1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2013 - 2018 Intel Corporation. */
3 
4 #include <linux/prefetch.h>
5 
6 #include "iavf.h"
7 #include "iavf_trace.h"
8 #include "iavf_prototype.h"
9 
10 static inline __le64 build_ctob(u32 td_cmd, u32 td_offset, unsigned int size,
11 				u32 td_tag)
12 {
13 	return cpu_to_le64(IAVF_TX_DESC_DTYPE_DATA |
14 			   ((u64)td_cmd  << IAVF_TXD_QW1_CMD_SHIFT) |
15 			   ((u64)td_offset << IAVF_TXD_QW1_OFFSET_SHIFT) |
16 			   ((u64)size  << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT) |
17 			   ((u64)td_tag  << IAVF_TXD_QW1_L2TAG1_SHIFT));
18 }
19 
20 #define IAVF_TXD_CMD (IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_RS)
21 
22 /**
23  * iavf_unmap_and_free_tx_resource - Release a Tx buffer
24  * @ring:      the ring that owns the buffer
25  * @tx_buffer: the buffer to free
26  **/
27 static void iavf_unmap_and_free_tx_resource(struct iavf_ring *ring,
28 					    struct iavf_tx_buffer *tx_buffer)
29 {
30 	if (tx_buffer->skb) {
31 		if (tx_buffer->tx_flags & IAVF_TX_FLAGS_FD_SB)
32 			kfree(tx_buffer->raw_buf);
33 		else
34 			dev_kfree_skb_any(tx_buffer->skb);
35 		if (dma_unmap_len(tx_buffer, len))
36 			dma_unmap_single(ring->dev,
37 					 dma_unmap_addr(tx_buffer, dma),
38 					 dma_unmap_len(tx_buffer, len),
39 					 DMA_TO_DEVICE);
40 	} else if (dma_unmap_len(tx_buffer, len)) {
41 		dma_unmap_page(ring->dev,
42 			       dma_unmap_addr(tx_buffer, dma),
43 			       dma_unmap_len(tx_buffer, len),
44 			       DMA_TO_DEVICE);
45 	}
46 
47 	tx_buffer->next_to_watch = NULL;
48 	tx_buffer->skb = NULL;
49 	dma_unmap_len_set(tx_buffer, len, 0);
50 	/* tx_buffer must be completely set up in the transmit path */
51 }
52 
53 /**
54  * iavf_clean_tx_ring - Free any empty Tx buffers
55  * @tx_ring: ring to be cleaned
56  **/
57 void iavf_clean_tx_ring(struct iavf_ring *tx_ring)
58 {
59 	unsigned long bi_size;
60 	u16 i;
61 
62 	/* ring already cleared, nothing to do */
63 	if (!tx_ring->tx_bi)
64 		return;
65 
66 	/* Free all the Tx ring sk_buffs */
67 	for (i = 0; i < tx_ring->count; i++)
68 		iavf_unmap_and_free_tx_resource(tx_ring, &tx_ring->tx_bi[i]);
69 
70 	bi_size = sizeof(struct iavf_tx_buffer) * tx_ring->count;
71 	memset(tx_ring->tx_bi, 0, bi_size);
72 
73 	/* Zero out the descriptor ring */
74 	memset(tx_ring->desc, 0, tx_ring->size);
75 
76 	tx_ring->next_to_use = 0;
77 	tx_ring->next_to_clean = 0;
78 
79 	if (!tx_ring->netdev)
80 		return;
81 
82 	/* cleanup Tx queue statistics */
83 	netdev_tx_reset_queue(txring_txq(tx_ring));
84 }
85 
86 /**
87  * iavf_free_tx_resources - Free Tx resources per queue
88  * @tx_ring: Tx descriptor ring for a specific queue
89  *
90  * Free all transmit software resources
91  **/
92 void iavf_free_tx_resources(struct iavf_ring *tx_ring)
93 {
94 	iavf_clean_tx_ring(tx_ring);
95 	kfree(tx_ring->tx_bi);
96 	tx_ring->tx_bi = NULL;
97 
98 	if (tx_ring->desc) {
99 		dma_free_coherent(tx_ring->dev, tx_ring->size,
100 				  tx_ring->desc, tx_ring->dma);
101 		tx_ring->desc = NULL;
102 	}
103 }
104 
105 /**
106  * iavf_get_tx_pending - how many Tx descriptors not processed
107  * @ring: the ring of descriptors
108  * @in_sw: is tx_pending being checked in SW or HW
109  *
110  * Since there is no access to the ring head register
111  * in XL710, we need to use our local copies
112  **/
113 u32 iavf_get_tx_pending(struct iavf_ring *ring, bool in_sw)
114 {
115 	u32 head, tail;
116 
117 	head = ring->next_to_clean;
118 	tail = readl(ring->tail);
119 
120 	if (head != tail)
121 		return (head < tail) ?
122 			tail - head : (tail + ring->count - head);
123 
124 	return 0;
125 }
126 
127 /**
128  * iavf_detect_recover_hung - Function to detect and recover hung_queues
129  * @vsi:  pointer to vsi struct with tx queues
130  *
131  * VSI has netdev and netdev has TX queues. This function is to check each of
132  * those TX queues if they are hung, trigger recovery by issuing SW interrupt.
133  **/
134 void iavf_detect_recover_hung(struct iavf_vsi *vsi)
135 {
136 	struct iavf_ring *tx_ring = NULL;
137 	struct net_device *netdev;
138 	unsigned int i;
139 	int packets;
140 
141 	if (!vsi)
142 		return;
143 
144 	if (test_bit(__IAVF_VSI_DOWN, vsi->state))
145 		return;
146 
147 	netdev = vsi->netdev;
148 	if (!netdev)
149 		return;
150 
151 	if (!netif_carrier_ok(netdev))
152 		return;
153 
154 	for (i = 0; i < vsi->back->num_active_queues; i++) {
155 		tx_ring = &vsi->back->tx_rings[i];
156 		if (tx_ring && tx_ring->desc) {
157 			/* If packet counter has not changed the queue is
158 			 * likely stalled, so force an interrupt for this
159 			 * queue.
160 			 *
161 			 * prev_pkt_ctr would be negative if there was no
162 			 * pending work.
163 			 */
164 			packets = tx_ring->stats.packets & INT_MAX;
165 			if (tx_ring->tx_stats.prev_pkt_ctr == packets) {
166 				iavf_force_wb(vsi, tx_ring->q_vector);
167 				continue;
168 			}
169 
170 			/* Memory barrier between read of packet count and call
171 			 * to iavf_get_tx_pending()
172 			 */
173 			smp_rmb();
174 			tx_ring->tx_stats.prev_pkt_ctr =
175 			  iavf_get_tx_pending(tx_ring, true) ? packets : -1;
176 		}
177 	}
178 }
179 
180 #define WB_STRIDE 4
181 
182 /**
183  * iavf_clean_tx_irq - Reclaim resources after transmit completes
184  * @vsi: the VSI we care about
185  * @tx_ring: Tx ring to clean
186  * @napi_budget: Used to determine if we are in netpoll
187  *
188  * Returns true if there's any budget left (e.g. the clean is finished)
189  **/
190 static bool iavf_clean_tx_irq(struct iavf_vsi *vsi,
191 			      struct iavf_ring *tx_ring, int napi_budget)
192 {
193 	int i = tx_ring->next_to_clean;
194 	struct iavf_tx_buffer *tx_buf;
195 	struct iavf_tx_desc *tx_desc;
196 	unsigned int total_bytes = 0, total_packets = 0;
197 	unsigned int budget = vsi->work_limit;
198 
199 	tx_buf = &tx_ring->tx_bi[i];
200 	tx_desc = IAVF_TX_DESC(tx_ring, i);
201 	i -= tx_ring->count;
202 
203 	do {
204 		struct iavf_tx_desc *eop_desc = tx_buf->next_to_watch;
205 
206 		/* if next_to_watch is not set then there is no work pending */
207 		if (!eop_desc)
208 			break;
209 
210 		/* prevent any other reads prior to eop_desc */
211 		smp_rmb();
212 
213 		iavf_trace(clean_tx_irq, tx_ring, tx_desc, tx_buf);
214 		/* if the descriptor isn't done, no work yet to do */
215 		if (!(eop_desc->cmd_type_offset_bsz &
216 		      cpu_to_le64(IAVF_TX_DESC_DTYPE_DESC_DONE)))
217 			break;
218 
219 		/* clear next_to_watch to prevent false hangs */
220 		tx_buf->next_to_watch = NULL;
221 
222 		/* update the statistics for this packet */
223 		total_bytes += tx_buf->bytecount;
224 		total_packets += tx_buf->gso_segs;
225 
226 		/* free the skb */
227 		napi_consume_skb(tx_buf->skb, napi_budget);
228 
229 		/* unmap skb header data */
230 		dma_unmap_single(tx_ring->dev,
231 				 dma_unmap_addr(tx_buf, dma),
232 				 dma_unmap_len(tx_buf, len),
233 				 DMA_TO_DEVICE);
234 
235 		/* clear tx_buffer data */
236 		tx_buf->skb = NULL;
237 		dma_unmap_len_set(tx_buf, len, 0);
238 
239 		/* unmap remaining buffers */
240 		while (tx_desc != eop_desc) {
241 			iavf_trace(clean_tx_irq_unmap,
242 				   tx_ring, tx_desc, tx_buf);
243 
244 			tx_buf++;
245 			tx_desc++;
246 			i++;
247 			if (unlikely(!i)) {
248 				i -= tx_ring->count;
249 				tx_buf = tx_ring->tx_bi;
250 				tx_desc = IAVF_TX_DESC(tx_ring, 0);
251 			}
252 
253 			/* unmap any remaining paged data */
254 			if (dma_unmap_len(tx_buf, len)) {
255 				dma_unmap_page(tx_ring->dev,
256 					       dma_unmap_addr(tx_buf, dma),
257 					       dma_unmap_len(tx_buf, len),
258 					       DMA_TO_DEVICE);
259 				dma_unmap_len_set(tx_buf, len, 0);
260 			}
261 		}
262 
263 		/* move us one more past the eop_desc for start of next pkt */
264 		tx_buf++;
265 		tx_desc++;
266 		i++;
267 		if (unlikely(!i)) {
268 			i -= tx_ring->count;
269 			tx_buf = tx_ring->tx_bi;
270 			tx_desc = IAVF_TX_DESC(tx_ring, 0);
271 		}
272 
273 		prefetch(tx_desc);
274 
275 		/* update budget accounting */
276 		budget--;
277 	} while (likely(budget));
278 
279 	i += tx_ring->count;
280 	tx_ring->next_to_clean = i;
281 	u64_stats_update_begin(&tx_ring->syncp);
282 	tx_ring->stats.bytes += total_bytes;
283 	tx_ring->stats.packets += total_packets;
284 	u64_stats_update_end(&tx_ring->syncp);
285 	tx_ring->q_vector->tx.total_bytes += total_bytes;
286 	tx_ring->q_vector->tx.total_packets += total_packets;
287 
288 	if (tx_ring->flags & IAVF_TXR_FLAGS_WB_ON_ITR) {
289 		/* check to see if there are < 4 descriptors
290 		 * waiting to be written back, then kick the hardware to force
291 		 * them to be written back in case we stay in NAPI.
292 		 * In this mode on X722 we do not enable Interrupt.
293 		 */
294 		unsigned int j = iavf_get_tx_pending(tx_ring, false);
295 
296 		if (budget &&
297 		    ((j / WB_STRIDE) == 0) && (j > 0) &&
298 		    !test_bit(__IAVF_VSI_DOWN, vsi->state) &&
299 		    (IAVF_DESC_UNUSED(tx_ring) != tx_ring->count))
300 			tx_ring->arm_wb = true;
301 	}
302 
303 	/* notify netdev of completed buffers */
304 	netdev_tx_completed_queue(txring_txq(tx_ring),
305 				  total_packets, total_bytes);
306 
307 #define TX_WAKE_THRESHOLD ((s16)(DESC_NEEDED * 2))
308 	if (unlikely(total_packets && netif_carrier_ok(tx_ring->netdev) &&
309 		     (IAVF_DESC_UNUSED(tx_ring) >= TX_WAKE_THRESHOLD))) {
310 		/* Make sure that anybody stopping the queue after this
311 		 * sees the new next_to_clean.
