1 // SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause)
2 /*
3  * Copyright(c) 2020 Intel Corporation.
4  *
5  */
6 
7 /*
8  * This file contains HFI1 support for IPOIB SDMA functionality
9  */
10 
11 #include <linux/log2.h>
12 #include <linux/circ_buf.h>
13 
14 #include "sdma.h"
15 #include "verbs.h"
16 #include "trace_ibhdrs.h"
17 #include "ipoib.h"
18 #include "trace_tx.h"
19 
20 /* Add a convenience helper */
21 #define CIRC_ADD(val, add, size) (((val) + (add)) & ((size) - 1))
22 #define CIRC_NEXT(val, size) CIRC_ADD(val, 1, size)
23 #define CIRC_PREV(val, size) CIRC_ADD(val, -1, size)
24 
25 struct ipoib_txparms {
26 	struct hfi1_devdata        *dd;
27 	struct rdma_ah_attr        *ah_attr;
28 	struct hfi1_ibport         *ibp;
29 	struct hfi1_ipoib_txq      *txq;
30 	union hfi1_ipoib_flow       flow;
31 	u32                         dqpn;
32 	u8                          hdr_dwords;
33 	u8                          entropy;
34 };
35 
36 static struct ipoib_txreq *
37 hfi1_txreq_from_idx(struct hfi1_ipoib_circ_buf *r, u32 idx)
38 {
39 	return (struct ipoib_txreq *)(r->items + (idx << r->shift));
40 }
41 
42 static u32 hfi1_ipoib_txreqs(const u64 sent, const u64 completed)
43 {
44 	return sent - completed;
45 }
46 
47 static u64 hfi1_ipoib_used(struct hfi1_ipoib_txq *txq)
48 {
49 	return hfi1_ipoib_txreqs(txq->tx_ring.sent_txreqs,
50 				 txq->tx_ring.complete_txreqs);
51 }
52 
53 static void hfi1_ipoib_stop_txq(struct hfi1_ipoib_txq *txq)
54 {
55 	trace_hfi1_txq_stop(txq);
56 	if (atomic_inc_return(&txq->tx_ring.stops) == 1)
57 		netif_stop_subqueue(txq->priv->netdev, txq->q_idx);
58 }
59 
60 static void hfi1_ipoib_wake_txq(struct hfi1_ipoib_txq *txq)
61 {
62 	trace_hfi1_txq_wake(txq);
63 	if (atomic_dec_and_test(&txq->tx_ring.stops))
64 		netif_wake_subqueue(txq->priv->netdev, txq->q_idx);
65 }
66 
67 static uint hfi1_ipoib_ring_hwat(struct hfi1_ipoib_txq *txq)
68 {
69 	return min_t(uint, txq->priv->netdev->tx_queue_len,
70 		     txq->tx_ring.max_items - 1);
71 }
72 
73 static uint hfi1_ipoib_ring_lwat(struct hfi1_ipoib_txq *txq)
74 {
75 	return min_t(uint, txq->priv->netdev->tx_queue_len,
76 		     txq->tx_ring.max_items) >> 1;
77 }
78 
79 static void hfi1_ipoib_check_queue_depth(struct hfi1_ipoib_txq *txq)
80 {
81 	++txq->tx_ring.sent_txreqs;
82 	if (hfi1_ipoib_used(txq) >= hfi1_ipoib_ring_hwat(txq) &&
83 	    !atomic_xchg(&txq->tx_ring.ring_full, 1)) {
84 		trace_hfi1_txq_full(txq);
85 		hfi1_ipoib_stop_txq(txq);
86 	}
87 }
88 
89 static void hfi1_ipoib_check_queue_stopped(struct hfi1_ipoib_txq *txq)
90 {
91 	struct net_device *dev = txq->priv->netdev;
92 
93 	/* If shutting down just return as queue state is irrelevant */
94 	if (unlikely(dev->reg_state != NETREG_REGISTERED))
95 		return;
96 
97 	/*
98 	 * When the queue has been drained to less than half full it will be
99 	 * restarted.
100 	 * The size of the txreq ring is fixed at initialization.
101 	 * The tx queue len can be adjusted upward while the interface is
102 	 * running.
103 	 * The tx queue len can be large enough to overflow the txreq_ring.
104 	 * Use the minimum of the current tx_queue_len or the rings max txreqs
105 	 * to protect against ring overflow.
