xref: /freebsd/sys/dev/ena/ena_netmap.c (revision 36712a94975f5bd0d26c85377283b49a2369c82f)
1 /*-
2  * BSD LICENSE
3  *
4  * Copyright (c) 2015-2019 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(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->tx_ring_size;
97 	na.num_rx_desc = adapter->rx_ring_size;
98 	na.num_tx_rings = adapter->num_queues;
99 	na.num_rx_rings = adapter->num_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(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(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(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_ASSERT(slot->buf_idx == 0, "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(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 	sx_xlock(&adapter->ioctl_sx);
281 	ENA_FLAG_CLEAR_ATOMIC(ENA_FLAG_TRIGGER_RESET, adapter);
282 	ena_down(adapter);
283 
284 	if (onoff) {
285 		ena_trace(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(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(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 	sx_unlock(&adapter->ioctl_sx);
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 		wmb();
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_trace(ENA_WARNING, "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 		device_printf(adapter->pdev, "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 
430 	/* There are no any offloads, as the netmap doesn't support them */
431 
432 	if (tx_ring->acum_pkts == DB_THRESHOLD ||
433 	    ena_com_is_doorbell_needed(ctx->io_sq, &ena_tx_ctx)) {
434 		wmb();
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_trace(ENA_NETMAP | ENA_DBG | ENA_TXPTH,
444 			    "Tx ring[%d] is out of space\n", tx_ring->que->id);
445 		} else {
446 			device_printf(adapter->pdev,
447 			    "Failed to prepare Tx bufs\n");
448 		}
449 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
450 
451 		ena_netmap_unmap_last_socket_chain(ctx, tx_info);
452 		return (rc);
453 	}
454 
455 	counter_enter();
456 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
457 	counter_u64_add_protected(tx_ring->tx_stats.bytes, packet_len);
458 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
459 	counter_u64_add_protected(adapter->hw_stats.tx_bytes, packet_len);
460 	counter_exit();
461 
462 	tx_info->tx_descs = nb_hw_desc;
463 
464 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
465 
466 	for (unsigned int i = 0; i < tx_info->num_of_bufs; i++)
467 		bus_dmamap_sync(adapter->tx_buf_tag,
468 		   tx_info->nm_info.map_seg[i], BUS_DMASYNC_PREWRITE);
469 
470 	return (0);
471 }
472 
473 static inline uint16_t
474 ena_netmap_count_slots(struct ena_netmap_ctx *ctx)
475 {
476 	uint16_t slots = 1;
477 	uint16_t nm = ctx->nm_i;
478 
479 	while ((ctx->slots[nm].flags & NS_MOREFRAG) != 0) {
480 		slots++;
481 		nm = nm_next(nm, ctx->lim);
482 	}
483 
484 	return slots;
485 }
486 
487 static inline uint16_t
488 ena_netmap_packet_len(struct netmap_slot *slots, u_int slot_index,
489     uint16_t limit)
490 {
491 	struct netmap_slot *nm_slot;
492 	uint16_t packet_size = 0;
493 
494 	do {
495 		nm_slot = &slots[slot_index];
496 		packet_size += nm_slot->len;
497 		slot_index = nm_next(slot_index, limit);
498 	} while ((nm_slot->flags & NS_MOREFRAG) != 0);
499 
500 	return packet_size;
501 }
502 
503 static int
504 ena_netmap_copy_data(struct netmap_adapter *na, struct netmap_slot *slots,
505     u_int slot_index, uint16_t limit, uint16_t bytes_to_copy, void *destination)
506 {
507 	struct netmap_slot *nm_slot;
508 	void *slot_vaddr;
509 	uint16_t packet_size;
510 	uint16_t data_amount;
511 
512 	packet_size = 0;
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 	int rc;
534 
535 	*vaddr = PNMB(na, slot, paddr);
536 	if (unlikely(vaddr == NULL)) {
537 		ena_trace(ENA_ALERT, "Slot address is NULL\n");
538 		return (EINVAL);
539 	}
540 
541 	rc = netmap_load_map(na, dmatag, dmamap, *vaddr);
542 	if (unlikely(rc != 0)) {
543 		ena_trace(ENA_ALERT, "Failed to map slot %d for DMA\n",
544 		    slot->buf_idx);