312 		 */
313 		smp_mb();
314 		if (__netif_subqueue_stopped(tx_ring->netdev,
315 					     tx_ring->queue_index) &&
316 		   !test_bit(__IAVF_VSI_DOWN, vsi->state)) {
317 			netif_wake_subqueue(tx_ring->netdev,
318 					    tx_ring->queue_index);
319 			++tx_ring->tx_stats.restart_queue;
320 		}
321 	}
322 
323 	return !!budget;
324 }
325 
326 /**
327  * iavf_enable_wb_on_itr - Arm hardware to do a wb, interrupts are not enabled
328  * @vsi: the VSI we care about
329  * @q_vector: the vector on which to enable writeback
330  *
331  **/
332 static void iavf_enable_wb_on_itr(struct iavf_vsi *vsi,
333 				  struct iavf_q_vector *q_vector)
334 {
335 	u16 flags = q_vector->tx.ring[0].flags;
336 	u32 val;
337 
338 	if (!(flags & IAVF_TXR_FLAGS_WB_ON_ITR))
339 		return;
340 
341 	if (q_vector->arm_wb_state)
342 		return;
343 
344 	val = IAVF_VFINT_DYN_CTLN1_WB_ON_ITR_MASK |
345 	      IAVF_VFINT_DYN_CTLN1_ITR_INDX_MASK; /* set noitr */
346 
347 	wr32(&vsi->back->hw,
348 	     IAVF_VFINT_DYN_CTLN1(q_vector->reg_idx), val);
349 	q_vector->arm_wb_state = true;
350 }
351 
352 /**
353  * iavf_force_wb - Issue SW Interrupt so HW does a wb
354  * @vsi: the VSI we care about
355  * @q_vector: the vector  on which to force writeback
356  *
357  **/
358 void iavf_force_wb(struct iavf_vsi *vsi, struct iavf_q_vector *q_vector)
359 {
360 	u32 val = IAVF_VFINT_DYN_CTLN1_INTENA_MASK |
361 		  IAVF_VFINT_DYN_CTLN1_ITR_INDX_MASK | /* set noitr */
362 		  IAVF_VFINT_DYN_CTLN1_SWINT_TRIG_MASK |
363 		  IAVF_VFINT_DYN_CTLN1_SW_ITR_INDX_ENA_MASK
364 		  /* allow 00 to be written to the index */;
365 
366 	wr32(&vsi->back->hw,
367 	     IAVF_VFINT_DYN_CTLN1(q_vector->reg_idx),
368 	     val);
369 }
370 
371 static inline bool iavf_container_is_rx(struct iavf_q_vector *q_vector,
372 					struct iavf_ring_container *rc)
373 {
374 	return &q_vector->rx == rc;
375 }
376 
377 static inline unsigned int iavf_itr_divisor(struct iavf_q_vector *q_vector)
378 {
379 	unsigned int divisor;
380 
381 	switch (q_vector->adapter->link_speed) {
382 	case VIRTCHNL_LINK_SPEED_40GB:
383 		divisor = IAVF_ITR_ADAPTIVE_MIN_INC * 1024;
384 		break;
385 	case VIRTCHNL_LINK_SPEED_25GB:
386 	case VIRTCHNL_LINK_SPEED_20GB:
387 		divisor = IAVF_ITR_ADAPTIVE_MIN_INC * 512;
388 		break;
389 	default:
390 	case VIRTCHNL_LINK_SPEED_10GB:
391 		divisor = IAVF_ITR_ADAPTIVE_MIN_INC * 256;
392 		break;
393 	case VIRTCHNL_LINK_SPEED_1GB:
394 	case VIRTCHNL_LINK_SPEED_100MB:
395 		divisor = IAVF_ITR_ADAPTIVE_MIN_INC * 32;
396 		break;
397 	}
398 
399 	return divisor;
400 }
401 
402 /**
403  * iavf_update_itr - update the dynamic ITR value based on statistics
404  * @q_vector: structure containing interrupt and ring information
405  * @rc: structure containing ring performance data
406  *
407  * Stores a new ITR value based on packets and byte
408  * counts during the last interrupt.  The advantage of per interrupt
409  * computation is faster updates and more accurate ITR for the current
410  * traffic pattern.  Constants in this function were computed
411  * based on theoretical maximum wire speed and thresholds were set based
412  * on testing data as well as attempting to minimize response time
413  * while increasing bulk throughput.
414  **/
415 static void iavf_update_itr(struct iavf_q_vector *q_vector,
416 			    struct iavf_ring_container *rc)
417 {
418 	unsigned int avg_wire_size, packets, bytes, itr;
419 	unsigned long next_update = jiffies;
420 
421 	/* If we don't have any rings just leave ourselves set for maximum
422 	 * possible latency so we take ourselves out of the equation.
423 	 */
424 	if (!rc->ring || !ITR_IS_DYNAMIC(rc->ring->itr_setting))
425 		return;
426 
427 	/* For Rx we want to push the delay up and default to low latency.
428 	 * for Tx we want to pull the delay down and default to high latency.
429 	 */
430 	itr = iavf_container_is_rx(q_vector, rc) ?
431 	      IAVF_ITR_ADAPTIVE_MIN_USECS | IAVF_ITR_ADAPTIVE_LATENCY :
432 	      IAVF_ITR_ADAPTIVE_MAX_USECS | IAVF_ITR_ADAPTIVE_LATENCY;
433 
434 	/* If we didn't update within up to 1 - 2 jiffies we can assume
435 	 * that either packets are coming in so slow there hasn't been
436 	 * any work, or that there is so much work that NAPI is dealing
437 	 * with interrupt moderation and we don't need to do anything.
438 	 */
439 	if (time_after(next_update, rc->next_update))
440 		goto clear_counts;
441 
442 	/* If itr_countdown is set it means we programmed an ITR within
443 	 * the last 4 interrupt cycles. This has a side effect of us
444 	 * potentially firing an early interrupt. In order to work around
445 	 * this we need to throw out any data received for a few
446 	 * interrupts following the update.
447 	 */
448 	if (q_vector->itr_countdown) {
449 		itr = rc->target_itr;
450 		goto clear_counts;
451 	}
452 
453 	packets = rc->total_packets;
454 	bytes = rc->total_bytes;
455 
456 	if (iavf_container_is_rx(q_vector, rc)) {
457 		/* If Rx there are 1 to 4 packets and bytes are less than
458 		 * 9000 assume insufficient data to use bulk rate limiting
459 		 * approach unless Tx is already in bulk rate limiting. We
460 		 * are likely latency driven.
461 		 */
462 		if (packets && packets < 4 && bytes < 9000 &&
463 		    (q_vector->tx.target_itr & IAVF_ITR_ADAPTIVE_LATENCY)) {
464 			itr = IAVF_ITR_ADAPTIVE_LATENCY;
465 			goto adjust_by_size;
466 		}
467 	} else if (packets < 4) {
468 		/* If we have Tx and Rx ITR maxed and Tx ITR is running in
469 		 * bulk mode and we are receiving 4 or fewer packets just
470 		 * reset the ITR_ADAPTIVE_LATENCY bit for latency mode so
471 		 * that the Rx can relax.
472 		 */
473 		if (rc->target_itr == IAVF_ITR_ADAPTIVE_MAX_USECS &&
474 		    (q_vector->rx.target_itr & IAVF_ITR_MASK) ==
475 		     IAVF_ITR_ADAPTIVE_MAX_USECS)
476 			goto clear_counts;
477 	} else if (packets > 32) {
478 		/* If we have processed over 32 packets in a single interrupt
479 		 * for Tx assume we need to switch over to "bulk" mode.
480 		 */
481 		rc->target_itr &= ~IAVF_ITR_ADAPTIVE_LATENCY;
482 	}
483 
484 	/* We have no packets to actually measure against. This means
485 	 * either one of the other queues on this vector is active or
486 	 * we are a Tx queue doing TSO with too high of an interrupt rate.
487 	 *
488 	 * Between 4 and 56 we can assume that our current interrupt delay
489 	 * is only slightly too low. As such we should increase it by a small
490 	 * fixed amount.
491 	 */
492 	if (packets < 56) {
493 		itr = rc->target_itr + IAVF_ITR_ADAPTIVE_MIN_INC;
494 		if ((itr & IAVF_ITR_MASK) > IAVF_ITR_ADAPTIVE_MAX_USECS) {
495 			itr &= IAVF_ITR_ADAPTIVE_LATENCY;
496 			itr += IAVF_ITR_ADAPTIVE_MAX_USECS;
497 		}
498 		goto clear_counts;
499 	}
500 
501 	if (packets <= 256) {
502 		itr = min(q_vector->tx.current_itr, q_vector->rx.current_itr);
503 		itr &= IAVF_ITR_MASK;
504 
505 		/* Between 56 and 112 is our "goldilocks" zone where we are
506 		 * working out "just right". Just report that our current
507 		 * ITR is good for us.
508 		 */
509 		if (packets <= 112)
510 			goto clear_counts;
511 
512 		/* If packet count is 128 or greater we are likely looking
513 		 * at a slight overrun of the delay we want. Try halving
514 		 * our delay to see if that will cut the number of packets
515 		 * in half per interrupt.
516 		 */
517 		itr /= 2;
518 		itr &= IAVF_ITR_MASK;
519 		if (itr < IAVF_ITR_ADAPTIVE_MIN_USECS)
520 			itr = IAVF_ITR_ADAPTIVE_MIN_USECS;
521 
522 		goto clear_counts;
523 	}
524 
525 	/* The paths below assume we are dealing with a bulk ITR since
526 	 * number of packets is greater than 256. We are just going to have
527 	 * to compute a value and try to bring the count under control,
528 	 * though for smaller packet sizes there isn't much we can do as
529 	 * NAPI polling will likely be kicking in sooner rather than later.
530 	 */
531 	itr = IAVF_ITR_ADAPTIVE_BULK;
532 
533 adjust_by_size:
534 	/* If packet counts are 256 or greater we can assume we have a gross
535 	 * overestimation of what the rate should be. Instead of trying to fine
536 	 * tune it just use the formula below to try and dial in an exact value
537 	 * give the current packet size of the frame.
538 	 */
539 	avg_wire_size = bytes / packets;
540 
541 	/* The following is a crude approximation of:
542 	 *  wmem_default / (size + overhead) = desired_pkts_per_int
543 	 *  rate / bits_per_byte / (size + ethernet overhead) = pkt_rate
544 	 *  (desired_pkt_rate / pkt_rate) * usecs_per_sec = ITR value
545 	 *
546 	 * Assuming wmem_default is 212992 and overhead is 640 bytes per
547 	 * packet, (256 skb, 64 headroom, 320 shared info), we can reduce the
548 	 * formula down to
549 	 *
550 	 *  (170 * (size + 24)) / (size + 640) = ITR
551 	 *
552 	 * We first do some math on the packet size and then finally bitshift
553 	 * by 8 after rounding up. We also have to account for PCIe link speed
554 	 * difference as ITR scales based on this.
555 	 */
556 	if (avg_wire_size <= 60) {
557 		/* Start at 250k ints/sec */
558 		avg_wire_size = 4096;
559 	} else if (avg_wire_size <= 380) {
560 		/* 250K ints/sec to 60K ints/sec */
561 		avg_wire_size *= 40;
562 		avg_wire_size += 1696;
563 	} else if (avg_wire_size <= 1084) {
564 		/* 60K ints/sec to 36K ints/sec */
565 		avg_wire_size *= 15;
566 		avg_wire_size += 11452;
567 	} else if (avg_wire_size <= 1980) {
568 		/* 36K ints/sec to 30K ints/sec */
569 		avg_wire_size *= 5;
570 		avg_wire_size += 22420;
571 	} else {
572 		/* plateau at a limit of 30K ints/sec */
573 		avg_wire_size = 32256;
574 	}
575 
576 	/* If we are in low latency mode halve our delay which doubles the
577 	 * rate to somewhere between 100K to 16K ints/sec
578 	 */
579 	if (itr & IAVF_ITR_ADAPTIVE_LATENCY)
580 		avg_wire_size /= 2;
581 
582 	/* Resultant value is 256 times larger than it needs to be. This
583 	 * gives us room to adjust the value as needed to either increase
584 	 * or decrease the value based on link speeds of 10G, 2.5G, 1G, etc.
585 	 *
586 	 * Use addition as we have already recorded the new latency flag
587 	 * for the ITR value.
588 	 */
589 	itr += DIV_ROUND_UP(avg_wire_size, iavf_itr_divisor(q_vector)) *
590 	       IAVF_ITR_ADAPTIVE_MIN_INC;
591 
592 	if ((itr & IAVF_ITR_MASK) > IAVF_ITR_ADAPTIVE_MAX_USECS) {
593 		itr &= IAVF_ITR_ADAPTIVE_LATENCY;
594 		itr += IAVF_ITR_ADAPTIVE_MAX_USECS;
595 	}
596 
597 clear_counts:
598 	/* write back value */
599 	rc->target_itr = itr;
600 
601 	/* next update should occur within next jiffy */
602 	rc->next_update = next_update + 1;
603 
604 	rc->total_bytes = 0;
605 	rc->total_packets = 0;
606 }
607 
608 /**
609  * iavf_setup_tx_descriptors - Allocate the Tx descriptors
610  * @tx_ring: the tx ring to set up
611  *
612  * Return 0 on success, negative on error
613  **/
614 int iavf_setup_tx_descriptors(struct iavf_ring *tx_ring)
615 {
616 	struct device *dev = tx_ring->dev;
617 	int bi_size;
618 
619 	if (!dev)
620 		return -ENOMEM;
621 
622 	/* warn if we are about to overwrite the pointer */
623 	WARN_ON(tx_ring->tx_bi);
624 	bi_size = sizeof(struct iavf_tx_buffer) * tx_ring->count;
625 	tx_ring->tx_bi = kzalloc(bi_size, GFP_KERNEL);
626 	if (!tx_ring->tx_bi)
627 		goto err;
628 
629 	/* round up to nearest 4K */
630 	tx_ring->size = tx_ring->count * sizeof(struct iavf_tx_desc);
631 	tx_ring->size = ALIGN(tx_ring->size, 4096);
632 	tx_ring->desc = dma_alloc_coherent(dev, tx_ring->size,
633 					   &tx_ring->dma, GFP_KERNEL);
634 	if (!tx_ring->desc) {
635 		dev_info(dev, "Unable to allocate memory for the Tx descriptor ring, size=%d\n",
636 			 tx_ring->size);
637 		goto err;
638 	}
639 
640 	tx_ring->next_to_use = 0;
641 	tx_ring->next_to_clean = 0;
642 	tx_ring->tx_stats.prev_pkt_ctr = -1;
643 	return 0;
644 
645 err:
646 	kfree(tx_ring->tx_bi);
647 	tx_ring->tx_bi = NULL;
648 	return -ENOMEM;
649 }
650 
651 /**
652  * iavf_clean_rx_ring - Free Rx buffers
653  * @rx_ring: ring to be cleaned
654  **/
655 void iavf_clean_rx_ring(struct iavf_ring *rx_ring)
656 {
657 	unsigned long bi_size;
658 	u16 i;
659 
660 	/* ring already cleared, nothing to do */
661 	if (!rx_ring->rx_bi)
662 		return;
663 
664 	if (rx_ring->skb) {
665 		dev_kfree_skb(rx_ring->skb);
666 		rx_ring->skb = NULL;
667 	}
668 
669 	/* Free all the Rx ring sk_buffs */
670 	for (i = 0; i < rx_ring->count; i++) {
671 		struct iavf_rx_buffer *rx_bi = &rx_ring->rx_bi[i];
672 
673 		if (!rx_bi->page)
674 			continue;
675 
676 		/* Invalidate cache lines that may have been written to by
677 		 * device so that we avoid corrupting memory.
