xref: /freebsd/sys/dev/ena/ena_netmap.c (revision 2e3507c25e42292b45a5482e116d278f5515d04d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2015-2023 Amazon.com, Inc. or its affiliates.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  *
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 #include <sys/cdefs.h>
31 #ifdef DEV_NETMAP
32 
33 #include "ena.h"
34 #include "ena_netmap.h"
35 
36 #define ENA_NETMAP_MORE_FRAMES		1
37 #define ENA_NETMAP_NO_MORE_FRAMES	0
38 #define ENA_MAX_FRAMES			16384
39 
40 struct ena_netmap_ctx {
41 	struct netmap_kring *kring;
42 	struct ena_adapter *adapter;
43 	struct netmap_adapter *na;
44 	struct netmap_slot *slots;
45 	struct ena_ring *ring;
46 	struct ena_com_io_cq *io_cq;
47 	struct ena_com_io_sq *io_sq;
48 	u_int nm_i;
49 	uint16_t nt;
50 	uint16_t lim;
51 };
52 
53 /* Netmap callbacks */
54 static int ena_netmap_reg(struct netmap_adapter *, int);
55 static int ena_netmap_txsync(struct netmap_kring *, int);
56 static int ena_netmap_rxsync(struct netmap_kring *, int);
57 
58 /* Helper functions */
59 static int ena_netmap_tx_frames(struct ena_netmap_ctx *);
60 static int ena_netmap_tx_frame(struct ena_netmap_ctx *);
61 static inline uint16_t ena_netmap_count_slots(struct ena_netmap_ctx *);
62 static inline uint16_t ena_netmap_packet_len(struct netmap_slot *, u_int,
63     uint16_t);
64 static int ena_netmap_copy_data(struct netmap_adapter *, struct netmap_slot *,
65     u_int, uint16_t, uint16_t, void *);
66 static int ena_netmap_map_single_slot(struct netmap_adapter *,
67     struct netmap_slot *, bus_dma_tag_t, bus_dmamap_t, void **, uint64_t *);
68 static int ena_netmap_tx_map_slots(struct ena_netmap_ctx *,
69     struct ena_tx_buffer *, void **, uint16_t *, uint16_t *);
70 static void ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *,
71     struct ena_tx_buffer *);
72 static void ena_netmap_tx_cleanup(struct ena_netmap_ctx *);
73 static uint16_t ena_netmap_tx_clean_one(struct ena_netmap_ctx *, uint16_t);
74 static inline int validate_tx_req_id(struct ena_ring *, uint16_t);
75 static int ena_netmap_rx_frames(struct ena_netmap_ctx *);
76 static int ena_netmap_rx_frame(struct ena_netmap_ctx *);
77 static int ena_netmap_rx_load_desc(struct ena_netmap_ctx *, uint16_t, int *);
78 static void ena_netmap_rx_cleanup(struct ena_netmap_ctx *);
79 static void ena_netmap_fill_ctx(struct netmap_kring *, struct ena_netmap_ctx *,
80     uint16_t);
81 
82 int
83 ena_netmap_attach(struct ena_adapter *adapter)
84 {
85 	struct netmap_adapter na;
86 
87 	ena_log_nm(adapter->pdev, INFO, "netmap attach\n");
88 
89 	bzero(&na, sizeof(na));
90 	na.na_flags = NAF_MOREFRAG;
91 	na.ifp = adapter->ifp;
92 	na.num_tx_desc = adapter->requested_tx_ring_size;
93 	na.num_rx_desc = adapter->requested_rx_ring_size;
94 	na.num_tx_rings = adapter->num_io_queues;
95 	na.num_rx_rings = adapter->num_io_queues;
96 	na.rx_buf_maxsize = adapter->buf_ring_size;
97 	na.nm_txsync = ena_netmap_txsync;
98 	na.nm_rxsync = ena_netmap_rxsync;
99 	na.nm_register = ena_netmap_reg;
100 
101 	return (netmap_attach(&na));
102 }
103 
104 int
105 ena_netmap_alloc_rx_slot(struct ena_adapter *adapter, struct ena_ring *rx_ring,
106     struct ena_rx_buffer *rx_info)
107 {
108 	struct netmap_adapter *na = NA(adapter->ifp);
109 	struct netmap_kring *kring;
110 	struct netmap_ring *ring;
111 	struct netmap_slot *slot;
112 	void *addr;
113 	uint64_t paddr;
114 	int nm_i, qid, head, lim, rc;
115 
116 	/* if previously allocated frag is not used */