106 	 */
107 	if (hfi1_ipoib_used(txq) < hfi1_ipoib_ring_lwat(txq) &&
108 	    atomic_xchg(&txq->tx_ring.ring_full, 0)) {
109 		trace_hfi1_txq_xmit_unstopped(txq);
110 		hfi1_ipoib_wake_txq(txq);
111 	}
112 }
113 
114 static void hfi1_ipoib_free_tx(struct ipoib_txreq *tx, int budget)
115 {
116 	struct hfi1_ipoib_dev_priv *priv = tx->txq->priv;
117 
118 	if (likely(!tx->sdma_status)) {
119 		dev_sw_netstats_tx_add(priv->netdev, 1, tx->skb->len);
120 	} else {
121 		++priv->netdev->stats.tx_errors;
122 		dd_dev_warn(priv->dd,
123 			    "%s: Status = 0x%x pbc 0x%llx txq = %d sde = %d\n",
124 			    __func__, tx->sdma_status,
125 			    le64_to_cpu(tx->sdma_hdr->pbc), tx->txq->q_idx,
126 			    tx->txq->sde->this_idx);
127 	}
128 
129 	napi_consume_skb(tx->skb, budget);
130 	tx->skb = NULL;
131 	sdma_txclean(priv->dd, &tx->txreq);
132 }
133 
134 static void hfi1_ipoib_drain_tx_ring(struct hfi1_ipoib_txq *txq)
135 {
136 	struct hfi1_ipoib_circ_buf *tx_ring = &txq->tx_ring;
137 	int i;
138 	struct ipoib_txreq *tx;
139 
140 	for (i = 0; i < tx_ring->max_items; i++) {
141 		tx = hfi1_txreq_from_idx(tx_ring, i);
142 		tx->complete = 0;
143 		dev_kfree_skb_any(tx->skb);
144 		tx->skb = NULL;
145 		sdma_txclean(txq->priv->dd, &tx->txreq);
146 	}
147 	tx_ring->head = 0;
148 	tx_ring->tail = 0;
149 	tx_ring->complete_txreqs = 0;
150 	tx_ring->sent_txreqs = 0;
151 	tx_ring->avail = hfi1_ipoib_ring_hwat(txq);
152 }
153 
154 static int hfi1_ipoib_poll_tx_ring(struct napi_struct *napi, int budget)
155 {
156 	struct hfi1_ipoib_txq *txq =
157 		container_of(napi, struct hfi1_ipoib_txq, napi);
158 	struct hfi1_ipoib_circ_buf *tx_ring = &txq->tx_ring;
159 	u32 head = tx_ring->head;
160 	u32 max_tx = tx_ring->max_items;
161 	int work_done;
162 	struct ipoib_txreq *tx =  hfi1_txreq_from_idx(tx_ring, head);
163 
164 	trace_hfi1_txq_poll(txq);
165 	for (work_done = 0; work_done < budget; work_done++) {
166 		/* See hfi1_ipoib_sdma_complete() */
167 		if (!smp_load_acquire(&tx->complete))
168 			break;
169 		tx->complete = 0;
170 		trace_hfi1_tx_produce(tx, head);
171 		hfi1_ipoib_free_tx(tx, budget);
172 		head = CIRC_NEXT(head, max_tx);
173 		tx =  hfi1_txreq_from_idx(tx_ring, head);
174 	}
175 	tx_ring->complete_txreqs += work_done;
176 
177 	/* Finished freeing tx items so store the head value. */
178 	smp_store_release(&tx_ring->head, head);
179 
180 	hfi1_ipoib_check_queue_stopped(txq);
181 
182 	if (work_done < budget)
183 		napi_complete_done(napi, work_done);
184 
185 	return work_done;
186 }
187 
188 static void hfi1_ipoib_sdma_complete(struct sdma_txreq *txreq, int status)
189 {
190 	struct ipoib_txreq *tx = container_of(txreq, struct ipoib_txreq, txreq);
191 
192 	trace_hfi1_txq_complete(tx->txq);
193 	tx->sdma_status = status;
194 	/* see hfi1_ipoib_poll_tx_ring */
195 	smp_store_release(&tx->complete, 1);
196 	napi_schedule_irqoff(&tx->txq->napi);
197 }
198 
199 static int hfi1_ipoib_build_ulp_payload(struct ipoib_txreq *tx,
200 					struct ipoib_txparms *txp)
201 {
202 	struct hfi1_devdata *dd = txp->dd;
203 	struct sdma_txreq *txreq = &tx->txreq;
204 	struct sk_buff *skb = tx->skb;
205 	int ret = 0;
206 	int i;
207 
208 	if (skb_headlen(skb)) {
209 		ret = sdma_txadd_kvaddr(dd, txreq, skb->data, skb_headlen(skb));
