xref: /freebsd/sys/dev/ena/ena_netmap.c (revision 38a52bd3b5cac3da6f7f6eef3dd050e6aa08ebb3)
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 *, struct netmap_slot *,
67     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 *, uint16_t);
76 static inline int validate_tx_req_id(struct ena_ring *, uint16_t);
77 static int ena_netmap_rx_frames(struct ena_netmap_ctx *);
78 static int ena_netmap_rx_frame(struct ena_netmap_ctx *);
79 static int ena_netmap_rx_load_desc(struct ena_netmap_ctx *, uint16_t, int *);
80 static void ena_netmap_rx_cleanup(struct ena_netmap_ctx *);
81 static void ena_netmap_fill_ctx(struct netmap_kring *, struct ena_netmap_ctx *,
82     uint16_t);
83 
84 int
85 ena_netmap_attach(struct ena_adapter *adapter)
86 {
87 	struct netmap_adapter na;
88 
89 	ena_log_nm(adapter->pdev, INFO, "netmap attach\n");
90 
91 	bzero(&na, sizeof(na));
92 	na.na_flags = NAF_MOREFRAG;
93 	na.ifp = adapter->ifp;
94 	na.num_tx_desc = adapter->requested_tx_ring_size;
95 	na.num_rx_desc = adapter->requested_rx_ring_size;
96 	na.num_tx_rings = adapter->num_io_queues;
97 	na.num_rx_rings = adapter->num_io_queues;
98 	na.rx_buf_maxsize = adapter->buf_ring_size;
99 	na.nm_txsync = ena_netmap_txsync;
100 	na.nm_rxsync = ena_netmap_rxsync;
101 	na.nm_register = ena_netmap_reg;
102 
103 	return (netmap_attach(&na));
104 }
105 
106 int
107 ena_netmap_alloc_rx_slot(struct ena_adapter *adapter, struct ena_ring *rx_ring,
108     struct ena_rx_buffer *rx_info)
109 {
110 	struct netmap_adapter *na = NA(adapter->ifp);
111 	struct netmap_kring *kring;
112 	struct netmap_ring *ring;
113 	struct netmap_slot *slot;
114 	void *addr;
115 	uint64_t paddr;
116 	int nm_i, qid, head, lim, rc;
117 
118 	/* if previously allocated frag is not used */
119 	if (unlikely(rx_info->netmap_buf_idx != 0))
120 		return (0);
121 
122 	qid = rx_ring->qid;
123 	kring = na->rx_rings[qid];
124 	nm_i = kring->nr_hwcur;
125 	head = kring->rhead;
126 
127 	ena_log_nm(adapter->pdev, DBG,
128 	    "nr_hwcur: %d, nr_hwtail: %d, rhead: %d, rcur: %d, rtail: %d\n",
129 	    kring->nr_hwcur, kring->nr_hwtail, kring->rhead, kring->rcur,
130 	    kring->rtail);
131 
132 	if ((nm_i == head) && rx_ring->initialized) {
133 		ena_log_nm(adapter->pdev, ERR,
134 		    "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, struct ena_ring *rx_ring,
173     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
366 			 * commonly 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_ring_tx_doorbell(tx_ring);
379 
380 		ctx->ring->next_to_use = ctx->nt;
381 		ctx->kring->nr_hwcur = ctx->nm_i;
382 	}
383 
384 	return (rc);
385 }
386 
387 static int
388 ena_netmap_tx_frame(struct ena_netmap_ctx *ctx)
389 {
390 	struct ena_com_tx_ctx ena_tx_ctx;
391 	struct ena_adapter *adapter;
392 	struct ena_ring *tx_ring;
393 	struct ena_tx_buffer *tx_info;
394 	uint16_t req_id;
395 	uint16_t header_len;
396 	uint16_t packet_len;
397 	int nb_hw_desc;
398 	int rc;
399 	void *push_hdr;
400 
401 	adapter = ctx->adapter;
402 	if (ena_netmap_count_slots(ctx) > adapter->max_tx_sgl_size) {
403 		ena_log_nm(adapter->pdev, WARN, "Too many slots per packet\n");