117 	if (unlikely(rx_info->netmap_buf_idx != 0))
118 		return (0);
119 
120 	qid = rx_ring->qid;
121 	kring = na->rx_rings[qid];
122 	nm_i = kring->nr_hwcur;
123 	head = kring->rhead;
124 
125 	ena_log_nm(adapter->pdev, DBG,
126 	    "nr_hwcur: %d, nr_hwtail: %d, rhead: %d, rcur: %d, rtail: %d\n",
127 	    kring->nr_hwcur, kring->nr_hwtail, kring->rhead, kring->rcur,
128 	    kring->rtail);
129 
130 	if ((nm_i == head) && rx_ring->initialized) {
131 		ena_log_nm(adapter->pdev, ERR,
132 		    "No free slots in netmap ring\n");
133 		return (ENOMEM);
134 	}
135 
136 	ring = kring->ring;
137 	if (ring == NULL) {
138 		ena_log_nm(adapter->pdev, ERR, "Rx ring %d is NULL\n", qid);
139 		return (EFAULT);
140 	}
141 	slot = &ring->slot[nm_i];
142 
143 	addr = PNMB(na, slot, &paddr);
144 	if (addr == NETMAP_BUF_BASE(na)) {
145 		ena_log_nm(adapter->pdev, ERR, "Bad buff in slot\n");
146 		return (EFAULT);
147 	}
148 
149 	rc = netmap_load_map(na, adapter->rx_buf_tag, rx_info->map, addr);
150 	if (rc != 0) {
151 		ena_log_nm(adapter->pdev, WARN, "DMA mapping error\n");
152 		return (rc);
153 	}
154 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map, BUS_DMASYNC_PREREAD);
155 
156 	rx_info->ena_buf.paddr = paddr;
157 	rx_info->ena_buf.len = ring->nr_buf_size;
158 	rx_info->mbuf = NULL;
159 	rx_info->netmap_buf_idx = slot->buf_idx;
160 
161 	slot->buf_idx = 0;
162 
163 	lim = kring->nkr_num_slots - 1;
164 	kring->nr_hwcur = nm_next(nm_i, lim);
165 
166 	return (0);
167 }
168 
169 void
170 ena_netmap_free_rx_slot(struct ena_adapter *adapter, struct ena_ring *rx_ring,
171     struct ena_rx_buffer *rx_info)
172 {
173 	struct netmap_adapter *na;
174 	struct netmap_kring *kring;
175 	struct netmap_slot *slot;
176 	int nm_i, qid, lim;
177 
178 	na = NA(adapter->ifp);
179 	if (na == NULL) {
180 		ena_log_nm(adapter->pdev, ERR, "netmap adapter is NULL\n");
181 		return;
182 	}
183 
184 	if (na->rx_rings == NULL) {
185 		ena_log_nm(adapter->pdev, ERR, "netmap rings are NULL\n");
186 		return;
187 	}
188 
189 	qid = rx_ring->qid;
190 	kring = na->rx_rings[qid];
191 	if (kring == NULL) {
192 		ena_log_nm(adapter->pdev, ERR,
193 		    "netmap kernel ring %d is NULL\n", qid);
194 		return;
195 	}
196 
197 	lim = kring->nkr_num_slots - 1;
198 	nm_i = nm_prev(kring->nr_hwcur, lim);
199 
200 	if (kring->nr_mode != NKR_NETMAP_ON)
201 		return;
202 
203 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map,
204 	    BUS_DMASYNC_POSTREAD);
205 	netmap_unload_map(na, adapter->rx_buf_tag, rx_info->map);
206 
207 	KASSERT(kring->ring != NULL, ("Netmap Rx ring is NULL\n"));
208 
209 	slot = &kring->ring->slot[nm_i];
210 
211 	ENA_WARN(slot->buf_idx != 0, adapter->ena_dev, "Overwrite slot buf\n");
212 	slot->buf_idx = rx_info->netmap_buf_idx;
213 	slot->flags = NS_BUF_CHANGED;
214 
215 	rx_info->netmap_buf_idx = 0;
216 	kring->nr_hwcur = nm_i;
217 }
218 
219 static bool
220 ena_ring_in_netmap(struct ena_adapter *adapter, int qid, enum txrx x)
221 {
222 	struct netmap_adapter *na;
223 	struct netmap_kring *kring;
224 
225 	if (if_getcapenable(adapter->ifp) & IFCAP_NETMAP) {
226 		na = NA(adapter->ifp);
227 		kring = (x == NR_RX) ? na->rx_rings[qid] : na->tx_rings[qid];
228 		if (kring->nr_mode == NKR_NETMAP_ON)
229 			return true;
230 	}
231 	return false;
232 }
233 
234 bool
235 ena_tx_ring_in_netmap(struct ena_adapter *adapter, int qid)
236 {
237 	return ena_ring_in_netmap(adapter, qid, NR_TX);
238 }
239 
240 bool
241 ena_rx_ring_in_netmap(struct ena_adapter *adapter, int qid)