210 		if (unlikely(ret))
211 			return ret;
212 	}
213 
214 	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
215 		const skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
216 
217 		ret = sdma_txadd_page(dd,
218 				      txreq,
219 				      skb_frag_page(frag),
220 				      frag->bv_offset,
221 				      skb_frag_size(frag));
222 		if (unlikely(ret))
223 			break;
224 	}
225 
226 	return ret;
227 }
228 
229 static int hfi1_ipoib_build_tx_desc(struct ipoib_txreq *tx,
230 				    struct ipoib_txparms *txp)
231 {
232 	struct hfi1_devdata *dd = txp->dd;
233 	struct sdma_txreq *txreq = &tx->txreq;
234 	struct hfi1_sdma_header *sdma_hdr = tx->sdma_hdr;
235 	u16 pkt_bytes =
236 		sizeof(sdma_hdr->pbc) + (txp->hdr_dwords << 2) + tx->skb->len;
237 	int ret;
238 
239 	ret = sdma_txinit(txreq, 0, pkt_bytes, hfi1_ipoib_sdma_complete);
240 	if (unlikely(ret))
241 		return ret;
242 
243 	/* add pbc + headers */
244 	ret = sdma_txadd_kvaddr(dd,
245 				txreq,
246 				sdma_hdr,
247 				sizeof(sdma_hdr->pbc) + (txp->hdr_dwords << 2));
248 	if (unlikely(ret))
249 		return ret;
250 
251 	/* add the ulp payload */
252 	return hfi1_ipoib_build_ulp_payload(tx, txp);
253 }
254 
255 static void hfi1_ipoib_build_ib_tx_headers(struct ipoib_txreq *tx,
256 					   struct ipoib_txparms *txp)
257 {
258 	struct hfi1_ipoib_dev_priv *priv = tx->txq->priv;
259 	struct hfi1_sdma_header *sdma_hdr = tx->sdma_hdr;
260 	struct sk_buff *skb = tx->skb;
261 	struct hfi1_pportdata *ppd = ppd_from_ibp(txp->ibp);
262 	struct rdma_ah_attr *ah_attr = txp->ah_attr;
263 	struct ib_other_headers *ohdr;
264 	struct ib_grh *grh;
265 	u16 dwords;
266 	u16 slid;
267 	u16 dlid;
268 	u16 lrh0;
269 	u32 bth0;
270 	u32 sqpn = (u32)(priv->netdev->dev_addr[1] << 16 |
271 			 priv->netdev->dev_addr[2] << 8 |
272 			 priv->netdev->dev_addr[3]);
273 	u16 payload_dwords;
274 	u8 pad_cnt;
275 
276 	pad_cnt = -skb->len & 3;
277 
278 	/* Includes ICRC */
279 	payload_dwords = ((skb->len + pad_cnt) >> 2) + SIZE_OF_CRC;
280 
281 	/* header size in dwords LRH+BTH+DETH = (8+12+8)/4. */
282 	txp->hdr_dwords = 7;
283 
284 	if (rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH) {
285 		grh = &sdma_hdr->hdr.ibh.u.l.grh;
286 		txp->hdr_dwords +=
287 			hfi1_make_grh(txp->ibp,
288 				      grh,
289 				      rdma_ah_read_grh(ah_attr),
290 				      txp->hdr_dwords - LRH_9B_DWORDS,
291 				      payload_dwords);
292 		lrh0 = HFI1_LRH_GRH;
293 		ohdr = &sdma_hdr->hdr.ibh.u.l.oth;
294 	} else {
295 		lrh0 = HFI1_LRH_BTH;
296 		ohdr = &sdma_hdr->hdr.ibh.u.oth;
297 	}
298 
299 	lrh0 |= (rdma_ah_get_sl(ah_attr) & 0xf) << 4;
300 	lrh0 |= (txp->flow.sc5 & 0xf) << 12;
301 
302 	dlid = opa_get_lid(rdma_ah_get_dlid(ah_attr), 9B);
303 	if (dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
304 		slid = be16_to_cpu(IB_LID_PERMISSIVE);
305 	} else {
306 		u16 lid = (u16)ppd->lid;
307 
308 		if (lid) {
309 			lid |= rdma_ah_get_path_bits(ah_attr) &
310 				((1 << ppd->lmc) - 1);
311 			slid = lid;
312 		} else {
313 			slid = be16_to_cpu(IB_LID_PERMISSIVE);
314 		}
315 	}
316 
317 	/* Includes ICRC */
318 	dwords = txp->hdr_dwords + payload_dwords;
319 
320 	/* Build the lrh */
321 	sdma_hdr->hdr.hdr_type = HFI1_PKT_TYPE_9B;
