
1495 	for (i = 0; i < fbn->num_napi; i++)
1496 		if (fbn->napi[i])
1497 			fbnic_free_napi_vector(fbn, fbn->napi[i]);
1498 }
1499 
1500 #define FBNIC_PAGE_POOL_FLAGS \
1501 	(PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV)
1502 
1503 static int fbnic_alloc_nv_page_pool(struct fbnic_net *fbn,
1504 				    struct fbnic_napi_vector *nv)
1505 {
1506 	struct page_pool_params pp_params = {
1507 		.order = 0,
1508 		.flags = FBNIC_PAGE_POOL_FLAGS,
1509 		.pool_size = (fbn->hpq_size + fbn->ppq_size) * nv->rxt_count,
1510 		.nid = NUMA_NO_NODE,
1511 		.dev = nv->dev,
1512 		.dma_dir = DMA_BIDIRECTIONAL,
1513 		.offset = 0,
1514 		.max_len = PAGE_SIZE,
1515 		.napi	= &nv->napi,
1516 		.netdev	= fbn->netdev,
1517 	};
1518 	struct page_pool *pp;
1519 
1520 	/* Page pool cannot exceed a size of 32768. This doesn't limit the
1521 	 * pages on the ring but the number we can have cached waiting on
1522 	 * the next use.
1523 	 *
1524 	 * TBD: Can this be reduced further? Would a multiple of
1525 	 * NAPI_POLL_WEIGHT possibly make more sense? The question is how
1526 	 * may pages do we need to hold in reserve to get the best return
1527 	 * without hogging too much system memory.
1528 	 */
1529 	if (pp_params.pool_size > 32768)
1530 		pp_params.pool_size = 32768;
1531 
1532 	pp = page_pool_create(&pp_params);
1533 	if (IS_ERR(pp))
1534 		return PTR_ERR(pp);
1535 
1536 	nv->page_pool = pp;
1537 
1538 	return 0;
1539 }
1540 
1541 static void fbnic_ring_init(struct fbnic_ring *ring, u32 __iomem *doorbell,
1542 			    int q_idx, u8 flags)
1543 {
1544 	u64_stats_init(&ring->stats.syncp);
1545 	ring->doorbell = doorbell;
1546 	ring->q_idx = q_idx;
1547 	ring->flags = flags;
1548 	ring->deferred_head = -1;
1549 }
1550 
1551 static int fbnic_alloc_napi_vector(struct fbnic_dev *fbd, struct fbnic_net *fbn,
1552 				   unsigned int v_count, unsigned int v_idx,
1553 				   unsigned int txq_count, unsigned int txq_idx,
1554 				   unsigned int rxq_count, unsigned int rxq_idx)
1555 {
1556 	int txt_count = txq_count, rxt_count = rxq_count;
1557 	u32 __iomem *uc_addr = fbd->uc_addr0;
1558 	int xdp_count = 0, qt_count, err;
1559 	struct fbnic_napi_vector *nv;
1560 	struct fbnic_q_triad *qt;
1561 	u32 __iomem *db;
1562 
1563 	/* We need to reserve at least one Tx Queue Triad for an XDP ring */
1564 	if (rxq_count) {
1565 		xdp_count = 1;
1566 		if (!txt_count)
1567 			txt_count = 1;
1568 	}
1569 
1570 	qt_count = txt_count + rxq_count;
1571 	if (!qt_count)
1572 		return -EINVAL;
1573 
1574 	/* If MMIO has already failed there are no rings to initialize */
1575 	if (!uc_addr)
1576 		return -EIO;
1577 
1578 	/* Allocate NAPI vector and queue triads */
1579 	nv = kzalloc(struct_size(nv, qt, qt_count), GFP_KERNEL);
1580 	if (!nv)
1581 		return -ENOMEM;
1582 
1583 	/* Record queue triad counts */
1584 	nv->txt_count = txt_count;
1585 	nv->rxt_count = rxt_count;
1586 
1587 	/* Provide pointer back to fbnic and MSI-X vectors */
1588 	nv->fbd = fbd;
1589 	nv->v_idx = v_idx;
1590 
1591 	/* Tie napi to netdev */
1592 	fbn->napi[fbnic_napi_idx(nv)] = nv;
1593 	netif_napi_add(fbn->netdev, &nv->napi, fbnic_poll);
1594 
1595 	/* Record IRQ to NAPI struct */
1596 	netif_napi_set_irq(&nv->napi,
1597 			   pci_irq_vector(to_pci_dev(fbd->dev), nv->v_idx));
1598 
1599 	/* Tie nv back to PCIe dev */
1600 	nv->dev = fbd->dev;
1601 
1602 	/* Allocate page pool */
1603 	if (rxq_count) {
1604 		err = fbnic_alloc_nv_page_pool(fbn, nv);
1605 		if (err)
1606 			goto napi_del;
1607 	}
1608 
1609 	/* Request the IRQ for napi vector */
1610 	err = fbnic_napi_request_irq(fbd, nv);
1611 	if (err)
1612 		goto pp_destroy;
1613 
1614 	/* Initialize queue triads */
1615 	qt = nv->qt;
1616 
1617 	while (txt_count) {
1618 		u8 flags = FBNIC_RING_F_CTX | FBNIC_RING_F_STATS;
1619 
1620 		/* Configure Tx queue */
1621 		db = &uc_addr[FBNIC_QUEUE(txq_idx) + FBNIC_QUEUE_TWQ0_TAIL];
1622 
1623 		/* Assign Tx queue to netdev if applicable */
1624 		if (txq_count > 0) {
1625 
1626 			fbnic_ring_init(&qt->sub0, db, txq_idx, flags);
1627 			fbn->tx[txq_idx] = &qt->sub0;
1628 			txq_count--;
1629 		} else {
1630 			fbnic_ring_init(&qt->sub0, db, 0,
1631 					FBNIC_RING_F_DISABLED);
1632 		}
1633 
1634 		/* Configure XDP queue */
1635 		db = &uc_addr[FBNIC_QUEUE(txq_idx) + FBNIC_QUEUE_TWQ1_TAIL];
1636 
1637 		/* Assign XDP queue to netdev if applicable
1638 		 *
1639 		 * The setup for this is in itself a bit different.
1640 		 * 1. We only need one XDP Tx queue per NAPI vector.
1641 		 * 2. We associate it to the first Rx queue index.
1642 		 * 3. The hardware side is associated based on the Tx Queue.