678 		 */
679 		dma_sync_single_range_for_cpu(rx_ring->dev,
680 					      rx_bi->dma,
681 					      rx_bi->page_offset,
682 					      rx_ring->rx_buf_len,
683 					      DMA_FROM_DEVICE);
684 
685 		/* free resources associated with mapping */
686 		dma_unmap_page_attrs(rx_ring->dev, rx_bi->dma,
687 				     iavf_rx_pg_size(rx_ring),
688 				     DMA_FROM_DEVICE,
689 				     IAVF_RX_DMA_ATTR);
690 
691 		__page_frag_cache_drain(rx_bi->page, rx_bi->pagecnt_bias);
692 
693 		rx_bi->page = NULL;
694 		rx_bi->page_offset = 0;
695 	}
696 
697 	bi_size = sizeof(struct iavf_rx_buffer) * rx_ring->count;
698 	memset(rx_ring->rx_bi, 0, bi_size);
699 
700 	/* Zero out the descriptor ring */
701 	memset(rx_ring->desc, 0, rx_ring->size);
702 
703 	rx_ring->next_to_alloc = 0;
704 	rx_ring->next_to_clean = 0;
705 	rx_ring->next_to_use = 0;
706 }
707 
708 /**
709  * iavf_free_rx_resources - Free Rx resources
710  * @rx_ring: ring to clean the resources from
711  *
712  * Free all receive software resources
713  **/
714 void iavf_free_rx_resources(struct iavf_ring *rx_ring)
715 {
716 	iavf_clean_rx_ring(rx_ring);
717 	kfree(rx_ring->rx_bi);
718 	rx_ring->rx_bi = NULL;
719 
720 	if (rx_ring->desc) {
721 		dma_free_coherent(rx_ring->dev, rx_ring->size,
722 				  rx_ring->desc, rx_ring->dma);
723 		rx_ring->desc = NULL;
724 	}
725 }
726 
727 /**
728  * iavf_setup_rx_descriptors - Allocate Rx descriptors
729  * @rx_ring: Rx descriptor ring (for a specific queue) to setup
730  *
731  * Returns 0 on success, negative on failure
732  **/
733 int iavf_setup_rx_descriptors(struct iavf_ring *rx_ring)
734 {
735 	struct device *dev = rx_ring->dev;
736 	int bi_size;
737 
738 	/* warn if we are about to overwrite the pointer */
739 	WARN_ON(rx_ring->rx_bi);
740 	bi_size = sizeof(struct iavf_rx_buffer) * rx_ring->count;
741 	rx_ring->rx_bi = kzalloc(bi_size, GFP_KERNEL);
742 	if (!rx_ring->rx_bi)
743 		goto err;
744 
745 	u64_stats_init(&rx_ring->syncp);
746 
747 	/* Round up to nearest 4K */
748 	rx_ring->size = rx_ring->count * sizeof(union iavf_32byte_rx_desc);
749 	rx_ring->size = ALIGN(rx_ring->size, 4096);
750 	rx_ring->desc = dma_alloc_coherent(dev, rx_ring->size,
751 					   &rx_ring->dma, GFP_KERNEL);
752 
753 	if (!rx_ring->desc) {
754 		dev_info(dev, "Unable to allocate memory for the Rx descriptor ring, size=%d\n",
755 			 rx_ring->size);
756 		goto err;
757 	}
758 
759 	rx_ring->next_to_alloc = 0;
760 	rx_ring->next_to_clean = 0;
761 	rx_ring->next_to_use = 0;
762 
763 	return 0;
764 err:
765 	kfree(rx_ring->rx_bi);
766 	rx_ring->rx_bi = NULL;
767 	return -ENOMEM;
768 }
769 
770 /**
771  * iavf_release_rx_desc - Store the new tail and head values
772  * @rx_ring: ring to bump
773  * @val: new head index
774  **/
775 static inline void iavf_release_rx_desc(struct iavf_ring *rx_ring, u32 val)
776 {
777 	rx_ring->next_to_use = val;
778 
779 	/* update next to alloc since we have filled the ring */
780 	rx_ring->next_to_alloc = val;
781 
782 	/* Force memory writes to complete before letting h/w
783 	 * know there are new descriptors to fetch.  (Only
784 	 * applicable for weak-ordered memory model archs,
785 	 * such as IA-64).
786 	 */
787 	wmb();
788 	writel(val, rx_ring->tail);
789 }
790 
791 /**
792  * iavf_rx_offset - Return expected offset into page to access data
793  * @rx_ring: Ring we are requesting offset of
794  *
795  * Returns the offset value for ring into the data buffer.
796  */
797 static inline unsigned int iavf_rx_offset(struct iavf_ring *rx_ring)
798 {
799 	return ring_uses_build_skb(rx_ring) ? IAVF_SKB_PAD : 0;
800 }
801 
802 /**
803  * iavf_alloc_mapped_page - recycle or make a new page
804  * @rx_ring: ring to use
805  * @bi: rx_buffer struct to modify
806  *
807  * Returns true if the page was successfully allocated or
808  * reused.
809  **/
810 static bool iavf_alloc_mapped_page(struct iavf_ring *rx_ring,
811 				   struct iavf_rx_buffer *bi)
812 {
813 	struct page *page = bi->page;
814 	dma_addr_t dma;
815 
816 	/* since we are recycling buffers we should seldom need to alloc */
817 	if (likely(page)) {
818 		rx_ring->rx_stats.page_reuse_count++;
819 		return true;
820 	}
821 
822 	/* alloc new page for storage */
823 	page = dev_alloc_pages(iavf_rx_pg_order(rx_ring));
824 	if (unlikely(!page)) {
825 		rx_ring->rx_stats.alloc_page_failed++;
826 		return false;
827 	}
828 
829 	/* map page for use */
830 	dma = dma_map_page_attrs(rx_ring->dev, page, 0,
831 				 iavf_rx_pg_size(rx_ring),
832 				 DMA_FROM_DEVICE,
833 				 IAVF_RX_DMA_ATTR);
834 
835 	/* if mapping failed free memory back to system since
836 	 * there isn't much point in holding memory we can't use
837 	 */
838 	if (dma_mapping_error(rx_ring->dev, dma)) {
839 		__free_pages(page, iavf_rx_pg_order(rx_ring));
840 		rx_ring->rx_stats.alloc_page_failed++;
841 		return false;
842 	}
843 
844 	bi->dma = dma;
845 	bi->page = page;
846 	bi->page_offset = iavf_rx_offset(rx_ring);
847 
848 	/* initialize pagecnt_bias to 1 representing we fully own page */
849 	bi->pagecnt_bias = 1;
850 
851 	return true;
852 }
853 
854 /**
855  * iavf_receive_skb - Send a completed packet up the stack
856  * @rx_ring:  rx ring in play
857  * @skb: packet to send up
858  * @vlan_tag: vlan tag for packet
859  **/
860 static void iavf_receive_skb(struct iavf_ring *rx_ring,
861 			     struct sk_buff *skb, u16 vlan_tag)
862 {
863 	struct iavf_q_vector *q_vector = rx_ring->q_vector;
864 
865 	if ((rx_ring->netdev->features & NETIF_F_HW_VLAN_CTAG_RX) &&
866 	    (vlan_tag & VLAN_VID_MASK))
867 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), vlan_tag);
868 
869 	napi_gro_receive(&q_vector->napi, skb);
870 }
871 
872 /**
873  * iavf_alloc_rx_buffers - Replace used receive buffers
874  * @rx_ring: ring to place buffers on
875  * @cleaned_count: number of buffers to replace
876  *
877  * Returns false if all allocations were successful, true if any fail
878  **/
879 bool iavf_alloc_rx_buffers(struct iavf_ring *rx_ring, u16 cleaned_count)
880 {
881 	u16 ntu = rx_ring->next_to_use;
882 	union iavf_rx_desc *rx_desc;
883 	struct iavf_rx_buffer *bi;
884 
885 	/* do nothing if no valid netdev defined */
886 	if (!rx_ring->netdev || !cleaned_count)
887 		return false;
888 
889 	rx_desc = IAVF_RX_DESC(rx_ring, ntu);
890 	bi = &rx_ring->rx_bi[ntu];
891 
892 	do {
893 		if (!iavf_alloc_mapped_page(rx_ring, bi))
894 			goto no_buffers;
895 
896 		/* sync the buffer for use by the device */
897 		dma_sync_single_range_for_device(rx_ring->dev, bi->dma,
898 						 bi->page_offset,
899 						 rx_ring->rx_buf_len,
900 						 DMA_FROM_DEVICE);
901 
902 		/* Refresh the desc even if buffer_addrs didn't change
903 		 * because each write-back erases this info.
904 		 */
905 		rx_desc->read.pkt_addr = cpu_to_le64(bi->dma + bi->page_offset);
906 
907 		rx_desc++;
908 		bi++;
909 		ntu++;
910 		if (unlikely(ntu == rx_ring->count)) {
911 			rx_desc = IAVF_RX_DESC(rx_ring, 0);
912 			bi = rx_ring->rx_bi;
913 			ntu = 0;
914 		}
915 
916 		/* clear the status bits for the next_to_use descriptor */
917 		rx_desc->wb.qword1.status_error_len = 0;
918 
919 		cleaned_count--;
920 	} while (cleaned_count);
921 
922 	if (rx_ring->next_to_use != ntu)
923 		iavf_release_rx_desc(rx_ring, ntu);
924 
925 	return false;
926 
927 no_buffers:
928 	if (rx_ring->next_to_use != ntu)
929 		iavf_release_rx_desc(rx_ring, ntu);
930 
931 	/* make sure to come back via polling to try again after
932 	 * allocation failure
933 	 */
934 	return true;
935 }
936 
937 /**
938  * iavf_rx_checksum - Indicate in skb if hw indicated a good cksum
939  * @vsi: the VSI we care about
940  * @skb: skb currently being received and modified
941  * @rx_desc: the receive descriptor
942  **/
943 static inline void iavf_rx_checksum(struct iavf_vsi *vsi,
944 				    struct sk_buff *skb,
945 				    union iavf_rx_desc *rx_desc)
946 {
947 	struct iavf_rx_ptype_decoded decoded;
948 	u32 rx_error, rx_status;
949 	bool ipv4, ipv6;
950 	u8 ptype;
951 	u64 qword;
952 
953 	qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
954 	ptype = (qword & IAVF_RXD_QW1_PTYPE_MASK) >> IAVF_RXD_QW1_PTYPE_SHIFT;
955 	rx_error = (qword & IAVF_RXD_QW1_ERROR_MASK) >>
956 		   IAVF_RXD_QW1_ERROR_SHIFT;
957 	rx_status = (qword & IAVF_RXD_QW1_STATUS_MASK) >>
958 		    IAVF_RXD_QW1_STATUS_SHIFT;
959 	decoded = decode_rx_desc_ptype(ptype);
960 
961 	skb->ip_summed = CHECKSUM_NONE;
962 
963 	skb_checksum_none_assert(skb);
964 
965 	/* Rx csum enabled and ip headers found? */
966 	if (!(vsi->netdev->features & NETIF_F_RXCSUM))
967 		return;
968 
969 	/* did the hardware decode the packet and checksum? */
970 	if (!(rx_status & BIT(IAVF_RX_DESC_STATUS_L3L4P_SHIFT)))
971 		return;
972 
973 	/* both known and outer_ip must be set for the below code to work */
974 	if (!(decoded.known && decoded.outer_ip))
975 		return;
976 
977 	ipv4 = (decoded.outer_ip == IAVF_RX_PTYPE_OUTER_IP) &&
978 	       (decoded.outer_ip_ver == IAVF_RX_PTYPE_OUTER_IPV4);
979 	ipv6 = (decoded.outer_ip == IAVF_RX_PTYPE_OUTER_IP) &&
980 	       (decoded.outer_ip_ver == IAVF_RX_PTYPE_OUTER_IPV6);
981 
982 	if (ipv4 &&
983 	    (rx_error & (BIT(IAVF_RX_DESC_ERROR_IPE_SHIFT) |
984 			 BIT(IAVF_RX_DESC_ERROR_EIPE_SHIFT))))
985 		goto checksum_fail;
986 
987 	/* likely incorrect csum if alternate IP extension headers found */
988 	if (ipv6 &&
989 	    rx_status & BIT(IAVF_RX_DESC_STATUS_IPV6EXADD_SHIFT))
990 		/* don't increment checksum err here, non-fatal err */
991 		return;
992 
993 	/* there was some L4 error, count error and punt packet to the stack */
994 	if (rx_error & BIT(IAVF_RX_DESC_ERROR_L4E_SHIFT))
995 		goto checksum_fail;
996 
997 	/* handle packets that were not able to be checksummed due
998 	 * to arrival speed, in this case the stack can compute
999 	 * the csum.