404 		return (EINVAL);
405 	}
406 
407 	tx_ring = ctx->ring;
408 
409 	req_id = tx_ring->free_tx_ids[ctx->nt];
410 	tx_info = &tx_ring->tx_buffer_info[req_id];
411 	tx_info->num_of_bufs = 0;
412 	tx_info->nm_info.sockets_used = 0;
413 
414 	rc = ena_netmap_tx_map_slots(ctx, tx_info, &push_hdr, &header_len,
415 	    &packet_len);
416 	if (unlikely(rc != 0)) {
417 		ena_log_nm(adapter->pdev, ERR, "Failed to map Tx slot\n");
418 		return (rc);
419 	}
420 
421 	bzero(&ena_tx_ctx, sizeof(struct ena_com_tx_ctx));
422 	ena_tx_ctx.ena_bufs = tx_info->bufs;
423 	ena_tx_ctx.push_header = push_hdr;
424 	ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
425 	ena_tx_ctx.req_id = req_id;
426 	ena_tx_ctx.header_len = header_len;
427 	ena_tx_ctx.meta_valid = adapter->disable_meta_caching;
428 
429 	/* There are no any offloads, as the netmap doesn't support them */
430 
431 	if (tx_ring->acum_pkts == ENA_DB_THRESHOLD ||
432 	    ena_com_is_doorbell_needed(ctx->io_sq, &ena_tx_ctx))
433 		ena_ring_tx_doorbell(tx_ring);
434 
435 	rc = ena_com_prepare_tx(ctx->io_sq, &ena_tx_ctx, &nb_hw_desc);
436 	if (unlikely(rc != 0)) {
437 		if (likely(rc == ENA_COM_NO_MEM)) {
438 			ena_log_nm(adapter->pdev, DBG,
439 			    "Tx ring[%d] is out of space\n", tx_ring->que->id);
440 		} else {
441 			ena_log_nm(adapter->pdev, ERR,
442 			    "Failed to prepare Tx bufs\n");
443 			ena_trigger_reset(adapter,
444 			    ENA_REGS_RESET_DRIVER_INVALID_STATE);
445 		}
446 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
447 
448 		ena_netmap_unmap_last_socket_chain(ctx, tx_info);
449 		return (rc);
450 	}
451 
452 	counter_enter();
453 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
454 	counter_u64_add_protected(tx_ring->tx_stats.bytes, packet_len);
455 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
456 	counter_u64_add_protected(adapter->hw_stats.tx_bytes, packet_len);
457 	counter_exit();
458 
459 	tx_info->tx_descs = nb_hw_desc;
460 
461 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
462 
463 	for (unsigned int i = 0; i < tx_info->num_of_bufs; i++)
464 		bus_dmamap_sync(adapter->tx_buf_tag,
465 		    tx_info->nm_info.map_seg[i], BUS_DMASYNC_PREWRITE);
466 
467 	return (0);
468 }
469 
470 static inline uint16_t
471 ena_netmap_count_slots(struct ena_netmap_ctx *ctx)
472 {
473 	uint16_t slots = 1;
474 	uint16_t nm = ctx->nm_i;
475 
476 	while ((ctx->slots[nm].flags & NS_MOREFRAG) != 0) {
477 		slots++;
478 		nm = nm_next(nm, ctx->lim);
479 	}
480 
481 	return slots;
482 }
483 
484 static inline uint16_t
485 ena_netmap_packet_len(struct netmap_slot *slots, u_int slot_index,
486     uint16_t limit)
487 {
488 	struct netmap_slot *nm_slot;
489 	uint16_t packet_size = 0;
490 
491 	do {
492 		nm_slot = &slots[slot_index];
493 		packet_size += nm_slot->len;
494 		slot_index = nm_next(slot_index, limit);
495 	} while ((nm_slot->flags & NS_MOREFRAG) != 0);
496 
497 	return packet_size;
498 }
499 
500 static int
501 ena_netmap_copy_data(struct netmap_adapter *na, struct netmap_slot *slots,
502     u_int slot_index, uint16_t limit, uint16_t bytes_to_copy, void *destination)
503 {