545 		return (EINVAL);
546 	}
547 
548 	return (0);
549 }
550 
551 static int
552 ena_netmap_tx_map_slots(struct ena_netmap_ctx *ctx,
553     struct ena_tx_buffer *tx_info, void **push_hdr, uint16_t *header_len,
554     uint16_t *packet_len)
555 {
556 	struct netmap_slot *slot;
557 	struct ena_com_buf *ena_buf;
558 	struct ena_adapter *adapter;
559 	struct ena_ring *tx_ring;
560 	struct ena_netmap_tx_info *nm_info;
561 	bus_dmamap_t *nm_maps;
562 	void *vaddr;
563 	uint64_t paddr;
564 	uint32_t *nm_buf_idx;
565 	uint32_t slot_head_len;
566 	uint32_t frag_len;
567 	uint32_t remaining_len;
568 	uint16_t push_len;
569 	uint16_t delta;
570 	int rc;
571 
572 	adapter = ctx->adapter;
573 	tx_ring = ctx->ring;
574 	ena_buf = tx_info->bufs;
575 	nm_info = &tx_info->nm_info;
576 	nm_maps = nm_info->map_seg;
577 	nm_buf_idx = nm_info->socket_buf_idx;
578 	slot = &ctx->slots[ctx->nm_i];
579 
580 	slot_head_len = slot->len;
581 	*packet_len = ena_netmap_packet_len(ctx->slots, ctx->nm_i, ctx->lim);
582 	remaining_len = *packet_len;
583 	delta = 0;
584 
585 	__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
586 	if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
587 		/*
588 		 * When the device is in LLQ mode, the driver will copy
589 		 * the header into the device memory space.
590 		 * The ena_com layer assumes that the header is in a linear
591 		 * memory space.
592 		 * This assumption might be wrong since part of the header
593 		 * can be in the fragmented buffers.
594 		 * First, check if header fits in the first slot. If not, copy
595 		 * it to separate buffer that will be holding linearized data.
596 		 */
597 		push_len = min_t(uint32_t, *packet_len,
598 		    tx_ring->tx_max_header_size);
599 		*header_len = push_len;
600 		/* If header is in linear space, just point to socket's data. */
601 		if (likely(push_len <= slot_head_len)) {
602 			*push_hdr = NMB(ctx->na, slot);
603 			if (unlikely(push_hdr == NULL)) {
604 				device_printf(adapter->pdev,
605 				    "Slot vaddress is NULL\n");
606 				return (EINVAL);
607 			}
608 		/*
609 		 * Otherwise, copy whole portion of header from multiple slots
610 		 * to intermediate buffer.
611 		 */
612 		} else {
613 			rc = ena_netmap_copy_data(ctx->na,
614 			    ctx->slots,
615 			    ctx->nm_i,
616 			    ctx->lim,
617 			    push_len,
618 			    tx_ring->push_buf_intermediate_buf);
619 			if (unlikely(rc)) {
620 				device_printf(adapter->pdev,
621 				    "Failed to copy data from slots to push_buf\n");
622 				return (EINVAL);
623 			}
624 
625 			*push_hdr = tx_ring->push_buf_intermediate_buf;
626 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
627 
628 			delta = push_len - slot_head_len;
629 		}
630 
631 		ena_trace(ENA_NETMAP | ENA_DBG | ENA_TXPTH,
632 		    "slot: %d header_buf->vaddr: %p push_len: %d\n",
633 		    slot->buf_idx, *push_hdr, push_len);
634 
635 		/*
636 		* If header was in linear memory space, map for the dma rest of the data
637 		* in the first mbuf of the mbuf chain.
638 		*/
639 		if (slot_head_len > push_len) {
640 			rc = ena_netmap_map_single_slot(ctx->na,
641 			    slot,
642 			    adapter->tx_buf_tag,
643 			    *nm_maps,
644 			    &vaddr,
645 			    &paddr);
646 			if (unlikely(rc != 0)) {
647 				device_printf(adapter->pdev,
648 				    "DMA mapping error\n");
649 				return (rc);
650 			}
651 			nm_maps++;
652 
653 			ena_buf->paddr = paddr + push_len;
654 			ena_buf->len = slot->len - push_len;
655 			ena_buf++;
656 
657 			tx_info->num_of_bufs++;
658 		}
659 
660 		remaining_len -= slot->len;
661 
662 		/* Save buf idx before advancing */
663 		*nm_buf_idx = slot->buf_idx;
664 		nm_buf_idx++;
665 		slot->buf_idx = 0;
666 
667 		/* Advance to the next socket */
668 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
669 		slot = &ctx->slots[ctx->nm_i];
670 		nm_info->sockets_used++;
671 
672 		/*
673 		 * If header is in non linear space (delta > 0), then skip mbufs
674 		 * containing header and map the last one containing both header
675 		 * and the packet data.