1000 	 */
1001 	if (rx_error & BIT(IAVF_RX_DESC_ERROR_PPRS_SHIFT))
1002 		return;
1003 
1004 	/* Only report checksum unnecessary for TCP, UDP, or SCTP */
1005 	switch (decoded.inner_prot) {
1006 	case IAVF_RX_PTYPE_INNER_PROT_TCP:
1007 	case IAVF_RX_PTYPE_INNER_PROT_UDP:
1008 	case IAVF_RX_PTYPE_INNER_PROT_SCTP:
1009 		skb->ip_summed = CHECKSUM_UNNECESSARY;
1010 		fallthrough;
1011 	default:
1012 		break;
1013 	}
1014 
1015 	return;
1016 
1017 checksum_fail:
1018 	vsi->back->hw_csum_rx_error++;
1019 }
1020 
1021 /**
1022  * iavf_ptype_to_htype - get a hash type
1023  * @ptype: the ptype value from the descriptor
1024  *
1025  * Returns a hash type to be used by skb_set_hash
1026  **/
1027 static inline int iavf_ptype_to_htype(u8 ptype)
1028 {
1029 	struct iavf_rx_ptype_decoded decoded = decode_rx_desc_ptype(ptype);
1030 
1031 	if (!decoded.known)
1032 		return PKT_HASH_TYPE_NONE;
1033 
1034 	if (decoded.outer_ip == IAVF_RX_PTYPE_OUTER_IP &&
1035 	    decoded.payload_layer == IAVF_RX_PTYPE_PAYLOAD_LAYER_PAY4)
1036 		return PKT_HASH_TYPE_L4;
1037 	else if (decoded.outer_ip == IAVF_RX_PTYPE_OUTER_IP &&
1038 		 decoded.payload_layer == IAVF_RX_PTYPE_PAYLOAD_LAYER_PAY3)
1039 		return PKT_HASH_TYPE_L3;
1040 	else
1041 		return PKT_HASH_TYPE_L2;
1042 }
1043 
1044 /**
1045  * iavf_rx_hash - set the hash value in the skb
1046  * @ring: descriptor ring
1047  * @rx_desc: specific descriptor
1048  * @skb: skb currently being received and modified
1049  * @rx_ptype: Rx packet type
1050  **/
1051 static inline void iavf_rx_hash(struct iavf_ring *ring,
1052 				union iavf_rx_desc *rx_desc,
1053 				struct sk_buff *skb,
1054 				u8 rx_ptype)
1055 {
1056 	u32 hash;
1057 	const __le64 rss_mask =
1058 		cpu_to_le64((u64)IAVF_RX_DESC_FLTSTAT_RSS_HASH <<
1059 			    IAVF_RX_DESC_STATUS_FLTSTAT_SHIFT);
1060 
1061 	if (ring->netdev->features & NETIF_F_RXHASH)
1062 		return;
1063 
1064 	if ((rx_desc->wb.qword1.status_error_len & rss_mask) == rss_mask) {
1065 		hash = le32_to_cpu(rx_desc->wb.qword0.hi_dword.rss);
1066 		skb_set_hash(skb, hash, iavf_ptype_to_htype(rx_ptype));
1067 	}
1068 }
1069 
1070 /**
1071  * iavf_process_skb_fields - Populate skb header fields from Rx descriptor
1072  * @rx_ring: rx descriptor ring packet is being transacted on
1073  * @rx_desc: pointer to the EOP Rx descriptor
1074  * @skb: pointer to current skb being populated
1075  * @rx_ptype: the packet type decoded by hardware
1076  *
1077  * This function checks the ring, descriptor, and packet information in
1078  * order to populate the hash, checksum, VLAN, protocol, and
1079  * other fields within the skb.
1080  **/
1081 static inline
1082 void iavf_process_skb_fields(struct iavf_ring *rx_ring,
1083 			     union iavf_rx_desc *rx_desc, struct sk_buff *skb,
1084 			     u8 rx_ptype)
1085 {
1086 	iavf_rx_hash(rx_ring, rx_desc, skb, rx_ptype);
1087 
1088 	iavf_rx_checksum(rx_ring->vsi, skb, rx_desc);
1089 
1090 	skb_record_rx_queue(skb, rx_ring->queue_index);
1091 
1092 	/* modifies the skb - consumes the enet header */
1093 	skb->protocol = eth_type_trans(skb, rx_ring->netdev);
1094 }
1095 
1096 /**
1097  * iavf_cleanup_headers - Correct empty headers
1098  * @rx_ring: rx descriptor ring packet is being transacted on
1099  * @skb: pointer to current skb being fixed
1100  *
1101  * Also address the case where we are pulling data in on pages only
1102  * and as such no data is present in the skb header.
1103  *
1104  * In addition if skb is not at least 60 bytes we need to pad it so that
1105  * it is large enough to qualify as a valid Ethernet frame.
1106  *
1107  * Returns true if an error was encountered and skb was freed.
1108  **/
1109 static bool iavf_cleanup_headers(struct iavf_ring *rx_ring, struct sk_buff *skb)
1110 {
1111 	/* if eth_skb_pad returns an error the skb was freed */
1112 	if (eth_skb_pad(skb))
1113 		return true;
1114 
1115 	return false;
1116 }
1117 
1118 /**
1119  * iavf_reuse_rx_page - page flip buffer and store it back on the ring
1120  * @rx_ring: rx descriptor ring to store buffers on
1121  * @old_buff: donor buffer to have page reused
1122  *
1123  * Synchronizes page for reuse by the adapter
1124  **/
1125 static void iavf_reuse_rx_page(struct iavf_ring *rx_ring,
1126 			       struct iavf_rx_buffer *old_buff)
1127 {
1128 	struct iavf_rx_buffer *new_buff;
1129 	u16 nta = rx_ring->next_to_alloc;
1130 
1131 	new_buff = &rx_ring->rx_bi[nta];
1132 
1133 	/* update, and store next to alloc */
1134 	nta++;
1135 	rx_ring->next_to_alloc = (nta < rx_ring->count) ? nta : 0;
1136 
1137 	/* transfer page from old buffer to new buffer */
1138 	new_buff->dma		= old_buff->dma;
1139 	new_buff->page		= old_buff->page;
1140 	new_buff->page_offset	= old_buff->page_offset;
1141 	new_buff->pagecnt_bias	= old_buff->pagecnt_bias;
1142 }
1143 
1144 /**
1145  * iavf_page_is_reusable - check if any reuse is possible
1146  * @page: page struct to check
1147  *
1148  * A page is not reusable if it was allocated under low memory
1149  * conditions, or it's not in the same NUMA node as this CPU.
1150  */
1151 static inline bool iavf_page_is_reusable(struct page *page)
1152 {
1153 	return (page_to_nid(page) == numa_mem_id()) &&
1154 		!page_is_pfmemalloc(page);
1155 }
1156 
1157 /**
1158  * iavf_can_reuse_rx_page - Determine if this page can be reused by
1159  * the adapter for another receive
1160  *
1161  * @rx_buffer: buffer containing the page
1162  *
1163  * If page is reusable, rx_buffer->page_offset is adjusted to point to
1164  * an unused region in the page.
1165  *
1166  * For small pages, @truesize will be a constant value, half the size
1167  * of the memory at page.  We'll attempt to alternate between high and
1168  * low halves of the page, with one half ready for use by the hardware
1169  * and the other half being consumed by the stack.  We use the page
1170  * ref count to determine whether the stack has finished consuming the
1171  * portion of this page that was passed up with a previous packet.  If
1172  * the page ref count is >1, we'll assume the "other" half page is
1173  * still busy, and this page cannot be reused.
1174  *
1175  * For larger pages, @truesize will be the actual space used by the
1176  * received packet (adjusted upward to an even multiple of the cache
1177  * line size).  This will advance through the page by the amount
1178  * actually consumed by the received packets while there is still
1179  * space for a buffer.  Each region of larger pages will be used at
1180  * most once, after which the page will not be reused.
1181  *
1182  * In either case, if the page is reusable its refcount is increased.
1183  **/
1184 static bool iavf_can_reuse_rx_page(struct iavf_rx_buffer *rx_buffer)
1185 {
1186 	unsigned int pagecnt_bias = rx_buffer->pagecnt_bias;
1187 	struct page *page = rx_buffer->page;
1188 
1189 	/* Is any reuse possible? */
1190 	if (unlikely(!iavf_page_is_reusable(page)))
1191 		return false;
1192 
1193 #if (PAGE_SIZE < 8192)
1194 	/* if we are only owner of page we can reuse it */
1195 	if (unlikely((page_count(page) - pagecnt_bias) > 1))
1196 		return false;
1197 #else
1198 #define IAVF_LAST_OFFSET \
1199 	(SKB_WITH_OVERHEAD(PAGE_SIZE) - IAVF_RXBUFFER_2048)
1200 	if (rx_buffer->page_offset > IAVF_LAST_OFFSET)
1201 		return false;
1202 #endif
1203 
1204 	/* If we have drained the page fragment pool we need to update
1205 	 * the pagecnt_bias and page count so that we fully restock the
1206 	 * number of references the driver holds.
1207 	 */
1208 	if (unlikely(!pagecnt_bias)) {
1209 		page_ref_add(page, USHRT_MAX);
1210 		rx_buffer->pagecnt_bias = USHRT_MAX;
1211 	}
1212 
1213 	return true;
1214 }
1215 
1216 /**
1217  * iavf_add_rx_frag - Add contents of Rx buffer to sk_buff
1218  * @rx_ring: rx descriptor ring to transact packets on
1219  * @rx_buffer: buffer containing page to add
1220  * @skb: sk_buff to place the data into
1221  * @size: packet length from rx_desc
1222  *
1223  * This function will add the data contained in rx_buffer->page to the skb.
1224  * It will just attach the page as a frag to the skb.
1225  *
1226  * The function will then update the page offset.
1227  **/
1228 static void iavf_add_rx_frag(struct iavf_ring *rx_ring,
1229 			     struct iavf_rx_buffer *rx_buffer,
1230 			     struct sk_buff *skb,
1231 			     unsigned int size)
1232 {
1233 #if (PAGE_SIZE < 8192)
1234 	unsigned int truesize = iavf_rx_pg_size(rx_ring) / 2;
1235 #else
1236 	unsigned int truesize = SKB_DATA_ALIGN(size + iavf_rx_offset(rx_ring));
1237 #endif
1238 
1239 	if (!size)
1240 		return;
1241 
1242 	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, rx_buffer->page,
1243 			rx_buffer->page_offset, size, truesize);
1244 
1245 	/* page is being used so we must update the page offset */
1246 #if (PAGE_SIZE < 8192)
1247 	rx_buffer->page_offset ^= truesize;
1248 #else
1249 	rx_buffer->page_offset += truesize;
1250 #endif
1251 }
1252 
1253 /**
1254  * iavf_get_rx_buffer - Fetch Rx buffer and synchronize data for use
1255  * @rx_ring: rx descriptor ring to transact packets on
1256  * @size: size of buffer to add to skb
1257  *
1258  * This function will pull an Rx buffer from the ring and synchronize it
1259  * for use by the CPU.
1260  */
1261 static struct iavf_rx_buffer *iavf_get_rx_buffer(struct iavf_ring *rx_ring,
1262 						 const unsigned int size)
1263 {
1264 	struct iavf_rx_buffer *rx_buffer;
1265 
1266 	if (!size)
1267 		return NULL;
1268 
1269 	rx_buffer = &rx_ring->rx_bi[rx_ring->next_to_clean];
1270 	prefetchw(rx_buffer->page);
1271 
1272 	/* we are reusing so sync this buffer for CPU use */
1273 	dma_sync_single_range_for_cpu(rx_ring->dev,
1274 				      rx_buffer->dma,
1275 				      rx_buffer->page_offset,
1276 				      size,
1277 				      DMA_FROM_DEVICE);
1278 
1279 	/* We have pulled a buffer for use, so decrement pagecnt_bias */
1280 	rx_buffer->pagecnt_bias--;
1281 
1282 	return rx_buffer;
1283 }
1284 
1285 /**
1286  * iavf_construct_skb - Allocate skb and populate it
1287  * @rx_ring: rx descriptor ring to transact packets on
1288  * @rx_buffer: rx buffer to pull data from
1289  * @size: size of buffer to add to skb
1290  *
1291  * This function allocates an skb.  It then populates it with the page
1292  * data from the current receive descriptor, taking care to set up the
1293  * skb correctly.
