xref: /freebsd/sys/dev/ena/ena_netmap.c (revision 924226fba12cc9a228c73b956e1b7fa24c60b055)
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_log_nm(adapter->pdev, INFO, "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_log_nm(adapter->pdev, 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_log_nm(adapter->pdev, ERR, "No free slots in netmap ring\n");
135 		return (ENOMEM);
136 	}
137 
138 	ring = kring->ring;
139 	if (ring == NULL) {
140 		ena_log_nm(adapter->pdev, ERR, "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 		ena_log_nm(adapter->pdev, ERR, "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_log_nm(adapter->pdev, WARN, "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 		ena_log_nm(adapter->pdev, ERR, "netmap adapter is NULL\n");
183 		return;
184 	}
185 
186 	if (na->rx_rings == NULL) {
187 		ena_log_nm(adapter->pdev, ERR, "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 		ena_log_nm(adapter->pdev, ERR,
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, adapter->ena_dev, "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_log_nm(adapter->pdev, INFO, "%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 	device_t pdev = adapter->pdev;
277 	struct netmap_kring *kring;
278 	enum txrx t;
279 	int rc, i;
280 
281 	ENA_LOCK_LOCK();
282 	ENA_FLAG_CLEAR_ATOMIC(ENA_FLAG_TRIGGER_RESET, adapter);
283 	ena_down(adapter);
284 
285 	if (onoff) {
286 		ena_log_nm(pdev, INFO, "netmap on\n");
287 		for_rx_tx(t) {
288 			for (i = 0; i <= nma_get_nrings(na, t); i++) {
289 				kring = NMR(na, t)[i];
290 				if (nm_kring_pending_on(kring)) {
291 					kring->nr_mode = NKR_NETMAP_ON;
292 				}
293 			}
294 		}
295 		nm_set_native_flags(na);
296 	} else {
297 		ena_log_nm(pdev, INFO, "netmap off\n");
298 		nm_clear_native_flags(na);
299 		for_rx_tx(t) {
300 			for (i = 0; i <= nma_get_nrings(na, t); i++) {
301 				kring = NMR(na, t)[i];
302 				if (nm_kring_pending_off(kring)) {
303 					kring->nr_mode = NKR_NETMAP_OFF;
304 				}
305 			}
306 		}
307 	}
308 
309 	rc = ena_up(adapter);
310 	if (rc != 0) {
311 		ena_log_nm(pdev, WARN, "ena_up failed with rc=%d\n", rc);
312 		adapter->reset_reason = ENA_REGS_RESET_DRIVER_INVALID_STATE;
313 		nm_clear_native_flags(na);
314 		ena_destroy_device(adapter, false);
315 		ENA_FLAG_SET_ATOMIC(ENA_FLAG_DEV_UP_BEFORE_RESET, adapter);
316 		rc = ena_restore_device(adapter);
317 	}
318 	ENA_LOCK_UNLOCK();
319 
320 	return (rc);
321 }
322 
323 static int
324 ena_netmap_txsync(struct netmap_kring *kring, int flags)
325 {
326 	struct ena_netmap_ctx ctx;
327 	int rc = 0;
328 
329 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_TXQ_IDX(kring->ring_id));
330 	ctx.ring = &ctx.adapter->tx_ring[kring->ring_id];
331 
332 	ENA_RING_MTX_LOCK(ctx.ring);
333 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_DEV_UP, ctx.adapter)))
334 		goto txsync_end;
335 
336 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
337 		goto txsync_end;
338 
339 	rc = ena_netmap_tx_frames(&ctx);
340 	ena_netmap_tx_cleanup(&ctx);
341 
342 txsync_end:
343 	ENA_RING_MTX_UNLOCK(ctx.ring);
344 	return (rc);
345 }
346 
347 static int
348 ena_netmap_tx_frames(struct ena_netmap_ctx *ctx)
349 {
350 	struct ena_ring *tx_ring = ctx->ring;
351 	int rc = 0;
352 
353 	ctx->nm_i = ctx->kring->nr_hwcur;
354 	ctx->nt = ctx->ring->next_to_use;
355 
356 	__builtin_prefetch(&ctx->slots[ctx->nm_i]);
357 
358 	while (ctx->nm_i != ctx->kring->rhead) {
359 		if ((rc = ena_netmap_tx_frame(ctx)) != 0) {
360 			/*
361 			* When there is no empty space in Tx ring, error is
362 			* still being returned. It should not be passed to the
363 			* netmap, as application knows current ring state from
364 			* netmap ring pointers. Returning error there could
365 			* cause application to exit, but the Tx ring is commonly
366 			* being full.