322 	hfi1_make_ib_hdr(&sdma_hdr->hdr.ibh, lrh0, dwords, dlid, slid);
323 
324 	/* Build the bth */
325 	bth0 = (IB_OPCODE_UD_SEND_ONLY << 24) | (pad_cnt << 20) | priv->pkey;
326 
327 	ohdr->bth[0] = cpu_to_be32(bth0);
328 	ohdr->bth[1] = cpu_to_be32(txp->dqpn);
329 	ohdr->bth[2] = cpu_to_be32(mask_psn((u32)txp->txq->tx_ring.sent_txreqs));
330 
331 	/* Build the deth */
332 	ohdr->u.ud.deth[0] = cpu_to_be32(priv->qkey);
333 	ohdr->u.ud.deth[1] = cpu_to_be32((txp->entropy <<
334 					  HFI1_IPOIB_ENTROPY_SHIFT) | sqpn);
335 
336 	/* Construct the pbc. */
337 	sdma_hdr->pbc =
338 		cpu_to_le64(create_pbc(ppd,
339 				       ib_is_sc5(txp->flow.sc5) <<
340 							      PBC_DC_INFO_SHIFT,
341 				       0,
342 				       sc_to_vlt(priv->dd, txp->flow.sc5),
343 				       dwords - SIZE_OF_CRC +
344 						(sizeof(sdma_hdr->pbc) >> 2)));
345 }
346 
347 static struct ipoib_txreq *hfi1_ipoib_send_dma_common(struct net_device *dev,
348 						      struct sk_buff *skb,
349 						      struct ipoib_txparms *txp)
350 {
351 	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
352 	struct hfi1_ipoib_txq *txq = txp->txq;
353 	struct ipoib_txreq *tx;
354 	struct hfi1_ipoib_circ_buf *tx_ring = &txq->tx_ring;
355 	u32 tail = tx_ring->tail;
356 	int ret;
357 
358 	if (unlikely(!tx_ring->avail)) {
359 		u32 head;
360 
361 		if (hfi1_ipoib_used(txq) >= hfi1_ipoib_ring_hwat(txq))
362 			/* This shouldn't happen with a stopped queue */
363 			return ERR_PTR(-ENOMEM);
364 		/* See hfi1_ipoib_poll_tx_ring() */
365 		head = smp_load_acquire(&tx_ring->head);
366 		tx_ring->avail =
367 			min_t(u32, hfi1_ipoib_ring_hwat(txq),
368 			      CIRC_CNT(head, tail, tx_ring->max_items));
369 	} else {
370 		tx_ring->avail--;
371 	}
372 	tx = hfi1_txreq_from_idx(tx_ring, tail);
373 	trace_hfi1_txq_alloc_tx(txq);
374 
375 	/* so that we can test if the sdma descriptors are there */
376 	tx->txreq.num_desc = 0;
377 	tx->txq = txq;
378 	tx->skb = skb;
379 	INIT_LIST_HEAD(&tx->txreq.list);
380 
381 	hfi1_ipoib_build_ib_tx_headers(tx, txp);
382 
383 	ret = hfi1_ipoib_build_tx_desc(tx, txp);
384 	if (likely(!ret)) {
385 		if (txq->flow.as_int != txp->flow.as_int) {
386 			txq->flow.tx_queue = txp->flow.tx_queue;
387 			txq->flow.sc5 = txp->flow.sc5;
388 			txq->sde =
389 				sdma_select_engine_sc(priv->dd,
390 						      txp->flow.tx_queue,
391 						      txp->flow.sc5);
392 			trace_hfi1_flow_switch(txq);
393 		}
394 
395 		return tx;
396 	}
397 
398 	sdma_txclean(priv->dd, &tx->txreq);
399 
400 	return ERR_PTR(ret);
401 }
402 
403 static int hfi1_ipoib_submit_tx_list(struct net_device *dev,
404 				     struct hfi1_ipoib_txq *txq)
405 {
406 	int ret;
407 	u16 count_out;
408 
409 	ret = sdma_send_txlist(txq->sde,
410 			       iowait_get_ib_work(&txq->wait),
411 			       &txq->tx_list,
412 			       &count_out);
413 	if (likely(!ret) || ret == -EBUSY || ret == -ECOMM)
414 		return ret;
415 
416 	dd_dev_warn(txq->priv->dd, "cannot send skb tx list, err %d.\n", ret);
417 
418 	return ret;
419 }
420 
421 static int hfi1_ipoib_flush_tx_list(struct net_device *dev,
422 				    struct hfi1_ipoib_txq *txq)
423 {
424 	int ret = 0;
425 
426 	if (!list_empty(&txq->tx_list)) {
427 		/* Flush the current list */
428 		ret = hfi1_ipoib_submit_tx_list(dev, txq);