1294  */
1295 static struct sk_buff *iavf_construct_skb(struct iavf_ring *rx_ring,
1296 					  struct iavf_rx_buffer *rx_buffer,
1297 					  unsigned int size)
1298 {
1299 	void *va;
1300 #if (PAGE_SIZE < 8192)
1301 	unsigned int truesize = iavf_rx_pg_size(rx_ring) / 2;
1302 #else
1303 	unsigned int truesize = SKB_DATA_ALIGN(size);
1304 #endif
1305 	unsigned int headlen;
1306 	struct sk_buff *skb;
1307 
1308 	if (!rx_buffer)
1309 		return NULL;
1310 	/* prefetch first cache line of first page */
1311 	va = page_address(rx_buffer->page) + rx_buffer->page_offset;
1312 	prefetch(va);
1313 #if L1_CACHE_BYTES < 128
1314 	prefetch(va + L1_CACHE_BYTES);
1315 #endif
1316 
1317 	/* allocate a skb to store the frags */
1318 	skb = __napi_alloc_skb(&rx_ring->q_vector->napi,
1319 			       IAVF_RX_HDR_SIZE,
1320 			       GFP_ATOMIC | __GFP_NOWARN);
1321 	if (unlikely(!skb))
1322 		return NULL;
1323 
1324 	/* Determine available headroom for copy */
1325 	headlen = size;
1326 	if (headlen > IAVF_RX_HDR_SIZE)
1327 		headlen = eth_get_headlen(skb->dev, va, IAVF_RX_HDR_SIZE);
1328 
1329 	/* align pull length to size of long to optimize memcpy performance */
1330 	memcpy(__skb_put(skb, headlen), va, ALIGN(headlen, sizeof(long)));
1331 
1332 	/* update all of the pointers */
1333 	size -= headlen;
1334 	if (size) {
1335 		skb_add_rx_frag(skb, 0, rx_buffer->page,
1336 				rx_buffer->page_offset + headlen,
1337 				size, truesize);
1338 
1339 		/* buffer is used by skb, update page_offset */
1340 #if (PAGE_SIZE < 8192)
1341 		rx_buffer->page_offset ^= truesize;
1342 #else
1343 		rx_buffer->page_offset += truesize;
1344 #endif
1345 	} else {
1346 		/* buffer is unused, reset bias back to rx_buffer */
1347 		rx_buffer->pagecnt_bias++;
1348 	}
1349 
1350 	return skb;
1351 }
1352 
1353 /**
1354  * iavf_build_skb - Build skb around an existing buffer
1355  * @rx_ring: Rx descriptor ring to transact packets on
1356  * @rx_buffer: Rx buffer to pull data from
1357  * @size: size of buffer to add to skb
1358  *
1359  * This function builds an skb around an existing Rx buffer, taking care
1360  * to set up the skb correctly and avoid any memcpy overhead.
1361  */
1362 static struct sk_buff *iavf_build_skb(struct iavf_ring *rx_ring,
1363 				      struct iavf_rx_buffer *rx_buffer,
1364 				      unsigned int size)
1365 {
1366 	void *va;
1367 #if (PAGE_SIZE < 8192)
1368 	unsigned int truesize = iavf_rx_pg_size(rx_ring) / 2;
1369 #else
1370 	unsigned int truesize = SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) +
1371 				SKB_DATA_ALIGN(IAVF_SKB_PAD + size);
1372 #endif
1373 	struct sk_buff *skb;
1374 
1375 	if (!rx_buffer)
1376 		return NULL;
1377 	/* prefetch first cache line of first page */
1378 	va = page_address(rx_buffer->page) + rx_buffer->page_offset;
1379 	prefetch(va);
1380 #if L1_CACHE_BYTES < 128
1381 	prefetch(va + L1_CACHE_BYTES);
1382 #endif
1383 	/* build an skb around the page buffer */
1384 	skb = build_skb(va - IAVF_SKB_PAD, truesize);
1385 	if (unlikely(!skb))
1386 		return NULL;
1387 
1388 	/* update pointers within the skb to store the data */
1389 	skb_reserve(skb, IAVF_SKB_PAD);
1390 	__skb_put(skb, size);
1391 
1392 	/* buffer is used by skb, update page_offset */
1393 #if (PAGE_SIZE < 8192)
1394 	rx_buffer->page_offset ^= truesize;
1395 #else
1396 	rx_buffer->page_offset += truesize;
1397 #endif
1398 
1399 	return skb;
1400 }
1401 
1402 /**
1403  * iavf_put_rx_buffer - Clean up used buffer and either recycle or free
1404  * @rx_ring: rx descriptor ring to transact packets on
1405  * @rx_buffer: rx buffer to pull data from
1406  *
1407  * This function will clean up the contents of the rx_buffer.  It will
1408  * either recycle the buffer or unmap it and free the associated resources.
1409  */
1410 static void iavf_put_rx_buffer(struct iavf_ring *rx_ring,
1411 			       struct iavf_rx_buffer *rx_buffer)
1412 {
1413 	if (!rx_buffer)
1414 		return;
1415 
1416 	if (iavf_can_reuse_rx_page(rx_buffer)) {
1417 		/* hand second half of page back to the ring */
1418 		iavf_reuse_rx_page(rx_ring, rx_buffer);
1419 		rx_ring->rx_stats.page_reuse_count++;
1420 	} else {
1421 		/* we are not reusing the buffer so unmap it */
1422 		dma_unmap_page_attrs(rx_ring->dev, rx_buffer->dma,
1423 				     iavf_rx_pg_size(rx_ring),
1424 				     DMA_FROM_DEVICE, IAVF_RX_DMA_ATTR);
1425 		__page_frag_cache_drain(rx_buffer->page,
1426 					rx_buffer->pagecnt_bias);
1427 	}
1428 
1429 	/* clear contents of buffer_info */
1430 	rx_buffer->page = NULL;
1431 }
1432 
1433 /**
1434  * iavf_is_non_eop - process handling of non-EOP buffers
1435  * @rx_ring: Rx ring being processed
1436  * @rx_desc: Rx descriptor for current buffer
1437  * @skb: Current socket buffer containing buffer in progress
1438  *
1439  * This function updates next to clean.  If the buffer is an EOP buffer
1440  * this function exits returning false, otherwise it will place the
1441  * sk_buff in the next buffer to be chained and return true indicating
1442  * that this is in fact a non-EOP buffer.
1443  **/
1444 static bool iavf_is_non_eop(struct iavf_ring *rx_ring,
1445 			    union iavf_rx_desc *rx_desc,
1446 			    struct sk_buff *skb)
1447 {
1448 	u32 ntc = rx_ring->next_to_clean + 1;
1449 
1450 	/* fetch, update, and store next to clean */
1451 	ntc = (ntc < rx_ring->count) ? ntc : 0;
1452 	rx_ring->next_to_clean = ntc;
1453 
1454 	prefetch(IAVF_RX_DESC(rx_ring, ntc));
1455 
1456 	/* if we are the last buffer then there is nothing else to do */
1457 #define IAVF_RXD_EOF BIT(IAVF_RX_DESC_STATUS_EOF_SHIFT)
1458 	if (likely(iavf_test_staterr(rx_desc, IAVF_RXD_EOF)))
1459 		return false;
1460 
1461 	rx_ring->rx_stats.non_eop_descs++;
1462 
1463 	return true;
1464 }
1465 
1466 /**
1467  * iavf_clean_rx_irq - Clean completed descriptors from Rx ring - bounce buf
1468  * @rx_ring: rx descriptor ring to transact packets on
1469  * @budget: Total limit on number of packets to process
1470  *
1471  * This function provides a "bounce buffer" approach to Rx interrupt
1472  * processing.  The advantage to this is that on systems that have
1473  * expensive overhead for IOMMU access this provides a means of avoiding
1474  * it by maintaining the mapping of the page to the system.
1475  *
1476  * Returns amount of work completed
1477  **/
1478 static int iavf_clean_rx_irq(struct iavf_ring *rx_ring, int budget)
1479 {
1480 	unsigned int total_rx_bytes = 0, total_rx_packets = 0;
1481 	struct sk_buff *skb = rx_ring->skb;
1482 	u16 cleaned_count = IAVF_DESC_UNUSED(rx_ring);
1483 	bool failure = false;
1484 
1485 	while (likely(total_rx_packets < (unsigned int)budget)) {
1486 		struct iavf_rx_buffer *rx_buffer;
1487 		union iavf_rx_desc *rx_desc;
1488 		unsigned int size;
1489 		u16 vlan_tag;
1490 		u8 rx_ptype;
1491 		u64 qword;
1492 
1493 		/* return some buffers to hardware, one at a time is too slow */
1494 		if (cleaned_count >= IAVF_RX_BUFFER_WRITE) {
1495 			failure = failure ||
1496 				  iavf_alloc_rx_buffers(rx_ring, cleaned_count);
1497 			cleaned_count = 0;
1498 		}
1499 
1500 		rx_desc = IAVF_RX_DESC(rx_ring, rx_ring->next_to_clean);
1501 
1502 		/* status_error_len will always be zero for unused descriptors
1503 		 * because it's cleared in cleanup, and overlaps with hdr_addr
1504 		 * which is always zero because packet split isn't used, if the
1505 		 * hardware wrote DD then the length will be non-zero
1506 		 */
1507 		qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
1508 
1509 		/* This memory barrier is needed to keep us from reading
1510 		 * any other fields out of the rx_desc until we have
1511 		 * verified the descriptor has been written back.
1512 		 */
1513 		dma_rmb();
1514 #define IAVF_RXD_DD BIT(IAVF_RX_DESC_STATUS_DD_SHIFT)
1515 		if (!iavf_test_staterr(rx_desc, IAVF_RXD_DD))
1516 			break;
1517 
1518 		size = (qword & IAVF_RXD_QW1_LENGTH_PBUF_MASK) >>
1519 		       IAVF_RXD_QW1_LENGTH_PBUF_SHIFT;
1520 
1521 		iavf_trace(clean_rx_irq, rx_ring, rx_desc, skb);
1522 		rx_buffer = iavf_get_rx_buffer(rx_ring, size);
1523 
1524 		/* retrieve a buffer from the ring */
1525 		if (skb)
1526 			iavf_add_rx_frag(rx_ring, rx_buffer, skb, size);
1527 		else if (ring_uses_build_skb(rx_ring))
1528 			skb = iavf_build_skb(rx_ring, rx_buffer, size);
1529 		else
1530 			skb = iavf_construct_skb(rx_ring, rx_buffer, size);
1531 
1532 		/* exit if we failed to retrieve a buffer */
1533 		if (!skb) {
1534 			rx_ring->rx_stats.alloc_buff_failed++;
1535 			if (rx_buffer)
1536 				rx_buffer->pagecnt_bias++;
1537 			break;
1538 		}
1539 
1540 		iavf_put_rx_buffer(rx_ring, rx_buffer);
1541 		cleaned_count++;
1542 
1543 		if (iavf_is_non_eop(rx_ring, rx_desc, skb))
1544 			continue;
1545 
1546 		/* ERR_MASK will only have valid bits if EOP set, and
1547 		 * what we are doing here is actually checking
1548 		 * IAVF_RX_DESC_ERROR_RXE_SHIFT, since it is the zeroth bit in
1549 		 * the error field
1550 		 */
1551 		if (unlikely(iavf_test_staterr(rx_desc, BIT(IAVF_RXD_QW1_ERROR_SHIFT)))) {
1552 			dev_kfree_skb_any(skb);
1553 			skb = NULL;
1554 			continue;
1555 		}
1556 
1557 		if (iavf_cleanup_headers(rx_ring, skb)) {
1558 			skb = NULL;
1559 			continue;
1560 		}
1561 
1562 		/* probably a little skewed due to removing CRC */
1563 		total_rx_bytes += skb->len;
1564 
1565 		qword = le64_to_cpu(rx_desc->wb.qword1.status_error_len);
1566 		rx_ptype = (qword & IAVF_RXD_QW1_PTYPE_MASK) >>
1567 			   IAVF_RXD_QW1_PTYPE_SHIFT;
1568 
1569 		/* populate checksum, VLAN, and protocol */
1570 		iavf_process_skb_fields(rx_ring, rx_desc, skb, rx_ptype);
1571 
1572 
1573 		vlan_tag = (qword & BIT(IAVF_RX_DESC_STATUS_L2TAG1P_SHIFT)) ?
1574 			   le16_to_cpu(rx_desc->wb.qword0.lo_dword.l2tag1) : 0;
1575 
1576 		iavf_trace(clean_rx_irq_rx, rx_ring, rx_desc, skb);
1577 		iavf_receive_skb(rx_ring, skb, vlan_tag);
1578 		skb = NULL;
1579 
1580 		/* update budget accounting */
1581 		total_rx_packets++;
1582 	}
1583 
1584 	rx_ring->skb = skb;
1585 
1586 	u64_stats_update_begin(&rx_ring->syncp);
1587 	rx_ring->stats.packets += total_rx_packets;
1588 	rx_ring->stats.bytes += total_rx_bytes;
1589 	u64_stats_update_end(&rx_ring->syncp);
1590 	rx_ring->q_vector->rx.total_packets += total_rx_packets;
1591 	rx_ring->q_vector->rx.total_bytes += total_rx_bytes;
1592 
1593 	/* guarantee a trip back through this routine if there was a failure */
1594 	return failure ? budget : (int)total_rx_packets;
1595 }
1596 
1597 static inline u32 iavf_buildreg_itr(const int type, u16 itr)
1598 {
1599 	u32 val;
1600 
1601 	/* We don't bother with setting the CLEARPBA bit as the data sheet
1602 	 * points out doing so is "meaningless since it was already
1603 	 * auto-cleared". The auto-clearing happens when the interrupt is
1604 	 * asserted.
1605 	 *
1606 	 * Hardware errata 28 for also indicates that writing to a
1607 	 * xxINT_DYN_CTLx CSR with INTENA_MSK (bit 31) set to 0 will clear
1608 	 * an event in the PBA anyway so we need to rely on the automask
1609 	 * to hold pending events for us until the interrupt is re-enabled
1610 	 *
1611 	 * The itr value is reported in microseconds, and the register
1612 	 * value is recorded in 2 microsecond units. For this reason we
1613 	 * only need to shift by the interval shift - 1 instead of the
1614 	 * full value.