367 			*/
368 			if (rc == ENA_COM_NO_MEM)
369 				rc = 0;
370 			break;
371 		}
372 		tx_ring->acum_pkts++;
373 	}
374 
375 	/* If any packet was sent... */
376 	if (likely(ctx->nm_i != ctx->kring->nr_hwcur)) {
377 		/* ...send the doorbell to the device. */
378 		ena_com_write_sq_doorbell(ctx->io_sq);
379 		counter_u64_add(ctx->ring->tx_stats.doorbells, 1);
380 		tx_ring->acum_pkts = 0;
381 
382 		ctx->ring->next_to_use = ctx->nt;
383 		ctx->kring->nr_hwcur = ctx->nm_i;
384 	}
385 
386 	return (rc);
387 }
388 
389 static int
390 ena_netmap_tx_frame(struct ena_netmap_ctx *ctx)
391 {
392 	struct ena_com_tx_ctx ena_tx_ctx;
393 	struct ena_adapter *adapter;
394 	struct ena_ring *tx_ring;
395 	struct ena_tx_buffer *tx_info;
396 	uint16_t req_id;
397 	uint16_t header_len;
398 	uint16_t packet_len;
399 	int nb_hw_desc;
400 	int rc;
401 	void *push_hdr;
402 
403 	adapter = ctx->adapter;
404 	if (ena_netmap_count_slots(ctx) > adapter->max_tx_sgl_size) {
405 		ena_log_nm(adapter->pdev, WARN, "Too many slots per packet\n");
406 		return (EINVAL);
407 	}
408 
409 	tx_ring = ctx->ring;
410 
411 	req_id = tx_ring->free_tx_ids[ctx->nt];
412 	tx_info = &tx_ring->tx_buffer_info[req_id];
413 	tx_info->num_of_bufs = 0;
414 	tx_info->nm_info.sockets_used = 0;
415 
416 	rc = ena_netmap_tx_map_slots(ctx, tx_info, &push_hdr, &header_len,
417 	    &packet_len);
418 	if (unlikely(rc != 0)) {
419 		ena_log_nm(adapter->pdev, ERR, "Failed to map Tx slot\n");
420 		return (rc);
421 	}
422 
423 	bzero(&ena_tx_ctx, sizeof(struct ena_com_tx_ctx));
424 	ena_tx_ctx.ena_bufs = tx_info->bufs;
425 	ena_tx_ctx.push_header = push_hdr;
426 	ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
427 	ena_tx_ctx.req_id = req_id;
428 	ena_tx_ctx.header_len = header_len;
429 	ena_tx_ctx.meta_valid = adapter->disable_meta_caching;
430 
431 	/* There are no any offloads, as the netmap doesn't support them */
432 
433 	if (tx_ring->acum_pkts == DB_THRESHOLD ||
434 	    ena_com_is_doorbell_needed(ctx->io_sq, &ena_tx_ctx)) {
435 		ena_com_write_sq_doorbell(ctx->io_sq);
436 		counter_u64_add(tx_ring->tx_stats.doorbells, 1);
437 		tx_ring->acum_pkts = 0;
438 	}
439 
440 	rc = ena_com_prepare_tx(ctx->io_sq, &ena_tx_ctx, &nb_hw_desc);
441 	if (unlikely(rc != 0)) {
442 		if (likely(rc == ENA_COM_NO_MEM)) {
443 			ena_log_nm(adapter->pdev, DBG,
444 			    "Tx ring[%d] is out of space\n", tx_ring->que->id);
445 		} else {
446 			ena_log_nm(adapter->pdev, ERR,
447 			    "Failed to prepare Tx bufs\n");
448 			ena_trigger_reset(adapter,
449 			    ENA_REGS_RESET_DRIVER_INVALID_STATE);
450 		}
451 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
452 
453 		ena_netmap_unmap_last_socket_chain(ctx, tx_info);
454 		return (rc);
455 	}
456 
457 	counter_enter();
458 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
459 	counter_u64_add_protected(tx_ring->tx_stats.bytes, packet_len);
460 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
461 	counter_u64_add_protected(adapter->hw_stats.tx_bytes, packet_len);
462 	counter_exit();
463 
464 	tx_info->tx_descs = nb_hw_desc;
465 
466 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
467 
468 	for (unsigned int i = 0; i < tx_info->num_of_bufs; i++)
469 		bus_dmamap_sync(adapter->tx_buf_tag,
470 		   tx_info->nm_info.map_seg[i], BUS_DMASYNC_PREWRITE);
471 
472 	return (0);
473 }
474 
475 static inline uint16_t
