xref: /freebsd/sys/dev/ena/ena_netmap.c (revision a7623790fb345e6dc986dfd31df0ace115e6f2e4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2015-2020 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 __FBSDID("$FreeBSD$");
32 
33 #ifdef DEV_NETMAP
34 
35 #include "ena.h"
36 #include "ena_netmap.h"
37 
38 #define ENA_NETMAP_MORE_FRAMES		1
39 #define ENA_NETMAP_NO_MORE_FRAMES	0
40 #define ENA_MAX_FRAMES			16384
41 
42 struct ena_netmap_ctx {
43 	struct netmap_kring *kring;
44 	struct ena_adapter *adapter;
45 	struct netmap_adapter *na;
46 	struct netmap_slot *slots;
47 	struct ena_ring *ring;
48 	struct ena_com_io_cq *io_cq;
49 	struct ena_com_io_sq *io_sq;
50 	u_int nm_i;
51 	uint16_t nt;
52 	uint16_t lim;
53 };
54 
55 /* Netmap callbacks */
56 static int ena_netmap_reg(struct netmap_adapter *, int);
57 static int ena_netmap_txsync(struct netmap_kring *, int);
58 static int ena_netmap_rxsync(struct netmap_kring *, int);
59 
60 /* Helper functions */
61 static int	ena_netmap_tx_frames(struct ena_netmap_ctx *);
62 static int	ena_netmap_tx_frame(struct ena_netmap_ctx *);
63 static inline uint16_t ena_netmap_count_slots(struct ena_netmap_ctx *);
64 static inline uint16_t ena_netmap_packet_len(struct netmap_slot *, u_int,
65     uint16_t);
66 static int	ena_netmap_copy_data(struct netmap_adapter *,
67     struct netmap_slot *, u_int, uint16_t, uint16_t, void *);
68 static int	ena_netmap_map_single_slot(struct netmap_adapter *,
69     struct netmap_slot *, bus_dma_tag_t, bus_dmamap_t, void **, uint64_t *);
70 static int	ena_netmap_tx_map_slots(struct ena_netmap_ctx *,
71     struct ena_tx_buffer *, void **, uint16_t *, uint16_t *);
72 static void	ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *,
73     struct ena_tx_buffer *);
74 static void	ena_netmap_tx_cleanup(struct ena_netmap_ctx *);
75 static uint16_t	ena_netmap_tx_clean_one(struct ena_netmap_ctx *,
76     uint16_t);
77 static inline int validate_tx_req_id(struct ena_ring *, uint16_t);
78 static int ena_netmap_rx_frames(struct ena_netmap_ctx *);
79 static int ena_netmap_rx_frame(struct ena_netmap_ctx *);
80 static int ena_netmap_rx_load_desc(struct ena_netmap_ctx *, uint16_t,
81     int *);
82 static void ena_netmap_rx_cleanup(struct ena_netmap_ctx *);
83 static void ena_netmap_fill_ctx(struct netmap_kring *,
84     struct ena_netmap_ctx *, uint16_t);
85 
86 int
87 ena_netmap_attach(struct ena_adapter *adapter)
88 {
89 	struct netmap_adapter na;
90 
91 	ena_trace(NULL, ENA_NETMAP, "netmap attach\n");
92 
93 	bzero(&na, sizeof(na));
94 	na.na_flags = NAF_MOREFRAG;
95 	na.ifp = adapter->ifp;
96 	na.num_tx_desc = adapter->requested_tx_ring_size;
97 	na.num_rx_desc = adapter->requested_rx_ring_size;
98 	na.num_tx_rings = adapter->num_io_queues;
99 	na.num_rx_rings = adapter->num_io_queues;
100 	na.rx_buf_maxsize = adapter->buf_ring_size;
101 	na.nm_txsync = ena_netmap_txsync;
102 	na.nm_rxsync = ena_netmap_rxsync;
103 	na.nm_register = ena_netmap_reg;
104 
105 	return (netmap_attach(&na));
106 }
107 
108 int
109 ena_netmap_alloc_rx_slot(struct ena_adapter *adapter,
110     struct ena_ring *rx_ring, struct ena_rx_buffer *rx_info)
111 {
112 	struct netmap_adapter *na = NA(adapter->ifp);
113 	struct netmap_kring *kring;
114 	struct netmap_ring *ring;
115 	struct netmap_slot *slot;
116 	void *addr;
117 	uint64_t paddr;
118 	int nm_i, qid, head, lim, rc;
119 
120 	/* if previously allocated frag is not used */
121 	if (unlikely(rx_info->netmap_buf_idx != 0))
122 		return (0);
123 
124 	qid = rx_ring->qid;
125 	kring = na->rx_rings[qid];
126 	nm_i = kring->nr_hwcur;
127 	head = kring->rhead;
128 
129 	ena_trace(NULL, ENA_NETMAP | ENA_DBG, "nr_hwcur: %d, nr_hwtail: %d, "
130 	    "rhead: %d, rcur: %d, rtail: %d\n", kring->nr_hwcur,
131 	    kring->nr_hwtail, kring->rhead, kring->rcur, kring->rtail);
132 
133 	if ((nm_i == head) && rx_ring->initialized) {
134 		ena_trace(NULL, ENA_NETMAP, "No free slots in netmap ring\n");
135 		return (ENOMEM);
136 	}
137 
138 	ring = kring->ring;
