xref: /src/sys/dev/ena/ena_netmap.c (revision 97e84c587d6f86aa883720296449b380adcf6915)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2015-2024 Amazon.com, Inc. or its affiliates.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  *
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29  */
30 #include <sys/cdefs.h>
31 #ifdef DEV_NETMAP
32 
33 #include "ena.h"
34 #include "ena_netmap.h"
35 
36 #define ENA_NETMAP_MORE_FRAMES		1
37 #define ENA_NETMAP_NO_MORE_FRAMES	0
38 #define ENA_MAX_FRAMES			16384
39 
40 struct ena_netmap_ctx {
41 	struct netmap_kring *kring;
42 	struct ena_adapter *adapter;
43 	struct netmap_adapter *na;
44 	struct netmap_slot *slots;
45 	struct ena_ring *ring;
46 	struct ena_com_io_cq *io_cq;
47 	struct ena_com_io_sq *io_sq;
48 	u_int nm_i;
49 	uint16_t nt;
50 	uint16_t lim;
51 };
52 
53 /* Netmap callbacks */
54 static int ena_netmap_reg(struct netmap_adapter *, int);
55 static int ena_netmap_txsync(struct netmap_kring *, int);
56 static int ena_netmap_rxsync(struct netmap_kring *, int);
57 
58 /* Helper functions */
59 static int ena_netmap_tx_frames(struct ena_netmap_ctx *);
60 static int ena_netmap_tx_frame(struct ena_netmap_ctx *);
61 static inline uint16_t ena_netmap_count_slots(struct ena_netmap_ctx *);
62 static inline uint16_t ena_netmap_packet_len(struct netmap_slot *, u_int,
63     uint16_t);
64 static int ena_netmap_copy_data(struct netmap_adapter *, struct netmap_slot *,
65     u_int, uint16_t, uint16_t, void *);
66 static int ena_netmap_map_single_slot(struct netmap_adapter *,
67     struct netmap_slot *, bus_dma_tag_t, bus_dmamap_t, void **, uint64_t *);
68 static int ena_netmap_tx_map_slots(struct ena_netmap_ctx *,
69     struct ena_tx_buffer *, void **, uint16_t *, uint16_t *);
70 static void ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *,
71     struct ena_tx_buffer *);
72 static void ena_netmap_tx_cleanup(struct ena_netmap_ctx *);
73 static uint16_t ena_netmap_tx_clean_one(struct ena_netmap_ctx *, uint16_t);
74 static int ena_netmap_rx_frames(struct ena_netmap_ctx *);
75 static int ena_netmap_rx_frame(struct ena_netmap_ctx *);
76 static int ena_netmap_rx_load_desc(struct ena_netmap_ctx *, uint16_t, int *);
77 static void ena_netmap_rx_cleanup(struct ena_netmap_ctx *);
78 static void ena_netmap_fill_ctx(struct netmap_kring *, struct ena_netmap_ctx *,
79     uint16_t);
80 
81 int
ena_netmap_attach(struct ena_adapter * adapter)82 ena_netmap_attach(struct ena_adapter *adapter)
83 {
84 	struct netmap_adapter na;
85 
86 	ena_log_nm(adapter->pdev, INFO, "netmap attach\n");
87 
88 	bzero(&na, sizeof(na));
89 	na.na_flags = NAF_MOREFRAG;
90 	na.ifp = adapter->ifp;
91 	na.num_tx_desc = adapter->requested_tx_ring_size;
92 	na.num_rx_desc = adapter->requested_rx_ring_size;
93 	na.num_tx_rings = adapter->num_io_queues;
94 	na.num_rx_rings = adapter->num_io_queues;
95 	na.rx_buf_maxsize = adapter->buf_ring_size;
96 	na.nm_txsync = ena_netmap_txsync;
97 	na.nm_rxsync = ena_netmap_rxsync;
98 	na.nm_register = ena_netmap_reg;
99 
100 	return (netmap_attach(&na));
101 }
102 
103 int
ena_netmap_alloc_rx_slot(struct ena_adapter * adapter,struct ena_ring * rx_ring,struct ena_rx_buffer * rx_info)104 ena_netmap_alloc_rx_slot(struct ena_adapter *adapter, struct ena_ring *rx_ring,
105     struct ena_rx_buffer *rx_info)
106 {
107 	struct netmap_adapter *na = NA(adapter->ifp);
108 	struct netmap_kring *kring;
109 	struct netmap_ring *ring;
110 	struct netmap_slot *slot;
111 	void *addr;
112 	uint64_t paddr;
113 	int nm_i, qid, head, lim, rc;
114 
115 	/* if previously allocated frag is not used */
116 	if (unlikely(rx_info->netmap_buf_idx != 0))
117 		return (0);
118 
119 	qid = rx_ring->qid;
120 	kring = na->rx_rings[qid];
121 	nm_i = kring->nr_hwcur;
122 	head = kring->rhead;
123 
124 	ena_log_nm(adapter->pdev, DBG,
125 	    "nr_hwcur: %d, nr_hwtail: %d, rhead: %d, rcur: %d, rtail: %d\n",
126 	    kring->nr_hwcur, kring->nr_hwtail, kring->rhead, kring->rcur,
127 	    kring->rtail);
128 
129 	if ((nm_i == head) && rx_ring->initialized) {
130 		ena_log_nm(adapter->pdev, ERR,
131 		    "No free slots in netmap ring\n");
132 		return (ENOMEM);
133 	}
134 
135 	ring = kring->ring;