1643 		 * 4. The netdev queue is offset by FBNIC_MAX_TXQs.
1644 		 */
1645 		if (xdp_count > 0) {
1646 			unsigned int xdp_idx = FBNIC_MAX_TXQS + rxq_idx;
1647 
1648 			fbnic_ring_init(&qt->sub1, db, xdp_idx, flags);
1649 			fbn->tx[xdp_idx] = &qt->sub1;
1650 			xdp_count--;
1651 		} else {
1652 			fbnic_ring_init(&qt->sub1, db, 0,
1653 					FBNIC_RING_F_DISABLED);
1654 		}
1655 
1656 		/* Configure Tx completion queue */
1657 		db = &uc_addr[FBNIC_QUEUE(txq_idx) + FBNIC_QUEUE_TCQ_HEAD];
1658 		fbnic_ring_init(&qt->cmpl, db, 0, 0);
1659 
1660 		/* Update Tx queue index */
1661 		txt_count--;
1662 		txq_idx += v_count;
1663 
1664 		/* Move to next queue triad */
1665 		qt++;
1666 	}
1667 
1668 	while (rxt_count) {
1669 		/* Configure header queue */
1670 		db = &uc_addr[FBNIC_QUEUE(rxq_idx) + FBNIC_QUEUE_BDQ_HPQ_TAIL];
1671 		fbnic_ring_init(&qt->sub0, db, 0, FBNIC_RING_F_CTX);
1672 
1673 		/* Configure payload queue */
1674 		db = &uc_addr[FBNIC_QUEUE(rxq_idx) + FBNIC_QUEUE_BDQ_PPQ_TAIL];
1675 		fbnic_ring_init(&qt->sub1, db, 0, FBNIC_RING_F_CTX);
1676 
1677 		/* Configure Rx completion queue */
1678 		db = &uc_addr[FBNIC_QUEUE(rxq_idx) + FBNIC_QUEUE_RCQ_HEAD];
1679 		fbnic_ring_init(&qt->cmpl, db, rxq_idx, FBNIC_RING_F_STATS);
1680 		fbn->rx[rxq_idx] = &qt->cmpl;
1681 
1682 		err = xdp_rxq_info_reg(&qt->xdp_rxq, fbn->netdev, rxq_idx,
1683 				       nv->napi.napi_id);
1684 		if (err)
1685 			goto free_ring_cur_qt;
1686 
1687 		/* Update Rx queue index */
1688 		rxt_count--;
1689 		rxq_idx += v_count;
1690 
1691 		/* Move to next queue triad */
1692 		qt++;
1693 	}
1694 
1695 	return 0;
1696 
1697 	while (rxt_count < nv->rxt_count) {
1698 		qt--;
1699 
1700 		xdp_rxq_info_unreg(&qt->xdp_rxq);
1701 free_ring_cur_qt:
1702 		fbnic_remove_rx_ring(fbn, &qt->sub0);
1703 		fbnic_remove_rx_ring(fbn, &qt->sub1);
1704 		fbnic_remove_rx_ring(fbn, &qt->cmpl);
1705 		rxt_count++;
1706 	}
1707 	while (txt_count < nv->txt_count) {
1708 		qt--;
1709 
1710 		fbnic_remove_tx_ring(fbn, &qt->sub0);
1711 		fbnic_remove_xdp_ring(fbn, &qt->sub1);
1712 		fbnic_remove_tx_ring(fbn, &qt->cmpl);
1713 
1714 		txt_count++;
1715 	}
1716 	fbnic_napi_free_irq(fbd, nv);
1717 pp_destroy:
1718 	page_pool_destroy(nv->page_pool);
1719 napi_del:
1720 	netif_napi_del(&nv->napi);
1721 	fbn->napi[fbnic_napi_idx(nv)] = NULL;
1722 	kfree(nv);
1723 	return err;
1724 }
1725 
1726 int fbnic_alloc_napi_vectors(struct fbnic_net *fbn)
1727 {
1728 	unsigned int txq_idx = 0, rxq_idx = 0, v_idx = FBNIC_NON_NAPI_VECTORS;
1729 	unsigned int num_tx = fbn->num_tx_queues;
1730 	unsigned int num_rx = fbn->num_rx_queues;
1731 	unsigned int num_napi = fbn->num_napi;
1732 	struct fbnic_dev *fbd = fbn->fbd;
1733 	int err;
1734 
1735 	/* Allocate 1 Tx queue per napi vector */
1736 	if (num_napi < FBNIC_MAX_TXQS && num_napi == num_tx + num_rx) {
1737 		while (num_tx) {
1738 			err = fbnic_alloc_napi_vector(fbd, fbn,
1739 						      num_napi, v_idx,
1740 						      1, txq_idx, 0, 0);
1741 			if (err)
1742 				goto free_vectors;
1743 
1744 			/* Update counts and index */
1745 			num_tx--;
1746 			txq_idx++;
1747 
1748 			v_idx++;
1749 		}
1750 	}
1751 
1752 	/* Allocate Tx/Rx queue pairs per vector, or allocate remaining Rx */
1753 	while (num_rx | num_tx) {
1754 		int tqpv = DIV_ROUND_UP(num_tx, num_napi - txq_idx);
1755 		int rqpv = DIV_ROUND_UP(num_rx, num_napi - rxq_idx);
1756 
1757 		err = fbnic_alloc_napi_vector(fbd, fbn, num_napi, v_idx,
1758 					      tqpv, txq_idx, rqpv, rxq_idx);
1759 		if (err)
1760 			goto free_vectors;
1761 
1762 		/* Update counts and index */
1763 		num_tx -= tqpv;
1764 		txq_idx++;
1765 
1766 		num_rx -= rqpv;
1767 		rxq_idx++;
1768 
1769 		v_idx++;
1770 	}
1771 
1772 	return 0;
1773 
1774 free_vectors:
1775 	fbnic_free_napi_vectors(fbn);
1776 
1777 	return -ENOMEM;
1778 }
1779 
1780 static void fbnic_free_ring_resources(struct device *dev,
1781 				      struct fbnic_ring *ring)
1782 {
1783 	kvfree(ring->buffer);
1784 	ring->buffer = NULL;
1785 
1786 	/* If size is not set there are no descriptors present */
1787 	if (!ring->size)
1788 		return;
1789 
1790 	dma_free_coherent(dev, ring->size, ring->desc, ring->dma);
1791 	ring->size_mask = 0;
1792 	ring->size = 0;
1793 }
1794 
1795 static int fbnic_alloc_tx_ring_desc(struct fbnic_net *fbn,
1796 				    struct fbnic_ring *txr)
1797 {
1798 	struct device *dev = fbn->netdev->dev.parent;
1799 	size_t size;
1800 
1801 	/* Round size up to nearest 4K */
1802 	size = ALIGN(array_size(sizeof(*txr->desc), fbn->txq_size), 4096);
1803 
1804 	txr->desc = dma_alloc_coherent(dev, size, &txr->dma,