504 	struct netmap_slot *nm_slot;
505 	void *slot_vaddr;
506 	uint16_t data_amount;
507 
508 	do {
509 		nm_slot = &slots[slot_index];
510 		slot_vaddr = NMB(na, nm_slot);
511 		if (unlikely(slot_vaddr == NULL))
512 			return (EINVAL);
513 
514 		data_amount = min_t(uint16_t, bytes_to_copy, nm_slot->len);
515 		memcpy(destination, slot_vaddr, data_amount);
516 		bytes_to_copy -= data_amount;
517 
518 		slot_index = nm_next(slot_index, limit);
519 	} while ((nm_slot->flags & NS_MOREFRAG) != 0 && bytes_to_copy > 0);
520 
521 	return (0);
522 }
523 
524 static int
525 ena_netmap_map_single_slot(struct netmap_adapter *na, struct netmap_slot *slot,
526     bus_dma_tag_t dmatag, bus_dmamap_t dmamap, void **vaddr, uint64_t *paddr)
527 {
528 	device_t pdev;
529 	int rc;
530 
531 	pdev = ((struct ena_adapter *)na->ifp->if_softc)->pdev;
532 
533 	*vaddr = PNMB(na, slot, paddr);
534 	if (unlikely(vaddr == NULL)) {
535 		ena_log_nm(pdev, ERR, "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_log_nm(pdev, ERR, "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 				ena_log_nm(adapter->pdev, ERR,
603 				    "Slot vaddress is NULL\n");
604 				return (EINVAL);
605 			}
606 		/*
607 		 * Otherwise, copy whole portion of header from multiple
608 		 * slots to intermediate buffer.
609 		 */
610 		} else {
611 			rc = ena_netmap_copy_data(ctx->na, ctx->slots,
612 			    ctx->nm_i, ctx->lim, push_len,
613 			    tx_ring->push_buf_intermediate_buf);
614 			if (unlikely(rc)) {
615 				ena_log_nm(adapter->pdev, ERR,
616 				    "Failed to copy data from slots to push_buf\n");
617 				return (EINVAL);
618 			}
619 
620 			*push_hdr = tx_ring->push_buf_intermediate_buf;
621 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
622 
623 			delta = push_len - slot_head_len;
624 		}
625 
626 		ena_log_nm(adapter->pdev, DBG,
627 		    "slot: %d header_buf->vaddr: %p push_len: %d\n",
628 		    slot->buf_idx, *push_hdr, push_len);
629 
630 		/*
631 		 * If header was in linear memory space, map for the dma rest of
632 		 * the data in the first mbuf of the mbuf chain.
633 		 */
634 		if (slot_head_len > push_len) {
635 			rc = ena_netmap_map_single_slot(ctx->na, slot,
636 			    adapter->tx_buf_tag, *nm_maps, &vaddr, &paddr);
637 			if (unlikely(rc != 0)) {
638 				ena_log_nm(adapter->pdev, ERR,
639 				    "DMA mapping error\n");
640 				return (rc);
641 			}
642 			nm_maps++;
643 
644 			ena_buf->paddr = paddr + push_len;
645 			ena_buf->len = slot->len - push_len;
646 			ena_buf++;
647 
648 			tx_info->num_of_bufs++;
649 		}
650 
651 		remaining_len -= slot->len;
652 
653 		/* Save buf idx before advancing */
654 		*nm_buf_idx = slot->buf_idx;
655 		nm_buf_idx++;
656 		slot->buf_idx = 0;
657 
658 		/* Advance to the next socket */
659 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
660 		slot = &ctx->slots[ctx->nm_i];
661 		nm_info->sockets_used++;
662 
663 		/*
664 		 * If header is in non linear space (delta > 0), then skip mbufs
665 		 * containing header and map the last one containing both header
666 		 * and the packet data.
667 		 * The first segment is already counted in.