139 	if (ring == NULL) {
140 		device_printf(adapter->pdev, "Rx ring %d is NULL\n", qid);
141 		return (EFAULT);
142 	}
143 	slot = &ring->slot[nm_i];
144 
145 	addr = PNMB(na, slot, &paddr);
146 	if (addr == NETMAP_BUF_BASE(na)) {
147 		device_printf(adapter->pdev, "Bad buff in slot\n");
148 		return (EFAULT);
149 	}
150 
151 	rc = netmap_load_map(na, adapter->rx_buf_tag, rx_info->map, addr);
152 	if (rc != 0) {
153 		ena_trace(NULL, ENA_WARNING, "DMA mapping error\n");
154 		return (rc);
155 	}
156 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map, BUS_DMASYNC_PREREAD);
157 
158 	rx_info->ena_buf.paddr = paddr;
159 	rx_info->ena_buf.len = ring->nr_buf_size;
160 	rx_info->mbuf = NULL;
161 	rx_info->netmap_buf_idx = slot->buf_idx;
162 
163 	slot->buf_idx = 0;
164 
165 	lim = kring->nkr_num_slots - 1;
166 	kring->nr_hwcur = nm_next(nm_i, lim);
167 
168 	return (0);
169 }
170 
171 void
172 ena_netmap_free_rx_slot(struct ena_adapter *adapter,
173     struct ena_ring *rx_ring, struct ena_rx_buffer *rx_info)
174 {
175 	struct netmap_adapter *na;
176 	struct netmap_kring *kring;
177 	struct netmap_slot *slot;
178 	int nm_i, qid, lim;
179 
180 	na = NA(adapter->ifp);
181 	if (na == NULL) {
182 		device_printf(adapter->pdev, "netmap adapter is NULL\n");
183 		return;
184 	}
185 
186 	if (na->rx_rings == NULL) {
187 		device_printf(adapter->pdev, "netmap rings are NULL\n");
188 		return;
189 	}
190 
191 	qid = rx_ring->qid;
192 	kring = na->rx_rings[qid];
193 	if (kring == NULL) {
194 		device_printf(adapter->pdev,
195 		    "netmap kernel ring %d is NULL\n", qid);
196 		return;
197 	}
198 
199 	lim = kring->nkr_num_slots - 1;
200 	nm_i = nm_prev(kring->nr_hwcur, lim);
201 
202 	if (kring->nr_mode != NKR_NETMAP_ON)
203 		return;
204 
205 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map,
206 	    BUS_DMASYNC_POSTREAD);
207 	netmap_unload_map(na, adapter->rx_buf_tag, rx_info->map);
208 
209 	KASSERT(kring->ring == NULL, ("Netmap Rx ring is NULL\n"));
210 
211 	slot = &kring->ring->slot[nm_i];
212 
213 	ENA_WARN(slot->buf_idx != 0, NULL, "Overwrite slot buf\n");
214 	slot->buf_idx = rx_info->netmap_buf_idx;
215 	slot->flags = NS_BUF_CHANGED;
216 
217 	rx_info->netmap_buf_idx = 0;
218 	kring->nr_hwcur = nm_i;
219 }
220 
221 static bool
222 ena_ring_in_netmap(struct ena_adapter *adapter, int qid, enum txrx x)
223 {
224 	struct netmap_adapter *na;
225 	struct netmap_kring *kring;
226 
227 	if (adapter->ifp->if_capenable & IFCAP_NETMAP) {
228 		na = NA(adapter->ifp);
229 		kring = (x == NR_RX) ? na->rx_rings[qid] : na->tx_rings[qid];
230 		if (kring->nr_mode == NKR_NETMAP_ON)
231 			return true;
232 	}
233 	return false;
234 }
235 
236 bool
237 ena_tx_ring_in_netmap(struct ena_adapter *adapter, int qid)
238 {
239 	return ena_ring_in_netmap(adapter, qid, NR_TX);
240 }
241 
242 bool
243 ena_rx_ring_in_netmap(struct ena_adapter *adapter, int qid)
244 {
245 	return ena_ring_in_netmap(adapter, qid, NR_RX);
246 }
247 
248 static void
249 ena_netmap_reset_ring(struct ena_adapter *adapter, int qid, enum txrx x)
250 {
251 	if (!ena_ring_in_netmap(adapter, qid, x))
252 		return;
253 
254 	netmap_reset(NA(adapter->ifp), x, qid, 0);
255 	ena_trace(NULL, ENA_NETMAP, "%s ring %d is in netmap mode\n",
256 	    (x == NR_TX) ? "Tx" : "Rx", qid);
257 }
258 
259 void
260 ena_netmap_reset_rx_ring(struct ena_adapter *adapter, int qid)
261 {
262 	ena_netmap_reset_ring(adapter, qid, NR_RX);
263 }
264 
265 void
266 ena_netmap_reset_tx_ring(struct ena_adapter *adapter, int qid)
267 {
268 	ena_netmap_reset_ring(adapter, qid, NR_TX);
269 }
270 
271 static int
272 ena_netmap_reg(struct netmap_adapter *na, int onoff)
273 {
274 	struct ifnet *ifp = na->ifp;
275 	struct ena_adapter* adapter = ifp->if_softc;
276 	struct netmap_kring *kring;
277 	enum txrx t;
278 	int rc, i;
279 
280 	ENA_LOCK_LOCK(adapter);
281 	ENA_FLAG_CLEAR_ATOMIC(ENA_FLAG_TRIGGER_RESET, adapter);
282 	ena_down(adapter);
283 
284 	if (onoff) {
285 		ena_trace(NULL, ENA_NETMAP, "netmap on\n");