1805 				       GFP_KERNEL | __GFP_NOWARN);
1806 	if (!txr->desc)
1807 		return -ENOMEM;
1808 
1809 	/* txq_size should be a power of 2, so mask is just that -1 */
1810 	txr->size_mask = fbn->txq_size - 1;
1811 	txr->size = size;
1812 
1813 	return 0;
1814 }
1815 
1816 static int fbnic_alloc_tx_ring_buffer(struct fbnic_ring *txr)
1817 {
1818 	size_t size = array_size(sizeof(*txr->tx_buf), txr->size_mask + 1);
1819 
1820 	txr->tx_buf = kvzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1821 
1822 	return txr->tx_buf ? 0 : -ENOMEM;
1823 }
1824 
1825 static int fbnic_alloc_tx_ring_resources(struct fbnic_net *fbn,
1826 					 struct fbnic_ring *txr)
1827 {
1828 	struct device *dev = fbn->netdev->dev.parent;
1829 	int err;
1830 
1831 	if (txr->flags & FBNIC_RING_F_DISABLED)
1832 		return 0;
1833 
1834 	err = fbnic_alloc_tx_ring_desc(fbn, txr);
1835 	if (err)
1836 		return err;
1837 
1838 	if (!(txr->flags & FBNIC_RING_F_CTX))
1839 		return 0;
1840 
1841 	err = fbnic_alloc_tx_ring_buffer(txr);
1842 	if (err)
1843 		goto free_desc;
1844 
1845 	return 0;
1846 
1847 free_desc:
1848 	fbnic_free_ring_resources(dev, txr);
1849 	return err;
1850 }
1851 
1852 static int fbnic_alloc_rx_ring_desc(struct fbnic_net *fbn,
1853 				    struct fbnic_ring *rxr)
1854 {
1855 	struct device *dev = fbn->netdev->dev.parent;
1856 	size_t desc_size = sizeof(*rxr->desc);
1857 	u32 rxq_size;
1858 	size_t size;
1859 
1860 	switch (rxr->doorbell - fbnic_ring_csr_base(rxr)) {
1861 	case FBNIC_QUEUE_BDQ_HPQ_TAIL:
1862 		rxq_size = fbn->hpq_size / FBNIC_BD_FRAG_COUNT;
1863 		desc_size *= FBNIC_BD_FRAG_COUNT;
1864 		break;
1865 	case FBNIC_QUEUE_BDQ_PPQ_TAIL:
1866 		rxq_size = fbn->ppq_size / FBNIC_BD_FRAG_COUNT;
1867 		desc_size *= FBNIC_BD_FRAG_COUNT;
1868 		break;
1869 	case FBNIC_QUEUE_RCQ_HEAD:
1870 		rxq_size = fbn->rcq_size;
1871 		break;
1872 	default:
1873 		return -EINVAL;
1874 	}
1875 
1876 	/* Round size up to nearest 4K */
1877 	size = ALIGN(array_size(desc_size, rxq_size), 4096);
1878 
1879 	rxr->desc = dma_alloc_coherent(dev, size, &rxr->dma,
1880 				       GFP_KERNEL | __GFP_NOWARN);
1881 	if (!rxr->desc)
1882 		return -ENOMEM;
1883 
1884 	/* rxq_size should be a power of 2, so mask is just that -1 */
1885 	rxr->size_mask = rxq_size - 1;
1886 	rxr->size = size;
1887 
1888 	return 0;
1889 }
1890 
1891 static int fbnic_alloc_rx_ring_buffer(struct fbnic_ring *rxr)
1892 {
1893 	size_t size = array_size(sizeof(*rxr->rx_buf), rxr->size_mask + 1);
1894 
1895 	if (rxr->flags & FBNIC_RING_F_CTX)
1896 		size = sizeof(*rxr->rx_buf) * (rxr->size_mask + 1);
1897 	else
1898 		size = sizeof(*rxr->pkt);
1899 
1900 	rxr->rx_buf = kvzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1901 
1902 	return rxr->rx_buf ? 0 : -ENOMEM;
1903 }
1904 
1905 static int fbnic_alloc_rx_ring_resources(struct fbnic_net *fbn,
1906 					 struct fbnic_ring *rxr)
1907 {
1908 	struct device *dev = fbn->netdev->dev.parent;
1909 	int err;
1910 
1911 	err = fbnic_alloc_rx_ring_desc(fbn, rxr);
1912 	if (err)
1913 		return err;
1914 
1915 	err = fbnic_alloc_rx_ring_buffer(rxr);
1916 	if (err)
1917 		goto free_desc;
1918 
1919 	return 0;
1920 
1921 free_desc:
1922 	fbnic_free_ring_resources(dev, rxr);
1923 	return err;
1924 }
1925 
1926 static void fbnic_free_qt_resources(struct fbnic_net *fbn,
1927 				    struct fbnic_q_triad *qt)
1928 {
1929 	struct device *dev = fbn->netdev->dev.parent;
1930 
1931 	fbnic_free_ring_resources(dev, &qt->cmpl);
1932 	fbnic_free_ring_resources(dev, &qt->sub1);
1933 	fbnic_free_ring_resources(dev, &qt->sub0);
1934 }
1935 
1936 static int fbnic_alloc_tx_qt_resources(struct fbnic_net *fbn,
1937 				       struct fbnic_q_triad *qt)
1938 {
1939 	struct device *dev = fbn->netdev->dev.parent;
1940 	int err;
1941 
1942 	err = fbnic_alloc_tx_ring_resources(fbn, &qt->sub0);
1943 	if (err)
1944 		return err;
1945 
1946 	err = fbnic_alloc_tx_ring_resources(fbn, &qt->sub1);
1947 	if (err)
1948 		goto free_sub0;
1949 
1950 	err = fbnic_alloc_tx_ring_resources(fbn, &qt->cmpl);
1951 	if (err)
1952 		goto free_sub1;
1953 
1954 	return 0;
1955 
1956 free_sub1:
1957 	fbnic_free_ring_resources(dev, &qt->sub1);
1958 free_sub0:
1959 	fbnic_free_ring_resources(dev, &qt->sub0);
1960 	return err;
1961 }
1962 
1963 static int fbnic_alloc_rx_qt_resources(struct fbnic_net *fbn,
1964 				       struct fbnic_q_triad *qt)
1965 {
1966 	struct device *dev = fbn->netdev->dev.parent;
1967 	int err;
1968 
1969 	err = fbnic_alloc_rx_ring_resources(fbn, &qt->sub0);
1970 	if (err)
1971 		return err;
1972 
