xref: /linux/drivers/net/ethernet/meta/fbnic/fbnic_txrx.c (revision 8f7aa3d3c7323f4ca2768a9e74ebbe359c4f8f88)
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 #define FBNIC_XMIT_NOUNMAP	((void *)1)
41 
42 static u32 __iomem *fbnic_ring_csr_base(const struct fbnic_ring *ring)
43 {
44 	unsigned long csr_base = (unsigned long)ring->doorbell;
45 
46 	csr_base &= ~(FBNIC_QUEUE_STRIDE * sizeof(u32) - 1);
47 
48 	return (u32 __iomem *)csr_base;
49 }
50 
51 static u32 fbnic_ring_rd32(struct fbnic_ring *ring, unsigned int csr)
52 {
53 	u32 __iomem *csr_base = fbnic_ring_csr_base(ring);
54 
55 	return readl(csr_base + csr);
56 }
57 
58 static void fbnic_ring_wr32(struct fbnic_ring *ring, unsigned int csr, u32 val)
59 {
60 	u32 __iomem *csr_base = fbnic_ring_csr_base(ring);
61 
62 	writel(val, csr_base + csr);
63 }
64 
65 /**
66  * fbnic_ts40_to_ns() - convert descriptor timestamp to PHC time
67  * @fbn: netdev priv of the FB NIC
68  * @ts40: timestamp read from a descriptor
69  *
70  * Return: u64 value of PHC time in nanoseconds
71  *
72  * Convert truncated 40 bit device timestamp as read from a descriptor
73  * to the full PHC time in nanoseconds.
74  */
75 static __maybe_unused u64 fbnic_ts40_to_ns(struct fbnic_net *fbn, u64 ts40)
76 {
77 	unsigned int s;
78 	u64 time_ns;
79 	s64 offset;
80 	u8 ts_top;
81 	u32 high;
82 
83 	do {
84 		s = u64_stats_fetch_begin(&fbn->time_seq);
85 		offset = READ_ONCE(fbn->time_offset);
86 	} while (u64_stats_fetch_retry(&fbn->time_seq, s));
87 
88 	high = READ_ONCE(fbn->time_high);
89 
90 	/* Bits 63..40 from periodic clock reads, 39..0 from ts40 */
91 	time_ns = (u64)(high >> 8) << 40 | ts40;
92 
93 	/* Compare bits 32-39 between periodic reads and ts40,
94 	 * see if HW clock may have wrapped since last read. We are sure
95 	 * that periodic reads are always at least ~1 minute behind, so
96 	 * this logic works perfectly fine.
97 	 */
98 	ts_top = ts40 >> 32;
99 	if (ts_top < (u8)high && (u8)high - ts_top > U8_MAX / 2)
100 		time_ns += 1ULL << 40;
101 
102 	return time_ns + offset;
103 }
104 
105 static unsigned int fbnic_desc_unused(struct fbnic_ring *ring)
106 {
107 	return (ring->head - ring->tail - 1) & ring->size_mask;
108 }
109 
110 static unsigned int fbnic_desc_used(struct fbnic_ring *ring)
111 {
112 	return (ring->tail - ring->head) & ring->size_mask;
113 }
114 
115 static struct netdev_queue *txring_txq(const struct net_device *dev,
116 				       const struct fbnic_ring *ring)
117 {
118 	return netdev_get_tx_queue(dev, ring->q_idx);
119 }
120 
121 static int fbnic_maybe_stop_tx(const struct net_device *dev,
122 			       struct fbnic_ring *ring,
123 			       const unsigned int size)
124 {
125 	struct netdev_queue *txq = txring_txq(dev, ring);
126 	int res;
127 
128 	res = netif_txq_maybe_stop(txq, fbnic_desc_unused(ring), size,
129 				   FBNIC_TX_DESC_WAKEUP);
130 	if (!res) {
131 		u64_stats_update_begin(&ring->stats.syncp);
132 		ring->stats.twq.stop++;
133 		u64_stats_update_end(&ring->stats.syncp);
134 	}
135 
136 	return !res;
137 }
138 
139 static bool fbnic_tx_sent_queue(struct sk_buff *skb, struct fbnic_ring *ring)
140 {
141 	struct netdev_queue *dev_queue = txring_txq(skb->dev, ring);
142 	unsigned int bytecount = FBNIC_XMIT_CB(skb)->bytecount;
143 	bool xmit_more = netdev_xmit_more();
144 
145 	/* TBD: Request completion more often if xmit_more becomes large */
146 
147 	return __netdev_tx_sent_queue(dev_queue, bytecount, xmit_more);
148 }
149 
150 static void fbnic_unmap_single_twd(struct device *dev, __le64 *twd)
151 {
152 	u64 raw_twd = le64_to_cpu(*twd);
153 	unsigned int len;
154 	dma_addr_t dma;
155 
156 	dma = FIELD_GET(FBNIC_TWD_ADDR_MASK, raw_twd);
157 	len = FIELD_GET(FBNIC_TWD_LEN_MASK, raw_twd);
158 
159 	dma_unmap_single(dev, dma, len, DMA_TO_DEVICE);
160 }
161 
162 static void fbnic_unmap_page_twd(struct device *dev, __le64 *twd)
163 {
164 	u64 raw_twd = le64_to_cpu(*twd);
165 	unsigned int len;
166 	dma_addr_t dma;
167 
168 	dma = FIELD_GET(FBNIC_TWD_ADDR_MASK, raw_twd);
169 	len = FIELD_GET(FBNIC_TWD_LEN_MASK, raw_twd);
170 
171 	dma_unmap_page(dev, dma, len, DMA_TO_DEVICE);
172 }
173 
174 #define FBNIC_TWD_TYPE(_type) \
175 	cpu_to_le64(FIELD_PREP(FBNIC_TWD_TYPE_MASK, FBNIC_TWD_TYPE_##_type))
176 
177 static bool fbnic_tx_tstamp(struct sk_buff *skb)
178 {
179 	struct fbnic_net *fbn;
180 
181 	if (!unlikely(skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP))
182 		return false;
183 
184 	fbn = netdev_priv(skb->dev);
185 	if (fbn->hwtstamp_config.tx_type == HWTSTAMP_TX_OFF)
186 		return false;
187 
188 	skb_shinfo(skb)->tx_flags |= SKBTX_IN_PROGRESS;
189 	FBNIC_XMIT_CB(skb)->flags |= FBNIC_XMIT_CB_TS;
190 	FBNIC_XMIT_CB(skb)->hw_head = -1;
191 
192 	return true;
193 }
194 
195 static bool
196 fbnic_tx_lso(struct fbnic_ring *ring, struct sk_buff *skb,
197 	     struct skb_shared_info *shinfo, __le64 *meta,
198 	     unsigned int *l2len, unsigned int *i3len)
199 {
200 	unsigned int l3_type, l4_type, l4len, hdrlen;
201 	unsigned char *l4hdr;
202 	__be16 payload_len;
203 
204 	if (unlikely(skb_cow_head(skb, 0)))
205 		return true;
206 
207 	if (shinfo->gso_type & SKB_GSO_PARTIAL) {
208 		l3_type = FBNIC_TWD_L3_TYPE_OTHER;
209 	} else if (!skb->encapsulation) {
210 		if (ip_hdr(skb)->version == 4)
211 			l3_type = FBNIC_TWD_L3_TYPE_IPV4;
212 		else
213 			l3_type = FBNIC_TWD_L3_TYPE_IPV6;
214 	} else {
215 		unsigned int o3len;
216 
217 		o3len = skb_inner_network_header(skb) - skb_network_header(skb);
218 		*i3len -= o3len;
219 		*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_L3_OHLEN_MASK,
220 						o3len / 2));
221 		l3_type = FBNIC_TWD_L3_TYPE_V6V6;
222 	}
223 
224 	l4hdr = skb_checksum_start(skb);
225 	payload_len = cpu_to_be16(skb->len - (l4hdr - skb->data));
226 
227 	if (shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)) {
228 		struct tcphdr *tcph = (struct tcphdr *)l4hdr;
229 
230 		l4_type = FBNIC_TWD_L4_TYPE_TCP;
231 		l4len = __tcp_hdrlen((struct tcphdr *)l4hdr);
232 		csum_replace_by_diff(&tcph->check, (__force __wsum)payload_len);
233 	} else {
234 		struct udphdr *udph = (struct udphdr *)l4hdr;
235 
236 		l4_type = FBNIC_TWD_L4_TYPE_UDP;
237 		l4len = sizeof(struct udphdr);
238 		csum_replace_by_diff(&udph->check, (__force __wsum)payload_len);
239 	}
240 
241 	hdrlen = (l4hdr - skb->data) + l4len;
242 	*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_L3_TYPE_MASK, l3_type) |
243 			     FIELD_PREP(FBNIC_TWD_L4_TYPE_MASK, l4_type) |
244 			     FIELD_PREP(FBNIC_TWD_L4_HLEN_MASK, l4len / 4) |
245 			     FIELD_PREP(FBNIC_TWD_MSS_MASK, shinfo->gso_size) |
246 			     FBNIC_TWD_FLAG_REQ_LSO);
247 
248 	FBNIC_XMIT_CB(skb)->bytecount += (shinfo->gso_segs - 1) * hdrlen;
249 	FBNIC_XMIT_CB(skb)->gso_segs = shinfo->gso_segs;
250 
251 	u64_stats_update_begin(&ring->stats.syncp);
252 	ring->stats.twq.lso += shinfo->gso_segs;
253 	u64_stats_update_end(&ring->stats.syncp);
254 
255 	return false;
256 }
257 
258 static bool
259 fbnic_tx_offloads(struct fbnic_ring *ring, struct sk_buff *skb, __le64 *meta)
260 {
261 	struct skb_shared_info *shinfo = skb_shinfo(skb);
262 	unsigned int l2len, i3len;
263 
264 	if (fbnic_tx_tstamp(skb))
265 		*meta |= cpu_to_le64(FBNIC_TWD_FLAG_REQ_TS);
266 
267 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL))
268 		return false;
269 
270 	l2len = skb_mac_header_len(skb);
271 	i3len = skb_checksum_start(skb) - skb_network_header(skb);
272 
273 	*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_CSUM_OFFSET_MASK,
274 					skb->csum_offset / 2));
275 
276 	if (shinfo->gso_size) {
277 		if (fbnic_tx_lso(ring, skb, shinfo, meta, &l2len, &i3len))
278 			return true;
279 	} else {
280 		*meta |= cpu_to_le64(FBNIC_TWD_FLAG_REQ_CSO);
281 		u64_stats_update_begin(&ring->stats.syncp);
282 		ring->stats.twq.csum_partial++;
283 		u64_stats_update_end(&ring->stats.syncp);
284 	}
285 
286 	*meta |= cpu_to_le64(FIELD_PREP(FBNIC_TWD_L2_HLEN_MASK, l2len / 2) |
287 			     FIELD_PREP(FBNIC_TWD_L3_IHLEN_MASK, i3len / 2));
288 	return false;
289 }
290 
291 static void
292 fbnic_rx_csum(u64 rcd, struct sk_buff *skb, struct fbnic_ring *rcq,
293 	      u64 *csum_cmpl, u64 *csum_none)
294 {
295 	skb_checksum_none_assert(skb);
296 
297 	if (unlikely(!(skb->dev->features & NETIF_F_RXCSUM))) {
298 		(*csum_none)++;
299 		return;
300 	}
301 
302 	if (FIELD_GET(FBNIC_RCD_META_L4_CSUM_UNNECESSARY, rcd)) {
303 		skb->ip_summed = CHECKSUM_UNNECESSARY;
304 	} else {
305 		u16 csum = FIELD_GET(FBNIC_RCD_META_L2_CSUM_MASK, rcd);
306 
307 		skb->ip_summed = CHECKSUM_COMPLETE;
308 		skb->csum = (__force __wsum)csum;
309 		(*csum_cmpl)++;
310 	}
311 }
312 
313 static bool
314 fbnic_tx_map(struct fbnic_ring *ring, struct sk_buff *skb, __le64 *meta)
315 {
316 	struct device *dev = skb->dev->dev.parent;
317 	unsigned int tail = ring->tail, first;
318 	unsigned int size, data_len;
319 	skb_frag_t *frag;
320 	bool is_net_iov;
321 	dma_addr_t dma;
322 	__le64 *twd;
323 
324 	ring->tx_buf[tail] = skb;
325 
326 	tail++;
327 	tail &= ring->size_mask;
328 	first = tail;
329 
330 	size = skb_headlen(skb);
331 	data_len = skb->data_len;
332 
333 	if (size > FIELD_MAX(FBNIC_TWD_LEN_MASK))
334 		goto dma_error;
335 
336 	is_net_iov = false;
337 	dma = dma_map_single(dev, skb->data, size, DMA_TO_DEVICE);
338 
339 	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
340 		twd = &ring->desc[tail];
341 
342 		if (dma_mapping_error(dev, dma))
343 			goto dma_error;
344 
345 		*twd = cpu_to_le64(FIELD_PREP(FBNIC_TWD_ADDR_MASK, dma) |
346 				   FIELD_PREP(FBNIC_TWD_LEN_MASK, size) |
347 				   FIELD_PREP(FBNIC_TWD_TYPE_MASK,
348 					      FBNIC_TWD_TYPE_AL));
349 		if (is_net_iov)
350 			ring->tx_buf[tail] = FBNIC_XMIT_NOUNMAP;
351 
352 		tail++;
353 		tail &= ring->size_mask;
354 
355 		if (!data_len)
356 			break;
357 
358 		size = skb_frag_size(frag);
359 		data_len -= size;
360 
361 		if (size > FIELD_MAX(FBNIC_TWD_LEN_MASK))
362 			goto dma_error;
363 
364 		is_net_iov = skb_frag_is_net_iov(frag);
365 		dma = skb_frag_dma_map(dev, frag, 0, size, DMA_TO_DEVICE);
366 	}
367 
368 	*twd |= FBNIC_TWD_TYPE(LAST_AL);
369 
370 	FBNIC_XMIT_CB(skb)->desc_count = ((twd - meta) + 1) & ring->size_mask;
371 
372 	ring->tail = tail;
373 
374 	/* Record SW timestamp */
375 	skb_tx_timestamp(skb);
376 
377 	/* Verify there is room for another packet */
378 	fbnic_maybe_stop_tx(skb->dev, ring, FBNIC_MAX_SKB_DESC);
379 
380 	if (fbnic_tx_sent_queue(skb, ring)) {
381 		*meta |= cpu_to_le64(FBNIC_TWD_FLAG_REQ_COMPLETION);
382 
383 		/* Force DMA writes to flush before writing to tail */
384 		dma_wmb();
385 
386 		writel(tail, ring->doorbell);
387 	}
388 
389 	return false;
390 dma_error:
391 	if (net_ratelimit())
392 		netdev_err(skb->dev, "TX DMA map failed\n");
393 
394 	while (tail != first) {
395 		tail--;
396 		tail &= ring->size_mask;
397 		twd = &ring->desc[tail];
398 		if (tail == first)
399 			fbnic_unmap_single_twd(dev, twd);
400 		else if (ring->tx_buf[tail] == FBNIC_XMIT_NOUNMAP)
401 			ring->tx_buf[tail] = NULL;
402 		else
403 			fbnic_unmap_page_twd(dev, twd);
404 	}
405 
406 	return true;
407 }
408 
409 #define FBNIC_MIN_FRAME_LEN	60
410 
411 static netdev_tx_t
412 fbnic_xmit_frame_ring(struct sk_buff *skb, struct fbnic_ring *ring)
413 {
414 	__le64 *meta = &ring->desc[ring->tail];
415 	u16 desc_needed;
416 
417 	if (skb_put_padto(skb, FBNIC_MIN_FRAME_LEN))
418 		goto err_count;
419 
420 	/* Need: 1 descriptor per page,
421 	 *       + 1 desc for skb_head,
422 	 *       + 2 desc for metadata and timestamp metadata
423 	 *       + 7 desc gap to keep tail from touching head
424 	 * otherwise try next time
425 	 */
426 	desc_needed = skb_shinfo(skb)->nr_frags + 10;
427 	if (fbnic_maybe_stop_tx(skb->dev, ring, desc_needed))
428 		return NETDEV_TX_BUSY;
429 
430 	*meta = cpu_to_le64(FBNIC_TWD_FLAG_DEST_MAC);
431 
432 	/* Write all members within DWORD to condense this into 2 4B writes */
433 	FBNIC_XMIT_CB(skb)->bytecount = skb->len;
434 	FBNIC_XMIT_CB(skb)->gso_segs = 1;
435 	FBNIC_XMIT_CB(skb)->desc_count = 0;
436 	FBNIC_XMIT_CB(skb)->flags = 0;
437 
438 	if (fbnic_tx_offloads(ring, skb, meta))
439 		goto err_free;
440 
441 	if (fbnic_tx_map(ring, skb, meta))
442 		goto err_free;
443 
444 	return NETDEV_TX_OK;
445 
446 err_free:
447 	dev_kfree_skb_any(skb);
448 err_count:
449 	u64_stats_update_begin(&ring->stats.syncp);
450 	ring->stats.dropped++;
451 	u64_stats_update_end(&ring->stats.syncp);
452 	return NETDEV_TX_OK;
453 }
454 
455 netdev_tx_t fbnic_xmit_frame(struct sk_buff *skb, struct net_device *dev)
456 {
457 	struct fbnic_net *fbn = netdev_priv(dev);
458 	unsigned int q_map = skb->queue_mapping;
459 
460 	return fbnic_xmit_frame_ring(skb, fbn->tx[q_map]);
461 }
462 
463 static netdev_features_t
464 fbnic_features_check_encap_gso(struct sk_buff *skb, struct net_device *dev,
465 			       netdev_features_t features, unsigned int l3len)
466 {
467 	netdev_features_t skb_gso_features;
468 	struct ipv6hdr *ip6_hdr;
469 	unsigned char l4_hdr;
470 	unsigned int start;
471 	__be16 frag_off;
472 
473 	/* Require MANGLEID for GSO_PARTIAL of IPv4.