136 	if (ring == NULL) {
137 		ena_log_nm(adapter->pdev, ERR, "Rx ring %d is NULL\n", qid);
138 		return (EFAULT);
139 	}
140 	slot = &ring->slot[nm_i];
141 
142 	addr = PNMB(na, slot, &paddr);
143 	if (addr == NETMAP_BUF_BASE(na)) {
144 		ena_log_nm(adapter->pdev, ERR, "Bad buff in slot\n");
145 		return (EFAULT);
146 	}
147 
148 	rc = netmap_load_map(na, adapter->rx_buf_tag, rx_info->map, addr);
149 	if (rc != 0) {
150 		ena_log_nm(adapter->pdev, WARN, "DMA mapping error\n");
151 		return (rc);
152 	}
153 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map, BUS_DMASYNC_PREREAD);
154 
155 	rx_info->ena_buf.paddr = paddr;
156 	rx_info->ena_buf.len = ring->nr_buf_size;
157 	rx_info->mbuf = NULL;
158 	rx_info->netmap_buf_idx = slot->buf_idx;
159 
160 	slot->buf_idx = 0;
161 
162 	lim = kring->nkr_num_slots - 1;
163 	kring->nr_hwcur = nm_next(nm_i, lim);
164 
165 	return (0);
166 }
167 
168 void
ena_netmap_free_rx_slot(struct ena_adapter * adapter,struct ena_ring * rx_ring,struct ena_rx_buffer * rx_info)169 ena_netmap_free_rx_slot(struct ena_adapter *adapter, struct ena_ring *rx_ring,
170     struct ena_rx_buffer *rx_info)
171 {
172 	struct netmap_adapter *na;
173 	struct netmap_kring *kring;
174 	struct netmap_slot *slot;
175 	int nm_i, qid, lim;
176 
177 	na = NA(adapter->ifp);
178 	if (na == NULL) {
179 		ena_log_nm(adapter->pdev, ERR, "netmap adapter is NULL\n");
180 		return;
181 	}
182 
183 	if (na->rx_rings == NULL) {
184 		ena_log_nm(adapter->pdev, ERR, "netmap rings are NULL\n");
185 		return;
186 	}
187 
188 	qid = rx_ring->qid;
189 	kring = na->rx_rings[qid];
190 	if (kring == NULL) {
191 		ena_log_nm(adapter->pdev, ERR,
192 		    "netmap kernel ring %d is NULL\n", qid);
193 		return;
194 	}
195 
196 	lim = kring->nkr_num_slots - 1;
197 	nm_i = nm_prev(kring->nr_hwcur, lim);
198 
199 	if (kring->nr_mode != NKR_NETMAP_ON)
200 		return;
201 
202 	bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map,
203 	    BUS_DMASYNC_POSTREAD);
204 	netmap_unload_map(na, adapter->rx_buf_tag, rx_info->map);
205 
206 	KASSERT(kring->ring != NULL, ("Netmap Rx ring is NULL\n"));
207 
208 	slot = &kring->ring->slot[nm_i];
209 
210 	ENA_WARN(slot->buf_idx != 0, adapter->ena_dev, "Overwrite slot buf\n");
211 	slot->buf_idx = rx_info->netmap_buf_idx;
212 	slot->flags = NS_BUF_CHANGED;
213 
214 	rx_info->netmap_buf_idx = 0;
215 	kring->nr_hwcur = nm_i;
216 }
217 
218 static bool
ena_ring_in_netmap(struct ena_adapter * adapter,int qid,enum txrx x)219 ena_ring_in_netmap(struct ena_adapter *adapter, int qid, enum txrx x)
220 {
221 	struct netmap_adapter *na;
222 	struct netmap_kring *kring;
223 
224 	if (if_getcapenable(adapter->ifp) & IFCAP_NETMAP) {
225 		na = NA(adapter->ifp);
226 		if (na->na_flags & NAF_NATIVE) {
227 			kring = (x == NR_RX) ? na->rx_rings[qid] : na->tx_rings[qid];
228 			if (kring->nr_mode == NKR_NETMAP_ON)
229 				return true;
230 		}
231 	}
232 	return false;
233 }
234 
235 bool
ena_tx_ring_in_netmap(struct ena_adapter * adapter,int qid)236 ena_tx_ring_in_netmap(struct ena_adapter *adapter, int qid)
237 {
238 	return ena_ring_in_netmap(adapter, qid, NR_TX);
239 }
240 
241 bool
ena_rx_ring_in_netmap(struct ena_adapter * adapter,int qid)242 ena_rx_ring_in_netmap(struct ena_adapter *adapter, int qid)
243 {
244 	return ena_ring_in_netmap(adapter, qid, NR_RX);
245 }
246 
247 static void
ena_netmap_reset_ring(struct ena_adapter * adapter,int qid,enum txrx x)248 ena_netmap_reset_ring(struct ena_adapter *adapter, int qid, enum txrx x)
249 {
250 	if (!ena_ring_in_netmap(adapter, qid, x))
251 		return;
252 
253 	netmap_reset(NA(adapter->ifp), x, qid, 0);
254 	ena_log_nm(adapter->pdev, INFO, "%s ring %d is in netmap mode\n",
255 	    (x == NR_TX) ? "Tx" : "Rx", qid);
256 }
257 
258 void
ena_netmap_reset_rx_ring(struct ena_adapter * adapter,int qid)259 ena_netmap_reset_rx_ring(struct ena_adapter *adapter, int qid)
260 {
261 	ena_netmap_reset_ring(adapter, qid, NR_RX);
262 }
263 
264 void
ena_netmap_reset_tx_ring(struct ena_adapter * adapter,int qid)265 ena_netmap_reset_tx_ring(struct ena_adapter *adapter, int qid)
266 {
267 	ena_netmap_reset_ring(adapter, qid, NR_TX);
268 }
269 
270 static int