242 {
243 	return ena_ring_in_netmap(adapter, qid, NR_RX);
244 }
245 
246 static void
247 ena_netmap_reset_ring(struct ena_adapter *adapter, int qid, enum txrx x)
248 {
249 	if (!ena_ring_in_netmap(adapter, qid, x))
250 		return;
251 
252 	netmap_reset(NA(adapter->ifp), x, qid, 0);
253 	ena_log_nm(adapter->pdev, INFO, "%s ring %d is in netmap mode\n",
254 	    (x == NR_TX) ? "Tx" : "Rx", qid);
255 }
256 
257 void
258 ena_netmap_reset_rx_ring(struct ena_adapter *adapter, int qid)
259 {
260 	ena_netmap_reset_ring(adapter, qid, NR_RX);
261 }
262 
263 void
264 ena_netmap_reset_tx_ring(struct ena_adapter *adapter, int qid)
265 {
266 	ena_netmap_reset_ring(adapter, qid, NR_TX);
267 }
268 
269 static int
270 ena_netmap_reg(struct netmap_adapter *na, int onoff)
271 {
272 	if_t ifp = na->ifp;
273 	struct ena_adapter *adapter = if_getsoftc(ifp);
274 	device_t pdev = adapter->pdev;
275 	struct netmap_kring *kring;
276 	enum txrx t;
277 	int rc, i;
278 
279 	ENA_LOCK_LOCK();
280 	ENA_FLAG_CLEAR_ATOMIC(ENA_FLAG_TRIGGER_RESET, adapter);
281 	ena_down(adapter);
282 
283 	if (onoff) {
284 		ena_log_nm(pdev, INFO, "netmap on\n");
285 		for_rx_tx(t) {
286 			for (i = 0; i <= nma_get_nrings(na, t); i++) {
287 				kring = NMR(na, t)[i];
288 				if (nm_kring_pending_on(kring)) {
289 					kring->nr_mode = NKR_NETMAP_ON;
290 				}
291 			}
292 		}
293 		nm_set_native_flags(na);
294 	} else {
295 		ena_log_nm(pdev, INFO, "netmap off\n");
296 		nm_clear_native_flags(na);
297 		for_rx_tx(t) {
298 			for (i = 0; i <= nma_get_nrings(na, t); i++) {
299 				kring = NMR(na, t)[i];
300 				if (nm_kring_pending_off(kring)) {
301 					kring->nr_mode = NKR_NETMAP_OFF;
302 				}
303 			}
304 		}
305 	}
306 
307 	rc = ena_up(adapter);
308 	if (rc != 0) {
309 		ena_log_nm(pdev, WARN, "ena_up failed with rc=%d\n", rc);
310 		adapter->reset_reason = ENA_REGS_RESET_DRIVER_INVALID_STATE;
311 		nm_clear_native_flags(na);
312 		ena_destroy_device(adapter, false);
313 		ENA_FLAG_SET_ATOMIC(ENA_FLAG_DEV_UP_BEFORE_RESET, adapter);
314 		rc = ena_restore_device(adapter);
315 	}
316 	ENA_LOCK_UNLOCK();
317 
318 	return (rc);
319 }
320 
321 static int
322 ena_netmap_txsync(struct netmap_kring *kring, int flags)
323 {
324 	struct ena_netmap_ctx ctx;
325 	int rc = 0;
326 
327 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_TXQ_IDX(kring->ring_id));
328 	ctx.ring = &ctx.adapter->tx_ring[kring->ring_id];
329 
330 	ENA_RING_MTX_LOCK(ctx.ring);
331 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEV_UP, ctx.adapter)))
332 		goto txsync_end;
333 
334 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
335 		goto txsync_end;
336 
337 	rc = ena_netmap_tx_frames(&ctx);
338 	ena_netmap_tx_cleanup(&ctx);
339 
340 txsync_end:
341 	ENA_RING_MTX_UNLOCK(ctx.ring);
342 	return (rc);
343 }
344 
345 static int
346 ena_netmap_tx_frames(struct ena_netmap_ctx *ctx)
347 {
348 	struct ena_ring *tx_ring = ctx->ring;
349 	int rc = 0;
350 
351 	ctx->nm_i = ctx->kring->nr_hwcur;
352 	ctx->nt = ctx->ring->next_to_use;
353 
354 	__builtin_prefetch(&ctx->slots[ctx->nm_i]);
355 
356 	while (ctx->nm_i != ctx->kring->rhead) {
357 		if ((rc = ena_netmap_tx_frame(ctx)) != 0) {
358 			/*
359 			 * When there is no empty space in Tx ring, error is
360 			 * still being returned. It should not be passed to the
361 			 * netmap, as application knows current ring state from
362 			 * netmap ring pointers. Returning error there could
363 			 * cause application to exit, but the Tx ring is
364 			 * commonly being full.