1973 	err = fbnic_alloc_rx_ring_resources(fbn, &qt->sub1);
1974 	if (err)
1975 		goto free_sub0;
1976 
1977 	err = fbnic_alloc_rx_ring_resources(fbn, &qt->cmpl);
1978 	if (err)
1979 		goto free_sub1;
1980 
1981 	return 0;
1982 
1983 free_sub1:
1984 	fbnic_free_ring_resources(dev, &qt->sub1);
1985 free_sub0:
1986 	fbnic_free_ring_resources(dev, &qt->sub0);
1987 	return err;
1988 }
1989 
1990 static void fbnic_free_nv_resources(struct fbnic_net *fbn,
1991 				    struct fbnic_napi_vector *nv)
1992 {
1993 	int i, j;
1994 
1995 	/* Free Tx Resources  */
1996 	for (i = 0; i < nv->txt_count; i++)
1997 		fbnic_free_qt_resources(fbn, &nv->qt[i]);
1998 
1999 	for (j = 0; j < nv->rxt_count; j++, i++) {
2000 		fbnic_free_qt_resources(fbn, &nv->qt[i]);
2001 		xdp_rxq_info_unreg_mem_model(&nv->qt[i].xdp_rxq);
2002 	}
2003 }
2004 
2005 static int fbnic_alloc_nv_resources(struct fbnic_net *fbn,
2006 				    struct fbnic_napi_vector *nv)
2007 {
2008 	int i, j, err;
2009 
2010 	/* Allocate Tx Resources */
2011 	for (i = 0; i < nv->txt_count; i++) {
2012 		err = fbnic_alloc_tx_qt_resources(fbn, &nv->qt[i]);
2013 		if (err)
2014 			goto free_qt_resources;
2015 	}
2016 
2017 	/* Allocate Rx Resources */
2018 	for (j = 0; j < nv->rxt_count; j++, i++) {
2019 		/* Register XDP memory model for completion queue */
2020 		err = xdp_reg_mem_model(&nv->qt[i].xdp_rxq.mem,
2021 					MEM_TYPE_PAGE_POOL,
2022 					nv->page_pool);
2023 		if (err)
2024 			goto xdp_unreg_mem_model;
2025 
2026 		err = fbnic_alloc_rx_qt_resources(fbn, &nv->qt[i]);
2027 		if (err)
2028 			goto xdp_unreg_cur_model;
2029 	}
2030 
2031 	return 0;
2032 
2033 xdp_unreg_mem_model:
2034 	while (j-- && i--) {
2035 		fbnic_free_qt_resources(fbn, &nv->qt[i]);
2036 xdp_unreg_cur_model:
2037 		xdp_rxq_info_unreg_mem_model(&nv->qt[i].xdp_rxq);
2038 	}
2039 free_qt_resources:
2040 	while (i--)
2041 		fbnic_free_qt_resources(fbn, &nv->qt[i]);
2042 	return err;
2043 }
2044 
2045 void fbnic_free_resources(struct fbnic_net *fbn)
2046 {
2047 	int i;
2048 
2049 	for (i = 0; i < fbn->num_napi; i++)
2050 		fbnic_free_nv_resources(fbn, fbn->napi[i]);
2051 }
2052 
2053 int fbnic_alloc_resources(struct fbnic_net *fbn)
2054 {
2055 	int i, err = -ENODEV;
2056 
2057 	for (i = 0; i < fbn->num_napi; i++) {
2058 		err = fbnic_alloc_nv_resources(fbn, fbn->napi[i]);
2059 		if (err)
2060 			goto free_resources;
2061 	}
2062 
2063 	return 0;
2064 
2065 free_resources:
2066 	while (i--)
2067 		fbnic_free_nv_resources(fbn, fbn->napi[i]);
2068 
2069 	return err;
2070 }
2071 
2072 static void fbnic_set_netif_napi(struct fbnic_napi_vector *nv)
2073 {
2074 	int i, j;
2075 
2076 	/* Associate Tx queue with NAPI */
2077 	for (i = 0; i < nv->txt_count; i++) {
2078 		struct fbnic_q_triad *qt = &nv->qt[i];
2079 
2080 		netif_queue_set_napi(nv->napi.dev, qt->sub0.q_idx,
2081 				     NETDEV_QUEUE_TYPE_TX, &nv->napi);
2082 	}
2083 
2084 	/* Associate Rx queue with NAPI */
2085 	for (j = 0; j < nv->rxt_count; j++, i++) {
2086 		struct fbnic_q_triad *qt = &nv->qt[i];
2087 
2088 		netif_queue_set_napi(nv->napi.dev, qt->cmpl.q_idx,
2089 				     NETDEV_QUEUE_TYPE_RX, &nv->napi);
2090 	}
2091 }
2092 
2093 static void fbnic_reset_netif_napi(struct fbnic_napi_vector *nv)
2094 {
2095 	int i, j;
2096 
2097 	/* Disassociate Tx queue from NAPI */
2098 	for (i = 0; i < nv->txt_count; i++) {
2099 		struct fbnic_q_triad *qt = &nv->qt[i];
2100 
2101 		netif_queue_set_napi(nv->napi.dev, qt->sub0.q_idx,
2102 				     NETDEV_QUEUE_TYPE_TX, NULL);
2103 	}
2104 
2105 	/* Disassociate Rx queue from NAPI */
2106 	for (j = 0; j < nv->rxt_count; j++, i++) {
2107 		struct fbnic_q_triad *qt = &nv->qt[i];
2108 
2109 		netif_queue_set_napi(nv->napi.dev, qt->cmpl.q_idx,
2110 				     NETDEV_QUEUE_TYPE_RX, NULL);
2111 	}
2112 }
2113 
2114 int fbnic_set_netif_queues(struct fbnic_net *fbn)
2115 {
2116 	int i, err;
2117 
2118 	err = netif_set_real_num_queues(fbn->netdev, fbn->num_tx_queues,
2119 					fbn->num_rx_queues);
2120 	if (err)
2121 		return err;
2122 
2123 	for (i = 0; i < fbn->num_napi; i++)
2124 		fbnic_set_netif_napi(fbn->napi[i]);
2125 
2126 	return 0;
2127 }
2128 
2129 void fbnic_reset_netif_queues(struct fbnic_net *fbn)
2130 {
2131 	int i;
2132 
2133 	for (i = 0; i < fbn->num_napi; i++)
2134 		fbnic_reset_netif_napi(fbn->napi[i]);
2135 }
2136 
2137 static void fbnic_disable_twq0(struct fbnic_ring *txr)
2138 {
2139 	u32 twq_ctl = fbnic_ring_rd32(txr, FBNIC_QUEUE_TWQ0_CTL);
2140 
2141 	twq_ctl &= ~FBNIC_QUEUE_TWQ_CTL_ENABLE;
2142 
2143 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TWQ0_CTL, twq_ctl);
2144 }