ena_netmap_reg(struct netmap_adapter * na,int onoff)271 ena_netmap_reg(struct netmap_adapter *na, int onoff)
272 {
273 	if_t ifp = na->ifp;
274 	struct ena_adapter *adapter = if_getsoftc(ifp);
275 	device_t pdev = adapter->pdev;
276 	struct netmap_kring *kring;
277 	enum txrx t;
278 	int rc, i;
279 
280 	ENA_LOCK_LOCK();
281 	ENA_FLAG_CLEAR_ATOMIC(ENA_FLAG_TRIGGER_RESET, adapter);
282 	ena_down(adapter);
283 
284 	if (onoff) {
285 		ena_log_nm(pdev, INFO, "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_log_nm(pdev, INFO, "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_log_nm(pdev, WARN, "ena_up failed with rc=%d\n", rc);
311 		adapter->reset_reason = ENA_REGS_RESET_DRIVER_INVALID_STATE;
312 		nm_clear_native_flags(na);
313 		ena_destroy_device(adapter, false);
314 		ENA_FLAG_SET_ATOMIC(ENA_FLAG_DEV_UP_BEFORE_RESET, adapter);
315 		rc = ena_restore_device(adapter);
316 	}
317 	ENA_LOCK_UNLOCK();
318 
319 	return (rc);
320 }
321 
322 static int
ena_netmap_txsync(struct netmap_kring * kring,int flags)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
ena_netmap_tx_frames(struct ena_netmap_ctx * ctx)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
365 			 * commonly being full.
366 			 */
367 			if (rc == ENA_COM_NO_MEM)
368 				rc = 0;
369 			break;
370 		}
371 		tx_ring->acum_pkts++;
372 	}
373 
374 	/* If any packet was sent... */
375 	if (likely(ctx->nm_i != ctx->kring->nr_hwcur)) {
376 		/* ...send the doorbell to the device. */
377 		ena_ring_tx_doorbell(tx_ring);
378 
379 		ctx->ring->next_to_use = ctx->nt;
380 		ctx->kring->nr_hwcur = ctx->nm_i;
381 	}
382 
383 	return (rc);
384 }
385 
386 static int
ena_netmap_tx_frame(struct ena_netmap_ctx * ctx)387 ena_netmap_tx_frame(struct ena_netmap_ctx *ctx)
388 {
389 	struct ena_com_tx_ctx ena_tx_ctx;
390 	struct ena_adapter *adapter;
391 	struct ena_ring *tx_ring;
392 	struct ena_tx_buffer *tx_info;
393 	uint16_t req_id;
394 	uint16_t header_len;
395 	uint16_t packet_len;
396 	int nb_hw_desc;
397 	int rc;
398 	void *push_hdr;
399 
400 	adapter = ctx->adapter;
401 	if (ena_netmap_count_slots(ctx) > adapter->max_tx_sgl_size) {
402 		ena_log_nm(adapter->pdev, WARN, "Too many slots per packet\n");
403 		return (EINVAL);
404 	}
405 
406 	tx_ring = ctx->ring;
407 
408 	req_id = tx_ring->free_tx_ids[ctx->nt];
409 	tx_info = &tx_ring->tx_buffer_info[req_id];
410 	tx_info->num_of_bufs = 0;
411 	tx_info->nm_info.sockets_used = 0;
412 
413 	rc = ena_netmap_tx_map_slots(ctx, tx_info, &push_hdr, &header_len,
414 	    &packet_len);
415 	if (unlikely(rc != 0)) {
416 		ena_log_nm(adapter->pdev, ERR, "Failed to map Tx slot\n");
417 		return (rc);
418 	}
419 
420 	bzero(&ena_tx_ctx, sizeof(struct ena_com_tx_ctx));
421 	ena_tx_ctx.ena_bufs = tx_info->bufs;
422 	ena_tx_ctx.push_header = push_hdr;
423 	ena_tx_ctx.num_bufs = tx_info->num_of_bufs;
424 	ena_tx_ctx.req_id = req_id;
425 	ena_tx_ctx.header_len = header_len;
426 	ena_tx_ctx.meta_valid = adapter->disable_meta_caching;
427 
428 	/* There are no any offloads, as the netmap doesn't support them */
429 
430 	if (tx_ring->acum_pkts == ENA_DB_THRESHOLD ||
431 	    ena_com_is_doorbell_needed(ctx->io_sq, &ena_tx_ctx))
432 		ena_ring_tx_doorbell(tx_ring);
433 
434 	rc = ena_com_prepare_tx(ctx->io_sq, &ena_tx_ctx, &nb_hw_desc);
435 	if (unlikely(rc != 0)) {
436 		if (likely(rc == ENA_COM_NO_MEM)) {
437 			ena_log_nm(adapter->pdev, DBG,
438 			    "Tx ring[%d] is out of space\n", tx_ring->que->id);
439 		} else {
440 			ena_log_nm(adapter->pdev, ERR,
441 			    "Failed to prepare Tx bufs\n");
442 			ena_trigger_reset(adapter,
443 			    ENA_REGS_RESET_DRIVER_INVALID_STATE);
444 		}
445 		counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1);
446 
447 		ena_netmap_unmap_last_socket_chain(ctx, tx_info);
448 		return (rc);
449 	}
450 
451 	counter_enter();
452 	counter_u64_add_protected(tx_ring->tx_stats.cnt, 1);
453 	counter_u64_add_protected(tx_ring->tx_stats.bytes, packet_len);
454 	counter_u64_add_protected(adapter->hw_stats.tx_packets, 1);
455 	counter_u64_add_protected(adapter->hw_stats.tx_bytes, packet_len);
456 	counter_exit();
457 
458 	tx_info->tx_descs = nb_hw_desc;
459 
460 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