365 			 */
366 			if (rc == ENA_COM_NO_MEM)
367 				rc = 0;
368 			break;
369 		}
370 		tx_ring->acum_pkts++;
371 	}
372 
373 	/* If any packet was sent... */
374 	if (likely(ctx->nm_i != ctx->kring->nr_hwcur)) {
375 		/* ...send the doorbell to the device. */
376 		ena_ring_tx_doorbell(tx_ring);
377 
378 		ctx->ring->next_to_use = ctx->nt;
379 		ctx->kring->nr_hwcur = ctx->nm_i;
380 	}
381 
382 	return (rc);
383 }
384 
385 static int
386 ena_netmap_tx_frame(struct ena_netmap_ctx *ctx)
387 {
388 	struct ena_com_tx_ctx ena_tx_ctx;
389 	struct ena_adapter *adapter;
390 	struct ena_ring *tx_ring;
391 	struct ena_tx_buffer *tx_info;
392 	uint16_t req_id;
393 	uint16_t header_len;
394 	uint16_t packet_len;
395 	int nb_hw_desc;
396 	int rc;
397 	void *push_hdr;
398 
399 	adapter = ctx->adapter;
400 	if (ena_netmap_count_slots(ctx) > adapter->max_tx_sgl_size) {
401 		ena_log_nm(adapter->pdev, WARN, "Too many slots per packet\n");
402 		return (EINVAL);
403 	}
404 
405 	tx_ring = ctx->ring;
406 
407 	req_id = tx_ring->free_tx_ids[ctx->nt];
408 	tx_info = &tx_ring->tx_buffer_info[req_id];
409 	tx_info->num_of_bufs = 0;
410 	tx_info->nm_info.sockets_used = 0;
411 
412 	rc = ena_netmap_tx_map_slots(ctx, tx_info, &push_hdr, &header_len,
413 	    &packet_len);
414 	if (unlikely(rc != 0)) {
415 		ena_log_nm(adapter->pdev, ERR, "Failed to map Tx slot\n");
416 		return (rc);
417 	}
418 
419 	bzero(&ena_tx_ctx, sizeof(struct ena_com_tx_ctx));
420 	ena_tx_ctx.ena_bufs = tx_info->bufs;
421 	ena_tx_ctx.push_header = push_hdr;
422 	ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
423 	ena_tx_ctx.req_id = req_id;
424 	ena_tx_ctx.header_len = header_len;
425 	ena_tx_ctx.meta_valid = adapter->disable_meta_caching;
426 
427 	/* There are no any offloads, as the netmap doesn't support them */
428 
429 	if (tx_ring->acum_pkts == ENA_DB_THRESHOLD ||
430 	    ena_com_is_doorbell_needed(ctx->io_sq, &ena_tx_ctx))
431 		ena_ring_tx_doorbell(tx_ring);
432 
433 	rc = ena_com_prepare_tx(ctx->io_sq, &ena_tx_ctx, &nb_hw_desc);
434 	if (unlikely(rc != 0)) {
435 		if (likely(rc == ENA_COM_NO_MEM)) {
436 			ena_log_nm(adapter->pdev, DBG,
437 			    "Tx ring[%d] is out of space\n", tx_ring->que->id);
438 		} else {
439 			ena_log_nm(adapter->pdev, ERR,
440 			    "Failed to prepare Tx bufs\n");
441 			ena_trigger_reset(adapter,
442 			    ENA_REGS_RESET_DRIVER_INVALID_STATE);
443 		}
444 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
445 
446 		ena_netmap_unmap_last_socket_chain(ctx, tx_info);
447 		return (rc);
448 	}
449 
450 	counter_enter();
451 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
452 	counter_u64_add_protected(tx_ring->tx_stats.bytes, packet_len);
453 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
454 	counter_u64_add_protected(adapter->hw_stats.tx_bytes, packet_len);
455 	counter_exit();
456 
457 	tx_info->tx_descs = nb_hw_desc;
458 
459 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
460 
461 	for (unsigned int i = 0; i < tx_info->num_of_bufs; i++)
462 		bus_dmamap_sync(adapter->tx_buf_tag,
463 		    tx_info->nm_info.map_seg[i], BUS_DMASYNC_PREWRITE);
464 
465 	return (0);
466 }
467 
468 static inline uint16_t
469 ena_netmap_count_slots(struct ena_netmap_ctx *ctx)
470 {
471 	uint16_t slots = 1;
472 	uint16_t nm = ctx->nm_i;
473 
474 	while ((ctx->slots[nm].flags & NS_MOREFRAG) != 0) {
475 		slots++;
476 		nm = nm_next(nm, ctx->lim);
477 	}
478 
479 	return slots;
480 }
481 
482 static inline uint16_t
483 ena_netmap_packet_len(struct netmap_slot *slots, u_int slot_index,
484     uint16_t limit)
485 {
486 	struct netmap_slot *nm_slot;
487 	uint16_t packet_size = 0;
488 
489 	do {
490 		nm_slot = &slots[slot_index];
491 		packet_size += nm_slot->len;
492 		slot_index = nm_next(slot_index, limit);
493 	} while ((nm_slot->flags & NS_MOREFRAG) != 0);
494 
495 	return packet_size;
496 }
497 
498 static int
499 ena_netmap_copy_data(struct netmap_adapter *na, struct netmap_slot *slots,
500     u_int slot_index, uint16_t limit, uint16_t bytes_to_copy, void *destination)
501 {
502 	struct netmap_slot *nm_slot;
503 	void *slot_vaddr;
504 	uint16_t data_amount;
505 
506 	do {
507 		nm_slot = &slots[slot_index];
508 		slot_vaddr = NMB(na, nm_slot);
509 		if (unlikely(slot_vaddr == NULL))
510 			return (EINVAL);
511 
512 		data_amount = min_t(uint16_t, bytes_to_copy, nm_slot->len);
513 		memcpy(destination, slot_vaddr, data_amount);
514 		bytes_to_copy -= data_amount;
515 
516 		slot_index = nm_next(slot_index, limit);
517 	} while ((nm_slot->flags & NS_MOREFRAG) != 0 && bytes_to_copy > 0);
518 
519 	return (0);
520 }
521 
522 static int
523 ena_netmap_map_single_slot(struct netmap_adapter *na, struct netmap_slot *slot,
524     bus_dma_tag_t dmatag, bus_dmamap_t dmamap, void **vaddr, uint64_t *paddr)