474 	 * In theory we could support TSO with single, innermost v4 header
475 	 * by pretending everything before it is L2, but that needs to be
476 	 * parsed case by case.. so leaving it for when the need arises.
477 	 */
478 	if (!(features & NETIF_F_TSO_MANGLEID))
479 		features &= ~NETIF_F_TSO;
480 
481 	skb_gso_features = skb_shinfo(skb)->gso_type;
482 	skb_gso_features <<= NETIF_F_GSO_SHIFT;
483 
484 	/* We'd only clear the native GSO features, so don't bother validating
485 	 * if the match can only be on those supported thru GSO_PARTIAL.
486 	 */
487 	if (!(skb_gso_features & FBNIC_TUN_GSO_FEATURES))
488 		return features;
489 
490 	/* We can only do IPv6-in-IPv6, not v4-in-v6. It'd be nice
491 	 * to fall back to partial for this, or any failure below.
492 	 * This is just an optimization, UDPv4 will be caught later on.
493 	 */
494 	if (skb_gso_features & NETIF_F_TSO)
495 		return features & ~FBNIC_TUN_GSO_FEATURES;
496 
497 	/* Inner headers multiple of 2 */
498 	if ((skb_inner_network_header(skb) - skb_network_header(skb)) % 2)
499 		return features & ~FBNIC_TUN_GSO_FEATURES;
500 
501 	/* Encapsulated GSO packet, make 100% sure it's IPv6-in-IPv6. */
502 	ip6_hdr = ipv6_hdr(skb);
503 	if (ip6_hdr->version != 6)
504 		return features & ~FBNIC_TUN_GSO_FEATURES;
505 
506 	l4_hdr = ip6_hdr->nexthdr;
507 	start = (unsigned char *)ip6_hdr - skb->data + sizeof(struct ipv6hdr);
508 	start = ipv6_skip_exthdr(skb, start, &l4_hdr, &frag_off);
509 	if (frag_off || l4_hdr != IPPROTO_IPV6 ||
510 	    skb->data + start != skb_inner_network_header(skb))
511 		return features & ~FBNIC_TUN_GSO_FEATURES;
512 
513 	return features;
514 }
515 
516 netdev_features_t
517 fbnic_features_check(struct sk_buff *skb, struct net_device *dev,
518 		     netdev_features_t features)
519 {
520 	unsigned int l2len, l3len;
521 
522 	if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL))
523 		return features;
524 
525 	l2len = skb_mac_header_len(skb);
526 	l3len = skb_checksum_start(skb) - skb_network_header(skb);
527 
528 	/* Check header lengths are multiple of 2.
529 	 * In case of 6in6 we support longer headers (IHLEN + OHLEN)
530 	 * but keep things simple for now, 512B is plenty.
531 	 */
532 	if ((l2len | l3len | skb->csum_offset) % 2 ||
533 	    !FIELD_FIT(FBNIC_TWD_L2_HLEN_MASK, l2len / 2) ||
534 	    !FIELD_FIT(FBNIC_TWD_L3_IHLEN_MASK, l3len / 2) ||
535 	    !FIELD_FIT(FBNIC_TWD_CSUM_OFFSET_MASK, skb->csum_offset / 2))
536 		return features & ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
537 
538 	if (likely(!skb->encapsulation) || !skb_is_gso(skb))
539 		return features;
540 
541 	return fbnic_features_check_encap_gso(skb, dev, features, l3len);
542 }
543 
544 static void fbnic_clean_twq0(struct fbnic_napi_vector *nv, int napi_budget,
545 			     struct fbnic_ring *ring, bool discard,
546 			     unsigned int hw_head)
547 {
548 	u64 total_bytes = 0, total_packets = 0, ts_lost = 0;
549 	unsigned int head = ring->head;
550 	struct netdev_queue *txq;
551 	unsigned int clean_desc;
552 
553 	clean_desc = (hw_head - head) & ring->size_mask;
554 
555 	while (clean_desc) {
556 		struct sk_buff *skb = ring->tx_buf[head];
557 		unsigned int desc_cnt;
558 
559 		desc_cnt = FBNIC_XMIT_CB(skb)->desc_count;
560 		if (desc_cnt > clean_desc)
561 			break;
562 
563 		if (unlikely(FBNIC_XMIT_CB(skb)->flags & FBNIC_XMIT_CB_TS)) {
564 			FBNIC_XMIT_CB(skb)->hw_head = hw_head;
565 			if (likely(!discard))
566 				break;
567 			ts_lost++;
568 		}
569 
570 		ring->tx_buf[head] = NULL;
571 
572 		clean_desc -= desc_cnt;
573 
574 		while (!(ring->desc[head] & FBNIC_TWD_TYPE(AL))) {
575 			head++;
576 			head &= ring->size_mask;
577 			desc_cnt--;
578 		}
579 
580 		fbnic_unmap_single_twd(nv->dev, &ring->desc[head]);
581 		head++;
582 		head &= ring->size_mask;
583 		desc_cnt--;
584 
585 		while (desc_cnt--) {
586 			if (ring->tx_buf[head] != FBNIC_XMIT_NOUNMAP)
587 				fbnic_unmap_page_twd(nv->dev,
588 						     &ring->desc[head]);
589 			else
590 				ring->tx_buf[head] = NULL;
591 			head++;
592 			head &= ring->size_mask;
593 		}
594 
595 		total_bytes += FBNIC_XMIT_CB(skb)->bytecount;
596 		total_packets += FBNIC_XMIT_CB(skb)->gso_segs;
597 
598 		napi_consume_skb(skb, napi_budget);
599 	}
600 
601 	if (!total_bytes)
602 		return;
603 
604 	ring->head = head;
605 
606 	txq = txring_txq(nv->napi.dev, ring);
607 
608 	if (unlikely(discard)) {
609 		u64_stats_update_begin(&ring->stats.syncp);
610 		ring->stats.dropped += total_packets;
611 		ring->stats.twq.ts_lost += ts_lost;
612 		u64_stats_update_end(&ring->stats.syncp);
613 
614 		netdev_tx_completed_queue(txq, total_packets, total_bytes);
615 		return;
616 	}
617 
618 	u64_stats_update_begin(&ring->stats.syncp);
619 	ring->stats.bytes += total_bytes;
620 	ring->stats.packets += total_packets;
621 	u64_stats_update_end(&ring->stats.syncp);
622 
623 	if (!netif_txq_completed_wake(txq, total_packets, total_bytes,
624 				      fbnic_desc_unused(ring),
625 				      FBNIC_TX_DESC_WAKEUP)) {
626 		u64_stats_update_begin(&ring->stats.syncp);
627 		ring->stats.twq.wake++;
628 		u64_stats_update_end(&ring->stats.syncp);
629 	}
630 }
631 
632 static void fbnic_clean_twq1(struct fbnic_napi_vector *nv, bool pp_allow_direct,
633 			     struct fbnic_ring *ring, bool discard,
634 			     unsigned int hw_head)
635 {
636 	u64 total_bytes = 0, total_packets = 0;
637 	unsigned int head = ring->head;
638 
639 	while (hw_head != head) {
640 		struct page *page;
641 		u64 twd;
642 
643 		if (unlikely(!(ring->desc[head] & FBNIC_TWD_TYPE(AL))))
644 			goto next_desc;
645 
646 		twd = le64_to_cpu(ring->desc[head]);
647 		page = ring->tx_buf[head];
648 
649 		/* TYPE_AL is 2, TYPE_LAST_AL is 3. So this trick gives
650 		 * us one increment per packet, with no branches.
651 		 */
652 		total_packets += FIELD_GET(FBNIC_TWD_TYPE_MASK, twd) -
653 				 FBNIC_TWD_TYPE_AL;
654 		total_bytes += FIELD_GET(FBNIC_TWD_LEN_MASK, twd);
655 
656 		page_pool_put_page(pp_page_to_nmdesc(page)->pp, page, -1,
657 				   pp_allow_direct);
658 next_desc:
659 		head++;
660 		head &= ring->size_mask;
661 	}
662 
663 	if (!total_bytes)
664 		return;
665 
666 	ring->head = head;
667 
668 	if (discard) {
669 		u64_stats_update_begin(&ring->stats.syncp);
670 		ring->stats.dropped += total_packets;
671 		u64_stats_update_end(&ring->stats.syncp);
672 		return;
673 	}
674 
675 	u64_stats_update_begin(&ring->stats.syncp);
676 	ring->stats.bytes += total_bytes;
677 	ring->stats.packets += total_packets;
678 	u64_stats_update_end(&ring->stats.syncp);
679 }
680 
681 static void fbnic_clean_tsq(struct fbnic_napi_vector *nv,
682 			    struct fbnic_ring *ring,
683 			    u64 tcd, int *ts_head, int *head0)
684 {
685 	struct skb_shared_hwtstamps hwtstamp;
686 	struct fbnic_net *fbn;
687 	struct sk_buff *skb;
688 	int head;
689 	u64 ns;
690 
691 	head = (*ts_head < 0) ? ring->head : *ts_head;
692 
693 	do {
694 		unsigned int desc_cnt;
695 
696 		if (head == ring->tail) {
697 			if (unlikely(net_ratelimit()))
698 				netdev_err(nv->napi.dev,
699 					   "Tx timestamp without matching packet\n");
700 			return;
701 		}
702 
703 		skb = ring->tx_buf[head];
704 		desc_cnt = FBNIC_XMIT_CB(skb)->desc_count;
705 
706 		head += desc_cnt;
707 		head &= ring->size_mask;
708 	} while (!(FBNIC_XMIT_CB(skb)->flags & FBNIC_XMIT_CB_TS));
709 
710 	fbn = netdev_priv(nv->napi.dev);
711 	ns = fbnic_ts40_to_ns(fbn, FIELD_GET(FBNIC_TCD_TYPE1_TS_MASK, tcd));
712 
713 	memset(&hwtstamp, 0, sizeof(hwtstamp));
714 	hwtstamp.hwtstamp = ns_to_ktime(ns);
715 
716 	*ts_head = head;
717 
718 	FBNIC_XMIT_CB(skb)->flags &= ~FBNIC_XMIT_CB_TS;
719 	if (*head0 < 0) {
720 		head = FBNIC_XMIT_CB(skb)->hw_head;
721 		if (head >= 0)
722 			*head0 = head;
723 	}
724 
725 	skb_tstamp_tx(skb, &hwtstamp);
726 	u64_stats_update_begin(&ring->stats.syncp);
727 	ring->stats.twq.ts_packets++;
728 	u64_stats_update_end(&ring->stats.syncp);
729 }
730 
731 static void fbnic_page_pool_init(struct fbnic_ring *ring, unsigned int idx,
732 				 netmem_ref netmem)
733 {
734 	struct fbnic_rx_buf *rx_buf = &ring->rx_buf[idx];
735 
736 	page_pool_fragment_netmem(netmem, FBNIC_PAGECNT_BIAS_MAX);
737 	rx_buf->pagecnt_bias = FBNIC_PAGECNT_BIAS_MAX;
738 	rx_buf->netmem = netmem;
739 }
740 
741 static struct page *
742 fbnic_page_pool_get_head(struct fbnic_q_triad *qt, unsigned int idx)
743 {
744 	struct fbnic_rx_buf *rx_buf = &qt->sub0.rx_buf[idx];
745 
746 	rx_buf->pagecnt_bias--;
747 
748 	/* sub0 is always fed system pages, from the NAPI-level page_pool */
749 	return netmem_to_page(rx_buf->netmem);
750 }
751 
752 static netmem_ref
753 fbnic_page_pool_get_data(struct fbnic_q_triad *qt, unsigned int idx)
754 {
755 	struct fbnic_rx_buf *rx_buf = &qt->sub1.rx_buf[idx];
756 
757 	rx_buf->pagecnt_bias--;
758 
759 	return rx_buf->netmem;
760 }
761 
762 static void fbnic_page_pool_drain(struct fbnic_ring *ring, unsigned int idx,
763 				  int budget)
764 {
765 	struct fbnic_rx_buf *rx_buf = &ring->rx_buf[idx];
766 	netmem_ref netmem = rx_buf->netmem;
767 
768 	if (!page_pool_unref_netmem(netmem, rx_buf->pagecnt_bias))
769 		page_pool_put_unrefed_netmem(ring->page_pool, netmem, -1,
770 					     !!budget);
771 
772 	rx_buf->netmem = 0;
773 }
774 
775 static void fbnic_clean_twq(struct fbnic_napi_vector *nv, int napi_budget,
776 			    struct fbnic_q_triad *qt, s32 ts_head, s32 head0,
777 			    s32 head1)
778 {
779 	if (head0 >= 0)
780 		fbnic_clean_twq0(nv, napi_budget, &qt->sub0, false, head0);
781 	else if (ts_head >= 0)
782 		fbnic_clean_twq0(nv, napi_budget, &qt->sub0, false, ts_head);
783 
784 	if (head1 >= 0) {
785 		qt->cmpl.deferred_head = -1;
786 		if (napi_budget)
787 			fbnic_clean_twq1(nv, true, &qt->sub1, false, head1);
788 		else
789 			qt->cmpl.deferred_head = head1;
790 	}
791 }
792 
793 static void
794 fbnic_clean_tcq(struct fbnic_napi_vector *nv, struct fbnic_q_triad *qt,
795 		int napi_budget)
796 {
797 	struct fbnic_ring *cmpl = &qt->cmpl;
798 	s32 head1 = cmpl->deferred_head;
799 	s32 head0 = -1, ts_head = -1;
800 	__le64 *raw_tcd, done;
801 	u32 head = cmpl->head;
802 
803 	done = (head & (cmpl->size_mask + 1)) ? 0 : cpu_to_le64(FBNIC_TCD_DONE);
804 	raw_tcd = &cmpl->desc[head & cmpl->size_mask];
805 
806 	/* Walk the completion queue collecting the heads reported by NIC */
807 	while ((*raw_tcd & cpu_to_le64(FBNIC_TCD_DONE)) == done) {
808 		u64 tcd;
809 
810 		dma_rmb();
811 
812 		tcd = le64_to_cpu(*raw_tcd);
813 
814 		switch (FIELD_GET(FBNIC_TCD_TYPE_MASK, tcd)) {
815 		case FBNIC_TCD_TYPE_0:
816 			if (tcd & FBNIC_TCD_TWQ1)
817 				head1 = FIELD_GET(FBNIC_TCD_TYPE0_HEAD1_MASK,
818 						  tcd);
819 			else
820 				head0 = FIELD_GET(FBNIC_TCD_TYPE0_HEAD0_MASK,
821 						  tcd);
822 			/* Currently all err status bits are related to
823 			 * timestamps and as those have yet to be added
824 			 * they are skipped for now.