429 
430 		if (unlikely(ret))
431 			if (ret != -EBUSY)
432 				++dev->stats.tx_carrier_errors;
433 	}
434 
435 	return ret;
436 }
437 
438 static int hfi1_ipoib_submit_tx(struct hfi1_ipoib_txq *txq,
439 				struct ipoib_txreq *tx)
440 {
441 	int ret;
442 
443 	ret = sdma_send_txreq(txq->sde,
444 			      iowait_get_ib_work(&txq->wait),
445 			      &tx->txreq,
446 			      txq->pkts_sent);
447 	if (likely(!ret)) {
448 		txq->pkts_sent = true;
449 		iowait_starve_clear(txq->pkts_sent, &txq->wait);
450 	}
451 
452 	return ret;
453 }
454 
455 static int hfi1_ipoib_send_dma_single(struct net_device *dev,
456 				      struct sk_buff *skb,
457 				      struct ipoib_txparms *txp)
458 {
459 	struct hfi1_ipoib_txq *txq = txp->txq;
460 	struct hfi1_ipoib_circ_buf *tx_ring;
461 	struct ipoib_txreq *tx;
462 	int ret;
463 
464 	tx = hfi1_ipoib_send_dma_common(dev, skb, txp);
465 	if (IS_ERR(tx)) {
466 		int ret = PTR_ERR(tx);
467 
468 		dev_kfree_skb_any(skb);
469 
470 		if (ret == -ENOMEM)
471 			++dev->stats.tx_errors;
472 		else
473 			++dev->stats.tx_carrier_errors;
474 
475 		return NETDEV_TX_OK;
476 	}
477 
478 	tx_ring = &txq->tx_ring;
479 	trace_hfi1_tx_consume(tx, tx_ring->tail);
480 	/* consume tx */
481 	smp_store_release(&tx_ring->tail, CIRC_NEXT(tx_ring->tail, tx_ring->max_items));
482 	ret = hfi1_ipoib_submit_tx(txq, tx);
483 	if (likely(!ret)) {
484 tx_ok:
485 		trace_sdma_output_ibhdr(txq->priv->dd,
486 					&tx->sdma_hdr->hdr,
487 					ib_is_sc5(txp->flow.sc5));
488 		hfi1_ipoib_check_queue_depth(txq);
489 		return NETDEV_TX_OK;
490 	}
491 
492 	txq->pkts_sent = false;
493 
494 	if (ret == -EBUSY || ret == -ECOMM)
495 		goto tx_ok;
496 
497 	/* mark complete and kick napi tx */
498 	smp_store_release(&tx->complete, 1);
499 	napi_schedule(&tx->txq->napi);
500 
501 	++dev->stats.tx_carrier_errors;
502 
503 	return NETDEV_TX_OK;
504 }
505 
506 static int hfi1_ipoib_send_dma_list(struct net_device *dev,
507 				    struct sk_buff *skb,
508 				    struct ipoib_txparms *txp)
509 {
510 	struct hfi1_ipoib_txq *txq = txp->txq;
511 	struct hfi1_ipoib_circ_buf *tx_ring;
512 	struct ipoib_txreq *tx;
513 
514 	/* Has the flow change ? */
515 	if (txq->flow.as_int != txp->flow.as_int) {
516 		int ret;
517 
518 		trace_hfi1_flow_flush(txq);
519 		ret = hfi1_ipoib_flush_tx_list(dev, txq);
520 		if (unlikely(ret)) {
521 			if (ret == -EBUSY)
522 				++dev->stats.tx_dropped;
523 			dev_kfree_skb_any(skb);
524 			return NETDEV_TX_OK;
525 		}
526 	}
527 	tx = hfi1_ipoib_send_dma_common(dev, skb, txp);
528 	if (IS_ERR(tx)) {
529 		int ret = PTR_ERR(tx);
530 
531 		dev_kfree_skb_any(skb);
532 
533 		if (ret == -ENOMEM)
534 			++dev->stats.tx_errors;
535 		else
536 			++dev->stats.tx_carrier_errors;
537 
538 		return NETDEV_TX_OK;
539 	}
540 
541 	tx_ring = &txq->tx_ring;
542 	trace_hfi1_tx_consume(tx, tx_ring->tail);
543 	/* consume tx */
544 	smp_store_release(&tx_ring->tail, CIRC_NEXT(tx_ring->tail, tx_ring->max_items));
545 	list_add_tail(&tx->txreq.list, &txq->tx_list);
546 
547 	hfi1_ipoib_check_queue_depth(txq);
548 
549 	trace_sdma_output_ibhdr(txq->priv->dd,
550 				&tx->sdma_hdr->hdr,
551 				ib_is_sc5(txp->flow.sc5));
552 
553 	if (!netdev_xmit_more())