525 {
526 	device_t pdev;
527 	int rc;
528 
529 	pdev = ((struct ena_adapter *)if_getsoftc(na->ifp))->pdev;
530 
531 	*vaddr = PNMB(na, slot, paddr);
532 	if (unlikely(vaddr == NULL)) {
533 		ena_log_nm(pdev, ERR, "Slot address is NULL\n");
534 		return (EINVAL);
535 	}
536 
537 	rc = netmap_load_map(na, dmatag, dmamap, *vaddr);
538 	if (unlikely(rc != 0)) {
539 		ena_log_nm(pdev, ERR, "Failed to map slot %d for DMA\n",
540 		    slot->buf_idx);
541 		return (EINVAL);
542 	}
543 
544 	return (0);
545 }
546 
547 static int
548 ena_netmap_tx_map_slots(struct ena_netmap_ctx *ctx,
549     struct ena_tx_buffer *tx_info, void **push_hdr, uint16_t *header_len,
550     uint16_t *packet_len)
551 {
552 	struct netmap_slot *slot;
553 	struct ena_com_buf *ena_buf;
554 	struct ena_adapter *adapter;
555 	struct ena_ring *tx_ring;
556 	struct ena_netmap_tx_info *nm_info;
557 	bus_dmamap_t *nm_maps;
558 	void *vaddr;
559 	uint64_t paddr;
560 	uint32_t *nm_buf_idx;
561 	uint32_t slot_head_len;
562 	uint32_t frag_len;
563 	uint32_t remaining_len;
564 	uint16_t push_len;
565 	uint16_t delta;
566 	int rc;
567 
568 	adapter = ctx->adapter;
569 	tx_ring = ctx->ring;
570 	ena_buf = tx_info->bufs;
571 	nm_info = &tx_info->nm_info;
572 	nm_maps = nm_info->map_seg;
573 	nm_buf_idx = nm_info->socket_buf_idx;
574 	slot = &ctx->slots[ctx->nm_i];
575 
576 	slot_head_len = slot->len;
577 	*packet_len = ena_netmap_packet_len(ctx->slots, ctx->nm_i, ctx->lim);
578 	remaining_len = *packet_len;
579 	delta = 0;
580 
581 	__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
582 	if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
583 		/*
584 		 * When the device is in LLQ mode, the driver will copy
585 		 * the header into the device memory space.
586 		 * The ena_com layer assumes that the header is in a linear
587 		 * memory space.
588 		 * This assumption might be wrong since part of the header
589 		 * can be in the fragmented buffers.
590 		 * First, check if header fits in the first slot. If not, copy
591 		 * it to separate buffer that will be holding linearized data.
592 		 */
593 		push_len = min_t(uint32_t, *packet_len,
594 		    tx_ring->tx_max_header_size);
595 		*header_len = push_len;
596 		/* If header is in linear space, just point to socket's data. */
597 		if (likely(push_len <= slot_head_len)) {
598 			*push_hdr = NMB(ctx->na, slot);
599 			if (unlikely(push_hdr == NULL)) {
600 				ena_log_nm(adapter->pdev, ERR,
601 				    "Slot vaddress is NULL\n");
602 				return (EINVAL);
603 			}
604 		/*
605 		 * Otherwise, copy whole portion of header from multiple
606 		 * slots to intermediate buffer.
607 		 */
608 		} else {
609 			rc = ena_netmap_copy_data(ctx->na, ctx->slots,
610 			    ctx->nm_i, ctx->lim, push_len,
611 			    tx_ring->push_buf_intermediate_buf);
612 			if (unlikely(rc)) {
613 				ena_log_nm(adapter->pdev, ERR,
614 				    "Failed to copy data from slots to push_buf\n");
615 				return (EINVAL);
616 			}
617 
618 			*push_hdr = tx_ring->push_buf_intermediate_buf;
619 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
620 
621 			delta = push_len - slot_head_len;
622 		}
623 
624 		ena_log_nm(adapter->pdev, DBG,
625 		    "slot: %d header_buf->vaddr: %p push_len: %d\n",
626 		    slot->buf_idx, *push_hdr, push_len);
627 
628 		/*
629 		 * If header was in linear memory space, map for the dma rest of
630 		 * the data in the first mbuf of the mbuf chain.
631 		 */
632 		if (slot_head_len > push_len) {
633 			rc = ena_netmap_map_single_slot(ctx->na, slot,
634 			    adapter->tx_buf_tag, *nm_maps, &vaddr, &paddr);
635 			if (unlikely(rc != 0)) {
636 				ena_log_nm(adapter->pdev, ERR,
637 				    "DMA mapping error\n");
638 				return (rc);
639 			}
640 			nm_maps++;
641 
642 			ena_buf->paddr = paddr + push_len;
643 			ena_buf->len = slot->len - push_len;
644 			ena_buf++;
645 
646 			tx_info->num_of_bufs++;
647 		}
648 
649 		remaining_len -= slot->len;
650 
651 		/* Save buf idx before advancing */
652 		*nm_buf_idx = slot->buf_idx;
653 		nm_buf_idx++;
654 		slot->buf_idx = 0;
655 
656 		/* Advance to the next socket */
657 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
658 		slot = &ctx->slots[ctx->nm_i];
659 		nm_info->sockets_used++;
660 
661 		/*
662 		 * If header is in non linear space (delta > 0), then skip mbufs
663 		 * containing header and map the last one containing both header
664 		 * and the packet data.
665 		 * The first segment is already counted in.
666 		 */
667 		while (delta > 0) {
668 			__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
669 			frag_len = slot->len;
670 
671 			/*
672 			 * If whole segment contains header just move to the
673 			 * next one and reduce delta.