1615 	 */
1616 	itr &= IAVF_ITR_MASK;
1617 
1618 	val = IAVF_VFINT_DYN_CTLN1_INTENA_MASK |
1619 	      (type << IAVF_VFINT_DYN_CTLN1_ITR_INDX_SHIFT) |
1620 	      (itr << (IAVF_VFINT_DYN_CTLN1_INTERVAL_SHIFT - 1));
1621 
1622 	return val;
1623 }
1624 
1625 /* a small macro to shorten up some long lines */
1626 #define INTREG IAVF_VFINT_DYN_CTLN1
1627 
1628 /* The act of updating the ITR will cause it to immediately trigger. In order
1629  * to prevent this from throwing off adaptive update statistics we defer the
1630  * update so that it can only happen so often. So after either Tx or Rx are
1631  * updated we make the adaptive scheme wait until either the ITR completely
1632  * expires via the next_update expiration or we have been through at least
1633  * 3 interrupts.
1634  */
1635 #define ITR_COUNTDOWN_START 3
1636 
1637 /**
1638  * iavf_update_enable_itr - Update itr and re-enable MSIX interrupt
1639  * @vsi: the VSI we care about
1640  * @q_vector: q_vector for which itr is being updated and interrupt enabled
1641  *
1642  **/
1643 static inline void iavf_update_enable_itr(struct iavf_vsi *vsi,
1644 					  struct iavf_q_vector *q_vector)
1645 {
1646 	struct iavf_hw *hw = &vsi->back->hw;
1647 	u32 intval;
1648 
1649 	/* These will do nothing if dynamic updates are not enabled */
1650 	iavf_update_itr(q_vector, &q_vector->tx);
1651 	iavf_update_itr(q_vector, &q_vector->rx);
1652 
1653 	/* This block of logic allows us to get away with only updating
1654 	 * one ITR value with each interrupt. The idea is to perform a
1655 	 * pseudo-lazy update with the following criteria.
1656 	 *
1657 	 * 1. Rx is given higher priority than Tx if both are in same state
1658 	 * 2. If we must reduce an ITR that is given highest priority.
1659 	 * 3. We then give priority to increasing ITR based on amount.
1660 	 */
1661 	if (q_vector->rx.target_itr < q_vector->rx.current_itr) {
1662 		/* Rx ITR needs to be reduced, this is highest priority */
1663 		intval = iavf_buildreg_itr(IAVF_RX_ITR,
1664 					   q_vector->rx.target_itr);
1665 		q_vector->rx.current_itr = q_vector->rx.target_itr;
1666 		q_vector->itr_countdown = ITR_COUNTDOWN_START;
1667 	} else if ((q_vector->tx.target_itr < q_vector->tx.current_itr) ||
1668 		   ((q_vector->rx.target_itr - q_vector->rx.current_itr) <
1669 		    (q_vector->tx.target_itr - q_vector->tx.current_itr))) {
1670 		/* Tx ITR needs to be reduced, this is second priority
1671 		 * Tx ITR needs to be increased more than Rx, fourth priority
1672 		 */
1673 		intval = iavf_buildreg_itr(IAVF_TX_ITR,
1674 					   q_vector->tx.target_itr);
1675 		q_vector->tx.current_itr = q_vector->tx.target_itr;
1676 		q_vector->itr_countdown = ITR_COUNTDOWN_START;
1677 	} else if (q_vector->rx.current_itr != q_vector->rx.target_itr) {
1678 		/* Rx ITR needs to be increased, third priority */
1679 		intval = iavf_buildreg_itr(IAVF_RX_ITR,
1680 					   q_vector->rx.target_itr);
1681 		q_vector->rx.current_itr = q_vector->rx.target_itr;
1682 		q_vector->itr_countdown = ITR_COUNTDOWN_START;
1683 	} else {
1684 		/* No ITR update, lowest priority */
1685 		intval = iavf_buildreg_itr(IAVF_ITR_NONE, 0);
1686 		if (q_vector->itr_countdown)
1687 			q_vector->itr_countdown--;
1688 	}
1689 
1690 	if (!test_bit(__IAVF_VSI_DOWN, vsi->state))
1691 		wr32(hw, INTREG(q_vector->reg_idx), intval);
1692 }
1693 
1694 /**
1695  * iavf_napi_poll - NAPI polling Rx/Tx cleanup routine
1696  * @napi: napi struct with our devices info in it
1697  * @budget: amount of work driver is allowed to do this pass, in packets
1698  *
1699  * This function will clean all queues associated with a q_vector.
1700  *
1701  * Returns the amount of work done
1702  **/
1703 int iavf_napi_poll(struct napi_struct *napi, int budget)
1704 {
1705 	struct iavf_q_vector *q_vector =
1706 			       container_of(napi, struct iavf_q_vector, napi);
1707 	struct iavf_vsi *vsi = q_vector->vsi;
1708 	struct iavf_ring *ring;
1709 	bool clean_complete = true;
1710 	bool arm_wb = false;
1711 	int budget_per_ring;
1712 	int work_done = 0;
1713 
1714 	if (test_bit(__IAVF_VSI_DOWN, vsi->state)) {
1715 		napi_complete(napi);
1716 		return 0;
1717 	}
1718 
1719 	/* Since the actual Tx work is minimal, we can give the Tx a larger
1720 	 * budget and be more aggressive about cleaning up the Tx descriptors.
1721 	 */
1722 	iavf_for_each_ring(ring, q_vector->tx) {
1723 		if (!iavf_clean_tx_irq(vsi, ring, budget)) {
1724 			clean_complete = false;
1725 			continue;
1726 		}
1727 		arm_wb |= ring->arm_wb;
1728 		ring->arm_wb = false;
1729 	}
1730 
1731 	/* Handle case where we are called by netpoll with a budget of 0 */
1732 	if (budget <= 0)
1733 		goto tx_only;
1734 
1735 	/* We attempt to distribute budget to each Rx queue fairly, but don't
1736 	 * allow the budget to go below 1 because that would exit polling early.
1737 	 */
1738 	budget_per_ring = max(budget/q_vector->num_ringpairs, 1);
1739 
1740 	iavf_for_each_ring(ring, q_vector->rx) {
1741 		int cleaned = iavf_clean_rx_irq(ring, budget_per_ring);
1742 
1743 		work_done += cleaned;
1744 		/* if we clean as many as budgeted, we must not be done */
1745 		if (cleaned >= budget_per_ring)
1746 			clean_complete = false;
1747 	}
1748 
1749 	/* If work not completed, return budget and polling will return */
1750 	if (!clean_complete) {
1751 		int cpu_id = smp_processor_id();
1752 
1753 		/* It is possible that the interrupt affinity has changed but,
1754 		 * if the cpu is pegged at 100%, polling will never exit while
1755 		 * traffic continues and the interrupt will be stuck on this
1756 		 * cpu.  We check to make sure affinity is correct before we
1757 		 * continue to poll, otherwise we must stop polling so the
1758 		 * interrupt can move to the correct cpu.
1759 		 */
1760 		if (!cpumask_test_cpu(cpu_id, &q_vector->affinity_mask)) {
1761 			/* Tell napi that we are done polling */
1762 			napi_complete_done(napi, work_done);
1763 
1764 			/* Force an interrupt */
1765 			iavf_force_wb(vsi, q_vector);
1766 
1767 			/* Return budget-1 so that polling stops */
1768 			return budget - 1;
1769 		}
1770 tx_only:
1771 		if (arm_wb) {
1772 			q_vector->tx.ring[0].tx_stats.tx_force_wb++;
1773 			iavf_enable_wb_on_itr(vsi, q_vector);
1774 		}
1775 		return budget;
1776 	}
1777 
1778 	if (vsi->back->flags & IAVF_TXR_FLAGS_WB_ON_ITR)
1779 		q_vector->arm_wb_state = false;
1780 
1781 	/* Exit the polling mode, but don't re-enable interrupts if stack might
1782 	 * poll us due to busy-polling
1783 	 */
1784 	if (likely(napi_complete_done(napi, work_done)))
1785 		iavf_update_enable_itr(vsi, q_vector);
1786 
1787 	return min(work_done, budget - 1);
1788 }
1789 
1790 /**
1791  * iavf_tx_prepare_vlan_flags - prepare generic TX VLAN tagging flags for HW
1792  * @skb:     send buffer
1793  * @tx_ring: ring to send buffer on
1794  * @flags:   the tx flags to be set
1795  *
1796  * Checks the skb and set up correspondingly several generic transmit flags
1797  * related to VLAN tagging for the HW, such as VLAN, DCB, etc.
1798  *
1799  * Returns error code indicate the frame should be dropped upon error and the
1800  * otherwise  returns 0 to indicate the flags has been set properly.
1801  **/
1802 static inline int iavf_tx_prepare_vlan_flags(struct sk_buff *skb,
1803 					     struct iavf_ring *tx_ring,
1804 					     u32 *flags)
1805 {
1806 	__be16 protocol = skb->protocol;
1807 	u32  tx_flags = 0;
1808 
1809 	if (protocol == htons(ETH_P_8021Q) &&
1810 	    !(tx_ring->netdev->features & NETIF_F_HW_VLAN_CTAG_TX)) {
1811 		/* When HW VLAN acceleration is turned off by the user the
1812 		 * stack sets the protocol to 8021q so that the driver
1813 		 * can take any steps required to support the SW only
1814 		 * VLAN handling.  In our case the driver doesn't need
1815 		 * to take any further steps so just set the protocol
1816 		 * to the encapsulated ethertype.
1817 		 */
1818 		skb->protocol = vlan_get_protocol(skb);
1819 		goto out;
1820 	}
1821 
1822 	/* if we have a HW VLAN tag being added, default to the HW one */
1823 	if (skb_vlan_tag_present(skb)) {
1824 		tx_flags |= skb_vlan_tag_get(skb) << IAVF_TX_FLAGS_VLAN_SHIFT;
1825 		tx_flags |= IAVF_TX_FLAGS_HW_VLAN;
1826 	/* else if it is a SW VLAN, check the next protocol and store the tag */
1827 	} else if (protocol == htons(ETH_P_8021Q)) {
1828 		struct vlan_hdr *vhdr, _vhdr;
1829 
1830 		vhdr = skb_header_pointer(skb, ETH_HLEN, sizeof(_vhdr), &_vhdr);
1831 		if (!vhdr)
1832 			return -EINVAL;
1833 
1834 		protocol = vhdr->h_vlan_encapsulated_proto;
1835 		tx_flags |= ntohs(vhdr->h_vlan_TCI) << IAVF_TX_FLAGS_VLAN_SHIFT;
1836 		tx_flags |= IAVF_TX_FLAGS_SW_VLAN;
1837 	}
1838 
1839 out:
1840 	*flags = tx_flags;
1841 	return 0;
1842 }
1843 
1844 /**
1845  * iavf_tso - set up the tso context descriptor
1846  * @first:    pointer to first Tx buffer for xmit
1847  * @hdr_len:  ptr to the size of the packet header
1848  * @cd_type_cmd_tso_mss: Quad Word 1
1849  *
1850  * Returns 0 if no TSO can happen, 1 if tso is going, or error
1851  **/
1852 static int iavf_tso(struct iavf_tx_buffer *first, u8 *hdr_len,
1853 		    u64 *cd_type_cmd_tso_mss)
1854 {
1855 	struct sk_buff *skb = first->skb;
1856 	u64 cd_cmd, cd_tso_len, cd_mss;
1857 	union {
1858 		struct iphdr *v4;
1859 		struct ipv6hdr *v6;
1860 		unsigned char *hdr;
1861 	} ip;
1862 	union {
1863 		struct tcphdr *tcp;
1864 		struct udphdr *udp;
1865 		unsigned char *hdr;
1866 	} l4;
1867 	u32 paylen, l4_offset;
1868 	u16 gso_segs, gso_size;
1869 	int err;
1870 
1871 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1872 		return 0;
1873 
1874 	if (!skb_is_gso(skb))
1875 		return 0;
1876 
1877 	err = skb_cow_head(skb, 0);
1878 	if (err < 0)
1879 		return err;
1880 
1881 	ip.hdr = skb_network_header(skb);
1882 	l4.hdr = skb_transport_header(skb);
1883 
1884 	/* initialize outer IP header fields */
1885 	if (ip.v4->version == 4) {
1886 		ip.v4->tot_len = 0;
1887 		ip.v4->check = 0;
1888 	} else {
1889 		ip.v6->payload_len = 0;
1890 	}
1891 
1892 	if (skb_shinfo(skb)->gso_type & (SKB_GSO_GRE |
1893 					 SKB_GSO_GRE_CSUM |
1894 					 SKB_GSO_IPXIP4 |
1895 					 SKB_GSO_IPXIP6 |
1896 					 SKB_GSO_UDP_TUNNEL |
1897 					 SKB_GSO_UDP_TUNNEL_CSUM)) {
1898 		if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL) &&
1899 		    (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM)) {
1900 			l4.udp->len = 0;
1901 
1902 			/* determine offset of outer transport header */
1903 			l4_offset = l4.hdr - skb->data;
1904 
1905 			/* remove payload length from outer checksum */
1906 			paylen = skb->len - l4_offset;
1907 			csum_replace_by_diff(&l4.udp->check,
1908 					     (__force __wsum)htonl(paylen));
1909 		}
1910 
1911 		/* reset pointers to inner headers */
1912 		ip.hdr = skb_inner_network_header(skb);
1913 		l4.hdr = skb_inner_transport_header(skb);
1914 
1915 		/* initialize inner IP header fields */
1916 		if (ip.v4->version == 4) {
1917 			ip.v4->tot_len = 0;
1918 			ip.v4->check = 0;
1919 		} else {
1920 			ip.v6->payload_len = 0;
1921 		}
1922 	}
1923 
1924 	/* determine offset of inner transport header */
1925 	l4_offset = l4.hdr - skb->data;
1926 
1927 	/* remove payload length from inner checksum */
1928 	paylen = skb->len - l4_offset;
1929 	csum_replace_by_diff(&l4.tcp->check, (__force __wsum)htonl(paylen));
1930 
1931 	/* compute length of segmentation header */
1932 	*hdr_len = (l4.tcp->doff * 4) + l4_offset;
1933 
1934 	/* pull values out of skb_shinfo */
1935 	gso_size = skb_shinfo(skb)->gso_size;
1936 	gso_segs = skb_shinfo(skb)->gso_segs;
1937 
1938 	/* update GSO size and bytecount with header size */
1939 	first->gso_segs = gso_segs;
1940 	first->bytecount += (first->gso_segs - 1) * *hdr_len;
1941 
1942 	/* find the field values */
1943 	cd_cmd = IAVF_TX_CTX_DESC_TSO;
1944 	cd_tso_len = skb->len - *hdr_len;
1945 	cd_mss = gso_size;
1946 	*cd_type_cmd_tso_mss |= (cd_cmd << IAVF_TXD_CTX_QW1_CMD_SHIFT) |
1947 				(cd_tso_len << IAVF_TXD_CTX_QW1_TSO_LEN_SHIFT) |
1948 				(cd_mss << IAVF_TXD_CTX_QW1_MSS_SHIFT);
1949 	return 1;
1950 }
1951 
1952 /**
1953  * iavf_tx_enable_csum - Enable Tx checksum offloads
1954  * @skb: send buffer
1955  * @tx_flags: pointer to Tx flags currently set
1956  * @td_cmd: Tx descriptor command bits to set
1957  * @td_offset: Tx descriptor header offsets to set
1958  * @tx_ring: Tx descriptor ring
1959  * @cd_tunneling: ptr to context desc bits
1960  **/
1961 static int iavf_tx_enable_csum(struct sk_buff *skb, u32 *tx_flags,
1962 			       u32 *td_cmd, u32 *td_offset,
1963 			       struct iavf_ring *tx_ring,
1964 			       u32 *cd_tunneling)
1965 {
1966 	union {
1967 		struct iphdr *v4;
1968 		struct ipv6hdr *v6;
1969 		unsigned char *hdr;
1970 	} ip;
1971 	union {
1972 		struct tcphdr *tcp;
1973 		struct udphdr *udp;
1974 		unsigned char *hdr;
1975 	} l4;
1976 	unsigned char *exthdr;
1977 	u32 offset, cmd = 0;
1978 	__be16 frag_off;
1979 	u8 l4_proto = 0;
1980 
1981 	if (skb->ip_summed != CHECKSUM_PARTIAL)
1982 		return 0;
1983 
1984 	ip.hdr = skb_network_header(skb);
1985 	l4.hdr = skb_transport_header(skb);
1986 
1987 	/* compute outer L2 header size */
1988 	offset = ((ip.hdr - skb->data) / 2) << IAVF_TX_DESC_LENGTH_MACLEN_SHIFT;
1989 
1990 	if (skb->encapsulation) {
1991 		u32 tunnel = 0;
1992 		/* define outer network header type */
1993 		if (*tx_flags & IAVF_TX_FLAGS_IPV4) {
1994 			tunnel |= (*tx_flags & IAVF_TX_FLAGS_TSO) ?