554 		(void)hfi1_ipoib_flush_tx_list(dev, txq);
555 
556 	return NETDEV_TX_OK;
557 }
558 
559 static u8 hfi1_ipoib_calc_entropy(struct sk_buff *skb)
560 {
561 	if (skb_transport_header_was_set(skb)) {
562 		u8 *hdr = (u8 *)skb_transport_header(skb);
563 
564 		return (hdr[0] ^ hdr[1] ^ hdr[2] ^ hdr[3]);
565 	}
566 
567 	return (u8)skb_get_queue_mapping(skb);
568 }
569 
570 int hfi1_ipoib_send(struct net_device *dev,
571 		    struct sk_buff *skb,
572 		    struct ib_ah *address,
573 		    u32 dqpn)
574 {
575 	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
576 	struct ipoib_txparms txp;
577 	struct rdma_netdev *rn = netdev_priv(dev);
578 
579 	if (unlikely(skb->len > rn->mtu + HFI1_IPOIB_ENCAP_LEN)) {
580 		dd_dev_warn(priv->dd, "packet len %d (> %d) too long to send, dropping\n",
581 			    skb->len,
582 			    rn->mtu + HFI1_IPOIB_ENCAP_LEN);
583 		++dev->stats.tx_dropped;
584 		++dev->stats.tx_errors;
585 		dev_kfree_skb_any(skb);
586 		return NETDEV_TX_OK;
587 	}
588 
589 	txp.dd = priv->dd;
590 	txp.ah_attr = &ibah_to_rvtah(address)->attr;
591 	txp.ibp = to_iport(priv->device, priv->port_num);
592 	txp.txq = &priv->txqs[skb_get_queue_mapping(skb)];
593 	txp.dqpn = dqpn;
594 	txp.flow.sc5 = txp.ibp->sl_to_sc[rdma_ah_get_sl(txp.ah_attr)];
595 	txp.flow.tx_queue = (u8)skb_get_queue_mapping(skb);
596 	txp.entropy = hfi1_ipoib_calc_entropy(skb);
597 
598 	if (netdev_xmit_more() || !list_empty(&txp.txq->tx_list))
599 		return hfi1_ipoib_send_dma_list(dev, skb, &txp);
600 
601 	return hfi1_ipoib_send_dma_single(dev, skb,  &txp);
602 }
603 
604 /*
605  * hfi1_ipoib_sdma_sleep - ipoib sdma sleep function
606  *
607  * This function gets called from sdma_send_txreq() when there are not enough
608  * sdma descriptors available to send the packet. It adds Tx queue's wait
609  * structure to sdma engine's dmawait list to be woken up when descriptors
610  * become available.
611  */
612 static int hfi1_ipoib_sdma_sleep(struct sdma_engine *sde,
613 				 struct iowait_work *wait,
614 				 struct sdma_txreq *txreq,
615 				 uint seq,
616 				 bool pkts_sent)
617 {
618 	struct hfi1_ipoib_txq *txq =
619 		container_of(wait->iow, struct hfi1_ipoib_txq, wait);
620 
621 	write_seqlock(&sde->waitlock);
622 
623 	if (likely(txq->priv->netdev->reg_state == NETREG_REGISTERED)) {
624 		if (sdma_progress(sde, seq, txreq)) {
625 			write_sequnlock(&sde->waitlock);
626 			return -EAGAIN;
627 		}
628 
629 		if (list_empty(&txreq->list))
630 			/* came from non-list submit */
631 			list_add_tail(&txreq->list, &txq->tx_list);
632 		if (list_empty(&txq->wait.list)) {
633 			struct hfi1_ibport *ibp = &sde->ppd->ibport_data;
634 
635 			if (!atomic_xchg(&txq->tx_ring.no_desc, 1)) {
636 				trace_hfi1_txq_queued(txq);
637 				hfi1_ipoib_stop_txq(txq);
638 			}
639 			ibp->rvp.n_dmawait++;
640 			iowait_queue(pkts_sent, wait->iow, &sde->dmawait);
641 		}
642 
643 		write_sequnlock(&sde->waitlock);
644 		return -EBUSY;
645 	}
646 
647 	write_sequnlock(&sde->waitlock);
648 	return -EINVAL;
649 }
650 
651 /*
652  * hfi1_ipoib_sdma_wakeup - ipoib sdma wakeup function
653  *
654  * This function gets called when SDMA descriptors becomes available and Tx
655  * queue's wait structure was previously added to sdma engine's dmawait list.