2145 
2146 static void fbnic_disable_twq1(struct fbnic_ring *txr)
2147 {
2148 	u32 twq_ctl = fbnic_ring_rd32(txr, FBNIC_QUEUE_TWQ1_CTL);
2149 
2150 	twq_ctl &= ~FBNIC_QUEUE_TWQ_CTL_ENABLE;
2151 
2152 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TWQ1_CTL, twq_ctl);
2153 }
2154 
2155 static void fbnic_disable_tcq(struct fbnic_ring *txr)
2156 {
2157 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TCQ_CTL, 0);
2158 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TIM_MASK, FBNIC_QUEUE_TIM_MASK_MASK);
2159 }
2160 
2161 static void fbnic_disable_bdq(struct fbnic_ring *hpq, struct fbnic_ring *ppq)
2162 {
2163 	u32 bdq_ctl = fbnic_ring_rd32(hpq, FBNIC_QUEUE_BDQ_CTL);
2164 
2165 	bdq_ctl &= ~FBNIC_QUEUE_BDQ_CTL_ENABLE;
2166 
2167 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_CTL, bdq_ctl);
2168 }
2169 
2170 static void fbnic_disable_rcq(struct fbnic_ring *rxr)
2171 {
2172 	fbnic_ring_wr32(rxr, FBNIC_QUEUE_RCQ_CTL, 0);
2173 	fbnic_ring_wr32(rxr, FBNIC_QUEUE_RIM_MASK, FBNIC_QUEUE_RIM_MASK_MASK);
2174 }
2175 
2176 void fbnic_napi_disable(struct fbnic_net *fbn)
2177 {
2178 	int i;
2179 
2180 	for (i = 0; i < fbn->num_napi; i++) {
2181 		napi_disable(&fbn->napi[i]->napi);
2182 
2183 		fbnic_nv_irq_disable(fbn->napi[i]);
2184 	}
2185 }
2186 
2187 void fbnic_disable(struct fbnic_net *fbn)
2188 {
2189 	struct fbnic_dev *fbd = fbn->fbd;
2190 	int i, j, t;
2191 
2192 	for (i = 0; i < fbn->num_napi; i++) {
2193 		struct fbnic_napi_vector *nv = fbn->napi[i];
2194 
2195 		/* Disable Tx queue triads */
2196 		for (t = 0; t < nv->txt_count; t++) {
2197 			struct fbnic_q_triad *qt = &nv->qt[t];
2198 
2199 			fbnic_disable_twq0(&qt->sub0);
2200 			fbnic_disable_twq1(&qt->sub1);
2201 			fbnic_disable_tcq(&qt->cmpl);
2202 		}
2203 
2204 		/* Disable Rx queue triads */
2205 		for (j = 0; j < nv->rxt_count; j++, t++) {
2206 			struct fbnic_q_triad *qt = &nv->qt[t];
2207 
2208 			fbnic_disable_bdq(&qt->sub0, &qt->sub1);
2209 			fbnic_disable_rcq(&qt->cmpl);
2210 		}
2211 	}
2212 
2213 	fbnic_wrfl(fbd);
2214 }
2215 
2216 static void fbnic_tx_flush(struct fbnic_dev *fbd)
2217 {
2218 	netdev_warn(fbd->netdev, "triggering Tx flush\n");
2219 
2220 	fbnic_rmw32(fbd, FBNIC_TMI_DROP_CTRL, FBNIC_TMI_DROP_CTRL_EN,
2221 		    FBNIC_TMI_DROP_CTRL_EN);
2222 }
2223 
2224 static void fbnic_tx_flush_off(struct fbnic_dev *fbd)
2225 {
2226 	fbnic_rmw32(fbd, FBNIC_TMI_DROP_CTRL, FBNIC_TMI_DROP_CTRL_EN, 0);
2227 }
2228 
2229 struct fbnic_idle_regs {
2230 	u32 reg_base;
2231 	u8 reg_cnt;
2232 };
2233 
2234 static bool fbnic_all_idle(struct fbnic_dev *fbd,
2235 			   const struct fbnic_idle_regs *regs,
2236 			   unsigned int nregs)
2237 {
2238 	unsigned int i, j;
2239 
2240 	for (i = 0; i < nregs; i++) {
2241 		for (j = 0; j < regs[i].reg_cnt; j++) {
2242 			if (fbnic_rd32(fbd, regs[i].reg_base + j) != ~0U)
2243 				return false;
2244 		}
2245 	}
2246 	return true;
2247 }
2248 
2249 static void fbnic_idle_dump(struct fbnic_dev *fbd,
2250 			    const struct fbnic_idle_regs *regs,
2251 			    unsigned int nregs, const char *dir, int err)
2252 {
2253 	unsigned int i, j;
2254 
2255 	netdev_err(fbd->netdev, "error waiting for %s idle %d\n", dir, err);
2256 	for (i = 0; i < nregs; i++)
2257 		for (j = 0; j < regs[i].reg_cnt; j++)
2258 			netdev_err(fbd->netdev, "0x%04x: %08x\n",
2259 				   regs[i].reg_base + j,
2260 				   fbnic_rd32(fbd, regs[i].reg_base + j));
2261 }
2262 
2263 int fbnic_wait_all_queues_idle(struct fbnic_dev *fbd, bool may_fail)
2264 {
2265 	static const struct fbnic_idle_regs tx[] = {
2266 		{ FBNIC_QM_TWQ_IDLE(0),	FBNIC_QM_TWQ_IDLE_CNT, },
2267 		{ FBNIC_QM_TQS_IDLE(0),	FBNIC_QM_TQS_IDLE_CNT, },
2268 		{ FBNIC_QM_TDE_IDLE(0),	FBNIC_QM_TDE_IDLE_CNT, },
2269 		{ FBNIC_QM_TCQ_IDLE(0),	FBNIC_QM_TCQ_IDLE_CNT, },
2270 	}, rx[] = {
2271 		{ FBNIC_QM_HPQ_IDLE(0),	FBNIC_QM_HPQ_IDLE_CNT, },
2272 		{ FBNIC_QM_PPQ_IDLE(0),	FBNIC_QM_PPQ_IDLE_CNT, },
2273 		{ FBNIC_QM_RCQ_IDLE(0),	FBNIC_QM_RCQ_IDLE_CNT, },
2274 	};
2275 	bool idle;
2276 	int err;
2277 
2278 	err = read_poll_timeout_atomic(fbnic_all_idle, idle, idle, 2, 500000,
2279 				       false, fbd, tx, ARRAY_SIZE(tx));
2280 	if (err == -ETIMEDOUT) {
2281 		fbnic_tx_flush(fbd);
2282 		err = read_poll_timeout_atomic(fbnic_all_idle, idle, idle,
2283 					       2, 500000, false,
2284 					       fbd, tx, ARRAY_SIZE(tx));
2285 		fbnic_tx_flush_off(fbd);
2286 	}
2287 	if (err) {
2288 		fbnic_idle_dump(fbd, tx, ARRAY_SIZE(tx), "Tx", err);
2289 		if (may_fail)
2290 			return err;
2291 	}
2292 
2293 	err = read_poll_timeout_atomic(fbnic_all_idle, idle, idle, 2, 500000,