668 		 */
669 		while (delta > 0) {
670 			__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
671 			frag_len = slot->len;
672 
673 			/*
674 			 * If whole segment contains header just move to the
675 			 * next one and reduce delta.
676 			 */
677 			if (unlikely(delta >= frag_len)) {
678 				delta -= frag_len;
679 			} else {
680 				/*
681 				 * Map the data and then assign it with the
682 				 * offsets
683 				 */
684 				rc = ena_netmap_map_single_slot(ctx->na, slot,
685 				    adapter->tx_buf_tag, *nm_maps, &vaddr,
686 				    &paddr);
687 				if (unlikely(rc != 0)) {
688 					ena_log_nm(adapter->pdev, ERR,
689 					    "DMA mapping error\n");
690 					goto error_map;
691 				}
692 				nm_maps++;
693 
694 				ena_buf->paddr = paddr + delta;
695 				ena_buf->len = slot->len - delta;
696 				ena_buf++;
697 
698 				tx_info->num_of_bufs++;
699 				delta = 0;
700 			}
701 
702 			remaining_len -= slot->len;
703 
704 			/* Save buf idx before advancing */
705 			*nm_buf_idx = slot->buf_idx;
706 			nm_buf_idx++;
707 			slot->buf_idx = 0;
708 
709 			/* Advance to the next socket */
710 			ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
711 			slot = &ctx->slots[ctx->nm_i];
712 			nm_info->sockets_used++;
713 		}
714 	} else {
715 		*push_hdr = NULL;
716 		/*
717 		 * header_len is just a hint for the device. Because netmap is
718 		 * not giving us any information about packet header length and
719 		 * it is not guaranteed that all packet headers will be in the
720 		 * 1st slot, setting header_len to 0 is making the device ignore
721 		 * this value and resolve header on it's own.
722 		 */
723 		*header_len = 0;
724 	}
725 
726 	/* Map all remaining data (regular routine for non-LLQ mode) */
727 	while (remaining_len > 0) {
728 		__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
729 
730 		rc = ena_netmap_map_single_slot(ctx->na, slot,
731 		    adapter->tx_buf_tag, *nm_maps, &vaddr, &paddr);
732 		if (unlikely(rc != 0)) {
733 			ena_log_nm(adapter->pdev, ERR, "DMA mapping error\n");
734 			goto error_map;
735 		}
736 		nm_maps++;
737 
738 		ena_buf->paddr = paddr;
739 		ena_buf->len = slot->len;
740 		ena_buf++;
741 
742 		tx_info->num_of_bufs++;
743 
744 		remaining_len -= slot->len;
745 
746 		/* Save buf idx before advancing */
747 		*nm_buf_idx = slot->buf_idx;
748 		nm_buf_idx++;
749 		slot->buf_idx = 0;
750 
751 		/* Advance to the next socket */
752 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
753 		slot = &ctx->slots[ctx->nm_i];
754 		nm_info->sockets_used++;
755 	}
756 
757 	return (0);
758 
759 error_map:
760 	ena_netmap_unmap_last_socket_chain(ctx, tx_info);
761 
762 	return (rc);
763 }
764 
765 static void
766 ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *ctx,
767     struct ena_tx_buffer *tx_info)
768 {
769 	struct ena_netmap_tx_info *nm_info;
770 	int n;
771 
772 	nm_info = &tx_info->nm_info;
773 
774 	/**
775 	 * As the used sockets must not be equal to the buffers used in the LLQ
776 	 * mode, they must be treated separately.
777 	 * First, unmap the DMA maps.