286 		for_rx_tx(t) {
287 			for (i = 0; i <= nma_get_nrings(na, t); i++) {
288 				kring = NMR(na, t)[i];
289 				if (nm_kring_pending_on(kring)) {
290 					kring->nr_mode = NKR_NETMAP_ON;
291 				}
292 			}
293 		}
294 		nm_set_native_flags(na);
295 	} else {
296 		ena_trace(NULL, ENA_NETMAP, "netmap off\n");
297 		nm_clear_native_flags(na);
298 		for_rx_tx(t) {
299 			for (i = 0; i <= nma_get_nrings(na, t); i++) {
300 				kring = NMR(na, t)[i];
301 				if (nm_kring_pending_off(kring)) {
302 					kring->nr_mode = NKR_NETMAP_OFF;
303 				}
304 			}
305 		}
306 	}
307 
308 	rc = ena_up(adapter);
309 	if (rc != 0) {
310 		ena_trace(NULL, ENA_WARNING, "ena_up failed with rc=%d\n", rc);
311 		adapter->reset_reason = ENA_REGS_RESET_DRIVER_INVALID_STATE;
312 		nm_clear_native_flags(na);
313 		ena_destroy_device(adapter, false);
314 		ENA_FLAG_SET_ATOMIC(ENA_FLAG_DEV_UP_BEFORE_RESET, adapter);
315 		rc = ena_restore_device(adapter);
316 	}
317 	ENA_LOCK_UNLOCK(adapter);
318 
319 	return (rc);
320 }
321 
322 static int
323 ena_netmap_txsync(struct netmap_kring *kring, int flags)
324 {
325 	struct ena_netmap_ctx ctx;
326 	int rc = 0;
327 
328 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_TXQ_IDX(kring->ring_id));
329 	ctx.ring = &ctx.adapter->tx_ring[kring->ring_id];
330 
331 	ENA_RING_MTX_LOCK(ctx.ring);
332 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEV_UP, ctx.adapter)))
333 		goto txsync_end;
334 
335 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
336 		goto txsync_end;
337 
338 	rc = ena_netmap_tx_frames(&ctx);
339 	ena_netmap_tx_cleanup(&ctx);
340 
341 txsync_end:
342 	ENA_RING_MTX_UNLOCK(ctx.ring);
343 	return (rc);
344 }
345 
346 static int
347 ena_netmap_tx_frames(struct ena_netmap_ctx *ctx)
348 {
349 	struct ena_ring *tx_ring = ctx->ring;
350 	int rc = 0;
351 
352 	ctx->nm_i = ctx->kring->nr_hwcur;
353 	ctx->nt = ctx->ring->next_to_use;
354 
355 	__builtin_prefetch(&ctx->slots[ctx->nm_i]);
356 
357 	while (ctx->nm_i != ctx->kring->rhead) {
358 		if ((rc = ena_netmap_tx_frame(ctx)) != 0) {
359 			/*
360 			* When there is no empty space in Tx ring, error is
361 			* still being returned. It should not be passed to the
362 			* netmap, as application knows current ring state from
363 			* netmap ring pointers. Returning error there could
364 			* cause application to exit, but the Tx ring is commonly
365 			* being full.
366 			*/
367 			if (rc == ENA_COM_NO_MEM)
368 				rc = 0;
369 			break;
370 		}
371 		tx_ring->acum_pkts++;
372 	}
373 
374 	/* If any packet was sent... */
375 	if (likely(ctx->nm_i != ctx->kring->nr_hwcur)) {
376 		/* ...send the doorbell to the device. */
377 		ena_com_write_sq_doorbell(ctx->io_sq);
378 		counter_u64_add(ctx->ring->tx_stats.doorbells, 1);
379 		tx_ring->acum_pkts = 0;
380 
381 		ctx->ring->next_to_use = ctx->nt;
382 		ctx->kring->nr_hwcur = ctx->nm_i;
383 	}
384 
385 	return (rc);
386 }
387 
388 static int
389 ena_netmap_tx_frame(struct ena_netmap_ctx *ctx)
390 {
391 	struct ena_com_tx_ctx ena_tx_ctx;
392 	struct ena_adapter *adapter;
393 	struct ena_ring *tx_ring;
394 	struct ena_tx_buffer *tx_info;
395 	uint16_t req_id;
396 	uint16_t header_len;
397 	uint16_t packet_len;
398 	int nb_hw_desc;
399 	int rc;
400 	void *push_hdr;
401 
402 	adapter = ctx->adapter;
403 	if (ena_netmap_count_slots(ctx) > adapter->max_tx_sgl_size) {
404 		ena_trace(NULL, ENA_WARNING, "Too many slots per packet\n");
405 		return (EINVAL);
406 	}
407 
408 	tx_ring = ctx->ring;
409 
410 	req_id = tx_ring->free_tx_ids[ctx->nt];
411 	tx_info = &tx_ring->tx_buffer_info[req_id];
412 	tx_info->num_of_bufs = 0;
413 	tx_info->nm_info.sockets_used = 0;
414 
415 	rc = ena_netmap_tx_map_slots(ctx, tx_info, &push_hdr, &header_len,
416 	    &packet_len);
417 	if (unlikely(rc != 0)) {
418 		device_printf(adapter->pdev, "Failed to map Tx slot\n");
419 		return (rc);
420 	}
421 
422 	bzero(&ena_tx_ctx, sizeof(struct ena_com_tx_ctx));
423 	ena_tx_ctx.ena_bufs = tx_info->bufs;
424 	ena_tx_ctx.push_header = push_hdr;