2294 				       false, fbd, rx, ARRAY_SIZE(rx));
2295 	if (err)
2296 		fbnic_idle_dump(fbd, rx, ARRAY_SIZE(rx), "Rx", err);
2297 	return err;
2298 }
2299 
2300 void fbnic_flush(struct fbnic_net *fbn)
2301 {
2302 	int i;
2303 
2304 	for (i = 0; i < fbn->num_napi; i++) {
2305 		struct fbnic_napi_vector *nv = fbn->napi[i];
2306 		int j, t;
2307 
2308 		/* Flush any processed Tx Queue Triads and drop the rest */
2309 		for (t = 0; t < nv->txt_count; t++) {
2310 			struct fbnic_q_triad *qt = &nv->qt[t];
2311 			struct netdev_queue *tx_queue;
2312 
2313 			/* Clean the work queues of unprocessed work */
2314 			fbnic_clean_twq0(nv, 0, &qt->sub0, true, qt->sub0.tail);
2315 			fbnic_clean_twq1(nv, false, &qt->sub1, true,
2316 					 qt->sub1.tail);
2317 
2318 			/* Reset completion queue descriptor ring */
2319 			memset(qt->cmpl.desc, 0, qt->cmpl.size);
2320 
2321 			/* Nothing else to do if Tx queue is disabled */
2322 			if (qt->sub0.flags & FBNIC_RING_F_DISABLED)
2323 				continue;
2324 
2325 			/* Reset BQL associated with Tx queue */
2326 			tx_queue = netdev_get_tx_queue(nv->napi.dev,
2327 						       qt->sub0.q_idx);
2328 			netdev_tx_reset_queue(tx_queue);
2329 		}
2330 
2331 		/* Flush any processed Rx Queue Triads and drop the rest */
2332 		for (j = 0; j < nv->rxt_count; j++, t++) {
2333 			struct fbnic_q_triad *qt = &nv->qt[t];
2334 
2335 			/* Clean the work queues of unprocessed work */
2336 			fbnic_clean_bdq(nv, 0, &qt->sub0, qt->sub0.tail);
2337 			fbnic_clean_bdq(nv, 0, &qt->sub1, qt->sub1.tail);
2338 
2339 			/* Reset completion queue descriptor ring */
2340 			memset(qt->cmpl.desc, 0, qt->cmpl.size);
2341 
2342 			fbnic_put_pkt_buff(nv, qt->cmpl.pkt, 0);
2343 			memset(qt->cmpl.pkt, 0, sizeof(struct fbnic_pkt_buff));
2344 		}
2345 	}
2346 }
2347 
2348 void fbnic_fill(struct fbnic_net *fbn)
2349 {
2350 	int i;
2351 
2352 	for (i = 0; i < fbn->num_napi; i++) {
2353 		struct fbnic_napi_vector *nv = fbn->napi[i];
2354 		int j, t;
2355 
2356 		/* Configure NAPI mapping and populate pages
2357 		 * in the BDQ rings to use for Rx
2358 		 */
2359 		for (j = 0, t = nv->txt_count; j < nv->rxt_count; j++, t++) {
2360 			struct fbnic_q_triad *qt = &nv->qt[t];
2361 
2362 			/* Populate the header and payload BDQs */
2363 			fbnic_fill_bdq(nv, &qt->sub0);
2364 			fbnic_fill_bdq(nv, &qt->sub1);
2365 		}
2366 	}
2367 }
2368 
2369 static void fbnic_enable_twq0(struct fbnic_ring *twq)
2370 {
2371 	u32 log_size = fls(twq->size_mask);
2372 
2373 	if (!twq->size_mask)
2374 		return;
2375 
2376 	/* Reset head/tail */
2377 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_CTL, FBNIC_QUEUE_TWQ_CTL_RESET);
2378 	twq->tail = 0;
2379 	twq->head = 0;
2380 
2381 	/* Store descriptor ring address and size */
2382 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_BAL, lower_32_bits(twq->dma));
2383 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_BAH, upper_32_bits(twq->dma));
2384 
2385 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2386 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_SIZE, log_size & 0xf);
2387 
2388 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_CTL, FBNIC_QUEUE_TWQ_CTL_ENABLE);
2389 }
2390 
2391 static void fbnic_enable_twq1(struct fbnic_ring *twq)
2392 {
2393 	u32 log_size = fls(twq->size_mask);
2394 
2395 	if (!twq->size_mask)
2396 		return;
2397 
2398 	/* Reset head/tail */
2399 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_CTL, FBNIC_QUEUE_TWQ_CTL_RESET);
2400 	twq->tail = 0;
2401 	twq->head = 0;
2402 
2403 	/* Store descriptor ring address and size */
2404 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_BAL, lower_32_bits(twq->dma));
2405 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_BAH, upper_32_bits(twq->dma));
2406 
2407 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2408 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_SIZE, log_size & 0xf);
2409 
2410 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_CTL, FBNIC_QUEUE_TWQ_CTL_ENABLE);
2411 }
2412 
2413 static void fbnic_enable_tcq(struct fbnic_napi_vector *nv,
2414 			     struct fbnic_ring *tcq)
2415 {
2416 	u32 log_size = fls(tcq->size_mask);
2417 
2418 	if (!tcq->size_mask)
2419 		return;
2420 
2421 	/* Reset head/tail */
2422 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_CTL, FBNIC_QUEUE_TCQ_CTL_RESET);
2423 	tcq->tail = 0;
2424 	tcq->head = 0;
2425 
2426 	/* Store descriptor ring address and size */
2427 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_BAL, lower_32_bits(tcq->dma));