825 			 */
826 			break;
827 		case FBNIC_TCD_TYPE_1:
828 			if (WARN_ON_ONCE(tcd & FBNIC_TCD_TWQ1))
829 				break;
830 
831 			fbnic_clean_tsq(nv, &qt->sub0, tcd, &ts_head, &head0);
832 			break;
833 		default:
834 			break;
835 		}
836 
837 		raw_tcd++;
838 		head++;
839 		if (!(head & cmpl->size_mask)) {
840 			done ^= cpu_to_le64(FBNIC_TCD_DONE);
841 			raw_tcd = &cmpl->desc[0];
842 		}
843 	}
844 
845 	/* Record the current head/tail of the queue */
846 	if (cmpl->head != head) {
847 		cmpl->head = head;
848 		writel(head & cmpl->size_mask, cmpl->doorbell);
849 	}
850 
851 	/* Unmap and free processed buffers */
852 	fbnic_clean_twq(nv, napi_budget, qt, ts_head, head0, head1);
853 }
854 
855 static void fbnic_clean_bdq(struct fbnic_ring *ring, unsigned int hw_head,
856 			    int napi_budget)
857 {
858 	unsigned int head = ring->head;
859 
860 	if (head == hw_head)
861 		return;
862 
863 	do {
864 		fbnic_page_pool_drain(ring, head, napi_budget);
865 
866 		head++;
867 		head &= ring->size_mask;
868 	} while (head != hw_head);
869 
870 	ring->head = head;
871 }
872 
873 static void fbnic_bd_prep(struct fbnic_ring *bdq, u16 id, netmem_ref netmem)
874 {
875 	__le64 *bdq_desc = &bdq->desc[id * FBNIC_BD_FRAG_COUNT];
876 	dma_addr_t dma = page_pool_get_dma_addr_netmem(netmem);
877 	u64 bd, i = FBNIC_BD_FRAG_COUNT;
878 
879 	bd = (FBNIC_BD_PAGE_ADDR_MASK & dma) |
880 	     FIELD_PREP(FBNIC_BD_PAGE_ID_MASK, id);
881 
882 	/* In the case that a page size is larger than 4K we will map a
883 	 * single page to multiple fragments. The fragments will be
884 	 * FBNIC_BD_FRAG_COUNT in size and the lower n bits will be use
885 	 * to indicate the individual fragment IDs.
886 	 */
887 	do {
888 		*bdq_desc = cpu_to_le64(bd);
889 		bd += FIELD_PREP(FBNIC_BD_DESC_ADDR_MASK, 1) |
890 		      FIELD_PREP(FBNIC_BD_DESC_ID_MASK, 1);
891 		bdq_desc++;
892 	} while (--i);
893 }
894 
895 static void fbnic_fill_bdq(struct fbnic_ring *bdq)
896 {
897 	unsigned int count = fbnic_desc_unused(bdq);
898 	unsigned int i = bdq->tail;
899 
900 	if (!count)
901 		return;
902 
903 	do {
904 		netmem_ref netmem;
905 
906 		netmem = page_pool_dev_alloc_netmems(bdq->page_pool);
907 		if (!netmem) {
908 			u64_stats_update_begin(&bdq->stats.syncp);
909 			bdq->stats.bdq.alloc_failed++;
910 			u64_stats_update_end(&bdq->stats.syncp);
911 
912 			break;
913 		}
914 
915 		fbnic_page_pool_init(bdq, i, netmem);
916 		fbnic_bd_prep(bdq, i, netmem);
917 
918 		i++;
919 		i &= bdq->size_mask;
920 
921 		count--;
922 	} while (count);
923 
924 	if (bdq->tail != i) {
925 		bdq->tail = i;
926 
927 		/* Force DMA writes to flush before writing to tail */
928 		dma_wmb();
929 
930 		writel(i, bdq->doorbell);
931 	}
932 }
933 
934 static unsigned int fbnic_hdr_pg_start(unsigned int pg_off)
935 {
936 	/* The headroom of the first header may be larger than FBNIC_RX_HROOM
937 	 * due to alignment. So account for that by just making the page
938 	 * offset 0 if we are starting at the first header.
939 	 */
940 	if (ALIGN(FBNIC_RX_HROOM, 128) > FBNIC_RX_HROOM &&
941 	    pg_off == ALIGN(FBNIC_RX_HROOM, 128))
942 		return 0;
943 
944 	return pg_off - FBNIC_RX_HROOM;
945 }
946 
947 static unsigned int fbnic_hdr_pg_end(unsigned int pg_off, unsigned int len)
948 {
949 	/* Determine the end of the buffer by finding the start of the next
950 	 * and then subtracting the headroom from that frame.
951 	 */
952 	pg_off += len + FBNIC_RX_TROOM + FBNIC_RX_HROOM;
953 
954 	return ALIGN(pg_off, 128) - FBNIC_RX_HROOM;
955 }
956 
957 static void fbnic_pkt_prepare(struct fbnic_napi_vector *nv, u64 rcd,
958 			      struct fbnic_pkt_buff *pkt,
959 			      struct fbnic_q_triad *qt)
960 {
961 	unsigned int hdr_pg_idx = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
962 	unsigned int hdr_pg_off = FIELD_GET(FBNIC_RCD_AL_BUFF_OFF_MASK, rcd);
963 	struct page *page = fbnic_page_pool_get_head(qt, hdr_pg_idx);
964 	unsigned int len = FIELD_GET(FBNIC_RCD_AL_BUFF_LEN_MASK, rcd);
965 	unsigned int frame_sz, hdr_pg_start, hdr_pg_end, headroom;
966 	unsigned char *hdr_start;
967 
968 	/* data_hard_start should always be NULL when this is called */
969 	WARN_ON_ONCE(pkt->buff.data_hard_start);
970 
971 	/* Short-cut the end calculation if we know page is fully consumed */
972 	hdr_pg_end = FIELD_GET(FBNIC_RCD_AL_PAGE_FIN, rcd) ?
973 		     FBNIC_BD_FRAG_SIZE : fbnic_hdr_pg_end(hdr_pg_off, len);
974 	hdr_pg_start = fbnic_hdr_pg_start(hdr_pg_off);
975 
976 	headroom = hdr_pg_off - hdr_pg_start + FBNIC_RX_PAD;
977 	frame_sz = hdr_pg_end - hdr_pg_start;
978 	xdp_init_buff(&pkt->buff, frame_sz, &qt->xdp_rxq);
979 	hdr_pg_start += (FBNIC_RCD_AL_BUFF_FRAG_MASK & rcd) *
980 			FBNIC_BD_FRAG_SIZE;
981 
982 	/* Sync DMA buffer */
983 	dma_sync_single_range_for_cpu(nv->dev, page_pool_get_dma_addr(page),
984 				      hdr_pg_start, frame_sz,
985 				      DMA_BIDIRECTIONAL);
986 
987 	/* Build frame around buffer */
988 	hdr_start = page_address(page) + hdr_pg_start;
989 	net_prefetch(pkt->buff.data);
990 	xdp_prepare_buff(&pkt->buff, hdr_start, headroom,
991 			 len - FBNIC_RX_PAD, true);
992 
993 	pkt->hwtstamp = 0;
994 	pkt->add_frag_failed = false;
995 }
996 
997 static void fbnic_add_rx_frag(struct fbnic_napi_vector *nv, u64 rcd,
998 			      struct fbnic_pkt_buff *pkt,
999 			      struct fbnic_q_triad *qt)
1000 {
1001 	unsigned int pg_idx = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
1002 	unsigned int pg_off = FIELD_GET(FBNIC_RCD_AL_BUFF_OFF_MASK, rcd);
1003 	unsigned int len = FIELD_GET(FBNIC_RCD_AL_BUFF_LEN_MASK, rcd);
1004 	netmem_ref netmem = fbnic_page_pool_get_data(qt, pg_idx);
1005 	unsigned int truesize;
1006 	bool added;
1007 
1008 	truesize = FIELD_GET(FBNIC_RCD_AL_PAGE_FIN, rcd) ?
1009 		   FBNIC_BD_FRAG_SIZE - pg_off : ALIGN(len, 128);
1010 
1011 	pg_off += (FBNIC_RCD_AL_BUFF_FRAG_MASK & rcd) *
1012 		  FBNIC_BD_FRAG_SIZE;
1013 
1014 	/* Sync DMA buffer */
1015 	page_pool_dma_sync_netmem_for_cpu(qt->sub1.page_pool, netmem,
1016 					  pg_off, truesize);
1017 
1018 	added = xdp_buff_add_frag(&pkt->buff, netmem, pg_off, len, truesize);
1019 	if (unlikely(!added)) {
1020 		pkt->add_frag_failed = true;
1021 		netdev_err_once(nv->napi.dev,
1022 				"Failed to add fragment to xdp_buff\n");
1023 	}
1024 }
1025 
1026 static void fbnic_put_pkt_buff(struct fbnic_q_triad *qt,
1027 			       struct fbnic_pkt_buff *pkt, int budget)
1028 {
1029 	struct page *page;
1030 
1031 	if (!pkt->buff.data_hard_start)
1032 		return;
1033 
1034 	if (xdp_buff_has_frags(&pkt->buff)) {
1035 		struct skb_shared_info *shinfo;
1036 		netmem_ref netmem;
1037 		int nr_frags;
1038 
1039 		shinfo = xdp_get_shared_info_from_buff(&pkt->buff);
1040 		nr_frags = shinfo->nr_frags;
1041 
1042 		while (nr_frags--) {
1043 			netmem = skb_frag_netmem(&shinfo->frags[nr_frags]);
1044 			page_pool_put_full_netmem(qt->sub1.page_pool, netmem,
1045 						  !!budget);
1046 		}
1047 	}
1048 
1049 	page = virt_to_page(pkt->buff.data_hard_start);
1050 	page_pool_put_full_page(qt->sub0.page_pool, page, !!budget);
1051 }
1052 
1053 static struct sk_buff *fbnic_build_skb(struct fbnic_napi_vector *nv,
1054 				       struct fbnic_pkt_buff *pkt)
1055 {
1056 	struct sk_buff *skb;
1057 
1058 	skb = xdp_build_skb_from_buff(&pkt->buff);
1059 	if (!skb)
1060 		return NULL;
1061 
1062 	/* Add timestamp if present */
1063 	if (pkt->hwtstamp)
1064 		skb_hwtstamps(skb)->hwtstamp = pkt->hwtstamp;
1065 
1066 	return skb;
1067 }
1068 
1069 static long fbnic_pkt_tx(struct fbnic_napi_vector *nv,
1070 			 struct fbnic_pkt_buff *pkt)
1071 {
1072 	struct fbnic_ring *ring = &nv->qt[0].sub1;
1073 	int size, offset, nsegs = 1, data_len = 0;
1074 	unsigned int tail = ring->tail;
1075 	struct skb_shared_info *shinfo;
1076 	skb_frag_t *frag = NULL;
1077 	struct page *page;
1078 	dma_addr_t dma;
1079 	__le64 *twd;
1080 
1081 	if (unlikely(xdp_buff_has_frags(&pkt->buff))) {
1082 		shinfo = xdp_get_shared_info_from_buff(&pkt->buff);
1083 		nsegs += shinfo->nr_frags;
1084 		data_len = shinfo->xdp_frags_size;
1085 		frag = &shinfo->frags[0];
1086 	}
1087 
1088 	if (fbnic_desc_unused(ring) < nsegs) {
1089 		u64_stats_update_begin(&ring->stats.syncp);
1090 		ring->stats.dropped++;
1091 		u64_stats_update_end(&ring->stats.syncp);
1092 		return -FBNIC_XDP_CONSUME;
1093 	}
1094 
1095 	page = virt_to_page(pkt->buff.data_hard_start);
1096 	offset = offset_in_page(pkt->buff.data);
1097 	dma = page_pool_get_dma_addr(page);
1098 
1099 	size = pkt->buff.data_end - pkt->buff.data;
1100 
1101 	while (nsegs--) {
1102 		dma_sync_single_range_for_device(nv->dev, dma, offset, size,
1103 						 DMA_BIDIRECTIONAL);
1104 		dma += offset;
1105 
1106 		ring->tx_buf[tail] = page;
1107 
1108 		twd = &ring->desc[tail];
1109 		*twd = cpu_to_le64(FIELD_PREP(FBNIC_TWD_ADDR_MASK, dma) |
1110 				   FIELD_PREP(FBNIC_TWD_LEN_MASK, size) |
1111 				   FIELD_PREP(FBNIC_TWD_TYPE_MASK,
1112 					      FBNIC_TWD_TYPE_AL));
1113 
1114 		tail++;
1115 		tail &= ring->size_mask;
1116 
1117 		if (!data_len)
1118 			break;
1119 
1120 		offset = skb_frag_off(frag);
1121 		page = skb_frag_page(frag);
1122 		dma = page_pool_get_dma_addr(page);
1123 
1124 		size = skb_frag_size(frag);
1125 		data_len -= size;
1126 		frag++;
1127 	}
1128 
1129 	*twd |= FBNIC_TWD_TYPE(LAST_AL);
1130 
1131 	ring->tail = tail;
1132 
1133 	return -FBNIC_XDP_TX;
1134 }
1135 
1136 static void fbnic_pkt_commit_tail(struct fbnic_napi_vector *nv,
1137 				  unsigned int pkt_tail)
1138 {
1139 	struct fbnic_ring *ring = &nv->qt[0].sub1;
1140 
1141 	/* Force DMA writes to flush before writing to tail */
1142 	dma_wmb();
1143 
1144 	writel(pkt_tail, ring->doorbell);
1145 }
1146 
1147 static struct sk_buff *fbnic_run_xdp(struct fbnic_napi_vector *nv,
1148 				     struct fbnic_pkt_buff *pkt)
1149 {
1150 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
1151 	struct bpf_prog *xdp_prog;
1152 	int act;
1153 
1154 	xdp_prog = READ_ONCE(fbn->xdp_prog);
1155 	if (!xdp_prog)
1156 		goto xdp_pass;
1157 
1158 	/* Should never happen, config paths enforce HDS threshold > MTU */
1159 	if (xdp_buff_has_frags(&pkt->buff) && !xdp_prog->aux->xdp_has_frags)
1160 		return ERR_PTR(-FBNIC_XDP_LEN_ERR);
1161 
1162 	act = bpf_prog_run_xdp(xdp_prog, &pkt->buff);
1163 	switch (act) {
1164 	case XDP_PASS:
1165 xdp_pass:
1166 		return fbnic_build_skb(nv, pkt);
1167 	case XDP_TX:
1168 		return ERR_PTR(fbnic_pkt_tx(nv, pkt));
1169 	default:
1170 		bpf_warn_invalid_xdp_action(nv->napi.dev, xdp_prog, act);
1171 		fallthrough;
1172 	case XDP_ABORTED:
1173 		trace_xdp_exception(nv->napi.dev, xdp_prog, act);
1174 		fallthrough;
1175 	case XDP_DROP:
1176 		break;
1177 	}
1178 
1179 	return ERR_PTR(-FBNIC_XDP_CONSUME);
1180 }
1181 
1182 static enum pkt_hash_types fbnic_skb_hash_type(u64 rcd)
1183 {
1184 	return (FBNIC_RCD_META_L4_TYPE_MASK & rcd) ? PKT_HASH_TYPE_L4 :
1185 	       (FBNIC_RCD_META_L3_TYPE_MASK & rcd) ? PKT_HASH_TYPE_L3 :
1186 						     PKT_HASH_TYPE_L2;
1187 }
1188 
1189 static void fbnic_rx_tstamp(struct fbnic_napi_vector *nv, u64 rcd,
1190 			    struct fbnic_pkt_buff *pkt)
1191 {
1192 	struct fbnic_net *fbn;
1193 	u64 ns, ts;
1194 
1195 	if (!FIELD_GET(FBNIC_RCD_OPT_META_TS, rcd))
1196 		return;
1197 
1198 	fbn = netdev_priv(nv->napi.dev);
1199 	ts = FIELD_GET(FBNIC_RCD_OPT_META_TS_MASK, rcd);
1200 	ns = fbnic_ts40_to_ns(fbn, ts);
1201 
1202 	/* Add timestamp to shared info */
1203 	pkt->hwtstamp = ns_to_ktime(ns);
1204 }
1205 
1206 static void fbnic_populate_skb_fields(struct fbnic_napi_vector *nv,