656  */
657 static void hfi1_ipoib_sdma_wakeup(struct iowait *wait, int reason)
658 {
659 	struct hfi1_ipoib_txq *txq =
660 		container_of(wait, struct hfi1_ipoib_txq, wait);
661 
662 	trace_hfi1_txq_wakeup(txq);
663 	if (likely(txq->priv->netdev->reg_state == NETREG_REGISTERED))
664 		iowait_schedule(wait, system_highpri_wq, WORK_CPU_UNBOUND);
665 }
666 
667 static void hfi1_ipoib_flush_txq(struct work_struct *work)
668 {
669 	struct iowait_work *ioww =
670 		container_of(work, struct iowait_work, iowork);
671 	struct iowait *wait = iowait_ioww_to_iow(ioww);
672 	struct hfi1_ipoib_txq *txq =
673 		container_of(wait, struct hfi1_ipoib_txq, wait);
674 	struct net_device *dev = txq->priv->netdev;
675 
676 	if (likely(dev->reg_state == NETREG_REGISTERED) &&
677 	    likely(!hfi1_ipoib_flush_tx_list(dev, txq)))
678 		if (atomic_xchg(&txq->tx_ring.no_desc, 0))
679 			hfi1_ipoib_wake_txq(txq);
680 }
681 
682 int hfi1_ipoib_txreq_init(struct hfi1_ipoib_dev_priv *priv)
683 {
684 	struct net_device *dev = priv->netdev;
685 	u32 tx_ring_size, tx_item_size;
686 	struct hfi1_ipoib_circ_buf *tx_ring;
687 	int i, j;
688 
689 	/*
690 	 * Ring holds 1 less than tx_ring_size
691 	 * Round up to next power of 2 in order to hold at least tx_queue_len
692 	 */
693 	tx_ring_size = roundup_pow_of_two(dev->tx_queue_len + 1);
694 	tx_item_size = roundup_pow_of_two(sizeof(struct ipoib_txreq));
695 
696 	priv->txqs = kcalloc_node(dev->num_tx_queues,
697 				  sizeof(struct hfi1_ipoib_txq),
698 				  GFP_KERNEL,
699 				  priv->dd->node);
700 	if (!priv->txqs)
701 		return -ENOMEM;
702 
703 	for (i = 0; i < dev->num_tx_queues; i++) {
704 		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
705 		struct ipoib_txreq *tx;
706 
707 		tx_ring = &txq->tx_ring;
708 		iowait_init(&txq->wait,
709 			    0,
710 			    hfi1_ipoib_flush_txq,
711 			    NULL,
712 			    hfi1_ipoib_sdma_sleep,
713 			    hfi1_ipoib_sdma_wakeup,
714 			    NULL,
715 			    NULL);
716 		txq->priv = priv;
717 		txq->sde = NULL;
718 		INIT_LIST_HEAD(&txq->tx_list);
719 		atomic_set(&txq->tx_ring.stops, 0);
720 		atomic_set(&txq->tx_ring.ring_full, 0);
721 		atomic_set(&txq->tx_ring.no_desc, 0);
722 		txq->q_idx = i;
723 		txq->flow.tx_queue = 0xff;
724 		txq->flow.sc5 = 0xff;
725 		txq->pkts_sent = false;
726 
727 		netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
728 					     priv->dd->node);
729 
730 		txq->tx_ring.items =
731 			kvzalloc_node(array_size(tx_ring_size, tx_item_size),
732 				      GFP_KERNEL, priv->dd->node);
733 		if (!txq->tx_ring.items)
734 			goto free_txqs;
735 
736 		txq->tx_ring.max_items = tx_ring_size;
737 		txq->tx_ring.shift = ilog2(tx_item_size);
738 		txq->tx_ring.avail = hfi1_ipoib_ring_hwat(txq);
739 		tx_ring = &txq->tx_ring;
740 		for (j = 0; j < tx_ring_size; j++)
741 			hfi1_txreq_from_idx(tx_ring, j)->sdma_hdr =
742 				kzalloc_node(sizeof(*tx->sdma_hdr),
743 					     GFP_KERNEL, priv->dd->node);
744 
745 		netif_tx_napi_add(dev, &txq->napi,
746 				  hfi1_ipoib_poll_tx_ring,
747 				  NAPI_POLL_WEIGHT);
748 	}
749 
750 	return 0;
751 
752 free_txqs:
753 	for (i--; i >= 0; i--) {
754 		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
755 
756 		netif_napi_del(&txq->napi);
757 		tx_ring = &txq->tx_ring;
758 		for (j = 0; j < tx_ring_size; j++)
759 			kfree(hfi1_txreq_from_idx(tx_ring, j)->sdma_hdr);
760 		kvfree(tx_ring->items);