476 ena_netmap_count_slots(struct ena_netmap_ctx *ctx)
477 {
478 	uint16_t slots = 1;
479 	uint16_t nm = ctx->nm_i;
480 
481 	while ((ctx->slots[nm].flags & NS_MOREFRAG) != 0) {
482 		slots++;
483 		nm = nm_next(nm, ctx->lim);
484 	}
485 
486 	return slots;
487 }
488 
489 static inline uint16_t
490 ena_netmap_packet_len(struct netmap_slot *slots, u_int slot_index,
491     uint16_t limit)
492 {
493 	struct netmap_slot *nm_slot;
494 	uint16_t packet_size = 0;
495 
496 	do {
497 		nm_slot = &slots[slot_index];
498 		packet_size += nm_slot->len;
499 		slot_index = nm_next(slot_index, limit);
500 	} while ((nm_slot->flags & NS_MOREFRAG) != 0);
501 
502 	return packet_size;
503 }
504 
505 static int
506 ena_netmap_copy_data(struct netmap_adapter *na, struct netmap_slot *slots,
507     u_int slot_index, uint16_t limit, uint16_t bytes_to_copy, void *destination)
508 {
509 	struct netmap_slot *nm_slot;
510 	void *slot_vaddr;
511 	uint16_t data_amount;
512 
513 	do {
514 		nm_slot = &slots[slot_index];
515 		slot_vaddr = NMB(na, nm_slot);
516 		if (unlikely(slot_vaddr == NULL))
517 			return (EINVAL);
518 
519 		data_amount = min_t(uint16_t, bytes_to_copy, nm_slot->len);
520 		memcpy(destination, slot_vaddr, data_amount);
521 		bytes_to_copy -= data_amount;
522 
523 		slot_index = nm_next(slot_index, limit);
524 	} while ((nm_slot->flags & NS_MOREFRAG) != 0 && bytes_to_copy > 0);
525 
526 	return (0);
527 }
528 
529 static int
530 ena_netmap_map_single_slot(struct netmap_adapter *na, struct netmap_slot *slot,
531     bus_dma_tag_t dmatag, bus_dmamap_t dmamap, void **vaddr, uint64_t *paddr)
532 {
533 	device_t pdev;
534 	int rc;
535 
536 	pdev = ((struct ena_adapter *)na->ifp->if_softc)->pdev;
537 
538 	*vaddr = PNMB(na, slot, paddr);
539 	if (unlikely(vaddr == NULL)) {
540 		ena_log_nm(pdev, ERR, "Slot address is NULL\n");
541 		return (EINVAL);
542 	}
543 
544 	rc = netmap_load_map(na, dmatag, dmamap, *vaddr);
545 	if (unlikely(rc != 0)) {
546 		ena_log_nm(pdev, ERR, "Failed to map slot %d for DMA\n",
547 		    slot->buf_idx);
548 		return (EINVAL);
549 	}
550 
551 	return (0);
552 }
553 
554 static int
555 ena_netmap_tx_map_slots(struct ena_netmap_ctx *ctx,
556     struct ena_tx_buffer *tx_info, void **push_hdr, uint16_t *header_len,
557     uint16_t *packet_len)
558 {
559 	struct netmap_slot *slot;
560 	struct ena_com_buf *ena_buf;
561 	struct ena_adapter *adapter;
562 	struct ena_ring *tx_ring;
563 	struct ena_netmap_tx_info *nm_info;
564 	bus_dmamap_t *nm_maps;
565 	void *vaddr;
566 	uint64_t paddr;
567 	uint32_t *nm_buf_idx;
568 	uint32_t slot_head_len;
569 	uint32_t frag_len;
570 	uint32_t remaining_len;
571 	uint16_t push_len;
572 	uint16_t delta;
573 	int rc;
574 
575 	adapter = ctx->adapter;
576 	tx_ring = ctx->ring;
577 	ena_buf = tx_info->bufs;
578 	nm_info = &tx_info->nm_info;
579 	nm_maps = nm_info->map_seg;
580 	nm_buf_idx = nm_info->socket_buf_idx;
581 	slot = &ctx->slots[ctx->nm_i];
582 
583 	slot_head_len = slot->len;
584 	*packet_len = ena_netmap_packet_len(ctx->slots, ctx->nm_i, ctx->lim);
585 	remaining_len = *packet_len;
586 	delta = 0;
587 
588 	__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
589 	if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
590 		/*
591 		 * When the device is in LLQ mode, the driver will copy
592 		 * the header into the device memory space.