674 			 */
675 			if (unlikely(delta >= frag_len)) {
676 				delta -= frag_len;
677 			} else {
678 				/*
679 				 * Map the data and then assign it with the
680 				 * offsets
681 				 */
682 				rc = ena_netmap_map_single_slot(ctx->na, slot,
683 				    adapter->tx_buf_tag, *nm_maps, &vaddr,
684 				    &paddr);
685 				if (unlikely(rc != 0)) {
686 					ena_log_nm(adapter->pdev, ERR,
687 					    "DMA mapping error\n");
688 					goto error_map;
689 				}
690 				nm_maps++;
691 
692 				ena_buf->paddr = paddr + delta;
693 				ena_buf->len = slot->len - delta;
694 				ena_buf++;
695 
696 				tx_info->num_of_bufs++;
697 				delta = 0;
698 			}
699 
700 			remaining_len -= slot->len;
701 
702 			/* Save buf idx before advancing */
703 			*nm_buf_idx = slot->buf_idx;
704 			nm_buf_idx++;
705 			slot->buf_idx = 0;
706 
707 			/* Advance to the next socket */
708 			ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
709 			slot = &ctx->slots[ctx->nm_i];
710 			nm_info->sockets_used++;
711 		}
712 	} else {
713 		*push_hdr = NULL;
714 		/*
715 		 * header_len is just a hint for the device. Because netmap is
716 		 * not giving us any information about packet header length and
717 		 * it is not guaranteed that all packet headers will be in the
718 		 * 1st slot, setting header_len to 0 is making the device ignore
719 		 * this value and resolve header on it's own.
720 		 */
721 		*header_len = 0;
722 	}
723 
724 	/* Map all remaining data (regular routine for non-LLQ mode) */
725 	while (remaining_len > 0) {
726 		__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
727 
728 		rc = ena_netmap_map_single_slot(ctx->na, slot,
729 		    adapter->tx_buf_tag, *nm_maps, &vaddr, &paddr);
730 		if (unlikely(rc != 0)) {
731 			ena_log_nm(adapter->pdev, ERR, "DMA mapping error\n");
732 			goto error_map;
733 		}
734 		nm_maps++;
735 
736 		ena_buf->paddr = paddr;
737 		ena_buf->len = slot->len;
738 		ena_buf++;
739 
740 		tx_info->num_of_bufs++;
741 
742 		remaining_len -= slot->len;
743 
744 		/* Save buf idx before advancing */
745 		*nm_buf_idx = slot->buf_idx;
746 		nm_buf_idx++;
747 		slot->buf_idx = 0;
748 
749 		/* Advance to the next socket */
750 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
751 		slot = &ctx->slots[ctx->nm_i];
752 		nm_info->sockets_used++;
753 	}
754 
755 	return (0);
756 
757 error_map:
758 	ena_netmap_unmap_last_socket_chain(ctx, tx_info);
759 
760 	return (rc);
761 }
762 
763 static void
764 ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *ctx,
765     struct ena_tx_buffer *tx_info)
766 {
767 	struct ena_netmap_tx_info *nm_info;
768 	int n;
769 
770 	nm_info = &tx_info->nm_info;
771 
772 	/**
773 	 * As the used sockets must not be equal to the buffers used in the LLQ
774 	 * mode, they must be treated separately.
775 	 * First, unmap the DMA maps.
776 	 */
777 	n = tx_info->num_of_bufs;
778 	while (n--) {
779 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
780 		    nm_info->map_seg[n]);
781 	}
782 	tx_info->num_of_bufs = 0;
783 
784 	/* Next, retain the sockets back to the userspace */
785 	n = nm_info->sockets_used;
786 	while (n--) {
787 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
788 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
789 		nm_info->socket_buf_idx[n] = 0;
790 		ctx->nm_i = nm_prev(ctx->nm_i, ctx->lim);
791 	}
792 	nm_info->sockets_used = 0;
793 }
794 
795 static void
796 ena_netmap_tx_cleanup(struct ena_netmap_ctx *ctx)
797 {
798 	uint16_t req_id;
799 	uint16_t total_tx_descs = 0;
800 
801 	ctx->nm_i = ctx->kring->nr_hwtail;
802 	ctx->nt = ctx->ring->next_to_clean;
803 
804 	/* Reclaim buffers for completed transmissions */
805 	while (ena_com_tx_comp_req_id_get(ctx->io_cq, &req_id) >= 0) {
806 		if (validate_tx_req_id(ctx->ring, req_id) != 0)
807 			break;
808 		total_tx_descs += ena_netmap_tx_clean_one(ctx, req_id);
809 	}
810 
811 	ctx->kring->nr_hwtail = ctx->nm_i;
812 
813 	if (total_tx_descs > 0) {
814 		/* acknowledge completion of sent packets */
815 		ctx->ring->next_to_clean = ctx->nt;
816 		ena_com_comp_ack(ctx->ring->ena_com_io_sq, total_tx_descs);
817 	}
818 }
819 
820 static uint16_t
821 ena_netmap_tx_clean_one(struct ena_netmap_ctx *ctx, uint16_t req_id)
822 {
823 	struct ena_tx_buffer *tx_info;
824 	struct ena_netmap_tx_info *nm_info;
825 	int n;
826 
827 	tx_info = &ctx->ring->tx_buffer_info[req_id];
828 	nm_info = &tx_info->nm_info;
829 
830 	/**
831 	 * As the used sockets must not be equal to the buffers used in the LLQ
832 	 * mode, they must be treated separately.