1207 				      u64 rcd, struct sk_buff *skb,
1208 				      struct fbnic_q_triad *qt,
1209 				      u64 *csum_cmpl, u64 *csum_none)
1210 {
1211 	struct net_device *netdev = nv->napi.dev;
1212 	struct fbnic_ring *rcq = &qt->cmpl;
1213 
1214 	fbnic_rx_csum(rcd, skb, rcq, csum_cmpl, csum_none);
1215 
1216 	if (netdev->features & NETIF_F_RXHASH)
1217 		skb_set_hash(skb,
1218 			     FIELD_GET(FBNIC_RCD_META_RSS_HASH_MASK, rcd),
1219 			     fbnic_skb_hash_type(rcd));
1220 
1221 	skb_record_rx_queue(skb, rcq->q_idx);
1222 }
1223 
1224 static bool fbnic_rcd_metadata_err(u64 rcd)
1225 {
1226 	return !!(FBNIC_RCD_META_UNCORRECTABLE_ERR_MASK & rcd);
1227 }
1228 
1229 static int fbnic_clean_rcq(struct fbnic_napi_vector *nv,
1230 			   struct fbnic_q_triad *qt, int budget)
1231 {
1232 	unsigned int packets = 0, bytes = 0, dropped = 0, alloc_failed = 0;
1233 	u64 csum_complete = 0, csum_none = 0, length_errors = 0;
1234 	s32 head0 = -1, head1 = -1, pkt_tail = -1;
1235 	struct fbnic_ring *rcq = &qt->cmpl;
1236 	struct fbnic_pkt_buff *pkt;
1237 	__le64 *raw_rcd, done;
1238 	u32 head = rcq->head;
1239 
1240 	done = (head & (rcq->size_mask + 1)) ? cpu_to_le64(FBNIC_RCD_DONE) : 0;
1241 	raw_rcd = &rcq->desc[head & rcq->size_mask];
1242 	pkt = rcq->pkt;
1243 
1244 	/* Walk the completion queue collecting the heads reported by NIC */
1245 	while (likely(packets < budget)) {
1246 		struct sk_buff *skb = ERR_PTR(-EINVAL);
1247 		u32 pkt_bytes;
1248 		u64 rcd;
1249 
1250 		if ((*raw_rcd & cpu_to_le64(FBNIC_RCD_DONE)) == done)
1251 			break;
1252 
1253 		dma_rmb();
1254 
1255 		rcd = le64_to_cpu(*raw_rcd);
1256 
1257 		switch (FIELD_GET(FBNIC_RCD_TYPE_MASK, rcd)) {
1258 		case FBNIC_RCD_TYPE_HDR_AL:
1259 			head0 = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
1260 			fbnic_pkt_prepare(nv, rcd, pkt, qt);
1261 
1262 			break;
1263 		case FBNIC_RCD_TYPE_PAY_AL:
1264 			head1 = FIELD_GET(FBNIC_RCD_AL_BUFF_PAGE_MASK, rcd);
1265 			fbnic_add_rx_frag(nv, rcd, pkt, qt);
1266 
1267 			break;
1268 		case FBNIC_RCD_TYPE_OPT_META:
1269 			/* Only type 0 is currently supported */
1270 			if (FIELD_GET(FBNIC_RCD_OPT_META_TYPE_MASK, rcd))
1271 				break;
1272 
1273 			fbnic_rx_tstamp(nv, rcd, pkt);
1274 
1275 			/* We currently ignore the action table index */
1276 			break;
1277 		case FBNIC_RCD_TYPE_META:
1278 			if (likely(!fbnic_rcd_metadata_err(rcd) &&
1279 				   !pkt->add_frag_failed)) {
1280 				pkt_bytes = xdp_get_buff_len(&pkt->buff);
1281 				skb = fbnic_run_xdp(nv, pkt);
1282 			}
1283 
1284 			/* Populate skb and invalidate XDP */
1285 			if (!IS_ERR_OR_NULL(skb)) {
1286 				fbnic_populate_skb_fields(nv, rcd, skb, qt,
1287 							  &csum_complete,
1288 							  &csum_none);
1289 				napi_gro_receive(&nv->napi, skb);
1290 			} else if (skb == ERR_PTR(-FBNIC_XDP_TX)) {
1291 				pkt_tail = nv->qt[0].sub1.tail;
1292 			} else if (PTR_ERR(skb) == -FBNIC_XDP_CONSUME) {
1293 				fbnic_put_pkt_buff(qt, pkt, 1);
1294 			} else {
1295 				if (!skb)
1296 					alloc_failed++;
1297 
1298 				if (skb == ERR_PTR(-FBNIC_XDP_LEN_ERR))
1299 					length_errors++;
1300 				else
1301 					dropped++;
1302 
1303 				fbnic_put_pkt_buff(qt, pkt, 1);
1304 				goto next_dont_count;
1305 			}
1306 
1307 			packets++;
1308 			bytes += pkt_bytes;
1309 next_dont_count:
1310 			pkt->buff.data_hard_start = NULL;
1311 
1312 			break;
1313 		}
1314 
1315 		raw_rcd++;
1316 		head++;
1317 		if (!(head & rcq->size_mask)) {
1318 			done ^= cpu_to_le64(FBNIC_RCD_DONE);
1319 			raw_rcd = &rcq->desc[0];
1320 		}
1321 	}
1322 
1323 	u64_stats_update_begin(&rcq->stats.syncp);
1324 	rcq->stats.packets += packets;
1325 	rcq->stats.bytes += bytes;
1326 	rcq->stats.dropped += dropped;
1327 	rcq->stats.rx.alloc_failed += alloc_failed;
1328 	rcq->stats.rx.csum_complete += csum_complete;
1329 	rcq->stats.rx.csum_none += csum_none;
1330 	rcq->stats.rx.length_errors += length_errors;
1331 	u64_stats_update_end(&rcq->stats.syncp);
1332 
1333 	if (pkt_tail >= 0)
1334 		fbnic_pkt_commit_tail(nv, pkt_tail);
1335 
1336 	/* Unmap and free processed buffers */
1337 	if (head0 >= 0)
1338 		fbnic_clean_bdq(&qt->sub0, head0, budget);
1339 	fbnic_fill_bdq(&qt->sub0);
1340 
1341 	if (head1 >= 0)
1342 		fbnic_clean_bdq(&qt->sub1, head1, budget);
1343 	fbnic_fill_bdq(&qt->sub1);
1344 
1345 	/* Record the current head/tail of the queue */
1346 	if (rcq->head != head) {
1347 		rcq->head = head;
1348 		writel(head & rcq->size_mask, rcq->doorbell);
1349 	}
1350 
1351 	return packets;
1352 }
1353 
1354 static void fbnic_nv_irq_disable(struct fbnic_napi_vector *nv)
1355 {
1356 	struct fbnic_dev *fbd = nv->fbd;
1357 	u32 v_idx = nv->v_idx;
1358 
1359 	fbnic_wr32(fbd, FBNIC_INTR_MASK_SET(v_idx / 32), 1 << (v_idx % 32));
1360 }
1361 
1362 static void fbnic_nv_irq_rearm(struct fbnic_napi_vector *nv)
1363 {
1364 	struct fbnic_dev *fbd = nv->fbd;
1365 	u32 v_idx = nv->v_idx;
1366 
1367 	fbnic_wr32(fbd, FBNIC_INTR_CQ_REARM(v_idx),
1368 		   FBNIC_INTR_CQ_REARM_INTR_UNMASK);
1369 }
1370 
1371 static int fbnic_poll(struct napi_struct *napi, int budget)
1372 {
1373 	struct fbnic_napi_vector *nv = container_of(napi,
1374 						    struct fbnic_napi_vector,
1375 						    napi);
1376 	int i, j, work_done = 0;
1377 
1378 	for (i = 0; i < nv->txt_count; i++)
1379 		fbnic_clean_tcq(nv, &nv->qt[i], budget);
1380 
1381 	for (j = 0; j < nv->rxt_count; j++, i++)
1382 		work_done += fbnic_clean_rcq(nv, &nv->qt[i], budget);
1383 
1384 	if (work_done >= budget)
1385 		return budget;
1386 
1387 	if (likely(napi_complete_done(napi, work_done)))
1388 		fbnic_nv_irq_rearm(nv);
1389 
1390 	return work_done;
1391 }
1392 
1393 irqreturn_t fbnic_msix_clean_rings(int __always_unused irq, void *data)
1394 {
1395 	struct fbnic_napi_vector *nv = *(void **)data;
1396 
1397 	napi_schedule_irqoff(&nv->napi);
1398 
1399 	return IRQ_HANDLED;
1400 }
1401 
1402 void fbnic_aggregate_ring_rx_counters(struct fbnic_net *fbn,
1403 				      struct fbnic_ring *rxr)
1404 {
1405 	struct fbnic_queue_stats *stats = &rxr->stats;
1406 
1407 	/* Capture stats from queues before dissasociating them */
1408 	fbn->rx_stats.bytes += stats->bytes;
1409 	fbn->rx_stats.packets += stats->packets;
1410 	fbn->rx_stats.dropped += stats->dropped;
1411 	fbn->rx_stats.rx.alloc_failed += stats->rx.alloc_failed;
1412 	fbn->rx_stats.rx.csum_complete += stats->rx.csum_complete;
1413 	fbn->rx_stats.rx.csum_none += stats->rx.csum_none;
1414 	fbn->rx_stats.rx.length_errors += stats->rx.length_errors;
1415 	/* Remember to add new stats here */
1416 	BUILD_BUG_ON(sizeof(fbn->rx_stats.rx) / 8 != 4);
1417 }
1418 
1419 void fbnic_aggregate_ring_bdq_counters(struct fbnic_net *fbn,
1420 				       struct fbnic_ring *bdq)
1421 {
1422 	struct fbnic_queue_stats *stats = &bdq->stats;
1423 
1424 	/* Capture stats from queues before dissasociating them */
1425 	fbn->bdq_stats.bdq.alloc_failed += stats->bdq.alloc_failed;
1426 	/* Remember to add new stats here */
1427 	BUILD_BUG_ON(sizeof(fbn->rx_stats.bdq) / 8 != 1);
1428 }
1429 
1430 void fbnic_aggregate_ring_tx_counters(struct fbnic_net *fbn,
1431 				      struct fbnic_ring *txr)
1432 {
1433 	struct fbnic_queue_stats *stats = &txr->stats;
1434 
1435 	/* Capture stats from queues before dissasociating them */
1436 	fbn->tx_stats.bytes += stats->bytes;
1437 	fbn->tx_stats.packets += stats->packets;
1438 	fbn->tx_stats.dropped += stats->dropped;
1439 	fbn->tx_stats.twq.csum_partial += stats->twq.csum_partial;
1440 	fbn->tx_stats.twq.lso += stats->twq.lso;
1441 	fbn->tx_stats.twq.ts_lost += stats->twq.ts_lost;
1442 	fbn->tx_stats.twq.ts_packets += stats->twq.ts_packets;
1443 	fbn->tx_stats.twq.stop += stats->twq.stop;
1444 	fbn->tx_stats.twq.wake += stats->twq.wake;
1445 	/* Remember to add new stats here */
1446 	BUILD_BUG_ON(sizeof(fbn->tx_stats.twq) / 8 != 6);
1447 }
1448 
1449 void fbnic_aggregate_ring_xdp_counters(struct fbnic_net *fbn,
1450 				       struct fbnic_ring *xdpr)
1451 {
1452 	struct fbnic_queue_stats *stats = &xdpr->stats;
1453 
1454 	if (!(xdpr->flags & FBNIC_RING_F_STATS))
1455 		return;
1456 
1457 	/* Capture stats from queues before dissasociating them */
1458 	fbn->tx_stats.dropped += stats->dropped;
1459 	fbn->tx_stats.bytes += stats->bytes;
1460 	fbn->tx_stats.packets += stats->packets;
1461 }
1462 
1463 static void fbnic_remove_tx_ring(struct fbnic_net *fbn,
1464 				 struct fbnic_ring *txr)
1465 {
1466 	if (!(txr->flags & FBNIC_RING_F_STATS))
1467 		return;
1468 
1469 	fbnic_aggregate_ring_tx_counters(fbn, txr);
1470 
1471 	/* Remove pointer to the Tx ring */
1472 	WARN_ON(fbn->tx[txr->q_idx] && fbn->tx[txr->q_idx] != txr);
1473 	fbn->tx[txr->q_idx] = NULL;
1474 }
1475 
1476 static void fbnic_remove_xdp_ring(struct fbnic_net *fbn,
1477 				  struct fbnic_ring *xdpr)
1478 {
1479 	if (!(xdpr->flags & FBNIC_RING_F_STATS))
1480 		return;
1481 
1482 	fbnic_aggregate_ring_xdp_counters(fbn, xdpr);
1483 
1484 	/* Remove pointer to the Tx ring */
1485 	WARN_ON(fbn->tx[xdpr->q_idx] && fbn->tx[xdpr->q_idx] != xdpr);
1486 	fbn->tx[xdpr->q_idx] = NULL;
1487 }
1488 
1489 static void fbnic_remove_rx_ring(struct fbnic_net *fbn,
1490 				 struct fbnic_ring *rxr)
1491 {
1492 	if (!(rxr->flags & FBNIC_RING_F_STATS))
1493 		return;
1494 
1495 	fbnic_aggregate_ring_rx_counters(fbn, rxr);
1496 
1497 	/* Remove pointer to the Rx ring */
1498 	WARN_ON(fbn->rx[rxr->q_idx] && fbn->rx[rxr->q_idx] != rxr);
1499 	fbn->rx[rxr->q_idx] = NULL;
1500 }
1501 
1502 static void fbnic_remove_bdq_ring(struct fbnic_net *fbn,
1503 				  struct fbnic_ring *bdq)
1504 {
1505 	if (!(bdq->flags & FBNIC_RING_F_STATS))
1506 		return;
1507 
1508 	fbnic_aggregate_ring_bdq_counters(fbn, bdq);
1509 }
1510 
1511 static void fbnic_free_qt_page_pools(struct fbnic_q_triad *qt)
1512 {
1513 	page_pool_destroy(qt->sub0.page_pool);
1514 	page_pool_destroy(qt->sub1.page_pool);
1515 }
1516 
1517 static void fbnic_free_napi_vector(struct fbnic_net *fbn,
1518 				   struct fbnic_napi_vector *nv)
1519 {
1520 	struct fbnic_dev *fbd = nv->fbd;
1521 	int i, j;
1522 
1523 	for (i = 0; i < nv->txt_count; i++) {
1524 		fbnic_remove_tx_ring(fbn, &nv->qt[i].sub0);
1525 		fbnic_remove_xdp_ring(fbn, &nv->qt[i].sub1);
1526 		fbnic_remove_tx_ring(fbn, &nv->qt[i].cmpl);
1527 	}
1528 
1529 	for (j = 0; j < nv->rxt_count; j++, i++) {
1530 		fbnic_remove_bdq_ring(fbn, &nv->qt[i].sub0);
1531 		fbnic_remove_bdq_ring(fbn, &nv->qt[i].sub1);
1532 		fbnic_remove_rx_ring(fbn, &nv->qt[i].cmpl);
1533 	}
1534 
1535 	fbnic_napi_free_irq(fbd, nv);
1536 	netif_napi_del_locked(&nv->napi);
1537 	fbn->napi[fbnic_napi_idx(nv)] = NULL;
1538 	kfree(nv);
1539 }
1540 
1541 void fbnic_free_napi_vectors(struct fbnic_net *fbn)
1542 {
1543 	int i;
1544 
1545 	for (i = 0; i < fbn->num_napi; i++)
1546 		if (fbn->napi[i])
1547 			fbnic_free_napi_vector(fbn, fbn->napi[i]);
1548 }
1549 
1550 static int
1551 fbnic_alloc_qt_page_pools(struct fbnic_net *fbn, struct fbnic_q_triad *qt,
1552 			  unsigned int rxq_idx)
1553 {
1554 	struct page_pool_params pp_params = {
1555 		.order = 0,
1556 		.flags = PP_FLAG_DMA_MAP |
1557 			 PP_FLAG_DMA_SYNC_DEV,
1558 		.pool_size = fbn->hpq_size + fbn->ppq_size,
1559 		.nid = NUMA_NO_NODE,
1560 		.dev = fbn->netdev->dev.parent,
1561 		.dma_dir = DMA_BIDIRECTIONAL,
1562 		.offset = 0,
1563 		.max_len = PAGE_SIZE,
1564 		.netdev	= fbn->netdev,
1565 		.queue_idx = rxq_idx,
1566 	};
1567 	struct page_pool *pp;
1568 
1569 	/* Page pool cannot exceed a size of 32768. This doesn't limit the
1570 	 * pages on the ring but the number we can have cached waiting on
1571 	 * the next use.