593 		 * The ena_com layer assumes that the header is in a linear
594 		 * memory space.
595 		 * This assumption might be wrong since part of the header
596 		 * can be in the fragmented buffers.
597 		 * First, check if header fits in the first slot. If not, copy
598 		 * it to separate buffer that will be holding linearized data.
599 		 */
600 		push_len = min_t(uint32_t, *packet_len,
601 		    tx_ring->tx_max_header_size);
602 		*header_len = push_len;
603 		/* If header is in linear space, just point to socket's data. */
604 		if (likely(push_len <= slot_head_len)) {
605 			*push_hdr = NMB(ctx->na, slot);
606 			if (unlikely(push_hdr == NULL)) {
607 				ena_log_nm(adapter->pdev, ERR,
608 				    "Slot vaddress is NULL\n");
609 				return (EINVAL);
610 			}
611 		/*
612 		 * Otherwise, copy whole portion of header from multiple slots
613 		 * to intermediate buffer.
614 		 */
615 		} else {
616 			rc = ena_netmap_copy_data(ctx->na,
617 			    ctx->slots,
618 			    ctx->nm_i,
619 			    ctx->lim,
620 			    push_len,
621 			    tx_ring->push_buf_intermediate_buf);
622 			if (unlikely(rc)) {
623 				ena_log_nm(adapter->pdev, ERR,
624 				    "Failed to copy data from slots to push_buf\n");
625 				return (EINVAL);
626 			}
627 
628 			*push_hdr = tx_ring->push_buf_intermediate_buf;
629 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
630 
631 			delta = push_len - slot_head_len;
632 		}
633 
634 		ena_log_nm(adapter->pdev, DBG,
635 		    "slot: %d header_buf->vaddr: %p push_len: %d\n",
636 		    slot->buf_idx, *push_hdr, push_len);
637 
638 		/*
639 		* If header was in linear memory space, map for the dma rest of the data
640 		* in the first mbuf of the mbuf chain.
641 		*/
642 		if (slot_head_len > push_len) {
643 			rc = ena_netmap_map_single_slot(ctx->na,
644 			    slot,
645 			    adapter->tx_buf_tag,
646 			    *nm_maps,
647 			    &vaddr,
648 			    &paddr);
649 			if (unlikely(rc != 0)) {
650 				ena_log_nm(adapter->pdev, ERR,
651 				    "DMA mapping error\n");
652 				return (rc);
653 			}
654 			nm_maps++;
655 
656 			ena_buf->paddr = paddr + push_len;
657 			ena_buf->len = slot->len - push_len;
658 			ena_buf++;
659 
660 			tx_info->num_of_bufs++;
661 		}
662 
663 		remaining_len -= slot->len;
664 
665 		/* Save buf idx before advancing */
666 		*nm_buf_idx = slot->buf_idx;
667 		nm_buf_idx++;
668 		slot->buf_idx = 0;
669 
670 		/* Advance to the next socket */
671 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
672 		slot = &ctx->slots[ctx->nm_i];
673 		nm_info->sockets_used++;
674 
675 		/*
676 		 * If header is in non linear space (delta > 0), then skip mbufs
677 		 * containing header and map the last one containing both header
678 		 * and the packet data.