461 
462 	for (unsigned int i = 0; i < tx_info->num_of_bufs; i++)
463 		bus_dmamap_sync(adapter->tx_buf_tag,
464 		    tx_info->nm_info.map_seg[i], BUS_DMASYNC_PREWRITE);
465 
466 	return (0);
467 }
468 
469 static inline uint16_t
ena_netmap_count_slots(struct ena_netmap_ctx * ctx)470 ena_netmap_count_slots(struct ena_netmap_ctx *ctx)
471 {
472 	uint16_t slots = 1;
473 	uint16_t nm = ctx->nm_i;
474 
475 	while ((ctx->slots[nm].flags & NS_MOREFRAG) != 0) {
476 		slots++;
477 		nm = nm_next(nm, ctx->lim);
478 	}
479 
480 	return slots;
481 }
482 
483 static inline uint16_t
ena_netmap_packet_len(struct netmap_slot * slots,u_int slot_index,uint16_t limit)484 ena_netmap_packet_len(struct netmap_slot *slots, u_int slot_index,
485     uint16_t limit)
486 {
487 	struct netmap_slot *nm_slot;
488 	uint16_t packet_size = 0;
489 
490 	do {
491 		nm_slot = &slots[slot_index];
492 		packet_size += nm_slot->len;
493 		slot_index = nm_next(slot_index, limit);
494 	} while ((nm_slot->flags & NS_MOREFRAG) != 0);
495 
496 	return packet_size;
497 }
498 
499 static int
ena_netmap_copy_data(struct netmap_adapter * na,struct netmap_slot * slots,u_int slot_index,uint16_t limit,uint16_t bytes_to_copy,void * destination)500 ena_netmap_copy_data(struct netmap_adapter *na, struct netmap_slot *slots,
501     u_int slot_index, uint16_t limit, uint16_t bytes_to_copy, void *destination)
502 {
503 	struct netmap_slot *nm_slot;
504 	void *slot_vaddr;
505 	uint16_t data_amount;
506 
507 	do {
508 		nm_slot = &slots[slot_index];
509 		slot_vaddr = NMB(na, nm_slot);
510 		if (unlikely(slot_vaddr == NULL))
511 			return (EINVAL);
512 
513 		data_amount = min_t(uint16_t, bytes_to_copy, nm_slot->len);
514 		memcpy(destination, slot_vaddr, data_amount);
515 		bytes_to_copy -= data_amount;
516 
517 		slot_index = nm_next(slot_index, limit);
518 	} while ((nm_slot->flags & NS_MOREFRAG) != 0 && bytes_to_copy > 0);
519 
520 	return (0);
521 }
522 
523 static int
ena_netmap_map_single_slot(struct netmap_adapter * na,struct netmap_slot * slot,bus_dma_tag_t dmatag,bus_dmamap_t dmamap,void ** vaddr,uint64_t * paddr)524 ena_netmap_map_single_slot(struct netmap_adapter *na, struct netmap_slot *slot,
525     bus_dma_tag_t dmatag, bus_dmamap_t dmamap, void **vaddr, uint64_t *paddr)
526 {
527 	device_t pdev;
528 	int rc;
529 
530 	pdev = ((struct ena_adapter *)if_getsoftc(na->ifp))->pdev;
531 
532 	*vaddr = PNMB(na, slot, paddr);
533 	if (unlikely(vaddr == NULL)) {
534 		ena_log_nm(pdev, ERR, "Slot address is NULL\n");
535 		return (EINVAL);
536 	}
537 
538 	rc = netmap_load_map(na, dmatag, dmamap, *vaddr);
539 	if (unlikely(rc != 0)) {
540 		ena_log_nm(pdev, ERR, "Failed to map slot %d for DMA\n",
541 		    slot->buf_idx);
542 		return (EINVAL);
543 	}
544 
545 	return (0);
546 }
547 
548 static int
ena_netmap_tx_map_slots(struct ena_netmap_ctx * ctx,struct ena_tx_buffer * tx_info,void ** push_hdr,uint16_t * header_len,uint16_t * packet_len)549 ena_netmap_tx_map_slots(struct ena_netmap_ctx *ctx,
550     struct ena_tx_buffer *tx_info, void **push_hdr, uint16_t *header_len,
551     uint16_t *packet_len)
552 {
553 	struct netmap_slot *slot;
554 	struct ena_com_buf *ena_buf;
555 	struct ena_adapter *adapter;
556 	struct ena_ring *tx_ring;
557 	struct ena_netmap_tx_info *nm_info;
558 	bus_dmamap_t *nm_maps;
559 	void *vaddr;
560 	uint64_t paddr;
561 	uint32_t *nm_buf_idx;
562 	uint32_t slot_head_len;
563 	uint32_t frag_len;
564 	uint32_t remaining_len;
565 	uint16_t push_len;
566 	uint16_t delta;
567 	int rc;
568 
569 	adapter = ctx->adapter;
570 	tx_ring = ctx->ring;
571 	ena_buf = tx_info->bufs;
572 	nm_info = &tx_info->nm_info;
573 	nm_maps = nm_info->map_seg;
574 	nm_buf_idx = nm_info->socket_buf_idx;
575 	slot = &ctx->slots[ctx->nm_i];
576 
577 	slot_head_len = slot->len;
578 	*packet_len = ena_netmap_packet_len(ctx->slots, ctx->nm_i, ctx->lim);
579 	remaining_len = *packet_len;
580 	delta = 0;
581 
582 	__builtin_prefetch(&ctx->slots[nm_next(ctx->nm_i, ctx->lim)]);
583 	if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) {
584 		/*
585 		 * When the device is in LLQ mode, the driver will copy
586 		 * the header into the device memory space.