833 	 * First, unmap the DMA maps.
834 	 */
835 	n = tx_info->num_of_bufs;
836 	for (n = 0; n < tx_info->num_of_bufs; n++) {
837 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
838 		    nm_info->map_seg[n]);
839 	}
840 	tx_info->num_of_bufs = 0;
841 
842 	/* Next, retain the sockets back to the userspace */
843 	for (n = 0; n < nm_info->sockets_used; n++) {
844 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
845 		ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0,
846 		    ctx->adapter->ena_dev, "Tx idx is not 0.\n");
847 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
848 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
849 		nm_info->socket_buf_idx[n] = 0;
850 	}
851 	nm_info->sockets_used = 0;
852 
853 	ctx->ring->free_tx_ids[ctx->nt] = req_id;
854 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->lim);
855 
856 	return tx_info->tx_descs;
857 }
858 
859 static inline int
860 validate_tx_req_id(struct ena_ring *tx_ring, uint16_t req_id)
861 {
862 	struct ena_adapter *adapter = tx_ring->adapter;
863 
864 	if (likely(req_id < tx_ring->ring_size))
865 		return (0);
866 
867 	ena_log_nm(adapter->pdev, WARN, "Invalid req_id %hu in qid %hu\n",
868 	    req_id, tx_ring->qid);
869 	counter_u64_add(tx_ring->tx_stats.bad_req_id, 1);
870 
871 	ena_trigger_reset(adapter, ENA_REGS_RESET_INV_TX_REQ_ID);
872 
873 	return (EFAULT);
874 }
875 
876 static int
877 ena_netmap_rxsync(struct netmap_kring *kring, int flags)
878 {
879 	struct ena_netmap_ctx ctx;
880 	int rc;
881 
882 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_RXQ_IDX(kring->ring_id));
883 	ctx.ring = &ctx.adapter->rx_ring[kring->ring_id];
884 
885 	if (ctx.kring->rhead > ctx.lim) {
886 		/* Probably not needed to release slots from RX ring. */
887 		return (netmap_ring_reinit(ctx.kring));
888 	}
889 
890 	if (unlikely((if_getdrvflags(ctx.na->ifp) & IFF_DRV_RUNNING) == 0))
891 		return (0);
892 
893 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
894 		return (0);
895 
896 	if ((rc = ena_netmap_rx_frames(&ctx)) != 0)
897 		return (rc);
898 
899 	ena_netmap_rx_cleanup(&ctx);
900 
901 	return (0);
902 }
903 
904 static inline int
905 ena_netmap_rx_frames(struct ena_netmap_ctx *ctx)
906 {
907 	int rc = 0;
908 	int frames_counter = 0;
909 
910 	ctx->nt = ctx->ring->next_to_clean;
911 	ctx->nm_i = ctx->kring->nr_hwtail;
912 
913 	while ((rc = ena_netmap_rx_frame(ctx)) == ENA_NETMAP_MORE_FRAMES) {
914 		frames_counter++;
915 		/* In case of multiple frames, it is not an error. */
916 		rc = 0;
917 		if (frames_counter > ENA_MAX_FRAMES) {
918 			ena_log_nm(ctx->adapter->pdev, ERR,
919 			    "Driver is stuck in the Rx loop\n");
920 			break;
921 		}
922 	};
923 
924 	ctx->kring->nr_hwtail = ctx->nm_i;
925 	ctx->kring->nr_kflags &= ~NKR_PENDINTR;
926 	ctx->ring->next_to_clean = ctx->nt;
927 
928 	return (rc);
929 }
930 
931 static inline int
932 ena_netmap_rx_frame(struct ena_netmap_ctx *ctx)
933 {
934 	struct ena_com_rx_ctx ena_rx_ctx;
935 	enum ena_regs_reset_reason_types reset_reason;
936 	int rc, len = 0;
937 	uint16_t buf, nm;
938 
939 	ena_rx_ctx.ena_bufs = ctx->ring->ena_bufs;
940 	ena_rx_ctx.max_bufs = ctx->adapter->max_rx_sgl_size;
941 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
942 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
943 
944 	rc = ena_com_rx_pkt(ctx->io_cq, ctx->io_sq, &ena_rx_ctx);
945 	if (unlikely(rc != 0)) {
946 		ena_log_nm(ctx->adapter->pdev, ERR,
947 		    "Failed to read pkt from the device with error: %d\n", rc);
948 		if (rc == ENA_COM_NO_SPACE) {
949 			counter_u64_add(ctx->ring->rx_stats.bad_desc_num, 1);
950 			reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
951 		} else {
952 			counter_u64_add(ctx->ring->rx_stats.bad_req_id, 1);
953 			reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID;
954 		}
955 		ena_trigger_reset(ctx->adapter, reset_reason);
956 		return (rc);
957 	}
958 	if (unlikely(ena_rx_ctx.descs == 0))
959 		return (ENA_NETMAP_NO_MORE_FRAMES);
960 
961 	ena_log_nm(ctx->adapter->pdev, DBG,
962 	    "Rx: q %d got packet from ena. descs #:"
963 	    " %d l3 proto %d l4 proto %d hash: %x\n",
964 	    ctx->ring->qid, ena_rx_ctx.descs, ena_rx_ctx.l3_proto,
965 	    ena_rx_ctx.l4_proto, ena_rx_ctx.hash);
966 
967 	for (buf = 0; buf < ena_rx_ctx.descs; buf++)
968 		if ((rc = ena_netmap_rx_load_desc(ctx, buf, &len)) != 0)
969 			break;
970 	/*
971 	 * ena_netmap_rx_load_desc doesn't know the number of descriptors.