425 	ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
426 	ena_tx_ctx.req_id = req_id;
427 	ena_tx_ctx.header_len = header_len;
428 
429 	/* There are no any offloads, as the netmap doesn't support them */
430 
431 	if (tx_ring->acum_pkts == DB_THRESHOLD ||
432 	    ena_com_is_doorbell_needed(ctx->io_sq, &ena_tx_ctx)) {
433 		ena_com_write_sq_doorbell(ctx->io_sq);
434 		counter_u64_add(tx_ring->tx_stats.doorbells, 1);
435 		tx_ring->acum_pkts = 0;
436 	}
437 
438 	rc = ena_com_prepare_tx(ctx->io_sq, &ena_tx_ctx, &nb_hw_desc);
439 	if (unlikely(rc != 0)) {
440 		if (likely(rc == ENA_COM_NO_MEM)) {
441 			ena_trace(NULL, ENA_NETMAP | ENA_DBG | ENA_TXPTH,
442 			    "Tx ring[%d] is out of space\n", tx_ring->que->id);
443 		} else {
444 			device_printf(adapter->pdev,
445 			    "Failed to prepare Tx bufs\n");
446 		}
447 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
448 
449 		ena_netmap_unmap_last_socket_chain(ctx, tx_info);
450 		return (rc);
451 	}
452 
453 	counter_enter();
454 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
455 	counter_u64_add_protected(tx_ring->tx_stats.bytes, packet_len);
456 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
457 	counter_u64_add_protected(adapter->hw_stats.tx_bytes, packet_len);
458 	counter_exit();
459 
460 	tx_info->tx_descs = nb_hw_desc;
461 
462 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
463 
464 	for (unsigned int i = 0; i < tx_info->num_of_bufs; i++)
465 		bus_dmamap_sync(adapter->tx_buf_tag,
466 		   tx_info->nm_info.map_seg[i], BUS_DMASYNC_PREWRITE);
467 
468 	return (0);
469 }
470 
471 static inline uint16_t
472 ena_netmap_count_slots(struct ena_netmap_ctx *ctx)
473 {
474 	uint16_t slots = 1;
475 	uint16_t nm = ctx->nm_i;
476 
477 	while ((ctx->slots[nm].flags & NS_MOREFRAG) != 0) {
478 		slots++;
479 		nm = nm_next(nm, ctx->lim);
480 	}
481 
482 	return slots;
483 }
484 
485 static inline uint16_t
486 ena_netmap_packet_len(struct netmap_slot *slots, u_int slot_index,
487     uint16_t limit)
488 {
489 	struct netmap_slot *nm_slot;
490 	uint16_t packet_size = 0;
491 
492 	do {
493 		nm_slot = &slots[slot_index];
494 		packet_size += nm_slot->len;
495 		slot_index = nm_next(slot_index, limit);
496 	} while ((nm_slot->flags & NS_MOREFRAG) != 0);
497 
498 	return packet_size;
499 }
500 
501 static int
502 ena_netmap_copy_data(struct netmap_adapter *na, struct netmap_slot *slots,
503     u_int slot_index, uint16_t limit, uint16_t bytes_to_copy, void *destination)
504 {
505 	struct netmap_slot *nm_slot;
506 	void *slot_vaddr;
507 	uint16_t packet_size;
508 	uint16_t data_amount;
509 
510 	packet_size = 0;
511 	do {
512 		nm_slot = &slots[slot_index];
513 		slot_vaddr = NMB(na, nm_slot);
514 		if (unlikely(slot_vaddr == NULL))
515 			return (EINVAL);
516 
517 		data_amount = min_t(uint16_t, bytes_to_copy, nm_slot->len);
518 		memcpy(destination, slot_vaddr, data_amount);
519 		bytes_to_copy -= data_amount;
520 
521 		slot_index = nm_next(slot_index, limit);
522 	} while ((nm_slot->flags & NS_MOREFRAG) != 0 && bytes_to_copy > 0);
523 
524 	return (0);
525 }
526 
527 static int
528 ena_netmap_map_single_slot(struct netmap_adapter *na, struct netmap_slot *slot,
529     bus_dma_tag_t dmatag, bus_dmamap_t dmamap, void **vaddr, uint64_t *paddr)
530 {
531 	int rc;
532 
533 	*vaddr = PNMB(na, slot, paddr);
534 	if (unlikely(vaddr == NULL)) {
535 		ena_trace(NULL, ENA_ALERT, "Slot address is NULL\n");
536 		return (EINVAL);
537 	}
538 
539 	rc = netmap_load_map(na, dmatag, dmamap, *vaddr);
540 	if (unlikely(rc != 0)) {
541 		ena_trace(NULL, ENA_ALERT, "Failed to map slot %d for DMA\n",
542 		    slot->buf_idx);
543 		return (EINVAL);
544 	}
545 
546 	return (0);
547 }
548 
549 static int
550 ena_netmap_tx_map_slots(struct ena_netmap_ctx *ctx,
551     struct ena_tx_buffer *tx_info, void **push_hdr, uint16_t *header_len,
552     uint16_t *packet_len)
553 {
554 	struct netmap_slot *slot;
555 	struct ena_com_buf *ena_buf;
556 	struct ena_adapter *adapter;