587 		 * The ena_com layer assumes that the header is in a linear
588 		 * memory space.
589 		 * This assumption might be wrong since part of the header
590 		 * can be in the fragmented buffers.
591 		 * First, check if header fits in the first slot. If not, copy
592 		 * it to separate buffer that will be holding linearized data.
593 		 */
594 		push_len = min_t(uint32_t, *packet_len,
595 		    tx_ring->tx_max_header_size);
596 		*header_len = push_len;
597 		/* If header is in linear space, just point to socket's data. */
598 		if (likely(push_len <= slot_head_len)) {
599 			*push_hdr = NMB(ctx->na, slot);
600 			if (unlikely(push_hdr == NULL)) {
601 				ena_log_nm(adapter->pdev, ERR,
602 				    "Slot vaddress is NULL\n");
603 				return (EINVAL);
604 			}
605 		/*
606 		 * Otherwise, copy whole portion of header from multiple
607 		 * slots to intermediate buffer.
608 		 */
609 		} else {
610 			rc = ena_netmap_copy_data(ctx->na, ctx->slots,
611 			    ctx->nm_i, ctx->lim, push_len,
612 			    tx_ring->push_buf_intermediate_buf);
613 			if (unlikely(rc)) {
614 				ena_log_nm(adapter->pdev, ERR,
615 				    "Failed to copy data from slots to push_buf\n");
616 				return (EINVAL);
617 			}
618 
619 			*push_hdr = tx_ring->push_buf_intermediate_buf;
620 			counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1);
621 
622 			delta = push_len - slot_head_len;
623 		}
624 
625 		ena_log_nm(adapter->pdev, DBG,
626 		    "slot: %d header_buf->vaddr: %p push_len: %d\n",
627 		    slot->buf_idx, *push_hdr, push_len);
628 
629 		/*
630 		 * If header was in linear memory space, map for the dma rest of
631 		 * the data in the first mbuf of the mbuf chain.
632 		 */
633 		if (slot_head_len > push_len) {
634 			rc = ena_netmap_map_single_slot(ctx->na, slot,
635 			    adapter->tx_buf_tag, *nm_maps, &vaddr, &paddr);
636 			if (unlikely(rc != 0)) {
637 				ena_log_nm(adapter->pdev, ERR,
638 				    "DMA mapping error\n");
639 				return (rc);
640 			}
641 			nm_maps++;
642 
643 			ena_buf->paddr = paddr + push_len;
644 			ena_buf->len = slot->len - push_len;
645 			ena_buf++;
646 
647 			tx_info->num_of_bufs++;
648 		}
649 
650 		remaining_len -= slot->len;
651 
652 		/* Save buf idx before advancing */
653 		*nm_buf_idx = slot->buf_idx;
654 		nm_buf_idx++;
655 		slot->buf_idx = 0;
656 
657 		/* Advance to the next socket */
658 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
659 		slot = &ctx->slots[ctx->nm_i];
660 		nm_info->sockets_used++;
661 
662 		/*
663 		 * If header is in non linear space (delta > 0), then skip mbufs
664 		 * containing header and map the last one containing both header
665 		 * and the packet data.
666 		 * The first segment is already counted in.
667 		 */
668 		while (delta > 0) {
669 			__builtin_prefetch(&ctx->slots[nm_next(ctx->nm_i, ctx->lim)]);
670 			frag_len = slot->len;
671 
672 			/*
673 			 * If whole segment contains header just move to the
674 			 * next one and reduce delta.
675 			 */
676 			if (unlikely(delta >= frag_len)) {
677 				delta -= frag_len;
678 			} else {
679 				/*
680 				 * Map the data and then assign it with the
681 				 * offsets
682 				 */
683 				rc = ena_netmap_map_single_slot(ctx->na, slot,
684 				    adapter->tx_buf_tag, *nm_maps, &vaddr,
685 				    &paddr);
686 				if (unlikely(rc != 0)) {
687 					ena_log_nm(adapter->pdev, ERR,
688 					    "DMA mapping error\n");
689 					goto error_map;
690 				}
691 				nm_maps++;
692 
693 				ena_buf->paddr = paddr + delta;
694 				ena_buf->len = slot->len - delta;
695 				ena_buf++;
696 
697 				tx_info->num_of_bufs++;
698 				delta = 0;
699 			}
700 
701 			remaining_len -= slot->len;
702 
703 			/* Save buf idx before advancing */
704 			*nm_buf_idx = slot->buf_idx;
705 			nm_buf_idx++;
706 			slot->buf_idx = 0;
707 
708 			/* Advance to the next socket */
709 			ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
710 			slot = &ctx->slots[ctx->nm_i];
711 			nm_info->sockets_used++;
712 		}
713 	} else {
714 		*push_hdr = NULL;
715 		/*
716 		 * header_len is just a hint for the device. Because netmap is
717 		 * not giving us any information about packet header length and
718 		 * it is not guaranteed that all packet headers will be in the
719 		 * 1st slot, setting header_len to 0 is making the device ignore
720 		 * this value and resolve header on it's own.