679 		 * The first segment is already counted in.
680 		 */
681 		while (delta > 0) {
682 			__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
683 			frag_len = slot->len;
684 
685 			/*
686 			 * If whole segment contains header just move to the
687 			 * next one and reduce delta.
688 			 */
689 			if (unlikely(delta >= frag_len)) {
690 				delta -= frag_len;
691 			} else {
692 				/*
693 				 * Map the data and then assign it with the
694 				 * offsets
695 				 */
696 				rc = ena_netmap_map_single_slot(ctx->na,
697 				    slot,
698 				    adapter->tx_buf_tag,
699 				    *nm_maps,
700 				    &vaddr,
701 				    &paddr);
702 				if (unlikely(rc != 0)) {
703 					ena_log_nm(adapter->pdev, ERR,
704 					    "DMA mapping error\n");
705 					goto error_map;
706 				}
707 				nm_maps++;
708 
709 				ena_buf->paddr = paddr + delta;
710 				ena_buf->len = slot->len - delta;
711 				ena_buf++;
712 
713 				tx_info->num_of_bufs++;
714 				delta = 0;
715 			}
716 
717 			remaining_len -= slot->len;
718 
719 			/* Save buf idx before advancing */
720 			*nm_buf_idx = slot->buf_idx;
721 			nm_buf_idx++;
722 			slot->buf_idx = 0;
723 
724 			/* Advance to the next socket */
725 			ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
726 			slot = &ctx->slots[ctx->nm_i];
727 			nm_info->sockets_used++;
728 		}
729 	} else {
730 		*push_hdr = NULL;
731 		/*
732 		* header_len is just a hint for the device. Because netmap is
733 		* not giving us any information about packet header length and
734 		* it is not guaranteed that all packet headers will be in the
735 		* 1st slot, setting header_len to 0 is making the device ignore
736 		* this value and resolve header on it's own.
737 		*/
738 		*header_len = 0;
739 	}
740 
741 	/* Map all remaining data (regular routine for non-LLQ mode) */
742 	while (remaining_len > 0) {
743 		__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
744 
745 		rc = ena_netmap_map_single_slot(ctx->na,
746 			    slot,
747 			    adapter->tx_buf_tag,
748 			    *nm_maps,
749 			    &vaddr,
750 			    &paddr);
751 		if (unlikely(rc != 0)) {
752 			ena_log_nm(adapter->pdev, ERR,
753 			    "DMA mapping error\n");
754 			goto error_map;
755 		}
756 		nm_maps++;
757 
758 		ena_buf->paddr = paddr;
759 		ena_buf->len = slot->len;
760 		ena_buf++;
761 
762 		tx_info->num_of_bufs++;
763 
764 		remaining_len -= slot->len;
765 
766 		/* Save buf idx before advancing */
767 		*nm_buf_idx = slot->buf_idx;
768 		nm_buf_idx++;
769 		slot->buf_idx = 0;
770 
771 		/* Advance to the next socket */
772 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
773 		slot = &ctx->slots[ctx->nm_i];
774 		nm_info->sockets_used++;
775 	}
776 
777 	return (0);
778 
779 error_map:
780 	ena_netmap_unmap_last_socket_chain(ctx, tx_info);
781 
782 	return (rc);
783 }
784 
785 static void
786 ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *ctx,
787     struct ena_tx_buffer *tx_info)
788 {
789 	struct ena_netmap_tx_info *nm_info;
790 	int n;
791 
792 	nm_info = &tx_info->nm_info;
793 
794 	/**
795 	 * As the used sockets must not be equal to the buffers used in the LLQ
796 	 * mode, they must be treated separately.
797 	 * First, unmap the DMA maps.