557 	struct ena_ring *tx_ring;
558 	struct ena_netmap_tx_info *nm_info;
559 	bus_dmamap_t *nm_maps;
560 	void *vaddr;
561 	uint64_t paddr;
562 	uint32_t *nm_buf_idx;
563 	uint32_t slot_head_len;
564 	uint32_t frag_len;
565 	uint32_t remaining_len;
566 	uint16_t push_len;
567 	uint16_t delta;
568 	int rc;
569 
570 	adapter = ctx->adapter;
571 	tx_ring = ctx->ring;
572 	ena_buf = tx_info->bufs;
573 	nm_info = &tx_info->nm_info;
574 	nm_maps = nm_info->map_seg;
575 	nm_buf_idx = nm_info->socket_buf_idx;
576 	slot = &ctx->slots[ctx->nm_i];
577 
578 	slot_head_len = slot->len;
579 	*packet_len = ena_netmap_packet_len(ctx->slots, ctx->nm_i, ctx->lim);
580 	remaining_len = *packet_len;
581 	delta = 0;
582 
583 	__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
584 	if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
585 		/*
586 		 * When the device is in LLQ mode, the driver will copy
587 		 * the header into the device memory space.
588 		 * The ena_com layer assumes that the header is in a linear
589 		 * memory space.
590 		 * This assumption might be wrong since part of the header
591 		 * can be in the fragmented buffers.
592 		 * First, check if header fits in the first slot. If not, copy
593 		 * it to separate buffer that will be holding linearized data.
594 		 */
595 		push_len = min_t(uint32_t, *packet_len,
596 		    tx_ring->tx_max_header_size);
597 		*header_len = push_len;
598 		/* If header is in linear space, just point to socket's data. */
599 		if (likely(push_len <= slot_head_len)) {
600 			*push_hdr = NMB(ctx->na, slot);
601 			if (unlikely(push_hdr == NULL)) {
602 				device_printf(adapter->pdev,
603 				    "Slot vaddress is NULL\n");
604 				return (EINVAL);
605 			}
606 		/*
607 		 * Otherwise, copy whole portion of header from multiple slots
608 		 * to intermediate buffer.
609 		 */
610 		} else {
611 			rc = ena_netmap_copy_data(ctx->na,
612 			    ctx->slots,
613 			    ctx->nm_i,
614 			    ctx->lim,
615 			    push_len,
616 			    tx_ring->push_buf_intermediate_buf);
617 			if (unlikely(rc)) {
618 				device_printf(adapter->pdev,
619 				    "Failed to copy data from slots to push_buf\n");
620 				return (EINVAL);
621 			}
622 
623 			*push_hdr = tx_ring->push_buf_intermediate_buf;
624 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
625 
626 			delta = push_len - slot_head_len;
627 		}
628 
629 		ena_trace(NULL, ENA_NETMAP | ENA_DBG | ENA_TXPTH,
630 		    "slot: %d header_buf->vaddr: %p push_len: %d\n",
631 		    slot->buf_idx, *push_hdr, push_len);
632 
633 		/*
634 		* If header was in linear memory space, map for the dma rest of the data
635 		* in the first mbuf of the mbuf chain.
636 		*/
637 		if (slot_head_len > push_len) {
638 			rc = ena_netmap_map_single_slot(ctx->na,
639 			    slot,
640 			    adapter->tx_buf_tag,
641 			    *nm_maps,
642 			    &vaddr,
643 			    &paddr);
644 			if (unlikely(rc != 0)) {
645 				device_printf(adapter->pdev,
646 				    "DMA mapping error\n");
647 				return (rc);
648 			}
649 			nm_maps++;
650 
651 			ena_buf->paddr = paddr + push_len;
652 			ena_buf->len = slot->len - push_len;
653 			ena_buf++;
654 
655 			tx_info->num_of_bufs++;
656 		}
657 
658 		remaining_len -= slot->len;
659 
660 		/* Save buf idx before advancing */
661 		*nm_buf_idx = slot->buf_idx;
662 		nm_buf_idx++;
663 		slot->buf_idx = 0;
664 
665 		/* Advance to the next socket */
666 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
667 		slot = &ctx->slots[ctx->nm_i];
668 		nm_info->sockets_used++;
669 
670 		/*
671 		 * If header is in non linear space (delta > 0), then skip mbufs
672 		 * containing header and map the last one containing both header
673 		 * and the packet data.
674 		 * The first segment is already counted in.
675 		 */
676 		while (delta > 0) {
677 			__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
678 			frag_len = slot->len;
679 
680 			/*
681 			 * If whole segment contains header just move to the
682 			 * next one and reduce delta.