721 		 */
722 		*header_len = 0;
723 	}
724 
725 	/* Map all remaining data (regular routine for non-LLQ mode) */
726 	while (remaining_len > 0) {
727 		__builtin_prefetch(&ctx->slots[nm_next(ctx->nm_i, ctx->lim)]);
728 
729 		rc = ena_netmap_map_single_slot(ctx->na, slot,
730 		    adapter->tx_buf_tag, *nm_maps, &vaddr, &paddr);
731 		if (unlikely(rc != 0)) {
732 			ena_log_nm(adapter->pdev, ERR, "DMA mapping error\n");
733 			goto error_map;
734 		}
735 		nm_maps++;
736 
737 		ena_buf->paddr = paddr;
738 		ena_buf->len = slot->len;
739 		ena_buf++;
740 
741 		tx_info->num_of_bufs++;
742 
743 		remaining_len -= slot->len;
744 
745 		/* Save buf idx before advancing */
746 		*nm_buf_idx = slot->buf_idx;
747 		nm_buf_idx++;
748 		slot->buf_idx = 0;
749 
750 		/* Advance to the next socket */
751 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
752 		slot = &ctx->slots[ctx->nm_i];
753 		nm_info->sockets_used++;
754 	}
755 
756 	return (0);
757 
758 error_map:
759 	ena_netmap_unmap_last_socket_chain(ctx, tx_info);
760 
761 	return (rc);
762 }
763 
764 static void
ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx * ctx,struct ena_tx_buffer * tx_info)765 ena_netmap_unmap_last_socket_chain(struct ena_netmap_ctx *ctx,
766     struct ena_tx_buffer *tx_info)
767 {
768 	struct ena_netmap_tx_info *nm_info;
769 	int n;
770 
771 	nm_info = &tx_info->nm_info;
772 
773 	/**
774 	 * As the used sockets must not be equal to the buffers used in the LLQ
775 	 * mode, they must be treated separately.
776 	 * First, unmap the DMA maps.
777 	 */
778 	n = tx_info->num_of_bufs;
779 	while (n--) {
780 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
781 		    nm_info->map_seg[n]);
782 	}
783 	tx_info->num_of_bufs = 0;
784 
785 	/* Next, retain the sockets back to the userspace */
786 	n = nm_info->sockets_used;
787 	while (n--) {
788 		ctx->nm_i = nm_prev(ctx->nm_i, ctx->lim);
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 	}
793 	nm_info->sockets_used = 0;
794 }
795 
796 static void
ena_netmap_tx_cleanup(struct ena_netmap_ctx * ctx)797 ena_netmap_tx_cleanup(struct ena_netmap_ctx *ctx)
798 {
799 	struct ena_ring *tx_ring = ctx->ring;
800 	int rc;
801 	uint16_t req_id;
802 	uint16_t total_tx_descs = 0;
803 
804 	ctx->nm_i = ctx->kring->nr_hwtail;
805 	ctx->nt = tx_ring->next_to_clean;
806 
807 	/* Reclaim buffers for completed transmissions */
808 	do {
809 		rc = ena_com_tx_comp_req_id_get(ctx->io_cq, &req_id);
810 		if(unlikely(rc == ENA_COM_TRY_AGAIN))
811 			break;
812 
813 		rc = validate_tx_req_id(tx_ring, req_id, rc);
814 		if(unlikely(rc != 0))
815 			break;
816 
817 		total_tx_descs += ena_netmap_tx_clean_one(ctx, req_id);
818 	} while (1);
819 
820 	ctx->kring->nr_hwtail = ctx->nm_i;
821 
822 	if (total_tx_descs > 0) {
823 		/* acknowledge completion of sent packets */
824 		tx_ring->next_to_clean = ctx->nt;
825 		ena_com_comp_ack(tx_ring->ena_com_io_sq, total_tx_descs);
826 	}
827 }
828 
829 static uint16_t
ena_netmap_tx_clean_one(struct ena_netmap_ctx * ctx,uint16_t req_id)830 ena_netmap_tx_clean_one(struct ena_netmap_ctx *ctx, uint16_t req_id)
831 {
832 	struct ena_tx_buffer *tx_info;
833 	struct ena_netmap_tx_info *nm_info;
834 	int n;
835 
836 	tx_info = &ctx->ring->tx_buffer_info[req_id];
837 	nm_info = &tx_info->nm_info;
838 
839 	/**
840 	 * As the used sockets must not be equal to the buffers used in the LLQ
841 	 * mode, they must be treated separately.
842 	 * First, unmap the DMA maps.