1572 	 *
1573 	 * TBD: Can this be reduced further? Would a multiple of
1574 	 * NAPI_POLL_WEIGHT possibly make more sense? The question is how
1575 	 * may pages do we need to hold in reserve to get the best return
1576 	 * without hogging too much system memory.
1577 	 */
1578 	if (pp_params.pool_size > 32768)
1579 		pp_params.pool_size = 32768;
1580 
1581 	pp = page_pool_create(&pp_params);
1582 	if (IS_ERR(pp))
1583 		return PTR_ERR(pp);
1584 
1585 	qt->sub0.page_pool = pp;
1586 	if (netif_rxq_has_unreadable_mp(fbn->netdev, rxq_idx)) {
1587 		pp_params.flags |= PP_FLAG_ALLOW_UNREADABLE_NETMEM;
1588 		pp_params.dma_dir = DMA_FROM_DEVICE;
1589 
1590 		pp = page_pool_create(&pp_params);
1591 		if (IS_ERR(pp))
1592 			goto err_destroy_sub0;
1593 	} else {
1594 		page_pool_get(pp);
1595 	}
1596 	qt->sub1.page_pool = pp;
1597 
1598 	return 0;
1599 
1600 err_destroy_sub0:
1601 	page_pool_destroy(pp);
1602 	return PTR_ERR(pp);
1603 }
1604 
1605 static void fbnic_ring_init(struct fbnic_ring *ring, u32 __iomem *doorbell,
1606 			    int q_idx, u8 flags)
1607 {
1608 	u64_stats_init(&ring->stats.syncp);
1609 	ring->doorbell = doorbell;
1610 	ring->q_idx = q_idx;
1611 	ring->flags = flags;
1612 	ring->deferred_head = -1;
1613 }
1614 
1615 static int fbnic_alloc_napi_vector(struct fbnic_dev *fbd, struct fbnic_net *fbn,
1616 				   unsigned int v_count, unsigned int v_idx,
1617 				   unsigned int txq_count, unsigned int txq_idx,
1618 				   unsigned int rxq_count, unsigned int rxq_idx)
1619 {
1620 	int txt_count = txq_count, rxt_count = rxq_count;
1621 	u32 __iomem *uc_addr = fbd->uc_addr0;
1622 	int xdp_count = 0, qt_count, err;
1623 	struct fbnic_napi_vector *nv;
1624 	struct fbnic_q_triad *qt;
1625 	u32 __iomem *db;
1626 
1627 	/* We need to reserve at least one Tx Queue Triad for an XDP ring */
1628 	if (rxq_count) {
1629 		xdp_count = 1;
1630 		if (!txt_count)
1631 			txt_count = 1;
1632 	}
1633 
1634 	qt_count = txt_count + rxq_count;
1635 	if (!qt_count)
1636 		return -EINVAL;
1637 
1638 	/* If MMIO has already failed there are no rings to initialize */
1639 	if (!uc_addr)
1640 		return -EIO;
1641 
1642 	/* Allocate NAPI vector and queue triads */
1643 	nv = kzalloc(struct_size(nv, qt, qt_count), GFP_KERNEL);
1644 	if (!nv)
1645 		return -ENOMEM;
1646 
1647 	/* Record queue triad counts */
1648 	nv->txt_count = txt_count;
1649 	nv->rxt_count = rxt_count;
1650 
1651 	/* Provide pointer back to fbnic and MSI-X vectors */
1652 	nv->fbd = fbd;
1653 	nv->v_idx = v_idx;
1654 
1655 	/* Tie napi to netdev */
1656 	fbn->napi[fbnic_napi_idx(nv)] = nv;
1657 	netif_napi_add_config_locked(fbn->netdev, &nv->napi, fbnic_poll,
1658 				     fbnic_napi_idx(nv));
1659 
1660 	/* Record IRQ to NAPI struct */
1661 	netif_napi_set_irq_locked(&nv->napi,
1662 				  pci_irq_vector(to_pci_dev(fbd->dev),
1663 						 nv->v_idx));
1664 
1665 	/* Tie nv back to PCIe dev */
1666 	nv->dev = fbd->dev;
1667 
1668 	/* Request the IRQ for napi vector */
1669 	err = fbnic_napi_request_irq(fbd, nv);
1670 	if (err)
1671 		goto napi_del;
1672 
1673 	/* Initialize queue triads */
1674 	qt = nv->qt;
1675 
1676 	while (txt_count) {
1677 		u8 flags = FBNIC_RING_F_CTX | FBNIC_RING_F_STATS;
1678 
1679 		/* Configure Tx queue */
1680 		db = &uc_addr[FBNIC_QUEUE(txq_idx) + FBNIC_QUEUE_TWQ0_TAIL];
1681 
1682 		/* Assign Tx queue to netdev if applicable */
1683 		if (txq_count > 0) {
1684 
1685 			fbnic_ring_init(&qt->sub0, db, txq_idx, flags);
1686 			fbn->tx[txq_idx] = &qt->sub0;
1687 			txq_count--;
1688 		} else {
1689 			fbnic_ring_init(&qt->sub0, db, 0,
1690 					FBNIC_RING_F_DISABLED);
1691 		}
1692 
1693 		/* Configure XDP queue */
1694 		db = &uc_addr[FBNIC_QUEUE(txq_idx) + FBNIC_QUEUE_TWQ1_TAIL];
1695 
1696 		/* Assign XDP queue to netdev if applicable
1697 		 *
1698 		 * The setup for this is in itself a bit different.
1699 		 * 1. We only need one XDP Tx queue per NAPI vector.
1700 		 * 2. We associate it to the first Rx queue index.
1701 		 * 3. The hardware side is associated based on the Tx Queue.
1702 		 * 4. The netdev queue is offset by FBNIC_MAX_TXQs.
1703 		 */
1704 		if (xdp_count > 0) {
1705 			unsigned int xdp_idx = FBNIC_MAX_TXQS + rxq_idx;
1706 
1707 			fbnic_ring_init(&qt->sub1, db, xdp_idx, flags);
1708 			fbn->tx[xdp_idx] = &qt->sub1;
1709 			xdp_count--;
1710 		} else {
1711 			fbnic_ring_init(&qt->sub1, db, 0,
1712 					FBNIC_RING_F_DISABLED);
1713 		}
1714 
1715 		/* Configure Tx completion queue */
1716 		db = &uc_addr[FBNIC_QUEUE(txq_idx) + FBNIC_QUEUE_TCQ_HEAD];
1717 		fbnic_ring_init(&qt->cmpl, db, 0, 0);
1718 
1719 		/* Update Tx queue index */
1720 		txt_count--;
1721 		txq_idx += v_count;
1722 
1723 		/* Move to next queue triad */
1724 		qt++;
1725 	}
1726 
1727 	while (rxt_count) {
1728 		/* Configure header queue */
1729 		db = &uc_addr[FBNIC_QUEUE(rxq_idx) + FBNIC_QUEUE_BDQ_HPQ_TAIL];
1730 		fbnic_ring_init(&qt->sub0, db, 0,
1731 				FBNIC_RING_F_CTX | FBNIC_RING_F_STATS);
1732 
1733 		/* Configure payload queue */
1734 		db = &uc_addr[FBNIC_QUEUE(rxq_idx) + FBNIC_QUEUE_BDQ_PPQ_TAIL];
1735 		fbnic_ring_init(&qt->sub1, db, 0,
1736 				FBNIC_RING_F_CTX | FBNIC_RING_F_STATS);
1737 
1738 		/* Configure Rx completion queue */
1739 		db = &uc_addr[FBNIC_QUEUE(rxq_idx) + FBNIC_QUEUE_RCQ_HEAD];
1740 		fbnic_ring_init(&qt->cmpl, db, rxq_idx, FBNIC_RING_F_STATS);
1741 		fbn->rx[rxq_idx] = &qt->cmpl;
1742 
1743 		/* Update Rx queue index */
1744 		rxt_count--;
1745 		rxq_idx += v_count;
1746 
1747 		/* Move to next queue triad */
1748 		qt++;
1749 	}
1750 
1751 	return 0;
1752 
1753 napi_del:
1754 	netif_napi_del_locked(&nv->napi);
1755 	fbn->napi[fbnic_napi_idx(nv)] = NULL;
1756 	kfree(nv);
1757 	return err;
1758 }
1759 
1760 int fbnic_alloc_napi_vectors(struct fbnic_net *fbn)
1761 {
1762 	unsigned int txq_idx = 0, rxq_idx = 0, v_idx = FBNIC_NON_NAPI_VECTORS;
1763 	unsigned int num_tx = fbn->num_tx_queues;
1764 	unsigned int num_rx = fbn->num_rx_queues;
1765 	unsigned int num_napi = fbn->num_napi;
1766 	struct fbnic_dev *fbd = fbn->fbd;
1767 	int err;
1768 
1769 	/* Allocate 1 Tx queue per napi vector */
1770 	if (num_napi < FBNIC_MAX_TXQS && num_napi == num_tx + num_rx) {
1771 		while (num_tx) {
1772 			err = fbnic_alloc_napi_vector(fbd, fbn,
1773 						      num_napi, v_idx,
1774 						      1, txq_idx, 0, 0);
1775 			if (err)
1776 				goto free_vectors;
1777 
1778 			/* Update counts and index */
1779 			num_tx--;
1780 			txq_idx++;
1781 
1782 			v_idx++;
1783 		}
1784 	}
1785 
1786 	/* Allocate Tx/Rx queue pairs per vector, or allocate remaining Rx */
1787 	while (num_rx | num_tx) {
1788 		int tqpv = DIV_ROUND_UP(num_tx, num_napi - txq_idx);
1789 		int rqpv = DIV_ROUND_UP(num_rx, num_napi - rxq_idx);
1790 
1791 		err = fbnic_alloc_napi_vector(fbd, fbn, num_napi, v_idx,
1792 					      tqpv, txq_idx, rqpv, rxq_idx);
1793 		if (err)
1794 			goto free_vectors;
1795 
1796 		/* Update counts and index */
1797 		num_tx -= tqpv;
1798 		txq_idx++;
1799 
1800 		num_rx -= rqpv;
1801 		rxq_idx++;
1802 
1803 		v_idx++;
1804 	}
1805 
1806 	return 0;
1807 
1808 free_vectors:
1809 	fbnic_free_napi_vectors(fbn);
1810 
1811 	return err;
1812 }
1813 
1814 static void fbnic_free_ring_resources(struct device *dev,
1815 				      struct fbnic_ring *ring)
1816 {
1817 	kvfree(ring->buffer);
1818 	ring->buffer = NULL;
1819 
1820 	/* If size is not set there are no descriptors present */
1821 	if (!ring->size)
1822 		return;
1823 
1824 	dma_free_coherent(dev, ring->size, ring->desc, ring->dma);
1825 	ring->size_mask = 0;
1826 	ring->size = 0;
1827 }
1828 
1829 static int fbnic_alloc_tx_ring_desc(struct fbnic_net *fbn,
1830 				    struct fbnic_ring *txr)
1831 {
1832 	struct device *dev = fbn->netdev->dev.parent;
1833 	size_t size;
1834 
1835 	/* Round size up to nearest 4K */
1836 	size = ALIGN(array_size(sizeof(*txr->desc), fbn->txq_size), 4096);
1837 
1838 	txr->desc = dma_alloc_coherent(dev, size, &txr->dma,
1839 				       GFP_KERNEL | __GFP_NOWARN);
1840 	if (!txr->desc)
1841 		return -ENOMEM;
1842 
1843 	/* txq_size should be a power of 2, so mask is just that -1 */
1844 	txr->size_mask = fbn->txq_size - 1;
1845 	txr->size = size;
1846 
1847 	return 0;
1848 }
1849 
1850 static int fbnic_alloc_tx_ring_buffer(struct fbnic_ring *txr)
1851 {
1852 	size_t size = array_size(sizeof(*txr->tx_buf), txr->size_mask + 1);
1853 
1854 	txr->tx_buf = kvzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1855 
1856 	return txr->tx_buf ? 0 : -ENOMEM;
1857 }
1858 
1859 static int fbnic_alloc_tx_ring_resources(struct fbnic_net *fbn,
1860 					 struct fbnic_ring *txr)
1861 {
1862 	struct device *dev = fbn->netdev->dev.parent;
1863 	int err;
1864 
1865 	if (txr->flags & FBNIC_RING_F_DISABLED)
1866 		return 0;
1867 
1868 	err = fbnic_alloc_tx_ring_desc(fbn, txr);
1869 	if (err)
1870 		return err;
1871 
1872 	if (!(txr->flags & FBNIC_RING_F_CTX))
1873 		return 0;
1874 
1875 	err = fbnic_alloc_tx_ring_buffer(txr);
1876 	if (err)
1877 		goto free_desc;
1878 
1879 	return 0;
1880 
1881 free_desc:
1882 	fbnic_free_ring_resources(dev, txr);
1883 	return err;
1884 }
1885 
1886 static int fbnic_alloc_rx_ring_desc(struct fbnic_net *fbn,
1887 				    struct fbnic_ring *rxr)
1888 {
1889 	struct device *dev = fbn->netdev->dev.parent;
1890 	size_t desc_size = sizeof(*rxr->desc);
1891 	u32 rxq_size;
1892 	size_t size;
1893 
1894 	switch (rxr->doorbell - fbnic_ring_csr_base(rxr)) {
1895 	case FBNIC_QUEUE_BDQ_HPQ_TAIL:
1896 		rxq_size = fbn->hpq_size / FBNIC_BD_FRAG_COUNT;
1897 		desc_size *= FBNIC_BD_FRAG_COUNT;
1898 		break;
1899 	case FBNIC_QUEUE_BDQ_PPQ_TAIL:
1900 		rxq_size = fbn->ppq_size / FBNIC_BD_FRAG_COUNT;
1901 		desc_size *= FBNIC_BD_FRAG_COUNT;
1902 		break;
1903 	case FBNIC_QUEUE_RCQ_HEAD:
1904 		rxq_size = fbn->rcq_size;
1905 		break;
1906 	default:
1907 		return -EINVAL;
1908 	}
1909 
1910 	/* Round size up to nearest 4K */
1911 	size = ALIGN(array_size(desc_size, rxq_size), 4096);
1912 
1913 	rxr->desc = dma_alloc_coherent(dev, size, &rxr->dma,
1914 				       GFP_KERNEL | __GFP_NOWARN);
1915 	if (!rxr->desc)
1916 		return -ENOMEM;
1917 
1918 	/* rxq_size should be a power of 2, so mask is just that -1 */
1919 	rxr->size_mask = rxq_size - 1;
1920 	rxr->size = size;
1921 
1922 	return 0;
1923 }
1924 
1925 static int fbnic_alloc_rx_ring_buffer(struct fbnic_ring *rxr)
1926 {
1927 	size_t size = array_size(sizeof(*rxr->rx_buf), rxr->size_mask + 1);
1928 
1929 	if (rxr->flags & FBNIC_RING_F_CTX)
1930 		size = sizeof(*rxr->rx_buf) * (rxr->size_mask + 1);
1931 	else
1932 		size = sizeof(*rxr->pkt);
1933 
1934 	rxr->rx_buf = kvzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1935 
1936 	return rxr->rx_buf ? 0 : -ENOMEM;
1937 }
1938 
1939 static int fbnic_alloc_rx_ring_resources(struct fbnic_net *fbn,
1940 					 struct fbnic_ring *rxr)
1941 {
1942 	struct device *dev = fbn->netdev->dev.parent;
1943 	int err;
1944 
1945 	err = fbnic_alloc_rx_ring_desc(fbn, rxr);
1946 	if (err)
1947 		return err;
1948 