683 			 */
684 			if (unlikely(delta >= frag_len)) {
685 				delta -= frag_len;
686 			} else {
687 				/*
688 				 * Map the data and then assign it with the
689 				 * offsets
690 				 */
691 				rc = ena_netmap_map_single_slot(ctx->na,
692 				    slot,
693 				    adapter->tx_buf_tag,
694 				    *nm_maps,
695 				    &vaddr,
696 				    &paddr);
697 				if (unlikely(rc != 0)) {
698 					device_printf(adapter->pdev,
699 					    "DMA mapping error\n");
700 					goto error_map;
701 				}
702 				nm_maps++;
703 
704 				ena_buf->paddr = paddr + delta;
705 				ena_buf->len = slot->len - delta;
706 				ena_buf++;
707 
708 				tx_info->num_of_bufs++;
709 				delta = 0;
710 			}
711 
712 			remaining_len -= slot->len;
713 
714 			/* Save buf idx before advancing */
715 			*nm_buf_idx = slot->buf_idx;
716 			nm_buf_idx++;
717 			slot->buf_idx = 0;
718 
719 			/* Advance to the next socket */
720 			ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
721 			slot = &ctx->slots[ctx->nm_i];
722 			nm_info->sockets_used++;
723 		}
724 	} else {
725 		*push_hdr = NULL;
726 		/*
727 		* header_len is just a hint for the device. Because netmap is
728 		* not giving us any information about packet header length and
729 		* it is not guaranteed that all packet headers will be in the
730 		* 1st slot, setting header_len to 0 is making the device ignore
731 		* this value and resolve header on it's own.
732 		*/
733 		*header_len = 0;
734 	}
735 
736 	/* Map all remaining data (regular routine for non-LLQ mode) */
737 	while (remaining_len > 0) {
738 		__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
739 
740 		rc = ena_netmap_map_single_slot(ctx->na,
741 			    slot,
742 			    adapter->tx_buf_tag,
743 			    *nm_maps,
744 			    &vaddr,
745 			    &paddr);
746 		if (unlikely(rc != 0)) {
747 			device_printf(adapter->pdev,
748 			    "DMA mapping error\n");
749 			goto error_map;
750 		}
751 		nm_maps++;
752 
753 		ena_buf->paddr = paddr;
754 		ena_buf->len = slot->len;
755 		ena_buf++;
756 
757 		tx_info->num_of_bufs++;
758 
759 		remaining_len -= slot->len;
760 
761 		/* Save buf idx before advancing */
762 		*nm_buf_idx = slot->buf_idx;
763 		nm_buf_idx++;
764 		slot->buf_idx = 0;
765 
766 		/* Advance to the next socket */
767 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
768 		slot = &ctx->slots[ctx->nm_i];
769 		nm_info->sockets_used++;
770 	}
771 
772 	return (0);
773 
774 error_map:
775 	ena_netmap_unmap_last_socket_chain(ctx, tx_info);
776 
777 	return (rc);
778 }
779 
780 static void
781 ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *ctx,
782     struct ena_tx_buffer *tx_info)
783 {
784 	struct ena_netmap_tx_info *nm_info;
785 	int n;
786 
787 	nm_info = &tx_info->nm_info;
788 
789 	/**
790 	 * As the used sockets must not be equal to the buffers used in the LLQ
791 	 * mode, they must be treated separately.
792 	 * First, unmap the DMA maps.
793 	 */
794 	n = tx_info->num_of_bufs;
795 	while (n--) {
796 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
797 		    nm_info->map_seg[n]);
798 	}
799 	tx_info->num_of_bufs = 0;
800 
801 	/* Next, retain the sockets back to the userspace */
802 	n = nm_info->sockets_used;
803 	while (n--) {
804 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
805 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
806 		nm_info->socket_buf_idx[n] = 0;
807 		ctx->nm_i = nm_prev(ctx->nm_i, ctx->lim);
808 	}
809 	nm_info->sockets_used = 0;
810 }
811 
812 static void
813 ena_netmap_tx_cleanup(struct ena_netmap_ctx *ctx)
814 {
815 	uint16_t req_id;
816 	uint16_t total_tx_descs = 0;
817 
818 	ctx->nm_i = ctx->kring->nr_hwtail;
819 	ctx->nt = ctx->ring->next_to_clean;
820 
821 	/* Reclaim buffers for completed transmissions */
822 	while (ena_com_tx_comp_req_id_get(ctx->io_cq, &req_id) >= 0) {
823 		if (validate_tx_req_id(ctx->ring, req_id) != 0)
824 			break;
825 		total_tx_descs += ena_netmap_tx_clean_one(ctx, req_id);
826 	}
827 
828 	ctx->kring->nr_hwtail = ctx->nm_i;
829 
830 	if (total_tx_descs > 0) {
831 		/* acknowledge completion of sent packets */
832 		ctx->ring->next_to_clean = ctx->nt;
833 		ena_com_comp_ack(ctx->ring->ena_com_io_sq, total_tx_descs);
834 		ena_com_update_dev_comp_head(ctx->ring->ena_com_io_cq);
835 	}
836 }
837 
838 static uint16_t
839 ena_netmap_tx_clean_one(struct ena_netmap_ctx *ctx, uint16_t req_id)
840 {
841 	struct ena_tx_buffer *tx_info;
842 	struct ena_netmap_tx_info *nm_info;
843 	int n;
844 
845 	tx_info = &ctx->ring->tx_buffer_info[req_id];
846 	nm_info = &tx_info->nm_info;
847 
848 	/**
849 	 * As the used sockets must not be equal to the buffers used in the LLQ
850 	 * mode, they must be treated separately.