798 	 */
799 	n = tx_info->num_of_bufs;
800 	while (n--) {
801 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
802 		    nm_info->map_seg[n]);
803 	}
804 	tx_info->num_of_bufs = 0;
805 
806 	/* Next, retain the sockets back to the userspace */
807 	n = nm_info->sockets_used;
808 	while (n--) {
809 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
810 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
811 		nm_info->socket_buf_idx[n] = 0;
812 		ctx->nm_i = nm_prev(ctx->nm_i, ctx->lim);
813 	}
814 	nm_info->sockets_used = 0;
815 }
816 
817 static void
818 ena_netmap_tx_cleanup(struct ena_netmap_ctx *ctx)
819 {
820 	uint16_t req_id;
821 	uint16_t total_tx_descs = 0;
822 
823 	ctx->nm_i = ctx->kring->nr_hwtail;
824 	ctx->nt = ctx->ring->next_to_clean;
825 
826 	/* Reclaim buffers for completed transmissions */
827 	while (ena_com_tx_comp_req_id_get(ctx->io_cq, &req_id) >= 0) {
828 		if (validate_tx_req_id(ctx->ring, req_id) != 0)
829 			break;
830 		total_tx_descs += ena_netmap_tx_clean_one(ctx, req_id);
831 	}
832 
833 	ctx->kring->nr_hwtail = ctx->nm_i;
834 
835 	if (total_tx_descs > 0) {
836 		/* acknowledge completion of sent packets */
837 		ctx->ring->next_to_clean = ctx->nt;
838 		ena_com_comp_ack(ctx->ring->ena_com_io_sq, total_tx_descs);
839 		ena_com_update_dev_comp_head(ctx->ring->ena_com_io_cq);
840 	}
841 }
842 
843 static uint16_t
844 ena_netmap_tx_clean_one(struct ena_netmap_ctx *ctx, uint16_t req_id)
845 {
846 	struct ena_tx_buffer *tx_info;
847 	struct ena_netmap_tx_info *nm_info;
848 	int n;
849 
850 	tx_info = &ctx->ring->tx_buffer_info[req_id];
851 	nm_info = &tx_info->nm_info;
852 
853 	/**
854 	 * As the used sockets must not be equal to the buffers used in the LLQ
855 	 * mode, they must be treated separately.
856 	 * First, unmap the DMA maps.
857 	 */
858 	n = tx_info->num_of_bufs;
859 	for (n = 0; n < tx_info->num_of_bufs; n++) {
860 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
861 		    nm_info->map_seg[n]);
862 	}
863 	tx_info->num_of_bufs = 0;
864 
865 	/* Next, retain the sockets back to the userspace */
866 	for (n = 0; n < nm_info->sockets_used; n++) {
867 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
868 		ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0,
869 		    ctx->adapter->ena_dev, "Tx idx is not 0.\n");
870 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
871 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
872 		nm_info->socket_buf_idx[n] = 0;
873 	}
874 	nm_info->sockets_used = 0;
875 
876 	ctx->ring->free_tx_ids[ctx->nt] = req_id;
877 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->lim);
878 
879 	return tx_info->tx_descs;
880 }
881 
882 static inline int
883 validate_tx_req_id(struct ena_ring *tx_ring, uint16_t req_id)
884 {
885 	struct ena_adapter *adapter = tx_ring->adapter;
886 
887 	if (likely(req_id < tx_ring->ring_size))
888 		return (0);
889 
890 	ena_log_nm(adapter->pdev, WARN, "Invalid req_id: %hu\n", req_id);
891 	counter_u64_add(tx_ring->tx_stats.bad_req_id, 1);
892 
893 	ena_trigger_reset(adapter, ENA_REGS_RESET_INV_TX_REQ_ID);
894 
895 	return (EFAULT);
896 }
897 
898 static int