1995 				  IAVF_TX_CTX_EXT_IP_IPV4 :
1996 				  IAVF_TX_CTX_EXT_IP_IPV4_NO_CSUM;
1997 
1998 			l4_proto = ip.v4->protocol;
1999 		} else if (*tx_flags & IAVF_TX_FLAGS_IPV6) {
2000 			tunnel |= IAVF_TX_CTX_EXT_IP_IPV6;
2001 
2002 			exthdr = ip.hdr + sizeof(*ip.v6);
2003 			l4_proto = ip.v6->nexthdr;
2004 			if (l4.hdr != exthdr)
2005 				ipv6_skip_exthdr(skb, exthdr - skb->data,
2006 						 &l4_proto, &frag_off);
2007 		}
2008 
2009 		/* define outer transport */
2010 		switch (l4_proto) {
2011 		case IPPROTO_UDP:
2012 			tunnel |= IAVF_TXD_CTX_UDP_TUNNELING;
2013 			*tx_flags |= IAVF_TX_FLAGS_VXLAN_TUNNEL;
2014 			break;
2015 		case IPPROTO_GRE:
2016 			tunnel |= IAVF_TXD_CTX_GRE_TUNNELING;
2017 			*tx_flags |= IAVF_TX_FLAGS_VXLAN_TUNNEL;
2018 			break;
2019 		case IPPROTO_IPIP:
2020 		case IPPROTO_IPV6:
2021 			*tx_flags |= IAVF_TX_FLAGS_VXLAN_TUNNEL;
2022 			l4.hdr = skb_inner_network_header(skb);
2023 			break;
2024 		default:
2025 			if (*tx_flags & IAVF_TX_FLAGS_TSO)
2026 				return -1;
2027 
2028 			skb_checksum_help(skb);
2029 			return 0;
2030 		}
2031 
2032 		/* compute outer L3 header size */
2033 		tunnel |= ((l4.hdr - ip.hdr) / 4) <<
2034 			  IAVF_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
2035 
2036 		/* switch IP header pointer from outer to inner header */
2037 		ip.hdr = skb_inner_network_header(skb);
2038 
2039 		/* compute tunnel header size */
2040 		tunnel |= ((ip.hdr - l4.hdr) / 2) <<
2041 			  IAVF_TXD_CTX_QW0_NATLEN_SHIFT;
2042 
2043 		/* indicate if we need to offload outer UDP header */
2044 		if ((*tx_flags & IAVF_TX_FLAGS_TSO) &&
2045 		    !(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL) &&
2046 		    (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM))
2047 			tunnel |= IAVF_TXD_CTX_QW0_L4T_CS_MASK;
2048 
2049 		/* record tunnel offload values */
2050 		*cd_tunneling |= tunnel;
2051 
2052 		/* switch L4 header pointer from outer to inner */
2053 		l4.hdr = skb_inner_transport_header(skb);
2054 		l4_proto = 0;
2055 
2056 		/* reset type as we transition from outer to inner headers */
2057 		*tx_flags &= ~(IAVF_TX_FLAGS_IPV4 | IAVF_TX_FLAGS_IPV6);
2058 		if (ip.v4->version == 4)
2059 			*tx_flags |= IAVF_TX_FLAGS_IPV4;
2060 		if (ip.v6->version == 6)
2061 			*tx_flags |= IAVF_TX_FLAGS_IPV6;
2062 	}
2063 
2064 	/* Enable IP checksum offloads */
2065 	if (*tx_flags & IAVF_TX_FLAGS_IPV4) {
2066 		l4_proto = ip.v4->protocol;
2067 		/* the stack computes the IP header already, the only time we
2068 		 * need the hardware to recompute it is in the case of TSO.
2069 		 */
2070 		cmd |= (*tx_flags & IAVF_TX_FLAGS_TSO) ?
2071 		       IAVF_TX_DESC_CMD_IIPT_IPV4_CSUM :
2072 		       IAVF_TX_DESC_CMD_IIPT_IPV4;
2073 	} else if (*tx_flags & IAVF_TX_FLAGS_IPV6) {
2074 		cmd |= IAVF_TX_DESC_CMD_IIPT_IPV6;
2075 
2076 		exthdr = ip.hdr + sizeof(*ip.v6);
2077 		l4_proto = ip.v6->nexthdr;
2078 		if (l4.hdr != exthdr)
2079 			ipv6_skip_exthdr(skb, exthdr - skb->data,
2080 					 &l4_proto, &frag_off);
2081 	}
2082 
2083 	/* compute inner L3 header size */
2084 	offset |= ((l4.hdr - ip.hdr) / 4) << IAVF_TX_DESC_LENGTH_IPLEN_SHIFT;
2085 
2086 	/* Enable L4 checksum offloads */
2087 	switch (l4_proto) {
2088 	case IPPROTO_TCP:
2089 		/* enable checksum offloads */
2090 		cmd |= IAVF_TX_DESC_CMD_L4T_EOFT_TCP;
2091 		offset |= l4.tcp->doff << IAVF_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
2092 		break;
2093 	case IPPROTO_SCTP:
2094 		/* enable SCTP checksum offload */
2095 		cmd |= IAVF_TX_DESC_CMD_L4T_EOFT_SCTP;
2096 		offset |= (sizeof(struct sctphdr) >> 2) <<
2097 			  IAVF_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
2098 		break;
2099 	case IPPROTO_UDP:
2100 		/* enable UDP checksum offload */
2101 		cmd |= IAVF_TX_DESC_CMD_L4T_EOFT_UDP;
2102 		offset |= (sizeof(struct udphdr) >> 2) <<
2103 			  IAVF_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
2104 		break;
2105 	default:
2106 		if (*tx_flags & IAVF_TX_FLAGS_TSO)
2107 			return -1;
2108 		skb_checksum_help(skb);
2109 		return 0;
2110 	}
2111 
2112 	*td_cmd |= cmd;
2113 	*td_offset |= offset;
2114 
2115 	return 1;
2116 }
2117 
2118 /**
2119  * iavf_create_tx_ctx Build the Tx context descriptor
2120  * @tx_ring:  ring to create the descriptor on
2121  * @cd_type_cmd_tso_mss: Quad Word 1
2122  * @cd_tunneling: Quad Word 0 - bits 0-31
2123  * @cd_l2tag2: Quad Word 0 - bits 32-63
2124  **/
2125 static void iavf_create_tx_ctx(struct iavf_ring *tx_ring,
2126 			       const u64 cd_type_cmd_tso_mss,
2127 			       const u32 cd_tunneling, const u32 cd_l2tag2)
2128 {
2129 	struct iavf_tx_context_desc *context_desc;
2130 	int i = tx_ring->next_to_use;
2131 
2132 	if ((cd_type_cmd_tso_mss == IAVF_TX_DESC_DTYPE_CONTEXT) &&
2133 	    !cd_tunneling && !cd_l2tag2)
2134 		return;
2135 
2136 	/* grab the next descriptor */
2137 	context_desc = IAVF_TX_CTXTDESC(tx_ring, i);
2138 
2139 	i++;
2140 	tx_ring->next_to_use = (i < tx_ring->count) ? i : 0;
2141 
2142 	/* cpu_to_le32 and assign to struct fields */
2143 	context_desc->tunneling_params = cpu_to_le32(cd_tunneling);
2144 	context_desc->l2tag2 = cpu_to_le16(cd_l2tag2);
2145 	context_desc->rsvd = cpu_to_le16(0);
2146 	context_desc->type_cmd_tso_mss = cpu_to_le64(cd_type_cmd_tso_mss);
2147 }
2148 
2149 /**
2150  * __iavf_chk_linearize - Check if there are more than 8 buffers per packet
2151  * @skb:      send buffer
2152  *
2153  * Note: Our HW can't DMA more than 8 buffers to build a packet on the wire
2154  * and so we need to figure out the cases where we need to linearize the skb.
2155  *
2156  * For TSO we need to count the TSO header and segment payload separately.
2157  * As such we need to check cases where we have 7 fragments or more as we
2158  * can potentially require 9 DMA transactions, 1 for the TSO header, 1 for
2159  * the segment payload in the first descriptor, and another 7 for the
2160  * fragments.
2161  **/
2162 bool __iavf_chk_linearize(struct sk_buff *skb)
2163 {
2164 	const skb_frag_t *frag, *stale;
2165 	int nr_frags, sum;
2166 
2167 	/* no need to check if number of frags is less than 7 */
2168 	nr_frags = skb_shinfo(skb)->nr_frags;
2169 	if (nr_frags < (IAVF_MAX_BUFFER_TXD - 1))
2170 		return false;
2171 
2172 	/* We need to walk through the list and validate that each group
2173 	 * of 6 fragments totals at least gso_size.
2174 	 */
2175 	nr_frags -= IAVF_MAX_BUFFER_TXD - 2;
2176 	frag = &skb_shinfo(skb)->frags[0];
2177 
2178 	/* Initialize size to the negative value of gso_size minus 1.  We
2179 	 * use this as the worst case scenerio in which the frag ahead
2180 	 * of us only provides one byte which is why we are limited to 6
2181 	 * descriptors for a single transmit as the header and previous
2182 	 * fragment are already consuming 2 descriptors.
2183 	 */
2184 	sum = 1 - skb_shinfo(skb)->gso_size;
2185 
2186 	/* Add size of frags 0 through 4 to create our initial sum */
2187 	sum += skb_frag_size(frag++);
2188 	sum += skb_frag_size(frag++);
2189 	sum += skb_frag_size(frag++);
2190 	sum += skb_frag_size(frag++);
2191 	sum += skb_frag_size(frag++);
2192 
2193 	/* Walk through fragments adding latest fragment, testing it, and
2194 	 * then removing stale fragments from the sum.
2195 	 */
2196 	for (stale = &skb_shinfo(skb)->frags[0];; stale++) {
2197 		int stale_size = skb_frag_size(stale);
2198 
2199 		sum += skb_frag_size(frag++);
2200 
2201 		/* The stale fragment may present us with a smaller
2202 		 * descriptor than the actual fragment size. To account
2203 		 * for that we need to remove all the data on the front and
2204 		 * figure out what the remainder would be in the last
2205 		 * descriptor associated with the fragment.