676 		 * The first segment is already counted in.
677 		 */
678 		while (delta > 0) {
679 			__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
680 			frag_len = slot->len;
681 
682 			/*
683 			 * If whole segment contains header just move to the
684 			 * next one and reduce delta.
685 			 */
686 			if (unlikely(delta >= frag_len)) {
687 				delta -= frag_len;
688 			} else {
689 				/*
690 				 * Map the data and then assign it with the
691 				 * offsets
692 				 */
693 				rc = ena_netmap_map_single_slot(ctx->na,
694 				    slot,
695 				    adapter->tx_buf_tag,
696 				    *nm_maps,
697 				    &vaddr,
698 				    &paddr);
699 				if (unlikely(rc != 0)) {
700 					device_printf(adapter->pdev,
701 					    "DMA mapping error\n");
702 					goto error_map;
703 				}
704 				nm_maps++;
705 
706 				ena_buf->paddr = paddr + delta;
707 				ena_buf->len = slot->len - delta;
708 				ena_buf++;
709 
710 				tx_info->num_of_bufs++;
711 				delta = 0;
712 			}
713 
714 			remaining_len -= slot->len;
715 
716 			/* Save buf idx before advancing */
717 			*nm_buf_idx = slot->buf_idx;
718 			nm_buf_idx++;
719 			slot->buf_idx = 0;
720 
721 			/* Advance to the next socket */
722 			ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
723 			slot = &ctx->slots[ctx->nm_i];
724 			nm_info->sockets_used++;
725 		}
726 	} else {
727 		*push_hdr = NULL;
728 		/*
729 		* header_len is just a hint for the device. Because netmap is
730 		* not giving us any information about packet header length and
731 		* it is not guaranteed that all packet headers will be in the
732 		* 1st slot, setting header_len to 0 is making the device ignore
733 		* this value and resolve header on it's own.
734 		*/
735 		*header_len = 0;
736 	}
737 
738 	/* Map all remaining data (regular routine for non-LLQ mode) */
739 	while (remaining_len > 0) {
740 		__builtin_prefetch(&ctx->slots[ctx->nm_i + 1]);
741 
742 		rc = ena_netmap_map_single_slot(ctx->na,
743 			    slot,
744 			    adapter->tx_buf_tag,
745 			    *nm_maps,
746 			    &vaddr,
747 			    &paddr);
748 		if (unlikely(rc != 0)) {
749 			device_printf(adapter->pdev,
750 			    "DMA mapping error\n");
751 			goto error_map;
752 		}
753 		nm_maps++;
754 
755 		ena_buf->paddr = paddr;
756 		ena_buf->len = slot->len;
757 		ena_buf++;
758 
759 		tx_info->num_of_bufs++;
760 
761 		remaining_len -= slot->len;
762 
763 		/* Save buf idx before advancing */
764 		*nm_buf_idx = slot->buf_idx;
765 		nm_buf_idx++;
766 		slot->buf_idx = 0;
767 
768 		/* Advance to the next socket */
769 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
770 		slot = &ctx->slots[ctx->nm_i];
771 		nm_info->sockets_used++;
772 	}
773 
774 	return (0);
775 
776 error_map:
777 	ena_netmap_unmap_last_socket_chain(ctx, tx_info);
778 
779 	return (rc);
780 }
781 
782 static void
783 ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *ctx,
784     struct ena_tx_buffer *tx_info)
785 {
786 	struct ena_netmap_tx_info *nm_info;
787 	int n;
788 
789 	nm_info = &tx_info->nm_info;
790 
791 	/**
792 	 * As the used sockets must not be equal to the buffers used in the LLQ
793 	 * mode, they must be treated separately.
794 	 * First, unmap the DMA maps.