851 	 * First, unmap the DMA maps.
852 	 */
853 	n = tx_info->num_of_bufs;
854 	for (n = 0; n < tx_info->num_of_bufs; n++) {
855 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
856 		    nm_info->map_seg[n]);
857 	}
858 	tx_info->num_of_bufs = 0;
859 
860 	/* Next, retain the sockets back to the userspace */
861 	for (n = 0; n < nm_info->sockets_used; n++) {
862 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
863 		ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0, NULL,
864 		    "Tx idx is not 0.\n");
865 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
866 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
867 		nm_info->socket_buf_idx[n] = 0;
868 	}
869 	nm_info->sockets_used = 0;
870 
871 	ctx->ring->free_tx_ids[ctx->nt] = req_id;
872 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->lim);
873 
874 	return tx_info->tx_descs;
875 }
876 
877 static inline int
878 validate_tx_req_id(struct ena_ring *tx_ring, uint16_t req_id)
879 {
880 	struct ena_adapter *adapter = tx_ring->adapter;
881 
882 	if (likely(req_id < tx_ring->ring_size))
883 		return (0);
884 
885 	ena_trace(NULL, ENA_WARNING, "Invalid req_id: %hu\n", req_id);
886 	counter_u64_add(tx_ring->tx_stats.bad_req_id, 1);
887 
888 	ena_trigger_reset(adapter, ENA_REGS_RESET_INV_TX_REQ_ID);
889 
890 	return (EFAULT);
891 }
892 
893 static int
894 ena_netmap_rxsync(struct netmap_kring *kring, int flags)
895 {
896 	struct ena_netmap_ctx ctx;
897 	int rc;
898 
899 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_RXQ_IDX(kring->ring_id));
900 	ctx.ring = &ctx.adapter->rx_ring[kring->ring_id];
901 
902 	if (ctx.kring->rhead > ctx.lim) {
903 		/* Probably not needed to release slots from RX ring. */
904 		return (netmap_ring_reinit(ctx.kring));
905 	}
906 
907 	if (unlikely((if_getdrvflags(ctx.na->ifp) & IFF_DRV_RUNNING) == 0))
908 		return (0);
909 
910 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
911 		return (0);
912 
913 	if ((rc = ena_netmap_rx_frames(&ctx)) != 0)
914 		return (rc);
915 
916 	ena_netmap_rx_cleanup(&ctx);
917 
918 	return (0);
919 }
920 
921 static inline int
922 ena_netmap_rx_frames(struct ena_netmap_ctx *ctx)
923 {
924 	int rc = 0;
925 	int frames_counter = 0;
926 
927 	ctx->nt = ctx->ring->next_to_clean;
928 	ctx->nm_i = ctx->kring->nr_hwtail;
929 
930 	while((rc = ena_netmap_rx_frame(ctx)) == ENA_NETMAP_MORE_FRAMES) {
931 		frames_counter++;
932 		/* In case of multiple frames, it is not an error. */
933 		rc = 0;
934 		if (frames_counter > ENA_MAX_FRAMES) {
935 			device_printf(ctx->adapter->pdev,
936 				"Driver is stuck in the Rx loop\n");
937 			break;
938 		}
939 	};
940 
941 	ctx->kring->nr_hwtail = ctx->nm_i;
942 	ctx->kring->nr_kflags &= ~NKR_PENDINTR;
943 	ctx->ring->next_to_clean = ctx->nt;
944 
945 	return (rc);
946 }
947 
948 static inline int
949 ena_netmap_rx_frame(struct ena_netmap_ctx *ctx)
950 {
951 	struct ena_com_rx_ctx ena_rx_ctx;
952 	enum ena_regs_reset_reason_types reset_reason;
953 	int rc, len = 0;
954 	uint16_t buf, nm;
955 
956 	ena_rx_ctx.ena_bufs = ctx->ring->ena_bufs;
957 	ena_rx_ctx.max_bufs = ctx->adapter->max_rx_sgl_size;
958 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
959 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
960 
961 	rc = ena_com_rx_pkt(ctx->io_cq, ctx->io_sq, &ena_rx_ctx);
962 	if (unlikely(rc != 0)) {
963 		ena_trace(NULL, ENA_ALERT,
964 		    "Failed to read pkt from the device with error: %d\n", rc);
965 		if (rc == ENA_COM_NO_SPACE) {
966 			counter_u64_add(ctx->ring->rx_stats.bad_desc_num, 1);
967 			reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
968 		} else {
969 			counter_u64_add(ctx->ring->rx_stats.bad_req_id, 1);
970 			reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID;
971 		}
972 		ena_trigger_reset(ctx->adapter, reset_reason);
973 		return (rc);
974 	}
975 	if (unlikely(ena_rx_ctx.descs == 0))
976 		return (ENA_NETMAP_NO_MORE_FRAMES);
977 
978         ena_trace(NULL, ENA_NETMAP | ENA_DBG, "Rx: q %d got packet from ena. descs #:"
979 	    " %d l3 proto %d l4 proto %d hash: %x\n", ctx->ring->qid,
980 	    ena_rx_ctx.descs, ena_rx_ctx.l3_proto, ena_rx_ctx.l4_proto,
981 	    ena_rx_ctx.hash);
982 
983 	for (buf = 0; buf < ena_rx_ctx.descs; buf++)
984 		if ((rc = ena_netmap_rx_load_desc(ctx, buf, &len)) != 0)
985 			break;
986 	/*
987 	 * ena_netmap_rx_load_desc doesn't know the number of descriptors.