1949 	err = fbnic_alloc_rx_ring_buffer(rxr);
1950 	if (err)
1951 		goto free_desc;
1952 
1953 	return 0;
1954 
1955 free_desc:
1956 	fbnic_free_ring_resources(dev, rxr);
1957 	return err;
1958 }
1959 
1960 static void fbnic_free_qt_resources(struct fbnic_net *fbn,
1961 				    struct fbnic_q_triad *qt)
1962 {
1963 	struct device *dev = fbn->netdev->dev.parent;
1964 
1965 	fbnic_free_ring_resources(dev, &qt->cmpl);
1966 	fbnic_free_ring_resources(dev, &qt->sub1);
1967 	fbnic_free_ring_resources(dev, &qt->sub0);
1968 
1969 	if (xdp_rxq_info_is_reg(&qt->xdp_rxq)) {
1970 		xdp_rxq_info_unreg_mem_model(&qt->xdp_rxq);
1971 		xdp_rxq_info_unreg(&qt->xdp_rxq);
1972 		fbnic_free_qt_page_pools(qt);
1973 	}
1974 }
1975 
1976 static int fbnic_alloc_tx_qt_resources(struct fbnic_net *fbn,
1977 				       struct fbnic_q_triad *qt)
1978 {
1979 	struct device *dev = fbn->netdev->dev.parent;
1980 	int err;
1981 
1982 	err = fbnic_alloc_tx_ring_resources(fbn, &qt->sub0);
1983 	if (err)
1984 		return err;
1985 
1986 	err = fbnic_alloc_tx_ring_resources(fbn, &qt->sub1);
1987 	if (err)
1988 		goto free_sub0;
1989 
1990 	err = fbnic_alloc_tx_ring_resources(fbn, &qt->cmpl);
1991 	if (err)
1992 		goto free_sub1;
1993 
1994 	return 0;
1995 
1996 free_sub1:
1997 	fbnic_free_ring_resources(dev, &qt->sub1);
1998 free_sub0:
1999 	fbnic_free_ring_resources(dev, &qt->sub0);
2000 	return err;
2001 }
2002 
2003 static int fbnic_alloc_rx_qt_resources(struct fbnic_net *fbn,
2004 				       struct fbnic_napi_vector *nv,
2005 				       struct fbnic_q_triad *qt)
2006 {
2007 	struct device *dev = fbn->netdev->dev.parent;
2008 	int err;
2009 
2010 	err = fbnic_alloc_qt_page_pools(fbn, qt, qt->cmpl.q_idx);
2011 	if (err)
2012 		return err;
2013 
2014 	err = xdp_rxq_info_reg(&qt->xdp_rxq, fbn->netdev, qt->sub0.q_idx,
2015 			       nv->napi.napi_id);
2016 	if (err)
2017 		goto free_page_pools;
2018 
2019 	err = xdp_rxq_info_reg_mem_model(&qt->xdp_rxq, MEM_TYPE_PAGE_POOL,
2020 					 qt->sub0.page_pool);
2021 	if (err)
2022 		goto unreg_rxq;
2023 
2024 	err = fbnic_alloc_rx_ring_resources(fbn, &qt->sub0);
2025 	if (err)
2026 		goto unreg_mm;
2027 
2028 	err = fbnic_alloc_rx_ring_resources(fbn, &qt->sub1);
2029 	if (err)
2030 		goto free_sub0;
2031 
2032 	err = fbnic_alloc_rx_ring_resources(fbn, &qt->cmpl);
2033 	if (err)
2034 		goto free_sub1;
2035 
2036 	return 0;
2037 
2038 free_sub1:
2039 	fbnic_free_ring_resources(dev, &qt->sub1);
2040 free_sub0:
2041 	fbnic_free_ring_resources(dev, &qt->sub0);
2042 unreg_mm:
2043 	xdp_rxq_info_unreg_mem_model(&qt->xdp_rxq);
2044 unreg_rxq:
2045 	xdp_rxq_info_unreg(&qt->xdp_rxq);
2046 free_page_pools:
2047 	fbnic_free_qt_page_pools(qt);
2048 	return err;
2049 }
2050 
2051 static void fbnic_free_nv_resources(struct fbnic_net *fbn,
2052 				    struct fbnic_napi_vector *nv)
2053 {
2054 	int i;
2055 
2056 	for (i = 0; i < nv->txt_count + nv->rxt_count; i++)
2057 		fbnic_free_qt_resources(fbn, &nv->qt[i]);
2058 }
2059 
2060 static int fbnic_alloc_nv_resources(struct fbnic_net *fbn,
2061 				    struct fbnic_napi_vector *nv)
2062 {
2063 	int i, j, err;
2064 
2065 	/* Allocate Tx Resources */
2066 	for (i = 0; i < nv->txt_count; i++) {
2067 		err = fbnic_alloc_tx_qt_resources(fbn, &nv->qt[i]);
2068 		if (err)
2069 			goto free_qt_resources;
2070 	}
2071 
2072 	/* Allocate Rx Resources */
2073 	for (j = 0; j < nv->rxt_count; j++, i++) {
2074 		err = fbnic_alloc_rx_qt_resources(fbn, nv, &nv->qt[i]);
2075 		if (err)
2076 			goto free_qt_resources;
2077 	}
2078 
2079 	return 0;
2080 
2081 free_qt_resources:
2082 	while (i--)
2083 		fbnic_free_qt_resources(fbn, &nv->qt[i]);
2084 	return err;
2085 }
2086 
2087 void fbnic_free_resources(struct fbnic_net *fbn)
2088 {
2089 	int i;
2090 
2091 	for (i = 0; i < fbn->num_napi; i++)
2092 		fbnic_free_nv_resources(fbn, fbn->napi[i]);
2093 }
2094 
2095 int fbnic_alloc_resources(struct fbnic_net *fbn)
2096 {
2097 	int i, err = -ENODEV;
2098 
2099 	for (i = 0; i < fbn->num_napi; i++) {
2100 		err = fbnic_alloc_nv_resources(fbn, fbn->napi[i]);
2101 		if (err)
2102 			goto free_resources;
2103 	}
2104 
2105 	return 0;
2106 
2107 free_resources:
2108 	while (i--)
2109 		fbnic_free_nv_resources(fbn, fbn->napi[i]);
2110 
2111 	return err;
2112 }
2113 
2114 static void fbnic_set_netif_napi(struct fbnic_napi_vector *nv)
2115 {
2116 	int i, j;
2117 
2118 	/* Associate Tx queue with NAPI */
2119 	for (i = 0; i < nv->txt_count; i++) {
2120 		struct fbnic_q_triad *qt = &nv->qt[i];
2121 
2122 		netif_queue_set_napi(nv->napi.dev, qt->sub0.q_idx,
2123 				     NETDEV_QUEUE_TYPE_TX, &nv->napi);
2124 	}
2125 
2126 	/* Associate Rx queue with NAPI */
2127 	for (j = 0; j < nv->rxt_count; j++, i++) {
2128 		struct fbnic_q_triad *qt = &nv->qt[i];
2129 
2130 		netif_queue_set_napi(nv->napi.dev, qt->cmpl.q_idx,
2131 				     NETDEV_QUEUE_TYPE_RX, &nv->napi);
2132 	}
2133 }
2134 
2135 static void fbnic_reset_netif_napi(struct fbnic_napi_vector *nv)
2136 {
2137 	int i, j;
2138 
2139 	/* Disassociate Tx queue from NAPI */
2140 	for (i = 0; i < nv->txt_count; i++) {
2141 		struct fbnic_q_triad *qt = &nv->qt[i];
2142 
2143 		netif_queue_set_napi(nv->napi.dev, qt->sub0.q_idx,
2144 				     NETDEV_QUEUE_TYPE_TX, NULL);
2145 	}
2146 
2147 	/* Disassociate Rx queue from NAPI */
2148 	for (j = 0; j < nv->rxt_count; j++, i++) {
2149 		struct fbnic_q_triad *qt = &nv->qt[i];
2150 
2151 		netif_queue_set_napi(nv->napi.dev, qt->cmpl.q_idx,
2152 				     NETDEV_QUEUE_TYPE_RX, NULL);
2153 	}
2154 }
2155 
2156 int fbnic_set_netif_queues(struct fbnic_net *fbn)
2157 {
2158 	int i, err;
2159 
2160 	err = netif_set_real_num_queues(fbn->netdev, fbn->num_tx_queues,
2161 					fbn->num_rx_queues);
2162 	if (err)
2163 		return err;
2164 
2165 	for (i = 0; i < fbn->num_napi; i++)
2166 		fbnic_set_netif_napi(fbn->napi[i]);
2167 
2168 	return 0;
2169 }
2170 
2171 void fbnic_reset_netif_queues(struct fbnic_net *fbn)
2172 {
2173 	int i;
2174 
2175 	for (i = 0; i < fbn->num_napi; i++)
2176 		fbnic_reset_netif_napi(fbn->napi[i]);
2177 }
2178 
2179 static void fbnic_disable_twq0(struct fbnic_ring *txr)
2180 {
2181 	u32 twq_ctl = fbnic_ring_rd32(txr, FBNIC_QUEUE_TWQ0_CTL);
2182 
2183 	twq_ctl &= ~FBNIC_QUEUE_TWQ_CTL_ENABLE;
2184 
2185 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TWQ0_CTL, twq_ctl);
2186 }
2187 
2188 static void fbnic_disable_twq1(struct fbnic_ring *txr)
2189 {
2190 	u32 twq_ctl = fbnic_ring_rd32(txr, FBNIC_QUEUE_TWQ1_CTL);
2191 
2192 	twq_ctl &= ~FBNIC_QUEUE_TWQ_CTL_ENABLE;
2193 
2194 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TWQ1_CTL, twq_ctl);
2195 }
2196 
2197 static void fbnic_disable_tcq(struct fbnic_ring *txr)
2198 {
2199 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TCQ_CTL, 0);
2200 	fbnic_ring_wr32(txr, FBNIC_QUEUE_TIM_MASK, FBNIC_QUEUE_TIM_MASK_MASK);
2201 }
2202 
2203 static void fbnic_disable_bdq(struct fbnic_ring *hpq, struct fbnic_ring *ppq)
2204 {
2205 	u32 bdq_ctl = fbnic_ring_rd32(hpq, FBNIC_QUEUE_BDQ_CTL);
2206 
2207 	bdq_ctl &= ~FBNIC_QUEUE_BDQ_CTL_ENABLE;
2208 
2209 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_CTL, bdq_ctl);
2210 }
2211 
2212 static void fbnic_disable_rcq(struct fbnic_ring *rxr)
2213 {
2214 	fbnic_ring_wr32(rxr, FBNIC_QUEUE_RCQ_CTL, 0);
2215 	fbnic_ring_wr32(rxr, FBNIC_QUEUE_RIM_MASK, FBNIC_QUEUE_RIM_MASK_MASK);
2216 }
2217 
2218 void fbnic_napi_disable(struct fbnic_net *fbn)
2219 {
2220 	int i;
2221 
2222 	for (i = 0; i < fbn->num_napi; i++) {
2223 		napi_disable_locked(&fbn->napi[i]->napi);
2224 
2225 		fbnic_nv_irq_disable(fbn->napi[i]);
2226 	}
2227 }
2228 
2229 static void __fbnic_nv_disable(struct fbnic_napi_vector *nv)
2230 {
2231 	int i, t;
2232 
2233 	/* Disable Tx queue triads */
2234 	for (t = 0; t < nv->txt_count; t++) {
2235 		struct fbnic_q_triad *qt = &nv->qt[t];
2236 
2237 		fbnic_disable_twq0(&qt->sub0);
2238 		fbnic_disable_twq1(&qt->sub1);
2239 		fbnic_disable_tcq(&qt->cmpl);
2240 	}
2241 
2242 	/* Disable Rx queue triads */
2243 	for (i = 0; i < nv->rxt_count; i++, t++) {
2244 		struct fbnic_q_triad *qt = &nv->qt[t];
2245 
2246 		fbnic_disable_bdq(&qt->sub0, &qt->sub1);
2247 		fbnic_disable_rcq(&qt->cmpl);
2248 	}
2249 }
2250 
2251 static void
2252 fbnic_nv_disable(struct fbnic_net *fbn, struct fbnic_napi_vector *nv)
2253 {
2254 	__fbnic_nv_disable(nv);
2255 	fbnic_wrfl(fbn->fbd);
2256 }
2257 
2258 void fbnic_disable(struct fbnic_net *fbn)
2259 {
2260 	struct fbnic_dev *fbd = fbn->fbd;
2261 	int i;
2262 
2263 	for (i = 0; i < fbn->num_napi; i++)
2264 		__fbnic_nv_disable(fbn->napi[i]);
2265 
2266 	fbnic_wrfl(fbd);
2267 }
2268 
2269 static void fbnic_tx_flush(struct fbnic_dev *fbd)
2270 {
2271 	netdev_warn(fbd->netdev, "triggering Tx flush\n");
2272 
2273 	fbnic_rmw32(fbd, FBNIC_TMI_DROP_CTRL, FBNIC_TMI_DROP_CTRL_EN,
2274 		    FBNIC_TMI_DROP_CTRL_EN);
2275 }
2276 
2277 static void fbnic_tx_flush_off(struct fbnic_dev *fbd)
2278 {
2279 	fbnic_rmw32(fbd, FBNIC_TMI_DROP_CTRL, FBNIC_TMI_DROP_CTRL_EN, 0);
2280 }
2281 
2282 struct fbnic_idle_regs {
2283 	u32 reg_base;
2284 	u8 reg_cnt;
2285 };
2286 
2287 static bool fbnic_all_idle(struct fbnic_dev *fbd,
2288 			   const struct fbnic_idle_regs *regs,
2289 			   unsigned int nregs)
2290 {
2291 	unsigned int i, j;
2292 
2293 	for (i = 0; i < nregs; i++) {
2294 		for (j = 0; j < regs[i].reg_cnt; j++) {
2295 			if (fbnic_rd32(fbd, regs[i].reg_base + j) != ~0U)
2296 				return false;
2297 		}
2298 	}
2299 	return true;
2300 }
2301 
2302 static void fbnic_idle_dump(struct fbnic_dev *fbd,
2303 			    const struct fbnic_idle_regs *regs,
2304 			    unsigned int nregs, const char *dir, int err)
2305 {
2306 	unsigned int i, j;
2307 
2308 	netdev_err(fbd->netdev, "error waiting for %s idle %d\n", dir, err);
2309 	for (i = 0; i < nregs; i++)
2310 		for (j = 0; j < regs[i].reg_cnt; j++)
2311 			netdev_err(fbd->netdev, "0x%04x: %08x\n",
2312 				   regs[i].reg_base + j,
2313 				   fbnic_rd32(fbd, regs[i].reg_base + j));
2314 }
2315 
2316 int fbnic_wait_all_queues_idle(struct fbnic_dev *fbd, bool may_fail)
2317 {
2318 	static const struct fbnic_idle_regs tx[] = {
2319 		{ FBNIC_QM_TWQ_IDLE(0),	FBNIC_QM_TWQ_IDLE_CNT, },
2320 		{ FBNIC_QM_TQS_IDLE(0),	FBNIC_QM_TQS_IDLE_CNT, },
2321 		{ FBNIC_QM_TDE_IDLE(0),	FBNIC_QM_TDE_IDLE_CNT, },
2322 		{ FBNIC_QM_TCQ_IDLE(0),	FBNIC_QM_TCQ_IDLE_CNT, },
2323 	}, rx[] = {
2324 		{ FBNIC_QM_HPQ_IDLE(0),	FBNIC_QM_HPQ_IDLE_CNT, },
2325 		{ FBNIC_QM_PPQ_IDLE(0),	FBNIC_QM_PPQ_IDLE_CNT, },
2326 		{ FBNIC_QM_RCQ_IDLE(0),	FBNIC_QM_RCQ_IDLE_CNT, },
2327 	};
2328 	bool idle;
2329 	int err;
2330 
2331 	err = read_poll_timeout_atomic(fbnic_all_idle, idle, idle, 2, 500000,
2332 				       false, fbd, tx, ARRAY_SIZE(tx));
2333 	if (err == -ETIMEDOUT) {
2334 		fbnic_tx_flush(fbd);
2335 		err = read_poll_timeout_atomic(fbnic_all_idle, idle, idle,
2336 					       2, 500000, false,
2337 					       fbd, tx, ARRAY_SIZE(tx));
2338 		fbnic_tx_flush_off(fbd);
2339 	}
2340 	if (err) {
2341 		fbnic_idle_dump(fbd, tx, ARRAY_SIZE(tx), "Tx", err);
2342 		if (may_fail)