795 	 */
796 	n = tx_info->num_of_bufs;
797 	while (n--) {
798 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
799 		    nm_info->map_seg[n]);
800 	}
801 	tx_info->num_of_bufs = 0;
802 
803 	/* Next, retain the sockets back to the userspace */
804 	n = nm_info->sockets_used;
805 	while (n--) {
806 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
807 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
808 		nm_info->socket_buf_idx[n] = 0;
809 		ctx->nm_i = nm_prev(ctx->nm_i, ctx->lim);
810 	}
811 	nm_info->sockets_used = 0;
812 }
813 
814 static void
815 ena_netmap_tx_cleanup(struct ena_netmap_ctx *ctx)
816 {
817 	uint16_t req_id;
818 	uint16_t total_tx_descs = 0;
819 
820 	ctx->nm_i = ctx->kring->nr_hwtail;
821 	ctx->nt = ctx->ring->next_to_clean;
822 
823 	/* Reclaim buffers for completed transmissions */
824 	while (ena_com_tx_comp_req_id_get(ctx->io_cq, &req_id) >= 0) {
825 		if (validate_tx_req_id(ctx->ring, req_id) != 0)
826 			break;
827 		total_tx_descs += ena_netmap_tx_clean_one(ctx, req_id);
828 	}
829 
830 	ctx->kring->nr_hwtail = ctx->nm_i;
831 
832 	if (total_tx_descs > 0) {
833 		/* acknowledge completion of sent packets */
834 		ctx->ring->next_to_clean = ctx->nt;
835 		ena_com_comp_ack(ctx->ring->ena_com_io_sq, total_tx_descs);
836 		ena_com_update_dev_comp_head(ctx->ring->ena_com_io_cq);
837 	}
838 }
839 
840 static uint16_t
841 ena_netmap_tx_clean_one(struct ena_netmap_ctx *ctx, uint16_t req_id)
842 {
843 	struct ena_tx_buffer *tx_info;
844 	struct ena_netmap_tx_info *nm_info;
845 	int n;
846 
847 	tx_info = &ctx->ring->tx_buffer_info[req_id];
848 	nm_info = &tx_info->nm_info;
849 
850 	/**
851 	 * As the used sockets must not be equal to the buffers used in the LLQ
852 	 * mode, they must be treated separately.
853 	 * First, unmap the DMA maps.
854 	 */
855 	n = tx_info->num_of_bufs;
856 	for (n = 0; n < tx_info->num_of_bufs; n++) {
857 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
858 		    nm_info->map_seg[n]);
859 	}
860 	tx_info->num_of_bufs = 0;
861 
862 	/* Next, retain the sockets back to the userspace */
863 	for (n = 0; n < nm_info->sockets_used; n++) {
864 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
865 		ENA_ASSERT(ctx->slots[ctx->nm_i].buf_idx == 0,
866 		    "Tx idx is not 0.\n");
867 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
868 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
869 		nm_info->socket_buf_idx[n] = 0;
870 	}
871 	nm_info->sockets_used = 0;
872 
873 	ctx->ring->free_tx_ids[ctx->nt] = req_id;
874 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->lim);
875 
876 	return tx_info->tx_descs;
877 }
878 
879 static inline int
880 validate_tx_req_id(struct ena_ring *tx_ring, uint16_t req_id)
881 {
882 	struct ena_adapter *adapter = tx_ring->adapter;
883 
884 	if (likely(req_id < tx_ring->ring_size))
885 		return (0);
886 
887 	ena_trace(ENA_WARNING, "Invalid req_id: %hu\n", req_id);
888 	counter_u64_add(tx_ring->tx_stats.bad_req_id, 1);
889 
890 	adapter->reset_reason = ENA_REGS_RESET_INV_TX_REQ_ID;
891 	ENA_FLAG_SET_ATOMIC(ENA_FLAG_TRIGGER_RESET, adapter);
892 
893 	return (EFAULT);
894 }
895 
896 static int
897 ena_netmap_rxsync(struct netmap_kring *kring, int flags)
898 {
899 	struct ena_netmap_ctx ctx;
900 	int rc;
901 
902 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_RXQ_IDX(kring->ring_id));
903 	ctx.ring = &ctx.adapter->rx_ring[kring->ring_id];
904 
905 	if (ctx.kring->rhead > ctx.lim) {
906 		/* Probably not needed to release slots from RX ring. */
907 		return (netmap_ring_reinit(ctx.kring));
908 	}
909 
910 	if (unlikely((if_getdrvflags(ctx.na->ifp) & IFF_DRV_RUNNING) == 0))
911 		return (0);
912 
913 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
914 		return (0);
915 
916 	if ((rc = ena_netmap_rx_frames(&ctx)) != 0)
917 		return (rc);
918 