988 	 * It just set flag NS_MOREFRAG to all slots, then here flag of
989 	 * last slot is cleared.
990 	 */
991 	ctx->slots[nm_prev(ctx->nm_i, ctx->lim)].flags = NS_BUF_CHANGED;
992 
993 	if (rc != 0) {
994 		goto rx_clear_desc;
995 	}
996 
997 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
998             ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
999 
1000 	counter_enter();
1001 	counter_u64_add_protected(ctx->ring->rx_stats.bytes, len);
1002 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_bytes, len);
1003 	counter_u64_add_protected(ctx->ring->rx_stats.cnt, 1);
1004 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_packets, 1);
1005 	counter_exit();
1006 
1007 	return (ENA_NETMAP_MORE_FRAMES);
1008 
1009 rx_clear_desc:
1010 	nm = ctx->nm_i;
1011 
1012 	/* Remove failed packet from ring */
1013 	while(buf--) {
1014 		ctx->slots[nm].flags = 0;
1015 		ctx->slots[nm].len = 0;
1016 		nm = nm_prev(nm, ctx->lim);
1017 	}
1018 
1019 	return (rc);
1020 }
1021 
1022 static inline int
1023 ena_netmap_rx_load_desc(struct ena_netmap_ctx *ctx, uint16_t buf, int *len)
1024 {
1025 	struct ena_rx_buffer *rx_info;
1026 	uint16_t req_id;
1027 
1028 	req_id = ctx->ring->ena_bufs[buf].req_id;
1029 	rx_info = &ctx->ring->rx_buffer_info[req_id];
1030 	bus_dmamap_sync(ctx->adapter->rx_buf_tag, rx_info->map,
1031 	    BUS_DMASYNC_POSTREAD);
1032 	netmap_unload_map(ctx->na, ctx->adapter->rx_buf_tag, rx_info->map);
1033 
1034 	ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0, NULL,
1035 	    "Rx idx is not 0.\n");
1036 
1037 	ctx->slots[ctx->nm_i].buf_idx = rx_info->netmap_buf_idx;
1038 	rx_info->netmap_buf_idx = 0;
1039 	/*
1040 	 * Set NS_MOREFRAG to all slots.
1041 	 * Then ena_netmap_rx_frame clears it from last one.
1042 	 */
1043 	ctx->slots[ctx->nm_i].flags |= NS_MOREFRAG | NS_BUF_CHANGED;
1044 	ctx->slots[ctx->nm_i].len = ctx->ring->ena_bufs[buf].len;
1045 	*len += ctx->slots[ctx->nm_i].len;
1046 	ctx->ring->free_rx_ids[ctx->nt] = req_id;
1047 	ena_trace(NULL, ENA_DBG, "rx_info %p, buf_idx %d, paddr %jx, nm: %d\n",
1048 	    rx_info, ctx->slots[ctx->nm_i].buf_idx,
1049 	    (uintmax_t)rx_info->ena_buf.paddr, ctx->nm_i);
1050 
1051 	ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
1052 	ctx->nt = ENA_RX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
1053 
1054 	return (0);
1055 }
1056 
1057 static inline void
1058 ena_netmap_rx_cleanup(struct ena_netmap_ctx *ctx)
1059 {
1060 	int refill_required;
1061 
1062 	refill_required = ctx->kring->rhead - ctx->kring->nr_hwcur;
1063 	if (ctx->kring->nr_hwcur != ctx->kring->nr_hwtail)
1064 		refill_required -= 1;
1065 
1066 	if (refill_required == 0)
1067 		return;
1068 	else if (refill_required < 0)
1069 		refill_required += ctx->kring->nkr_num_slots;
1070 
1071 	ena_refill_rx_bufs(ctx->ring, refill_required);
1072 }
1073 
1074 static inline void
1075 ena_netmap_fill_ctx(struct netmap_kring *kring, struct ena_netmap_ctx *ctx,
1076     uint16_t ena_qid)
1077 {
1078 	ctx->kring = kring;
1079 	ctx->na = kring->na;
1080 	ctx->adapter = ctx->na->ifp->if_softc;
1081 	ctx->lim = kring->nkr_num_slots - 1;
1082 	ctx->io_cq = &ctx->adapter->ena_dev->io_cq_queues[ena_qid];
1083 	ctx->io_sq = &ctx->adapter->ena_dev->io_sq_queues[ena_qid];
1084 	ctx->slots = kring->ring->slot;
1085 }
1086 
1087 void
1088 ena_netmap_unload(struct ena_adapter *adapter, bus_dmamap_t map)
1089 {
1090 	struct netmap_adapter *na = NA(adapter->ifp);
1091 
1092 	netmap_unload_map(na, adapter->tx_buf_tag, map);
1093 }
1094 
1095 #endif /* DEV_NETMAP */
1096