2428 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_BAH, upper_32_bits(tcq->dma));
2429 
2430 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2431 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_SIZE, log_size & 0xf);
2432 
2433 	/* Store interrupt information for the completion queue */
2434 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TIM_CTL, nv->v_idx);
2435 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TIM_THRESHOLD, tcq->size_mask / 2);
2436 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TIM_MASK, 0);
2437 
2438 	/* Enable queue */
2439 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_CTL, FBNIC_QUEUE_TCQ_CTL_ENABLE);
2440 }
2441 
2442 static void fbnic_enable_bdq(struct fbnic_ring *hpq, struct fbnic_ring *ppq)
2443 {
2444 	u32 bdq_ctl = FBNIC_QUEUE_BDQ_CTL_ENABLE;
2445 	u32 log_size;
2446 
2447 	/* Reset head/tail */
2448 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_CTL, FBNIC_QUEUE_BDQ_CTL_RESET);
2449 	ppq->tail = 0;
2450 	ppq->head = 0;
2451 	hpq->tail = 0;
2452 	hpq->head = 0;
2453 
2454 	log_size = fls(hpq->size_mask);
2455 
2456 	/* Store descriptor ring address and size */
2457 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_HPQ_BAL, lower_32_bits(hpq->dma));
2458 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_HPQ_BAH, upper_32_bits(hpq->dma));
2459 
2460 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2461 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_HPQ_SIZE, log_size & 0xf);
2462 
2463 	if (!ppq->size_mask)
2464 		goto write_ctl;
2465 
2466 	log_size = fls(ppq->size_mask);
2467 
2468 	/* Add enabling of PPQ to BDQ control */
2469 	bdq_ctl |= FBNIC_QUEUE_BDQ_CTL_PPQ_ENABLE;
2470 
2471 	/* Store descriptor ring address and size */
2472 	fbnic_ring_wr32(ppq, FBNIC_QUEUE_BDQ_PPQ_BAL, lower_32_bits(ppq->dma));
2473 	fbnic_ring_wr32(ppq, FBNIC_QUEUE_BDQ_PPQ_BAH, upper_32_bits(ppq->dma));
2474 	fbnic_ring_wr32(ppq, FBNIC_QUEUE_BDQ_PPQ_SIZE, log_size & 0xf);
2475 
2476 write_ctl:
2477 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_CTL, bdq_ctl);
2478 }
2479 
2480 static void fbnic_config_drop_mode_rcq(struct fbnic_napi_vector *nv,
2481 				       struct fbnic_ring *rcq)
2482 {
2483 	u32 drop_mode, rcq_ctl;
2484 
2485 	drop_mode = FBNIC_QUEUE_RDE_CTL0_DROP_IMMEDIATE;
2486 
2487 	/* Specify packet layout */
2488 	rcq_ctl = FIELD_PREP(FBNIC_QUEUE_RDE_CTL0_DROP_MODE_MASK, drop_mode) |
2489 	    FIELD_PREP(FBNIC_QUEUE_RDE_CTL0_MIN_HROOM_MASK, FBNIC_RX_HROOM) |
2490 	    FIELD_PREP(FBNIC_QUEUE_RDE_CTL0_MIN_TROOM_MASK, FBNIC_RX_TROOM);
2491 
2492 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RDE_CTL0, rcq_ctl);
2493 }
2494 
2495 static void fbnic_config_rim_threshold(struct fbnic_ring *rcq, u16 nv_idx, u32 rx_desc)
2496 {
2497 	u32 threshold;
2498 
2499 	/* Set the threhsold to half the ring size if rx_frames
2500 	 * is not configured
2501 	 */
2502 	threshold = rx_desc ? : rcq->size_mask / 2;
2503 
2504 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RIM_CTL, nv_idx);
2505 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RIM_THRESHOLD, threshold);
2506 }
2507 
2508 void fbnic_config_txrx_usecs(struct fbnic_napi_vector *nv, u32 arm)
2509 {
2510 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2511 	struct fbnic_dev *fbd = nv->fbd;
2512 	u32 val = arm;
2513 
2514 	val |= FIELD_PREP(FBNIC_INTR_CQ_REARM_RCQ_TIMEOUT, fbn->rx_usecs) |
2515 	       FBNIC_INTR_CQ_REARM_RCQ_TIMEOUT_UPD_EN;
2516 	val |= FIELD_PREP(FBNIC_INTR_CQ_REARM_TCQ_TIMEOUT, fbn->tx_usecs) |
2517 	       FBNIC_INTR_CQ_REARM_TCQ_TIMEOUT_UPD_EN;
2518 
2519 	fbnic_wr32(fbd, FBNIC_INTR_CQ_REARM(nv->v_idx), val);
2520 }
2521 
2522 void fbnic_config_rx_frames(struct fbnic_napi_vector *nv)
2523 {
2524 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2525 	int i;
2526 
2527 	for (i = nv->txt_count; i < nv->rxt_count + nv->txt_count; i++) {
2528 		struct fbnic_q_triad *qt = &nv->qt[i];
2529 
2530 		fbnic_config_rim_threshold(&qt->cmpl, nv->v_idx,
2531 					   fbn->rx_max_frames *
2532 					   FBNIC_MIN_RXD_PER_FRAME);
2533 	}
2534 }
2535 
2536 static void fbnic_enable_rcq(struct fbnic_napi_vector *nv,
2537 			     struct fbnic_ring *rcq)
2538 {
2539 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2540 	u32 log_size = fls(rcq->size_mask);
2541 	u32 hds_thresh = fbn->hds_thresh;
2542 	u32 rcq_ctl = 0;
2543 
2544 	fbnic_config_drop_mode_rcq(nv, rcq);
2545 
2546 	/* Force lower bound on MAX_HEADER_BYTES. Below this, all frames should
2547 	 * be split at L4. It would also result in the frames being split at
2548 	 * L2/L3 depending on the frame size.