899 ena_netmap_rxsync(struct netmap_kring *kring, int flags)
900 {
901 	struct ena_netmap_ctx ctx;
902 	int rc;
903 
904 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_RXQ_IDX(kring->ring_id));
905 	ctx.ring = &ctx.adapter->rx_ring[kring->ring_id];
906 
907 	if (ctx.kring->rhead > ctx.lim) {
908 		/* Probably not needed to release slots from RX ring. */
909 		return (netmap_ring_reinit(ctx.kring));
910 	}
911 
912 	if (unlikely((if_getdrvflags(ctx.na->ifp) & IFF_DRV_RUNNING) == 0))
913 		return (0);
914 
915 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
916 		return (0);
917 
918 	if ((rc = ena_netmap_rx_frames(&ctx)) != 0)
919 		return (rc);
920 
921 	ena_netmap_rx_cleanup(&ctx);
922 
923 	return (0);
924 }
925 
926 static inline int
927 ena_netmap_rx_frames(struct ena_netmap_ctx *ctx)
928 {
929 	int rc = 0;
930 	int frames_counter = 0;
931 
932 	ctx->nt = ctx->ring->next_to_clean;
933 	ctx->nm_i = ctx->kring->nr_hwtail;
934 
935 	while((rc = ena_netmap_rx_frame(ctx)) == ENA_NETMAP_MORE_FRAMES) {
936 		frames_counter++;
937 		/* In case of multiple frames, it is not an error. */
938 		rc = 0;
939 		if (frames_counter > ENA_MAX_FRAMES) {
940 			ena_log_nm(ctx->adapter->pdev, ERR,
941 				"Driver is stuck in the Rx loop\n");
942 			break;
943 		}
944 	};
945 
946 	ctx->kring->nr_hwtail = ctx->nm_i;
947 	ctx->kring->nr_kflags &= ~NKR_PENDINTR;
948 	ctx->ring->next_to_clean = ctx->nt;
949 
950 	return (rc);
951 }
952 
953 static inline int
954 ena_netmap_rx_frame(struct ena_netmap_ctx *ctx)
955 {
956 	struct ena_com_rx_ctx ena_rx_ctx;
957 	enum ena_regs_reset_reason_types reset_reason;
958 	int rc, len = 0;
959 	uint16_t buf, nm;
960 
961 	ena_rx_ctx.ena_bufs = ctx->ring->ena_bufs;
962 	ena_rx_ctx.max_bufs = ctx->adapter->max_rx_sgl_size;
963 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
964 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
965 
966 	rc = ena_com_rx_pkt(ctx->io_cq, ctx->io_sq, &ena_rx_ctx);
967 	if (unlikely(rc != 0)) {
968 		ena_log_nm(ctx->adapter->pdev, ERR,
969 		    "Failed to read pkt from the device with error: %d\n", rc);
970 		if (rc == ENA_COM_NO_SPACE) {
971 			counter_u64_add(ctx->ring->rx_stats.bad_desc_num, 1);
972 			reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
973 		} else {
974 			counter_u64_add(ctx->ring->rx_stats.bad_req_id, 1);
975 			reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID;
976 		}
977 		ena_trigger_reset(ctx->adapter, reset_reason);
978 		return (rc);
979 	}
980 	if (unlikely(ena_rx_ctx.descs == 0))
981 		return (ENA_NETMAP_NO_MORE_FRAMES);
982 
983         ena_log_nm(ctx->adapter->pdev, DBG,
984 	    "Rx: q %d got packet from ena. descs #:"
985 	    " %d l3 proto %d l4 proto %d hash: %x\n", ctx->ring->qid,
986 	    ena_rx_ctx.descs, ena_rx_ctx.l3_proto, ena_rx_ctx.l4_proto,
987 	    ena_rx_ctx.hash);
988 
989 	for (buf = 0; buf < ena_rx_ctx.descs; buf++)
990 		if ((rc = ena_netmap_rx_load_desc(ctx, buf, &len)) != 0)
991 			break;
992 	/*
993 	 * ena_netmap_rx_load_desc doesn't know the number of descriptors.
994 	 * It just set flag NS_MOREFRAG to all slots, then here flag of
995 	 * last slot is cleared.