778 	 */
779 	n = tx_info->num_of_bufs;
780 	while (n--) {
781 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
782 		    nm_info->map_seg[n]);
783 	}
784 	tx_info->num_of_bufs = 0;
785 
786 	/* Next, retain the sockets back to the userspace */
787 	n = nm_info->sockets_used;
788 	while (n--) {
789 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
790 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
791 		nm_info->socket_buf_idx[n] = 0;
792 		ctx->nm_i = nm_prev(ctx->nm_i, ctx->lim);
793 	}
794 	nm_info->sockets_used = 0;
795 }
796 
797 static void
798 ena_netmap_tx_cleanup(struct ena_netmap_ctx *ctx)
799 {
800 	uint16_t req_id;
801 	uint16_t total_tx_descs = 0;
802 
803 	ctx->nm_i = ctx->kring->nr_hwtail;
804 	ctx->nt = ctx->ring->next_to_clean;
805 
806 	/* Reclaim buffers for completed transmissions */
807 	while (ena_com_tx_comp_req_id_get(ctx->io_cq, &req_id) >= 0) {
808 		if (validate_tx_req_id(ctx->ring, req_id) != 0)
809 			break;
810 		total_tx_descs += ena_netmap_tx_clean_one(ctx, req_id);
811 	}
812 
813 	ctx->kring->nr_hwtail = ctx->nm_i;
814 
815 	if (total_tx_descs > 0) {
816 		/* acknowledge completion of sent packets */
817 		ctx->ring->next_to_clean = ctx->nt;
818 		ena_com_comp_ack(ctx->ring->ena_com_io_sq, total_tx_descs);
819 		ena_com_update_dev_comp_head(ctx->ring->ena_com_io_cq);
820 	}
821 }
822 
823 static uint16_t
824 ena_netmap_tx_clean_one(struct ena_netmap_ctx *ctx, uint16_t req_id)
825 {
826 	struct ena_tx_buffer *tx_info;
827 	struct ena_netmap_tx_info *nm_info;
828 	int n;
829 
830 	tx_info = &ctx->ring->tx_buffer_info[req_id];
831 	nm_info = &tx_info->nm_info;
832 
833 	/**
834 	 * As the used sockets must not be equal to the buffers used in the LLQ
835 	 * mode, they must be treated separately.
836 	 * First, unmap the DMA maps.
837 	 */
838 	n = tx_info->num_of_bufs;
839 	for (n = 0; n < tx_info->num_of_bufs; n++) {
840 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
841 		    nm_info->map_seg[n]);
842 	}
843 	tx_info->num_of_bufs = 0;
844 
845 	/* Next, retain the sockets back to the userspace */
846 	for (n = 0; n < nm_info->sockets_used; n++) {
847 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
848 		ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0,
849 		    ctx->adapter->ena_dev, "Tx idx is not 0.\n");
850 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
851 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
852 		nm_info->socket_buf_idx[n] = 0;
853 	}
854 	nm_info->sockets_used = 0;
855 
856 	ctx->ring->free_tx_ids[ctx->nt] = req_id;
857 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->lim);
858 
859 	return tx_info->tx_descs;
860 }
861 
862 static inline int
863 validate_tx_req_id(struct ena_ring *tx_ring, uint16_t req_id)
864 {
865 	struct ena_adapter *adapter = tx_ring->adapter;
866 
867 	if (likely(req_id < tx_ring->ring_size))
868 		return (0);
869 
870 	ena_log_nm(adapter->pdev, WARN, "Invalid req_id %hu in qid %hu\n",
871 	    req_id, tx_ring->qid);
872 	counter_u64_add(tx_ring->tx_stats.bad_req_id, 1);
873 
874 	ena_trigger_reset(adapter, ENA_REGS_RESET_INV_TX_REQ_ID);
875 
876 	return (EFAULT);
877 }
878 