2206 		 */
2207 		if (stale_size > IAVF_MAX_DATA_PER_TXD) {
2208 			int align_pad = -(skb_frag_off(stale)) &
2209 					(IAVF_MAX_READ_REQ_SIZE - 1);
2210 
2211 			sum -= align_pad;
2212 			stale_size -= align_pad;
2213 
2214 			do {
2215 				sum -= IAVF_MAX_DATA_PER_TXD_ALIGNED;
2216 				stale_size -= IAVF_MAX_DATA_PER_TXD_ALIGNED;
2217 			} while (stale_size > IAVF_MAX_DATA_PER_TXD);
2218 		}
2219 
2220 		/* if sum is negative we failed to make sufficient progress */
2221 		if (sum < 0)
2222 			return true;
2223 
2224 		if (!nr_frags--)
2225 			break;
2226 
2227 		sum -= stale_size;
2228 	}
2229 
2230 	return false;
2231 }
2232 
2233 /**
2234  * __iavf_maybe_stop_tx - 2nd level check for tx stop conditions
2235  * @tx_ring: the ring to be checked
2236  * @size:    the size buffer we want to assure is available
2237  *
2238  * Returns -EBUSY if a stop is needed, else 0
2239  **/
2240 int __iavf_maybe_stop_tx(struct iavf_ring *tx_ring, int size)
2241 {
2242 	netif_stop_subqueue(tx_ring->netdev, tx_ring->queue_index);
2243 	/* Memory barrier before checking head and tail */
2244 	smp_mb();
2245 
2246 	/* Check again in a case another CPU has just made room available. */
2247 	if (likely(IAVF_DESC_UNUSED(tx_ring) < size))
2248 		return -EBUSY;
2249 
2250 	/* A reprieve! - use start_queue because it doesn't call schedule */
2251 	netif_start_subqueue(tx_ring->netdev, tx_ring->queue_index);
2252 	++tx_ring->tx_stats.restart_queue;
2253 	return 0;
2254 }
2255 
2256 /**
2257  * iavf_tx_map - Build the Tx descriptor
2258  * @tx_ring:  ring to send buffer on
2259  * @skb:      send buffer
2260  * @first:    first buffer info buffer to use
2261  * @tx_flags: collected send information
2262  * @hdr_len:  size of the packet header
2263  * @td_cmd:   the command field in the descriptor
2264  * @td_offset: offset for checksum or crc
2265  **/
2266 static inline void iavf_tx_map(struct iavf_ring *tx_ring, struct sk_buff *skb,
2267 			       struct iavf_tx_buffer *first, u32 tx_flags,
2268 			       const u8 hdr_len, u32 td_cmd, u32 td_offset)
2269 {
2270 	unsigned int data_len = skb->data_len;
2271 	unsigned int size = skb_headlen(skb);
2272 	skb_frag_t *frag;
2273 	struct iavf_tx_buffer *tx_bi;
2274 	struct iavf_tx_desc *tx_desc;
2275 	u16 i = tx_ring->next_to_use;
2276 	u32 td_tag = 0;
2277 	dma_addr_t dma;
2278 
2279 	if (tx_flags & IAVF_TX_FLAGS_HW_VLAN) {
2280 		td_cmd |= IAVF_TX_DESC_CMD_IL2TAG1;
2281 		td_tag = (tx_flags & IAVF_TX_FLAGS_VLAN_MASK) >>
2282 			 IAVF_TX_FLAGS_VLAN_SHIFT;
2283 	}
2284 
2285 	first->tx_flags = tx_flags;
2286 
2287 	dma = dma_map_single(tx_ring->dev, skb->data, size, DMA_TO_DEVICE);
2288 
2289 	tx_desc = IAVF_TX_DESC(tx_ring, i);
2290 	tx_bi = first;
2291 
2292 	for (frag = &skb_shinfo(skb)->frags[0];; frag++) {
2293 		unsigned int max_data = IAVF_MAX_DATA_PER_TXD_ALIGNED;
2294 
2295 		if (dma_mapping_error(tx_ring->dev, dma))
2296 			goto dma_error;
2297 
2298 		/* record length, and DMA address */
2299 		dma_unmap_len_set(tx_bi, len, size);
2300 		dma_unmap_addr_set(tx_bi, dma, dma);
2301 
2302 		/* align size to end of page */
2303 		max_data += -dma & (IAVF_MAX_READ_REQ_SIZE - 1);
2304 		tx_desc->buffer_addr = cpu_to_le64(dma);
2305 
2306 		while (unlikely(size > IAVF_MAX_DATA_PER_TXD)) {
2307 			tx_desc->cmd_type_offset_bsz =
2308 				build_ctob(td_cmd, td_offset,
2309 					   max_data, td_tag);
2310 
2311 			tx_desc++;
2312 			i++;
2313 
2314 			if (i == tx_ring->count) {
2315 				tx_desc = IAVF_TX_DESC(tx_ring, 0);
2316 				i = 0;
2317 			}
2318 
2319 			dma += max_data;
2320 			size -= max_data;
2321 
2322 			max_data = IAVF_MAX_DATA_PER_TXD_ALIGNED;
2323 			tx_desc->buffer_addr = cpu_to_le64(dma);
2324 		}
2325 
2326 		if (likely(!data_len))
2327 			break;
2328 
2329 		tx_desc->cmd_type_offset_bsz = build_ctob(td_cmd, td_offset,
2330 							  size, td_tag);
2331 
2332 		tx_desc++;
2333 		i++;
2334 
2335 		if (i == tx_ring->count) {
2336 			tx_desc = IAVF_TX_DESC(tx_ring, 0);
2337 			i = 0;
2338 		}
2339 
2340 		size = skb_frag_size(frag);
2341 		data_len -= size;
2342 
2343 		dma = skb_frag_dma_map(tx_ring->dev, frag, 0, size,
2344 				       DMA_TO_DEVICE);
2345 
2346 		tx_bi = &tx_ring->tx_bi[i];
2347 	}
2348 
2349 	netdev_tx_sent_queue(txring_txq(tx_ring), first->bytecount);
2350 
2351 	i++;
2352 	if (i == tx_ring->count)
2353 		i = 0;
2354 
2355 	tx_ring->next_to_use = i;
2356 
2357 	iavf_maybe_stop_tx(tx_ring, DESC_NEEDED);
2358 
2359 	/* write last descriptor with RS and EOP bits */
2360 	td_cmd |= IAVF_TXD_CMD;
2361 	tx_desc->cmd_type_offset_bsz =
2362 			build_ctob(td_cmd, td_offset, size, td_tag);
2363 
2364 	skb_tx_timestamp(skb);
2365 
2366 	/* Force memory writes to complete before letting h/w know there
2367 	 * are new descriptors to fetch.
2368 	 *
2369 	 * We also use this memory barrier to make certain all of the
2370 	 * status bits have been updated before next_to_watch is written.
2371 	 */
2372 	wmb();
2373 
2374 	/* set next_to_watch value indicating a packet is present */
2375 	first->next_to_watch = tx_desc;
2376 
2377 	/* notify HW of packet */
2378 	if (netif_xmit_stopped(txring_txq(tx_ring)) || !netdev_xmit_more()) {
2379 		writel(i, tx_ring->tail);
2380 	}
2381 
2382 	return;
2383 
2384 dma_error:
2385 	dev_info(tx_ring->dev, "TX DMA map failed\n");
2386 
2387 	/* clear dma mappings for failed tx_bi map */
2388 	for (;;) {
2389 		tx_bi = &tx_ring->tx_bi[i];
2390 		iavf_unmap_and_free_tx_resource(tx_ring, tx_bi);
2391 		if (tx_bi == first)
2392 			break;
2393 		if (i == 0)
2394 			i = tx_ring->count;
2395 		i--;
2396 	}
2397 
2398 	tx_ring->next_to_use = i;
2399 }
2400 
2401 /**
2402  * iavf_xmit_frame_ring - Sends buffer on Tx ring
2403  * @skb:     send buffer
2404  * @tx_ring: ring to send buffer on
2405  *
2406  * Returns NETDEV_TX_OK if sent, else an error code
2407  **/
2408 static netdev_tx_t iavf_xmit_frame_ring(struct sk_buff *skb,
2409 					struct iavf_ring *tx_ring)
2410 {
2411 	u64 cd_type_cmd_tso_mss = IAVF_TX_DESC_DTYPE_CONTEXT;
2412 	u32 cd_tunneling = 0, cd_l2tag2 = 0;
2413 	struct iavf_tx_buffer *first;
2414 	u32 td_offset = 0;
2415 	u32 tx_flags = 0;
2416 	__be16 protocol;
2417 	u32 td_cmd = 0;
2418 	u8 hdr_len = 0;
2419 	int tso, count;
2420 
2421 	/* prefetch the data, we'll need it later */
2422 	prefetch(skb->data);
2423 
2424 	iavf_trace(xmit_frame_ring, skb, tx_ring);
2425 
2426 	count = iavf_xmit_descriptor_count(skb);
2427 	if (iavf_chk_linearize(skb, count)) {
2428 		if (__skb_linearize(skb)) {
2429 			dev_kfree_skb_any(skb);
2430 			return NETDEV_TX_OK;
2431 		}
2432 		count = iavf_txd_use_count(skb->len);
2433 		tx_ring->tx_stats.tx_linearize++;
2434 	}
2435 
2436 	/* need: 1 descriptor per page * PAGE_SIZE/IAVF_MAX_DATA_PER_TXD,
2437 	 *       + 1 desc for skb_head_len/IAVF_MAX_DATA_PER_TXD,
2438 	 *       + 4 desc gap to avoid the cache line where head is,
2439 	 *       + 1 desc for context descriptor,
2440 	 * otherwise try next time
2441 	 */
2442 	if (iavf_maybe_stop_tx(tx_ring, count + 4 + 1)) {
2443 		tx_ring->tx_stats.tx_busy++;
2444 		return NETDEV_TX_BUSY;
2445 	}
2446 
2447 	/* record the location of the first descriptor for this packet */
2448 	first = &tx_ring->tx_bi[tx_ring->next_to_use];
2449 	first->skb = skb;
2450 	first->bytecount = skb->len;
2451 	first->gso_segs = 1;
2452 
2453 	/* prepare the xmit flags */
2454 	if (iavf_tx_prepare_vlan_flags(skb, tx_ring, &tx_flags))
2455 		goto out_drop;
2456 
2457 	/* obtain protocol of skb */
2458 	protocol = vlan_get_protocol(skb);
2459 
2460 	/* setup IPv4/IPv6 offloads */
2461 	if (protocol == htons(ETH_P_IP))
2462 		tx_flags |= IAVF_TX_FLAGS_IPV4;
2463 	else if (protocol == htons(ETH_P_IPV6))
2464 		tx_flags |= IAVF_TX_FLAGS_IPV6;
2465 
2466 	tso = iavf_tso(first, &hdr_len, &cd_type_cmd_tso_mss);
2467 
2468 	if (tso < 0)
2469 		goto out_drop;
2470 	else if (tso)
2471 		tx_flags |= IAVF_TX_FLAGS_TSO;
2472 
2473 	/* Always offload the checksum, since it's in the data descriptor */
2474 	tso = iavf_tx_enable_csum(skb, &tx_flags, &td_cmd, &td_offset,
2475 				  tx_ring, &cd_tunneling);
2476 	if (tso < 0)
2477 		goto out_drop;
2478 
2479 	/* always enable CRC insertion offload */
2480 	td_cmd |= IAVF_TX_DESC_CMD_ICRC;
2481 
2482 	iavf_create_tx_ctx(tx_ring, cd_type_cmd_tso_mss,
2483 			   cd_tunneling, cd_l2tag2);
2484 
2485 	iavf_tx_map(tx_ring, skb, first, tx_flags, hdr_len,
2486 		    td_cmd, td_offset);
2487 
2488 	return NETDEV_TX_OK;
2489 
2490 out_drop:
2491 	iavf_trace(xmit_frame_ring_drop, first->skb, tx_ring);
2492 	dev_kfree_skb_any(first->skb);
2493 	first->skb = NULL;
2494 	return NETDEV_TX_OK;
2495 }
2496 
2497 /**
2498  * iavf_xmit_frame - Selects the correct VSI and Tx queue to send buffer
2499  * @skb:    send buffer
2500  * @netdev: network interface device structure
2501  *
2502  * Returns NETDEV_TX_OK if sent, else an error code
2503  **/
2504 netdev_tx_t iavf_xmit_frame(struct sk_buff *skb, struct net_device *netdev)
2505 {
2506 	struct iavf_adapter *adapter = netdev_priv(netdev);
2507 	struct iavf_ring *tx_ring = &adapter->tx_rings[skb->queue_mapping];
2508 
2509 	/* hardware can't handle really short frames, hardware padding works
2510 	 * beyond this point
2511 	 */
2512 	if (unlikely(skb->len < IAVF_MIN_TX_LEN)) {
2513 		if (skb_pad(skb, IAVF_MIN_TX_LEN - skb->len))
2514 			return NETDEV_TX_OK;
2515 		skb->len = IAVF_MIN_TX_LEN;
2516 		skb_set_tail_pointer(skb, IAVF_MIN_TX_LEN);
2517 	}
2518 
2519 	return iavf_xmit_frame_ring(skb, tx_ring);
2520 }
2521