761 	}
762 
763 	kfree(priv->txqs);
764 	priv->txqs = NULL;
765 	return -ENOMEM;
766 }
767 
768 static void hfi1_ipoib_drain_tx_list(struct hfi1_ipoib_txq *txq)
769 {
770 	struct sdma_txreq *txreq;
771 	struct sdma_txreq *txreq_tmp;
772 
773 	list_for_each_entry_safe(txreq, txreq_tmp, &txq->tx_list, list) {
774 		struct ipoib_txreq *tx =
775 			container_of(txreq, struct ipoib_txreq, txreq);
776 
777 		list_del(&txreq->list);
778 		sdma_txclean(txq->priv->dd, &tx->txreq);
779 		dev_kfree_skb_any(tx->skb);
780 		tx->skb = NULL;
781 		txq->tx_ring.complete_txreqs++;
782 	}
783 
784 	if (hfi1_ipoib_used(txq))
785 		dd_dev_warn(txq->priv->dd,
786 			    "txq %d not empty found %u requests\n",
787 			    txq->q_idx,
788 			    hfi1_ipoib_txreqs(txq->tx_ring.sent_txreqs,
789 					      txq->tx_ring.complete_txreqs));
790 }
791 
792 void hfi1_ipoib_txreq_deinit(struct hfi1_ipoib_dev_priv *priv)
793 {
794 	int i, j;
795 
796 	for (i = 0; i < priv->netdev->num_tx_queues; i++) {
797 		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
798 		struct hfi1_ipoib_circ_buf *tx_ring = &txq->tx_ring;
799 
800 		iowait_cancel_work(&txq->wait);
801 		iowait_sdma_drain(&txq->wait);
802 		hfi1_ipoib_drain_tx_list(txq);
803 		netif_napi_del(&txq->napi);
804 		hfi1_ipoib_drain_tx_ring(txq);
805 		for (j = 0; j < tx_ring->max_items; j++)
806 			kfree(hfi1_txreq_from_idx(tx_ring, j)->sdma_hdr);
807 		kvfree(tx_ring->items);
808 	}
809 
810 	kfree(priv->txqs);
811 	priv->txqs = NULL;
812 }
813 
814 void hfi1_ipoib_napi_tx_enable(struct net_device *dev)
815 {
816 	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
817 	int i;
818 
819 	for (i = 0; i < dev->num_tx_queues; i++) {
820 		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
821 
822 		napi_enable(&txq->napi);
823 	}
824 }
825 
826 void hfi1_ipoib_napi_tx_disable(struct net_device *dev)
827 {
828 	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
829 	int i;
830 
831 	for (i = 0; i < dev->num_tx_queues; i++) {
832 		struct hfi1_ipoib_txq *txq = &priv->txqs[i];
833 
834 		napi_disable(&txq->napi);
835 		hfi1_ipoib_drain_tx_ring(txq);
836 	}
837 }
838 
839 void hfi1_ipoib_tx_timeout(struct net_device *dev, unsigned int q)
840 {
841 	struct hfi1_ipoib_dev_priv *priv = hfi1_ipoib_priv(dev);
842 	struct hfi1_ipoib_txq *txq = &priv->txqs[q];
843 
844 	dd_dev_info(priv->dd, "timeout txq %p q %u stopped %u stops %d no_desc %d ring_full %d\n",
845 		    txq, q,
846 		    __netif_subqueue_stopped(dev, txq->q_idx),
847 		    atomic_read(&txq->tx_ring.stops),
848 		    atomic_read(&txq->tx_ring.no_desc),
849 		    atomic_read(&txq->tx_ring.ring_full));
850 	dd_dev_info(priv->dd, "sde %p engine %u\n",
851 		    txq->sde,
852 		    txq->sde ? txq->sde->this_idx : 0);
853 	dd_dev_info(priv->dd, "flow %x\n", txq->flow.as_int);
854 	dd_dev_info(priv->dd, "sent %llu completed %llu used %llu\n",
855 		    txq->tx_ring.sent_txreqs, txq->tx_ring.complete_txreqs,
856 		    hfi1_ipoib_used(txq));
857 	dd_dev_info(priv->dd, "tx_queue_len %u max_items %u\n",
858 		    dev->tx_queue_len, txq->tx_ring.max_items);
859 	dd_dev_info(priv->dd, "head %u tail %u\n",
860 		    txq->tx_ring.head, txq->tx_ring.tail);
861 	dd_dev_info(priv->dd, "wait queued %u\n",
862 		    !list_empty(&txq->wait.list));
863 	dd_dev_info(priv->dd, "tx_list empty %u\n",
864 		    list_empty(&txq->tx_list));
865 }
866 
867