879 static int
880 ena_netmap_rxsync(struct netmap_kring *kring, int flags)
881 {
882 	struct ena_netmap_ctx ctx;
883 	int rc;
884 
885 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_RXQ_IDX(kring->ring_id));
886 	ctx.ring = &ctx.adapter->rx_ring[kring->ring_id];
887 
888 	if (ctx.kring->rhead > ctx.lim) {
889 		/* Probably not needed to release slots from RX ring. */
890 		return (netmap_ring_reinit(ctx.kring));
891 	}
892 
893 	if (unlikely((if_getdrvflags(ctx.na->ifp) & IFF_DRV_RUNNING) == 0))
894 		return (0);
895 
896 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
897 		return (0);
898 
899 	if ((rc = ena_netmap_rx_frames(&ctx)) != 0)
900 		return (rc);
901 
902 	ena_netmap_rx_cleanup(&ctx);
903 
904 	return (0);
905 }
906 
907 static inline int
908 ena_netmap_rx_frames(struct ena_netmap_ctx *ctx)
909 {
910 	int rc = 0;
911 	int frames_counter = 0;
912 
913 	ctx->nt = ctx->ring->next_to_clean;
914 	ctx->nm_i = ctx->kring->nr_hwtail;
915 
916 	while ((rc = ena_netmap_rx_frame(ctx)) == ENA_NETMAP_MORE_FRAMES) {
917 		frames_counter++;
918 		/* In case of multiple frames, it is not an error. */
919 		rc = 0;
920 		if (frames_counter > ENA_MAX_FRAMES) {
921 			ena_log_nm(ctx->adapter->pdev, ERR,
922 			    "Driver is stuck in the Rx loop\n");
923 			break;
924 		}
925 	};
926 
927 	ctx->kring->nr_hwtail = ctx->nm_i;
928 	ctx->kring->nr_kflags &= ~NKR_PENDINTR;
929 	ctx->ring->next_to_clean = ctx->nt;
930 
931 	return (rc);
932 }
933 
934 static inline int
935 ena_netmap_rx_frame(struct ena_netmap_ctx *ctx)
936 {
937 	struct ena_com_rx_ctx ena_rx_ctx;
938 	enum ena_regs_reset_reason_types reset_reason;
939 	int rc, len = 0;
940 	uint16_t buf, nm;
941 
942 	ena_rx_ctx.ena_bufs = ctx->ring->ena_bufs;
943 	ena_rx_ctx.max_bufs = ctx->adapter->max_rx_sgl_size;
944 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
945 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
946 
947 	rc = ena_com_rx_pkt(ctx->io_cq, ctx->io_sq, &ena_rx_ctx);
948 	if (unlikely(rc != 0)) {
949 		ena_log_nm(ctx->adapter->pdev, ERR,
950 		    "Failed to read pkt from the device with error: %d\n", rc);
951 		if (rc == ENA_COM_NO_SPACE) {
952 			counter_u64_add(ctx->ring->rx_stats.bad_desc_num, 1);
953 			reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
954 		} else {
955 			counter_u64_add(ctx->ring->rx_stats.bad_req_id, 1);
956 			reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID;
957 		}
958 		ena_trigger_reset(ctx->adapter, reset_reason);
959 		return (rc);
960 	}
961 	if (unlikely(ena_rx_ctx.descs == 0))
962 		return (ENA_NETMAP_NO_MORE_FRAMES);
963 
964 	ena_log_nm(ctx->adapter->pdev, DBG,
965 	    "Rx: q %d got packet from ena. descs #:"
966 	    " %d l3 proto %d l4 proto %d hash: %x\n",
967 	    ctx->ring->qid, ena_rx_ctx.descs, ena_rx_ctx.l3_proto,
968 	    ena_rx_ctx.l4_proto, ena_rx_ctx.hash);
969 
970 	for (buf = 0; buf < ena_rx_ctx.descs; buf++)
971 		if ((rc = ena_netmap_rx_load_desc(ctx, buf, &len)) != 0)
972 			break;
973 	/*
974 	 * ena_netmap_rx_load_desc doesn't know the number of descriptors.
975 	 * It just set flag NS_MOREFRAG to all slots, then here flag of
976 	 * last slot is cleared.