2549 	 */
2550 	if (fbn->hds_thresh < FBNIC_HDR_BYTES_MIN) {
2551 		rcq_ctl = FBNIC_QUEUE_RDE_CTL0_EN_HDR_SPLIT;
2552 		hds_thresh = FBNIC_HDR_BYTES_MIN;
2553 	}
2554 
2555 	rcq_ctl |= FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_PADLEN_MASK, FBNIC_RX_PAD) |
2556 		   FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_MAX_HDR_MASK, hds_thresh) |
2557 		   FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_PAYLD_OFF_MASK,
2558 			      FBNIC_RX_PAYLD_OFFSET) |
2559 		   FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_PAYLD_PG_CL_MASK,
2560 			      FBNIC_RX_PAYLD_PG_CL);
2561 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RDE_CTL1, rcq_ctl);
2562 
2563 	/* Reset head/tail */
2564 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_CTL, FBNIC_QUEUE_RCQ_CTL_RESET);
2565 	rcq->head = 0;
2566 	rcq->tail = 0;
2567 
2568 	/* Store descriptor ring address and size */
2569 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_BAL, lower_32_bits(rcq->dma));
2570 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_BAH, upper_32_bits(rcq->dma));
2571 
2572 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2573 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_SIZE, log_size & 0xf);
2574 
2575 	/* Store interrupt information for the completion queue */
2576 	fbnic_config_rim_threshold(rcq, nv->v_idx, fbn->rx_max_frames *
2577 						   FBNIC_MIN_RXD_PER_FRAME);
2578 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RIM_MASK, 0);
2579 
2580 	/* Enable queue */
2581 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_CTL, FBNIC_QUEUE_RCQ_CTL_ENABLE);
2582 }
2583 
2584 void fbnic_enable(struct fbnic_net *fbn)
2585 {
2586 	struct fbnic_dev *fbd = fbn->fbd;
2587 	int i;
2588 
2589 	for (i = 0; i < fbn->num_napi; i++) {
2590 		struct fbnic_napi_vector *nv = fbn->napi[i];
2591 		int j, t;
2592 
2593 		/* Setup Tx Queue Triads */
2594 		for (t = 0; t < nv->txt_count; t++) {
2595 			struct fbnic_q_triad *qt = &nv->qt[t];
2596 
2597 			fbnic_enable_twq0(&qt->sub0);
2598 			fbnic_enable_twq1(&qt->sub1);
2599 			fbnic_enable_tcq(nv, &qt->cmpl);
2600 		}
2601 
2602 		/* Setup Rx Queue Triads */
2603 		for (j = 0; j < nv->rxt_count; j++, t++) {
2604 			struct fbnic_q_triad *qt = &nv->qt[t];
2605 
2606 			fbnic_enable_bdq(&qt->sub0, &qt->sub1);
2607 			fbnic_config_drop_mode_rcq(nv, &qt->cmpl);
2608 			fbnic_enable_rcq(nv, &qt->cmpl);
2609 		}
2610 	}
2611 
2612 	fbnic_wrfl(fbd);
2613 }
2614 
2615 static void fbnic_nv_irq_enable(struct fbnic_napi_vector *nv)
2616 {
2617 	fbnic_config_txrx_usecs(nv, FBNIC_INTR_CQ_REARM_INTR_UNMASK);
2618 }
2619 
2620 void fbnic_napi_enable(struct fbnic_net *fbn)
2621 {
2622 	u32 irqs[FBNIC_MAX_MSIX_VECS / 32] = {};
2623 	struct fbnic_dev *fbd = fbn->fbd;
2624 	int i;
2625 
2626 	for (i = 0; i < fbn->num_napi; i++) {
2627 		struct fbnic_napi_vector *nv = fbn->napi[i];
2628 
2629 		napi_enable(&nv->napi);
2630 
2631 		fbnic_nv_irq_enable(nv);
2632 
2633 		/* Record bit used for NAPI IRQs so we can
2634 		 * set the mask appropriately
2635 		 */
2636 		irqs[nv->v_idx / 32] |= BIT(nv->v_idx % 32);
2637 	}
2638 
2639 	/* Force the first interrupt on the device to guarantee
2640 	 * that any packets that may have been enqueued during the
2641 	 * bringup are processed.
2642 	 */
2643 	for (i = 0; i < ARRAY_SIZE(irqs); i++) {
2644 		if (!irqs[i])
2645 			continue;
2646 		fbnic_wr32(fbd, FBNIC_INTR_SET(i), irqs[i]);
2647 	}
2648 
2649 	fbnic_wrfl(fbd);
2650 }
2651 
2652 void fbnic_napi_depletion_check(struct net_device *netdev)
2653 {
2654 	struct fbnic_net *fbn = netdev_priv(netdev);
2655 	u32 irqs[FBNIC_MAX_MSIX_VECS / 32] = {};
2656 	struct fbnic_dev *fbd = fbn->fbd;
2657 	int i, j, t;
2658 
2659 	for (i = 0; i < fbn->num_napi; i++) {
2660 		struct fbnic_napi_vector *nv = fbn->napi[i];
2661 
2662 		/* Find RQs which are completely out of pages */
2663 		for (t = nv->txt_count, j = 0; j < nv->rxt_count; j++, t++) {
2664 			/* Assume 4 pages is always enough to fit a packet
2665 			 * and therefore generate a completion and an IRQ.
2666 			 */
2667 			if (fbnic_desc_used(&nv->qt[t].sub0) < 4 ||
2668 			    fbnic_desc_used(&nv->qt[t].sub1) < 4)
2669 				irqs[nv->v_idx / 32] |= BIT(nv->v_idx % 32);
2670 		}
2671 	}
2672 
2673 	for (i = 0; i < ARRAY_SIZE(irqs); i++) {
2674 		if (!irqs[i])
2675 			continue;
2676 		fbnic_wr32(fbd, FBNIC_INTR_MASK_CLEAR(i), irqs[i]);
2677 		fbnic_wr32(fbd, FBNIC_INTR_SET(i), irqs[i]);
2678 	}
2679 
2680 	fbnic_wrfl(fbd);
2681 }
2682