2343 			return err;
2344 	}
2345 
2346 	err = read_poll_timeout_atomic(fbnic_all_idle, idle, idle, 2, 500000,
2347 				       false, fbd, rx, ARRAY_SIZE(rx));
2348 	if (err)
2349 		fbnic_idle_dump(fbd, rx, ARRAY_SIZE(rx), "Rx", err);
2350 	return err;
2351 }
2352 
2353 static int
2354 fbnic_wait_queue_idle(struct fbnic_net *fbn, bool rx, unsigned int idx)
2355 {
2356 	static const unsigned int tx_regs[] = {
2357 		FBNIC_QM_TWQ_IDLE(0), FBNIC_QM_TQS_IDLE(0),
2358 		FBNIC_QM_TDE_IDLE(0), FBNIC_QM_TCQ_IDLE(0),
2359 	}, rx_regs[] = {
2360 		FBNIC_QM_HPQ_IDLE(0), FBNIC_QM_PPQ_IDLE(0),
2361 		FBNIC_QM_RCQ_IDLE(0),
2362 	};
2363 	struct fbnic_dev *fbd = fbn->fbd;
2364 	unsigned int val, mask, off;
2365 	const unsigned int *regs;
2366 	unsigned int reg_cnt;
2367 	int i, err;
2368 
2369 	regs = rx ? rx_regs : tx_regs;
2370 	reg_cnt = rx ? ARRAY_SIZE(rx_regs) : ARRAY_SIZE(tx_regs);
2371 
2372 	off = idx / 32;
2373 	mask = BIT(idx % 32);
2374 
2375 	for (i = 0; i < reg_cnt; i++) {
2376 		err = read_poll_timeout_atomic(fbnic_rd32, val, val & mask,
2377 					       2, 500000, false,
2378 					       fbd, regs[i] + off);
2379 		if (err) {
2380 			netdev_err(fbd->netdev,
2381 				   "wait for queue %s%d idle failed 0x%04x(%d): %08x (mask: %08x)\n",
2382 				   rx ? "Rx" : "Tx", idx, regs[i] + off, i,
2383 				   val, mask);
2384 			return err;
2385 		}
2386 	}
2387 
2388 	return 0;
2389 }
2390 
2391 static void fbnic_nv_flush(struct fbnic_napi_vector *nv)
2392 {
2393 	int j, t;
2394 
2395 	/* Flush any processed Tx Queue Triads and drop the rest */
2396 	for (t = 0; t < nv->txt_count; t++) {
2397 		struct fbnic_q_triad *qt = &nv->qt[t];
2398 		struct netdev_queue *tx_queue;
2399 
2400 		/* Clean the work queues of unprocessed work */
2401 		fbnic_clean_twq0(nv, 0, &qt->sub0, true, qt->sub0.tail);
2402 		fbnic_clean_twq1(nv, false, &qt->sub1, true,
2403 				 qt->sub1.tail);
2404 
2405 		/* Reset completion queue descriptor ring */
2406 		memset(qt->cmpl.desc, 0, qt->cmpl.size);
2407 
2408 		/* Nothing else to do if Tx queue is disabled */
2409 		if (qt->sub0.flags & FBNIC_RING_F_DISABLED)
2410 			continue;
2411 
2412 		/* Reset BQL associated with Tx queue */
2413 		tx_queue = netdev_get_tx_queue(nv->napi.dev,
2414 					       qt->sub0.q_idx);
2415 		netdev_tx_reset_queue(tx_queue);
2416 	}
2417 
2418 	/* Flush any processed Rx Queue Triads and drop the rest */
2419 	for (j = 0; j < nv->rxt_count; j++, t++) {
2420 		struct fbnic_q_triad *qt = &nv->qt[t];
2421 
2422 		/* Clean the work queues of unprocessed work */
2423 		fbnic_clean_bdq(&qt->sub0, qt->sub0.tail, 0);
2424 		fbnic_clean_bdq(&qt->sub1, qt->sub1.tail, 0);
2425 
2426 		/* Reset completion queue descriptor ring */
2427 		memset(qt->cmpl.desc, 0, qt->cmpl.size);
2428 
2429 		fbnic_put_pkt_buff(qt, qt->cmpl.pkt, 0);
2430 		memset(qt->cmpl.pkt, 0, sizeof(struct fbnic_pkt_buff));
2431 	}
2432 }
2433 
2434 void fbnic_flush(struct fbnic_net *fbn)
2435 {
2436 	int i;
2437 
2438 	for (i = 0; i < fbn->num_napi; i++)
2439 		fbnic_nv_flush(fbn->napi[i]);
2440 }
2441 
2442 static void fbnic_nv_fill(struct fbnic_napi_vector *nv)
2443 {
2444 	int j, t;
2445 
2446 	/* Configure NAPI mapping and populate pages
2447 	 * in the BDQ rings to use for Rx
2448 	 */
2449 	for (j = 0, t = nv->txt_count; j < nv->rxt_count; j++, t++) {
2450 		struct fbnic_q_triad *qt = &nv->qt[t];
2451 
2452 		/* Populate the header and payload BDQs */
2453 		fbnic_fill_bdq(&qt->sub0);
2454 		fbnic_fill_bdq(&qt->sub1);
2455 	}
2456 }
2457 
2458 void fbnic_fill(struct fbnic_net *fbn)
2459 {
2460 	int i;
2461 
2462 	for (i = 0; i < fbn->num_napi; i++)
2463 		fbnic_nv_fill(fbn->napi[i]);
2464 }
2465 
2466 static void fbnic_enable_twq0(struct fbnic_ring *twq)
2467 {
2468 	u32 log_size = fls(twq->size_mask);
2469 
2470 	if (!twq->size_mask)
2471 		return;
2472 
2473 	/* Reset head/tail */
2474 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_CTL, FBNIC_QUEUE_TWQ_CTL_RESET);
2475 	twq->tail = 0;
2476 	twq->head = 0;
2477 
2478 	/* Store descriptor ring address and size */
2479 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_BAL, lower_32_bits(twq->dma));
2480 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_BAH, upper_32_bits(twq->dma));
2481 
2482 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2483 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_SIZE, log_size & 0xf);
2484 
2485 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ0_CTL, FBNIC_QUEUE_TWQ_CTL_ENABLE);
2486 }
2487 
2488 static void fbnic_enable_twq1(struct fbnic_ring *twq)
2489 {
2490 	u32 log_size = fls(twq->size_mask);
2491 
2492 	if (!twq->size_mask)
2493 		return;
2494 
2495 	/* Reset head/tail */
2496 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_CTL, FBNIC_QUEUE_TWQ_CTL_RESET);
2497 	twq->tail = 0;
2498 	twq->head = 0;
2499 
2500 	/* Store descriptor ring address and size */
2501 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_BAL, lower_32_bits(twq->dma));
2502 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_BAH, upper_32_bits(twq->dma));
2503 
2504 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2505 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_SIZE, log_size & 0xf);
2506 
2507 	fbnic_ring_wr32(twq, FBNIC_QUEUE_TWQ1_CTL, FBNIC_QUEUE_TWQ_CTL_ENABLE);
2508 }
2509 
2510 static void fbnic_enable_tcq(struct fbnic_napi_vector *nv,
2511 			     struct fbnic_ring *tcq)
2512 {
2513 	u32 log_size = fls(tcq->size_mask);
2514 
2515 	if (!tcq->size_mask)
2516 		return;
2517 
2518 	/* Reset head/tail */
2519 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_CTL, FBNIC_QUEUE_TCQ_CTL_RESET);
2520 	tcq->tail = 0;
2521 	tcq->head = 0;
2522 
2523 	/* Store descriptor ring address and size */
2524 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_BAL, lower_32_bits(tcq->dma));
2525 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_BAH, upper_32_bits(tcq->dma));
2526 
2527 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2528 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_SIZE, log_size & 0xf);
2529 
2530 	/* Store interrupt information for the completion queue */
2531 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TIM_CTL, nv->v_idx);
2532 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TIM_THRESHOLD, tcq->size_mask / 2);
2533 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TIM_MASK, 0);
2534 
2535 	/* Enable queue */
2536 	fbnic_ring_wr32(tcq, FBNIC_QUEUE_TCQ_CTL, FBNIC_QUEUE_TCQ_CTL_ENABLE);
2537 }
2538 
2539 static void fbnic_enable_bdq(struct fbnic_ring *hpq, struct fbnic_ring *ppq)
2540 {
2541 	u32 bdq_ctl = FBNIC_QUEUE_BDQ_CTL_ENABLE;
2542 	u32 log_size;
2543 
2544 	/* Reset head/tail */
2545 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_CTL, FBNIC_QUEUE_BDQ_CTL_RESET);
2546 	ppq->tail = 0;
2547 	ppq->head = 0;
2548 	hpq->tail = 0;
2549 	hpq->head = 0;
2550 
2551 	log_size = fls(hpq->size_mask);
2552 
2553 	/* Store descriptor ring address and size */
2554 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_HPQ_BAL, lower_32_bits(hpq->dma));
2555 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_HPQ_BAH, upper_32_bits(hpq->dma));
2556 
2557 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2558 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_HPQ_SIZE, log_size & 0xf);
2559 
2560 	if (!ppq->size_mask)
2561 		goto write_ctl;
2562 
2563 	log_size = fls(ppq->size_mask);
2564 
2565 	/* Add enabling of PPQ to BDQ control */
2566 	bdq_ctl |= FBNIC_QUEUE_BDQ_CTL_PPQ_ENABLE;
2567 
2568 	/* Store descriptor ring address and size */
2569 	fbnic_ring_wr32(ppq, FBNIC_QUEUE_BDQ_PPQ_BAL, lower_32_bits(ppq->dma));
2570 	fbnic_ring_wr32(ppq, FBNIC_QUEUE_BDQ_PPQ_BAH, upper_32_bits(ppq->dma));
2571 	fbnic_ring_wr32(ppq, FBNIC_QUEUE_BDQ_PPQ_SIZE, log_size & 0xf);
2572 
2573 write_ctl:
2574 	fbnic_ring_wr32(hpq, FBNIC_QUEUE_BDQ_CTL, bdq_ctl);
2575 }
2576 
2577 static void fbnic_config_drop_mode_rcq(struct fbnic_napi_vector *nv,
2578 				       struct fbnic_ring *rcq, bool tx_pause)
2579 {
2580 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2581 	u32 drop_mode, rcq_ctl;
2582 
2583 	if (!tx_pause && fbn->num_rx_queues > 1)
2584 		drop_mode = FBNIC_QUEUE_RDE_CTL0_DROP_IMMEDIATE;
2585 	else
2586 		drop_mode = FBNIC_QUEUE_RDE_CTL0_DROP_NEVER;
2587 
2588 	/* Specify packet layout */
2589 	rcq_ctl = FIELD_PREP(FBNIC_QUEUE_RDE_CTL0_DROP_MODE_MASK, drop_mode) |
2590 	    FIELD_PREP(FBNIC_QUEUE_RDE_CTL0_MIN_HROOM_MASK, FBNIC_RX_HROOM) |
2591 	    FIELD_PREP(FBNIC_QUEUE_RDE_CTL0_MIN_TROOM_MASK, FBNIC_RX_TROOM);
2592 
2593 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RDE_CTL0, rcq_ctl);
2594 }
2595 
2596 void fbnic_config_drop_mode(struct fbnic_net *fbn, bool tx_pause)
2597 {
2598 	int i, t;
2599 
2600 	for (i = 0; i < fbn->num_napi; i++) {
2601 		struct fbnic_napi_vector *nv = fbn->napi[i];
2602 
2603 		for (t = 0; t < nv->rxt_count; t++) {
2604 			struct fbnic_q_triad *qt = &nv->qt[nv->txt_count + t];
2605 
2606 			fbnic_config_drop_mode_rcq(nv, &qt->cmpl, tx_pause);
2607 		}
2608 	}
2609 }
2610 
2611 static void fbnic_config_rim_threshold(struct fbnic_ring *rcq, u16 nv_idx, u32 rx_desc)
2612 {
2613 	u32 threshold;
2614 
2615 	/* Set the threhsold to half the ring size if rx_frames
2616 	 * is not configured
2617 	 */
2618 	threshold = rx_desc ? : rcq->size_mask / 2;
2619 
2620 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RIM_CTL, nv_idx);
2621 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RIM_THRESHOLD, threshold);
2622 }
2623 
2624 void fbnic_config_txrx_usecs(struct fbnic_napi_vector *nv, u32 arm)
2625 {
2626 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2627 	struct fbnic_dev *fbd = nv->fbd;
2628 	u32 val = arm;
2629 
2630 	val |= FIELD_PREP(FBNIC_INTR_CQ_REARM_RCQ_TIMEOUT, fbn->rx_usecs) |
2631 	       FBNIC_INTR_CQ_REARM_RCQ_TIMEOUT_UPD_EN;
2632 	val |= FIELD_PREP(FBNIC_INTR_CQ_REARM_TCQ_TIMEOUT, fbn->tx_usecs) |
2633 	       FBNIC_INTR_CQ_REARM_TCQ_TIMEOUT_UPD_EN;
2634 
2635 	fbnic_wr32(fbd, FBNIC_INTR_CQ_REARM(nv->v_idx), val);
2636 }
2637 
2638 void fbnic_config_rx_frames(struct fbnic_napi_vector *nv)
2639 {
2640 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2641 	int i;
2642 
2643 	for (i = nv->txt_count; i < nv->rxt_count + nv->txt_count; i++) {
2644 		struct fbnic_q_triad *qt = &nv->qt[i];
2645 
2646 		fbnic_config_rim_threshold(&qt->cmpl, nv->v_idx,
2647 					   fbn->rx_max_frames *
2648 					   FBNIC_MIN_RXD_PER_FRAME);
2649 	}
2650 }
2651 
2652 static void fbnic_enable_rcq(struct fbnic_napi_vector *nv,
2653 			     struct fbnic_ring *rcq)
2654 {
2655 	struct fbnic_net *fbn = netdev_priv(nv->napi.dev);
2656 	u32 log_size = fls(rcq->size_mask);
2657 	u32 hds_thresh = fbn->hds_thresh;
2658 	u32 rcq_ctl = 0;
2659 
2660 	fbnic_config_drop_mode_rcq(nv, rcq, fbn->tx_pause);
2661 
2662 	/* Force lower bound on MAX_HEADER_BYTES. Below this, all frames should
2663 	 * be split at L4. It would also result in the frames being split at
2664 	 * L2/L3 depending on the frame size.