977 	 */
978 	ctx->slots[nm_prev(ctx->nm_i, ctx->lim)].flags = NS_BUF_CHANGED;
979 
980 	if (rc != 0) {
981 		goto rx_clear_desc;
982 	}
983 
984 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
985 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
986 
987 	counter_enter();
988 	counter_u64_add_protected(ctx->ring->rx_stats.bytes, len);
989 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_bytes, len);
990 	counter_u64_add_protected(ctx->ring->rx_stats.cnt, 1);
991 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_packets, 1);
992 	counter_exit();
993 
994 	return (ENA_NETMAP_MORE_FRAMES);
995 
996 rx_clear_desc:
997 	nm = ctx->nm_i;
998 
999 	/* Remove failed packet from ring */
1000 	while (buf--) {
1001 		ctx->slots[nm].flags = 0;
1002 		ctx->slots[nm].len = 0;
1003 		nm = nm_prev(nm, ctx->lim);
1004 	}
1005 
1006 	return (rc);
1007 }
1008 
1009 static inline int
1010 ena_netmap_rx_load_desc(struct ena_netmap_ctx *ctx, uint16_t buf, int *len)
1011 {
1012 	struct ena_rx_buffer *rx_info;
1013 	uint16_t req_id;
1014 
1015 	req_id = ctx->ring->ena_bufs[buf].req_id;
1016 	rx_info = &ctx->ring->rx_buffer_info[req_id];
1017 	bus_dmamap_sync(ctx->adapter->rx_buf_tag, rx_info->map,
1018 	    BUS_DMASYNC_POSTREAD);
1019 	netmap_unload_map(ctx->na, ctx->adapter->rx_buf_tag, rx_info->map);
1020 
1021 	ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0, ctx->adapter->ena_dev,
1022 	    "Rx idx is not 0.\n");
1023 
1024 	ctx->slots[ctx->nm_i].buf_idx = rx_info->netmap_buf_idx;
1025 	rx_info->netmap_buf_idx = 0;
1026 	/*
1027 	 * Set NS_MOREFRAG to all slots.
1028 	 * Then ena_netmap_rx_frame clears it from last one.
1029 	 */
1030 	ctx->slots[ctx->nm_i].flags |= NS_MOREFRAG | NS_BUF_CHANGED;
1031 	ctx->slots[ctx->nm_i].len = ctx->ring->ena_bufs[buf].len;
1032 	*len += ctx->slots[ctx->nm_i].len;
1033 	ctx->ring->free_rx_ids[ctx->nt] = req_id;
1034 	ena_log_nm(ctx->adapter->pdev, DBG,
1035 	    "rx_info %p, buf_idx %d, paddr %jx, nm: %d\n", rx_info,
1036 	    ctx->slots[ctx->nm_i].buf_idx, (uintmax_t)rx_info->ena_buf.paddr,
1037 	    ctx->nm_i);
1038 
1039 	ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
1040 	ctx->nt = ENA_RX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
1041 
1042 	return (0);
1043 }
1044 
1045 static inline void
1046 ena_netmap_rx_cleanup(struct ena_netmap_ctx *ctx)
1047 {
1048 	int refill_required;
1049 
1050 	refill_required = ctx->kring->rhead - ctx->kring->nr_hwcur;
1051 	if (ctx->kring->nr_hwcur != ctx->kring->nr_hwtail)
1052 		refill_required -= 1;
1053 
1054 	if (refill_required == 0)
1055 		return;
1056 	else if (refill_required < 0)
1057 		refill_required += ctx->kring->nkr_num_slots;
1058 
1059 	ena_refill_rx_bufs(ctx->ring, refill_required);
1060 }
1061 
1062 static inline void
1063 ena_netmap_fill_ctx(struct netmap_kring *kring, struct ena_netmap_ctx *ctx,
1064     uint16_t ena_qid)
1065 {
1066 	ctx->kring = kring;
1067 	ctx->na = kring->na;
1068 	ctx->adapter = ctx->na->ifp->if_softc;
1069 	ctx->lim = kring->nkr_num_slots - 1;
1070 	ctx->io_cq = &ctx->adapter->ena_dev->io_cq_queues[ena_qid];
1071 	ctx->io_sq = &ctx->adapter->ena_dev->io_sq_queues[ena_qid];
1072 	ctx->slots = kring->ring->slot;
1073 }
1074 
1075 void
1076 ena_netmap_unload(struct ena_adapter *adapter, bus_dmamap_t map)
1077 {
1078 	struct netmap_adapter *na = NA(adapter->ifp);
1079 
1080 	netmap_unload_map(na, adapter->tx_buf_tag, map);
1081 }
1082 
1083 #endif /* DEV_NETMAP */
1084