843 	 */
844 	n = tx_info->num_of_bufs;
845 	for (n = 0; n < tx_info->num_of_bufs; n++) {
846 		netmap_unload_map(ctx->na, ctx->adapter->tx_buf_tag,
847 		    nm_info->map_seg[n]);
848 	}
849 	tx_info->num_of_bufs = 0;
850 
851 	/* Next, retain the sockets back to the userspace */
852 	for (n = 0; n < nm_info->sockets_used; n++) {
853 		ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
854 		ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0,
855 		    ctx->adapter->ena_dev, "Tx idx is not 0.\n");
856 		ctx->slots[ctx->nm_i].buf_idx = nm_info->socket_buf_idx[n];
857 		ctx->slots[ctx->nm_i].flags = NS_BUF_CHANGED;
858 		nm_info->socket_buf_idx[n] = 0;
859 	}
860 	nm_info->sockets_used = 0;
861 
862 	ctx->ring->free_tx_ids[ctx->nt] = req_id;
863 	ctx->nt = ENA_TX_RING_IDX_NEXT(ctx->nt, ctx->lim);
864 
865 	return tx_info->tx_descs;
866 }
867 
868 static int
ena_netmap_rxsync(struct netmap_kring * kring,int flags)869 ena_netmap_rxsync(struct netmap_kring *kring, int flags)
870 {
871 	struct ena_netmap_ctx ctx;
872 	int rc;
873 
874 	ena_netmap_fill_ctx(kring, &ctx, ENA_IO_RXQ_IDX(kring->ring_id));
875 	ctx.ring = &ctx.adapter->rx_ring[kring->ring_id];
876 
877 	if (ctx.kring->rhead > ctx.lim) {
878 		/* Probably not needed to release slots from RX ring. */
879 		return (netmap_ring_reinit(ctx.kring));
880 	}
881 
882 	if (unlikely((if_getdrvflags(ctx.na->ifp) & IFF_DRV_RUNNING) == 0))
883 		return (0);
884 
885 	if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, ctx.adapter)))
886 		return (0);
887 
888 	if ((rc = ena_netmap_rx_frames(&ctx)) != 0)
889 		return (rc);
890 
891 	ena_netmap_rx_cleanup(&ctx);
892 
893 	return (0);
894 }
895 
896 static inline int
ena_netmap_rx_frames(struct ena_netmap_ctx * ctx)897 ena_netmap_rx_frames(struct ena_netmap_ctx *ctx)
898 {
899 	int rc = 0;
900 	int frames_counter = 0;
901 
902 	ctx->nt = ctx->ring->next_to_clean;
903 	ctx->nm_i = ctx->kring->nr_hwtail;
904 
905 	while ((rc = ena_netmap_rx_frame(ctx)) == ENA_NETMAP_MORE_FRAMES) {
906 		frames_counter++;
907 		/* In case of multiple frames, it is not an error. */
908 		rc = 0;
909 		if (frames_counter > ENA_MAX_FRAMES) {
910 			ena_log_nm(ctx->adapter->pdev, ERR,
911 			    "Driver is stuck in the Rx loop\n");
912 			break;
913 		}
914 	};
915 
916 	ctx->kring->nr_hwtail = ctx->nm_i;
917 	ctx->kring->nr_kflags &= ~NKR_PENDINTR;
918 	ctx->ring->next_to_clean = ctx->nt;
919 
920 	return (rc);
921 }
922 
923 static inline int
ena_netmap_rx_frame(struct ena_netmap_ctx * ctx)924 ena_netmap_rx_frame(struct ena_netmap_ctx *ctx)
925 {
926 	struct ena_com_rx_ctx ena_rx_ctx;
927 	enum ena_regs_reset_reason_types reset_reason;
928 	int rc, len = 0;
929 	uint16_t buf, nm;
930 
931 	ena_rx_ctx.ena_bufs = ctx->ring->ena_bufs;
932 	ena_rx_ctx.max_bufs = ctx->adapter->max_rx_sgl_size;
933 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
934 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD);
935 
936 	rc = ena_com_rx_pkt(ctx->io_cq, ctx->io_sq, &ena_rx_ctx);
937 	if (unlikely(rc != 0)) {
938 		ena_log_nm(ctx->adapter->pdev, ERR,
939 		    "Failed to read pkt from the device with error: %d\n", rc);
940 		if (rc == ENA_COM_NO_SPACE) {
941 			counter_u64_add(ctx->ring->rx_stats.bad_desc_num, 1);
942 			reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS;
943 		} else if (rc == ENA_COM_FAULT) {
944 			reset_reason = ENA_REGS_RESET_RX_DESCRIPTOR_MALFORMED;
945 		} else {
946 			counter_u64_add(ctx->ring->rx_stats.bad_req_id, 1);
947 			reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID;
948 		}
949 		ena_trigger_reset(ctx->adapter, reset_reason);
950 		return (rc);
951 	}
952 	if (unlikely(ena_rx_ctx.descs == 0))
953 		return (ENA_NETMAP_NO_MORE_FRAMES);
954 
955 	ena_log_nm(ctx->adapter->pdev, DBG,
956 	    "Rx: q %d got packet from ena. descs #:"
957 	    " %d l3 proto %d l4 proto %d hash: %x\n",
958 	    ctx->ring->qid, ena_rx_ctx.descs, ena_rx_ctx.l3_proto,
959 	    ena_rx_ctx.l4_proto, ena_rx_ctx.hash);
960 
961 	for (buf = 0; buf < ena_rx_ctx.descs; buf++)
962 		if ((rc = ena_netmap_rx_load_desc(ctx, buf, &len)) != 0)
963 			break;
964 	/*
965 	 * ena_netmap_rx_load_desc doesn't know the number of descriptors.
966 	 * It just set flag NS_MOREFRAG to all slots, then here flag of
967 	 * last slot is cleared.