996 	 */
997 	ctx->slots[nm_prev(ctx->nm_i, ctx->lim)].flags = NS_BUF_CHANGED;
998 
999 	if (rc != 0) {
1000 		goto rx_clear_desc;
1001 	}
1002 
1003 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
1004             ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
1005 
1006 	counter_enter();
1007 	counter_u64_add_protected(ctx->ring->rx_stats.bytes, len);
1008 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_bytes, len);
1009 	counter_u64_add_protected(ctx->ring->rx_stats.cnt, 1);
1010 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_packets, 1);
1011 	counter_exit();
1012 
1013 	return (ENA_NETMAP_MORE_FRAMES);
1014 
1015 rx_clear_desc:
1016 	nm = ctx->nm_i;
1017 
1018 	/* Remove failed packet from ring */
1019 	while(buf--) {
1020 		ctx->slots[nm].flags = 0;
1021 		ctx->slots[nm].len = 0;
1022 		nm = nm_prev(nm, ctx->lim);
1023 	}
1024 
1025 	return (rc);
1026 }
1027 
1028 static inline int
1029 ena_netmap_rx_load_desc(struct ena_netmap_ctx *ctx, uint16_t buf, int *len)
1030 {
1031 	struct ena_rx_buffer *rx_info;
1032 	uint16_t req_id;
1033 
1034 	req_id = ctx->ring->ena_bufs[buf].req_id;
1035 	rx_info = &ctx->ring->rx_buffer_info[req_id];
1036 	bus_dmamap_sync(ctx->adapter->rx_buf_tag, rx_info->map,
1037 	    BUS_DMASYNC_POSTREAD);
1038 	netmap_unload_map(ctx->na, ctx->adapter->rx_buf_tag, rx_info->map);
1039 
1040 	ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0, ctx->adapter->ena_dev,
1041 	    "Rx idx is not 0.\n");
1042 
1043 	ctx->slots[ctx->nm_i].buf_idx = rx_info->netmap_buf_idx;
1044 	rx_info->netmap_buf_idx = 0;
1045 	/*
1046 	 * Set NS_MOREFRAG to all slots.
1047 	 * Then ena_netmap_rx_frame clears it from last one.
1048 	 */
1049 	ctx->slots[ctx->nm_i].flags |= NS_MOREFRAG | NS_BUF_CHANGED;
1050 	ctx->slots[ctx->nm_i].len = ctx->ring->ena_bufs[buf].len;
1051 	*len += ctx->slots[ctx->nm_i].len;
1052 	ctx->ring->free_rx_ids[ctx->nt] = req_id;
1053 	ena_log_nm(ctx->adapter->pdev, DBG, "rx_info %p, buf_idx %d, paddr %jx, nm: %d\n",
1054 	    rx_info, ctx->slots[ctx->nm_i].buf_idx,
1055 	    (uintmax_t)rx_info->ena_buf.paddr, ctx->nm_i);
1056 
1057 	ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
1058 	ctx->nt = ENA_RX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
1059 
1060 	return (0);
1061 }
1062 
1063 static inline void
1064 ena_netmap_rx_cleanup(struct ena_netmap_ctx *ctx)
1065 {
1066 	int refill_required;
1067 
1068 	refill_required = ctx->kring->rhead - ctx->kring->nr_hwcur;
1069 	if (ctx->kring->nr_hwcur != ctx->kring->nr_hwtail)
1070 		refill_required -= 1;
1071 
1072 	if (refill_required == 0)
1073 		return;
1074 	else if (refill_required < 0)
1075 		refill_required += ctx->kring->nkr_num_slots;
1076 
1077 	ena_refill_rx_bufs(ctx->ring, refill_required);
1078 }
1079 
1080 static inline void
1081 ena_netmap_fill_ctx(struct netmap_kring *kring, struct ena_netmap_ctx *ctx,
1082     uint16_t ena_qid)
1083 {
1084 	ctx->kring = kring;
1085 	ctx->na = kring->na;
1086 	ctx->adapter = ctx->na->ifp->if_softc;
1087 	ctx->lim = kring->nkr_num_slots - 1;
1088 	ctx->io_cq = &ctx->adapter->ena_dev->io_cq_queues[ena_qid];
1089 	ctx->io_sq = &ctx->adapter->ena_dev->io_sq_queues[ena_qid];
1090 	ctx->slots = kring->ring->slot;
1091 }
1092 
1093 void
1094 ena_netmap_unload(struct ena_adapter *adapter, bus_dmamap_t map)
1095 {
1096 	struct netmap_adapter *na = NA(adapter->ifp);
1097 
1098 	netmap_unload_map(na, adapter->tx_buf_tag, map);
1099 }
1100 
1101 #endif /* DEV_NETMAP */
1102