972 	 * It just set flag NS_MOREFRAG to all slots, then here flag of
973 	 * last slot is cleared.
974 	 */
975 	ctx->slots[nm_prev(ctx->nm_i, ctx->lim)].flags = NS_BUF_CHANGED;
976 
977 	if (rc != 0) {
978 		goto rx_clear_desc;
979 	}
980 
981 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
982 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
983 
984 	counter_enter();
985 	counter_u64_add_protected(ctx->ring->rx_stats.bytes, len);
986 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_bytes, len);
987 	counter_u64_add_protected(ctx->ring->rx_stats.cnt, 1);
988 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_packets, 1);
989 	counter_exit();
990 
991 	return (ENA_NETMAP_MORE_FRAMES);
992 
993 rx_clear_desc:
994 	nm = ctx->nm_i;
995 
996 	/* Remove failed packet from ring */
997 	while (buf--) {
998 		ctx->slots[nm].flags = 0;
999 		ctx->slots[nm].len = 0;
1000 		nm = nm_prev(nm, ctx->lim);
1001 	}
1002 
1003 	return (rc);
1004 }
1005 
1006 static inline int
1007 ena_netmap_rx_load_desc(struct ena_netmap_ctx *ctx, uint16_t buf, int *len)
1008 {
1009 	struct ena_rx_buffer *rx_info;
1010 	uint16_t req_id;
1011 
1012 	req_id = ctx->ring->ena_bufs[buf].req_id;
1013 	rx_info = &ctx->ring->rx_buffer_info[req_id];
1014 	bus_dmamap_sync(ctx->adapter->rx_buf_tag, rx_info->map,
1015 	    BUS_DMASYNC_POSTREAD);
1016 	netmap_unload_map(ctx->na, ctx->adapter->rx_buf_tag, rx_info->map);
1017 
1018 	ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0, ctx->adapter->ena_dev,
1019 	    "Rx idx is not 0.\n");
1020 
1021 	ctx->slots[ctx->nm_i].buf_idx = rx_info->netmap_buf_idx;
1022 	rx_info->netmap_buf_idx = 0;
1023 	/*
1024 	 * Set NS_MOREFRAG to all slots.
1025 	 * Then ena_netmap_rx_frame clears it from last one.
1026 	 */
1027 	ctx->slots[ctx->nm_i].flags |= NS_MOREFRAG | NS_BUF_CHANGED;
1028 	ctx->slots[ctx->nm_i].len = ctx->ring->ena_bufs[buf].len;
1029 	*len += ctx->slots[ctx->nm_i].len;
1030 	ctx->ring->free_rx_ids[ctx->nt] = req_id;
1031 	ena_log_nm(ctx->adapter->pdev, DBG,
1032 	    "rx_info %p, buf_idx %d, paddr %jx, nm: %d\n", rx_info,
1033 	    ctx->slots[ctx->nm_i].buf_idx, (uintmax_t)rx_info->ena_buf.paddr,
1034 	    ctx->nm_i);
1035 
1036 	ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
1037 	ctx->nt = ENA_RX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
1038 
1039 	return (0);
1040 }
1041 
1042 static inline void
1043 ena_netmap_rx_cleanup(struct ena_netmap_ctx *ctx)
1044 {
1045 	int refill_required;
1046 
1047 	refill_required = ctx->kring->rhead - ctx->kring->nr_hwcur;
1048 	if (ctx->kring->nr_hwcur != ctx->kring->nr_hwtail)
1049 		refill_required -= 1;
1050 
1051 	if (refill_required == 0)
1052 		return;
1053 	else if (refill_required < 0)
1054 		refill_required += ctx->kring->nkr_num_slots;
1055 
1056 	ena_refill_rx_bufs(ctx->ring, refill_required);
1057 }
1058 
1059 static inline void
1060 ena_netmap_fill_ctx(struct netmap_kring *kring, struct ena_netmap_ctx *ctx,
1061     uint16_t ena_qid)
1062 {
1063 	ctx->kring = kring;
1064 	ctx->na = kring->na;
1065 	ctx->adapter = if_getsoftc(ctx->na->ifp);
1066 	ctx->lim = kring->nkr_num_slots - 1;
1067 	ctx->io_cq = &ctx->adapter->ena_dev->io_cq_queues[ena_qid];
1068 	ctx->io_sq = &ctx->adapter->ena_dev->io_sq_queues[ena_qid];
1069 	ctx->slots = kring->ring->slot;
1070 }
1071 
1072 void
1073 ena_netmap_unload(struct ena_adapter *adapter, bus_dmamap_t map)
1074 {
1075 	struct netmap_adapter *na = NA(adapter->ifp);
1076 
1077 	netmap_unload_map(na, adapter->tx_buf_tag, map);
1078 }
1079 
1080 #endif /* DEV_NETMAP */
1081