919 	ena_netmap_rx_cleanup(&ctx);
920 
921 	return (0);
922 }
923 
924 static inline int
925 ena_netmap_rx_frames(struct ena_netmap_ctx *ctx)
926 {
927 	int rc = 0;
928 	int frames_counter = 0;
929 
930 	ctx->nt = ctx->ring->next_to_clean;
931 	ctx->nm_i = ctx->kring->nr_hwtail;
932 
933 	while((rc = ena_netmap_rx_frame(ctx)) == ENA_NETMAP_MORE_FRAMES) {
934 		frames_counter++;
935 		/* In case of multiple frames, it is not an error. */
936 		rc = 0;
937 		if (frames_counter > ENA_MAX_FRAMES) {
938 			device_printf(ctx->adapter->pdev,
939 				"Driver is stuck in the Rx loop\n");
940 			break;
941 		}
942 	};
943 
944 	ctx->kring->nr_hwtail = ctx->nm_i;
945 	ctx->kring->nr_kflags &= ~NKR_PENDINTR;
946 	ctx->ring->next_to_clean = ctx->nt;
947 
948 	return (rc);
949 }
950 
951 static inline int
952 ena_netmap_rx_frame(struct ena_netmap_ctx *ctx)
953 {
954 	struct ena_com_rx_ctx ena_rx_ctx;
955 	int rc, len = 0;
956 	uint16_t buf, nm;
957 
958 	ena_rx_ctx.ena_bufs = ctx->ring->ena_bufs;
959 	ena_rx_ctx.max_bufs = ctx->adapter->max_rx_sgl_size;
960 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
961 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
962 
963 	rc = ena_com_rx_pkt(ctx->io_cq, ctx->io_sq, &ena_rx_ctx);
964 	if (unlikely(rc != 0)) {
965 		ena_trace(ENA_ALERT, "Too many desc from the device.\n");
966 		counter_u64_add(ctx->ring->rx_stats.bad_desc_num, 1);
967 		ctx->adapter->reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
968 		ENA_FLAG_SET_ATOMIC(ENA_FLAG_TRIGGER_RESET, ctx->adapter);
969 		return (rc);
970 	}
971 	if (unlikely(ena_rx_ctx.descs == 0))
972 		return (ENA_NETMAP_NO_MORE_FRAMES);
973 
974         ena_trace(ENA_NETMAP | ENA_DBG, "Rx: q %d got packet from ena. descs #:"
975 	    " %d l3 proto %d l4 proto %d hash: %x\n", ctx->ring->qid,
976 	    ena_rx_ctx.descs, ena_rx_ctx.l3_proto, ena_rx_ctx.l4_proto,
977 	    ena_rx_ctx.hash);
978 
979 	for (buf = 0; buf < ena_rx_ctx.descs; buf++)
980 		if ((rc = ena_netmap_rx_load_desc(ctx, buf, &len)) != 0)
981 			break;
982 	/*
983 	 * ena_netmap_rx_load_desc doesn't know the number of descriptors.
984 	 * It just set flag NS_MOREFRAG to all slots, then here flag of
985 	 * last slot is cleared.
986 	 */
987 	ctx->slots[nm_prev(ctx->nm_i, ctx->lim)].flags = NS_BUF_CHANGED;
988 
989 	if (rc != 0) {
990 		goto rx_clear_desc;
991 	}
992 
993 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
994             ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
995 
996 	counter_enter();
997 	counter_u64_add_protected(ctx->ring->rx_stats.bytes, len);
998 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_bytes, len);
999 	counter_u64_add_protected(ctx->ring->rx_stats.cnt, 1);
1000 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_packets, 1);
1001 	counter_exit();
1002 
1003 	return (ENA_NETMAP_MORE_FRAMES);
1004 
1005 rx_clear_desc:
1006 	nm = ctx->nm_i;
1007 
1008 	/* Remove failed packet from ring */
1009 	while(buf--) {
1010 		ctx->slots[nm].flags = 0;
1011 		ctx->slots[nm].len = 0;
1012 		nm = nm_prev(nm, ctx->lim);
1013 	}
1014 
1015 	return (rc);
1016 }
1017 
1018 static inline int
1019 ena_netmap_rx_load_desc(struct ena_netmap_ctx *ctx, uint16_t buf, int *len)
1020 {
1021 	struct ena_rx_buffer *rx_info;
1022 	uint16_t req_id;
1023 	int rc;
1024 
1025 	req_id = ctx->ring->ena_bufs[buf].req_id;
1026 	rc = validate_rx_req_id(ctx->ring, req_id);
1027 	if (unlikely(rc != 0))
1028 		return (rc);
1029 
1030 	rx_info = &ctx->ring->rx_buffer_info[req_id];
1031 	bus_dmamap_sync(ctx->adapter->rx_buf_tag, rx_info->map,
1032 	    BUS_DMASYNC_POSTREAD);
1033 	netmap_unload_map(ctx->na, ctx->adapter->rx_buf_tag, rx_info->map);
1034 
1035 	ENA_ASSERT(ctx->slots[ctx->nm_i].buf_idx == 0, "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(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