968 	 */
969 	ctx->slots[nm_prev(ctx->nm_i, ctx->lim)].flags &= ~NS_MOREFRAG;
970 
971 	if (rc != 0) {
972 		goto rx_clear_desc;
973 	}
974 
975 	bus_dmamap_sync(ctx->io_cq->cdesc_addr.mem_handle.tag,
976 	    ctx->io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD);
977 
978 	counter_enter();
979 	counter_u64_add_protected(ctx->ring->rx_stats.bytes, len);
980 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_bytes, len);
981 	counter_u64_add_protected(ctx->ring->rx_stats.cnt, 1);
982 	counter_u64_add_protected(ctx->adapter->hw_stats.rx_packets, 1);
983 	counter_exit();
984 
985 	return (ENA_NETMAP_MORE_FRAMES);
986 
987 rx_clear_desc:
988 	nm = ctx->nm_i;
989 
990 	/* Remove failed packet from ring */
991 	while (buf--) {
992 		ctx->slots[nm].flags = 0;
993 		ctx->slots[nm].len = 0;
994 		nm = nm_prev(nm, ctx->lim);
995 	}
996 
997 	return (rc);
998 }
999 
1000 static inline int
ena_netmap_rx_load_desc(struct ena_netmap_ctx * ctx,uint16_t buf,int * len)1001 ena_netmap_rx_load_desc(struct ena_netmap_ctx *ctx, uint16_t buf, int *len)
1002 {
1003 	struct ena_rx_buffer *rx_info;
1004 	uint16_t req_id;
1005 
1006 	req_id = ctx->ring->ena_bufs[buf].req_id;
1007 	rx_info = &ctx->ring->rx_buffer_info[req_id];
1008 	bus_dmamap_sync(ctx->adapter->rx_buf_tag, rx_info->map,
1009 	    BUS_DMASYNC_POSTREAD);
1010 	netmap_unload_map(ctx->na, ctx->adapter->rx_buf_tag, rx_info->map);
1011 
1012 	ENA_WARN(ctx->slots[ctx->nm_i].buf_idx != 0, ctx->adapter->ena_dev,
1013 	    "Rx idx is not 0.\n");
1014 
1015 	ctx->slots[ctx->nm_i].buf_idx = rx_info->netmap_buf_idx;
1016 	rx_info->netmap_buf_idx = 0;
1017 	/*
1018 	 * Set NS_MOREFRAG to all slots.
1019 	 * Then ena_netmap_rx_frame clears it from last one.
1020 	 */
1021 	ctx->slots[ctx->nm_i].flags |= NS_MOREFRAG | NS_BUF_CHANGED;
1022 	ctx->slots[ctx->nm_i].len = ctx->ring->ena_bufs[buf].len;
1023 	*len += ctx->slots[ctx->nm_i].len;
1024 	ctx->ring->free_rx_ids[ctx->nt] = req_id;
1025 	ena_log_nm(ctx->adapter->pdev, DBG,
1026 	    "rx_info %p, buf_idx %d, paddr %jx, nm: %d\n", rx_info,
1027 	    ctx->slots[ctx->nm_i].buf_idx, (uintmax_t)rx_info->ena_buf.paddr,
1028 	    ctx->nm_i);
1029 
1030 	ctx->nm_i = nm_next(ctx->nm_i, ctx->lim);
1031 	ctx->nt = ENA_RX_RING_IDX_NEXT(ctx->nt, ctx->ring->ring_size);
1032 
1033 	return (0);
1034 }
1035 
1036 static inline void
ena_netmap_rx_cleanup(struct ena_netmap_ctx * ctx)1037 ena_netmap_rx_cleanup(struct ena_netmap_ctx *ctx)
1038 {
1039 	int refill_required;
1040 
1041 	refill_required = ctx->kring->rhead - ctx->kring->nr_hwcur;
1042 	if (ctx->kring->nr_hwcur != ctx->kring->nr_hwtail)
1043 		refill_required -= 1;
1044 
1045 	if (refill_required == 0)
1046 		return;
1047 	else if (refill_required < 0)
1048 		refill_required += ctx->kring->nkr_num_slots;
1049 
1050 	ena_refill_rx_bufs(ctx->ring, refill_required);
1051 }
1052 
1053 static inline void
ena_netmap_fill_ctx(struct netmap_kring * kring,struct ena_netmap_ctx * ctx,uint16_t ena_qid)1054 ena_netmap_fill_ctx(struct netmap_kring *kring, struct ena_netmap_ctx *ctx,
1055     uint16_t ena_qid)
1056 {
1057 	ctx->kring = kring;
1058 	ctx->na = kring->na;
1059 	ctx->adapter = if_getsoftc(ctx->na->ifp);
1060 	ctx->lim = kring->nkr_num_slots - 1;
1061 	ctx->io_cq = &ctx->adapter->ena_dev->io_cq_queues[ena_qid];
1062 	ctx->io_sq = &ctx->adapter->ena_dev->io_sq_queues[ena_qid];
1063 	ctx->slots = kring->ring->slot;
1064 }
1065 
1066 void
ena_netmap_unload(struct ena_adapter * adapter,bus_dmamap_t map)1067 ena_netmap_unload(struct ena_adapter *adapter, bus_dmamap_t map)
1068 {
1069 	struct netmap_adapter *na = NA(adapter->ifp);
1070 
1071 	netmap_unload_map(na, adapter->tx_buf_tag, map);
1072 }
1073 
1074 #endif /* DEV_NETMAP */
1075