2665 	 */
2666 	if (fbn->hds_thresh < FBNIC_HDR_BYTES_MIN) {
2667 		rcq_ctl = FBNIC_QUEUE_RDE_CTL0_EN_HDR_SPLIT;
2668 		hds_thresh = FBNIC_HDR_BYTES_MIN;
2669 	}
2670 
2671 	rcq_ctl |= FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_PADLEN_MASK, FBNIC_RX_PAD) |
2672 		   FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_MAX_HDR_MASK, hds_thresh) |
2673 		   FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_PAYLD_OFF_MASK,
2674 			      FBNIC_RX_PAYLD_OFFSET) |
2675 		   FIELD_PREP(FBNIC_QUEUE_RDE_CTL1_PAYLD_PG_CL_MASK,
2676 			      FBNIC_RX_PAYLD_PG_CL);
2677 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RDE_CTL1, rcq_ctl);
2678 
2679 	/* Reset head/tail */
2680 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_CTL, FBNIC_QUEUE_RCQ_CTL_RESET);
2681 	rcq->head = 0;
2682 	rcq->tail = 0;
2683 
2684 	/* Store descriptor ring address and size */
2685 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_BAL, lower_32_bits(rcq->dma));
2686 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_BAH, upper_32_bits(rcq->dma));
2687 
2688 	/* Write lower 4 bits of log size as 64K ring size is 0 */
2689 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_SIZE, log_size & 0xf);
2690 
2691 	/* Store interrupt information for the completion queue */
2692 	fbnic_config_rim_threshold(rcq, nv->v_idx, fbn->rx_max_frames *
2693 						   FBNIC_MIN_RXD_PER_FRAME);
2694 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RIM_MASK, 0);
2695 
2696 	/* Enable queue */
2697 	fbnic_ring_wr32(rcq, FBNIC_QUEUE_RCQ_CTL, FBNIC_QUEUE_RCQ_CTL_ENABLE);
2698 }
2699 
2700 static void __fbnic_nv_enable(struct fbnic_napi_vector *nv)
2701 {
2702 	int j, t;
2703 
2704 	/* Setup Tx Queue Triads */
2705 	for (t = 0; t < nv->txt_count; t++) {
2706 		struct fbnic_q_triad *qt = &nv->qt[t];
2707 
2708 		fbnic_enable_twq0(&qt->sub0);
2709 		fbnic_enable_twq1(&qt->sub1);
2710 		fbnic_enable_tcq(nv, &qt->cmpl);
2711 	}
2712 
2713 	/* Setup Rx Queue Triads */
2714 	for (j = 0; j < nv->rxt_count; j++, t++) {
2715 		struct fbnic_q_triad *qt = &nv->qt[t];
2716 
2717 		page_pool_enable_direct_recycling(qt->sub0.page_pool,
2718 						  &nv->napi);
2719 		page_pool_enable_direct_recycling(qt->sub1.page_pool,
2720 						  &nv->napi);
2721 
2722 		fbnic_enable_bdq(&qt->sub0, &qt->sub1);
2723 		fbnic_enable_rcq(nv, &qt->cmpl);
2724 	}
2725 }
2726 
2727 static void fbnic_nv_enable(struct fbnic_net *fbn, struct fbnic_napi_vector *nv)
2728 {
2729 	__fbnic_nv_enable(nv);
2730 	fbnic_wrfl(fbn->fbd);
2731 }
2732 
2733 void fbnic_enable(struct fbnic_net *fbn)
2734 {
2735 	struct fbnic_dev *fbd = fbn->fbd;
2736 	int i;
2737 
2738 	for (i = 0; i < fbn->num_napi; i++)
2739 		__fbnic_nv_enable(fbn->napi[i]);
2740 
2741 	fbnic_wrfl(fbd);
2742 }
2743 
2744 static void fbnic_nv_irq_enable(struct fbnic_napi_vector *nv)
2745 {
2746 	fbnic_config_txrx_usecs(nv, FBNIC_INTR_CQ_REARM_INTR_UNMASK);
2747 }
2748 
2749 void fbnic_napi_enable(struct fbnic_net *fbn)
2750 {
2751 	u32 irqs[FBNIC_MAX_MSIX_VECS / 32] = {};
2752 	struct fbnic_dev *fbd = fbn->fbd;
2753 	int i;
2754 
2755 	for (i = 0; i < fbn->num_napi; i++) {
2756 		struct fbnic_napi_vector *nv = fbn->napi[i];
2757 
2758 		napi_enable_locked(&nv->napi);
2759 
2760 		fbnic_nv_irq_enable(nv);
2761 
2762 		/* Record bit used for NAPI IRQs so we can
2763 		 * set the mask appropriately
2764 		 */
2765 		irqs[nv->v_idx / 32] |= BIT(nv->v_idx % 32);
2766 	}
2767 
2768 	/* Force the first interrupt on the device to guarantee
2769 	 * that any packets that may have been enqueued during the
2770 	 * bringup are processed.
2771 	 */
2772 	for (i = 0; i < ARRAY_SIZE(irqs); i++) {
2773 		if (!irqs[i])
2774 			continue;
2775 		fbnic_wr32(fbd, FBNIC_INTR_SET(i), irqs[i]);
2776 	}
2777 
2778 	fbnic_wrfl(fbd);
2779 }
2780 
2781 void fbnic_napi_depletion_check(struct net_device *netdev)
2782 {
2783 	struct fbnic_net *fbn = netdev_priv(netdev);
2784 	u32 irqs[FBNIC_MAX_MSIX_VECS / 32] = {};
2785 	struct fbnic_dev *fbd = fbn->fbd;
2786 	int i, j, t;
2787 
2788 	for (i = 0; i < fbn->num_napi; i++) {
2789 		struct fbnic_napi_vector *nv = fbn->napi[i];
2790 
2791 		/* Find RQs which are completely out of pages */
2792 		for (t = nv->txt_count, j = 0; j < nv->rxt_count; j++, t++) {
2793 			/* Assume 4 pages is always enough to fit a packet
2794 			 * and therefore generate a completion and an IRQ.
2795 			 */
2796 			if (fbnic_desc_used(&nv->qt[t].sub0) < 4 ||
2797 			    fbnic_desc_used(&nv->qt[t].sub1) < 4)
2798 				irqs[nv->v_idx / 32] |= BIT(nv->v_idx % 32);
2799 		}
2800 	}
2801 
2802 	for (i = 0; i < ARRAY_SIZE(irqs); i++) {
2803 		if (!irqs[i])
2804 			continue;
2805 		fbnic_wr32(fbd, FBNIC_INTR_MASK_CLEAR(i), irqs[i]);
2806 		fbnic_wr32(fbd, FBNIC_INTR_SET(i), irqs[i]);
2807 	}
2808 
2809 	fbnic_wrfl(fbd);
2810 }
2811 
2812 static int fbnic_queue_mem_alloc(struct net_device *dev, void *qmem, int idx)
2813 {
2814 	struct fbnic_net *fbn = netdev_priv(dev);
2815 	const struct fbnic_q_triad *real;
2816 	struct fbnic_q_triad *qt = qmem;
2817 	struct fbnic_napi_vector *nv;
2818 
2819 	if (!netif_running(dev))
2820 		return fbnic_alloc_qt_page_pools(fbn, qt, idx);
2821 
2822 	real = container_of(fbn->rx[idx], struct fbnic_q_triad, cmpl);
2823 	nv = fbn->napi[idx % fbn->num_napi];
2824 
2825 	fbnic_ring_init(&qt->sub0, real->sub0.doorbell, real->sub0.q_idx,
2826 			real->sub0.flags);
2827 	fbnic_ring_init(&qt->sub1, real->sub1.doorbell, real->sub1.q_idx,
2828 			real->sub1.flags);
2829 	fbnic_ring_init(&qt->cmpl, real->cmpl.doorbell, real->cmpl.q_idx,
2830 			real->cmpl.flags);
2831 
2832 	return fbnic_alloc_rx_qt_resources(fbn, nv, qt);
2833 }
2834 
2835 static void fbnic_queue_mem_free(struct net_device *dev, void *qmem)
2836 {
2837 	struct fbnic_net *fbn = netdev_priv(dev);
2838 	struct fbnic_q_triad *qt = qmem;
2839 
2840 	if (!netif_running(dev))
2841 		fbnic_free_qt_page_pools(qt);
2842 	else
2843 		fbnic_free_qt_resources(fbn, qt);
2844 }
2845 
2846 static void __fbnic_nv_restart(struct fbnic_net *fbn,
2847 			       struct fbnic_napi_vector *nv)
2848 {
2849 	struct fbnic_dev *fbd = fbn->fbd;
2850 	int i;
2851 
2852 	fbnic_nv_enable(fbn, nv);
2853 	fbnic_nv_fill(nv);
2854 
2855 	napi_enable_locked(&nv->napi);
2856 	fbnic_nv_irq_enable(nv);
2857 	fbnic_wr32(fbd, FBNIC_INTR_SET(nv->v_idx / 32), BIT(nv->v_idx % 32));
2858 	fbnic_wrfl(fbd);
2859 
2860 	for (i = 0; i < nv->txt_count; i++)
2861 		netif_wake_subqueue(fbn->netdev, nv->qt[i].sub0.q_idx);
2862 }
2863 
2864 static int fbnic_queue_start(struct net_device *dev, void *qmem, int idx)
2865 {
2866 	struct fbnic_net *fbn = netdev_priv(dev);
2867 	struct fbnic_napi_vector *nv;
2868 	struct fbnic_q_triad *real;
2869 
2870 	real = container_of(fbn->rx[idx], struct fbnic_q_triad, cmpl);
2871 	nv = fbn->napi[idx % fbn->num_napi];
2872 
2873 	fbnic_aggregate_ring_bdq_counters(fbn, &real->sub0);
2874 	fbnic_aggregate_ring_bdq_counters(fbn, &real->sub1);
2875 	fbnic_aggregate_ring_rx_counters(fbn, &real->cmpl);
2876 
2877 	memcpy(real, qmem, sizeof(*real));
2878 
2879 	__fbnic_nv_restart(fbn, nv);
2880 
2881 	return 0;
2882 }
2883 
2884 static int fbnic_queue_stop(struct net_device *dev, void *qmem, int idx)
2885 {
2886 	struct fbnic_net *fbn = netdev_priv(dev);
2887 	const struct fbnic_q_triad *real;
2888 	struct fbnic_napi_vector *nv;
2889 	int i, t;
2890 	int err;
2891 
2892 	real = container_of(fbn->rx[idx], struct fbnic_q_triad, cmpl);
2893 	nv = fbn->napi[idx % fbn->num_napi];
2894 
2895 	napi_disable_locked(&nv->napi);
2896 	fbnic_nv_irq_disable(nv);
2897 
2898 	for (i = 0; i < nv->txt_count; i++)
2899 		netif_stop_subqueue(dev, nv->qt[i].sub0.q_idx);
2900 	fbnic_nv_disable(fbn, nv);
2901 
2902 	for (t = 0; t < nv->txt_count + nv->rxt_count; t++) {
2903 		err = fbnic_wait_queue_idle(fbn, t >= nv->txt_count,
2904 					    nv->qt[t].sub0.q_idx);
2905 		if (err)
2906 			goto err_restart;
2907 	}
2908 
2909 	fbnic_synchronize_irq(fbn->fbd, nv->v_idx);
2910 	fbnic_nv_flush(nv);
2911 
2912 	page_pool_disable_direct_recycling(real->sub0.page_pool);
2913 	page_pool_disable_direct_recycling(real->sub1.page_pool);
2914 
2915 	memcpy(qmem, real, sizeof(*real));
2916 
2917 	return 0;
2918 
2919 err_restart:
2920 	__fbnic_nv_restart(fbn, nv);
2921 	return err;
2922 }
2923 
2924 const struct netdev_queue_mgmt_ops fbnic_queue_mgmt_ops = {
2925 	.ndo_queue_mem_size	= sizeof(struct fbnic_q_triad),
2926 	.ndo_queue_mem_alloc	= fbnic_queue_mem_alloc,
2927 	.ndo_queue_mem_free	= fbnic_queue_mem_free,
2928 	.ndo_queue_start	= fbnic_queue_start,
2929 	.ndo_queue_stop		= fbnic_queue_stop,
2930 };
2931