xref: /linux/drivers/net/ethernet/sfc/tx.c (revision c0c914eca7f251c70facc37dfebeaf176601918d)
1 /****************************************************************************
2  * Driver for Solarflare network controllers and boards
3  * Copyright 2005-2006 Fen Systems Ltd.
4  * Copyright 2005-2013 Solarflare Communications Inc.
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published
8  * by the Free Software Foundation, incorporated herein by reference.
9  */
10 
11 #include <linux/pci.h>
12 #include <linux/tcp.h>
13 #include <linux/ip.h>
14 #include <linux/in.h>
15 #include <linux/ipv6.h>
16 #include <linux/slab.h>
17 #include <net/ipv6.h>
18 #include <linux/if_ether.h>
19 #include <linux/highmem.h>
20 #include <linux/cache.h>
21 #include "net_driver.h"
22 #include "efx.h"
23 #include "io.h"
24 #include "nic.h"
25 #include "workarounds.h"
26 #include "ef10_regs.h"
27 
28 #ifdef EFX_USE_PIO
29 
30 #define EFX_PIOBUF_SIZE_MAX ER_DZ_TX_PIOBUF_SIZE
31 #define EFX_PIOBUF_SIZE_DEF ALIGN(256, L1_CACHE_BYTES)
32 unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF;
33 
34 #endif /* EFX_USE_PIO */
35 
36 static inline unsigned int
37 efx_tx_queue_get_insert_index(const struct efx_tx_queue *tx_queue)
38 {
39 	return tx_queue->insert_count & tx_queue->ptr_mask;
40 }
41 
42 static inline struct efx_tx_buffer *
43 __efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue)
44 {
45 	return &tx_queue->buffer[efx_tx_queue_get_insert_index(tx_queue)];
46 }
47 
48 static inline struct efx_tx_buffer *
49 efx_tx_queue_get_insert_buffer(const struct efx_tx_queue *tx_queue)
50 {
51 	struct efx_tx_buffer *buffer =
52 		__efx_tx_queue_get_insert_buffer(tx_queue);
53 
54 	EFX_BUG_ON_PARANOID(buffer->len);
55 	EFX_BUG_ON_PARANOID(buffer->flags);
56 	EFX_BUG_ON_PARANOID(buffer->unmap_len);
57 
58 	return buffer;
59 }
60 
61 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
62 			       struct efx_tx_buffer *buffer,
63 			       unsigned int *pkts_compl,
64 			       unsigned int *bytes_compl)
65 {
66 	if (buffer->unmap_len) {
67 		struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
68 		dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
69 		if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
70 			dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
71 					 DMA_TO_DEVICE);
72 		else
73 			dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
74 				       DMA_TO_DEVICE);
75 		buffer->unmap_len = 0;
76 	}
77 
78 	if (buffer->flags & EFX_TX_BUF_SKB) {
79 		(*pkts_compl)++;
80 		(*bytes_compl) += buffer->skb->len;
81 		dev_consume_skb_any((struct sk_buff *)buffer->skb);
82 		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
83 			   "TX queue %d transmission id %x complete\n",
84 			   tx_queue->queue, tx_queue->read_count);
85 	} else if (buffer->flags & EFX_TX_BUF_HEAP) {
86 		kfree(buffer->heap_buf);
87 	}
88 
89 	buffer->len = 0;
90 	buffer->flags = 0;
91 }
92 
93 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
94 			       struct sk_buff *skb);
95 
96 static inline unsigned
97 efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr)
98 {
99 	/* Depending on the NIC revision, we can use descriptor
100 	 * lengths up to 8K or 8K-1.  However, since PCI Express
101 	 * devices must split read requests at 4K boundaries, there is
102 	 * little benefit from using descriptors that cross those
103 	 * boundaries and we keep things simple by not doing so.
104 	 */
105 	unsigned len = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1;
106 
107 	/* Work around hardware bug for unaligned buffers. */
108 	if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf))
109 		len = min_t(unsigned, len, 512 - (dma_addr & 0xf));
110 
111 	return len;
112 }
113 
114 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
115 {
116 	/* Header and payload descriptor for each output segment, plus
117 	 * one for every input fragment boundary within a segment
118 	 */
119 	unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
120 
121 	/* Possibly one more per segment for the alignment workaround,
122 	 * or for option descriptors
123 	 */
124 	if (EFX_WORKAROUND_5391(efx) || efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
125 		max_descs += EFX_TSO_MAX_SEGS;
126 
127 	/* Possibly more for PCIe page boundaries within input fragments */
128 	if (PAGE_SIZE > EFX_PAGE_SIZE)
129 		max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
130 				   DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE));
131 
132 	return max_descs;
133 }
134 
135 static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
136 {
137 	/* We need to consider both queues that the net core sees as one */
138 	struct efx_tx_queue *txq2 = efx_tx_queue_partner(txq1);
139 	struct efx_nic *efx = txq1->efx;
140 	unsigned int fill_level;
141 
142 	fill_level = max(txq1->insert_count - txq1->old_read_count,
143 			 txq2->insert_count - txq2->old_read_count);
144 	if (likely(fill_level < efx->txq_stop_thresh))
145 		return;
146 
147 	/* We used the stale old_read_count above, which gives us a
148 	 * pessimistic estimate of the fill level (which may even
149 	 * validly be >= efx->txq_entries).  Now try again using
150 	 * read_count (more likely to be a cache miss).
151 	 *
152 	 * If we read read_count and then conditionally stop the
153 	 * queue, it is possible for the completion path to race with
154 	 * us and complete all outstanding descriptors in the middle,
155 	 * after which there will be no more completions to wake it.
156 	 * Therefore we stop the queue first, then read read_count
157 	 * (with a memory barrier to ensure the ordering), then
158 	 * restart the queue if the fill level turns out to be low
159 	 * enough.
160 	 */
161 	netif_tx_stop_queue(txq1->core_txq);
162 	smp_mb();
163 	txq1->old_read_count = ACCESS_ONCE(txq1->read_count);
164 	txq2->old_read_count = ACCESS_ONCE(txq2->read_count);
165 
166 	fill_level = max(txq1->insert_count - txq1->old_read_count,
167 			 txq2->insert_count - txq2->old_read_count);
168 	EFX_BUG_ON_PARANOID(fill_level >= efx->txq_entries);
169 	if (likely(fill_level < efx->txq_stop_thresh)) {
170 		smp_mb();
171 		if (likely(!efx->loopback_selftest))
172 			netif_tx_start_queue(txq1->core_txq);
173 	}
174 }
175 
176 #ifdef EFX_USE_PIO
177 
178 struct efx_short_copy_buffer {
179 	int used;
180 	u8 buf[L1_CACHE_BYTES];
181 };
182 
183 /* Copy to PIO, respecting that writes to PIO buffers must be dword aligned.
184  * Advances piobuf pointer. Leaves additional data in the copy buffer.
185  */
186 static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf,
187 				    u8 *data, int len,
188 				    struct efx_short_copy_buffer *copy_buf)
189 {
190 	int block_len = len & ~(sizeof(copy_buf->buf) - 1);
191 
192 	__iowrite64_copy(*piobuf, data, block_len >> 3);
193 	*piobuf += block_len;
194 	len -= block_len;
195 
196 	if (len) {
197 		data += block_len;
198 		BUG_ON(copy_buf->used);
199 		BUG_ON(len > sizeof(copy_buf->buf));
200 		memcpy(copy_buf->buf, data, len);
201 		copy_buf->used = len;
202 	}
203 }
204 
205 /* Copy to PIO, respecting dword alignment, popping data from copy buffer first.
206  * Advances piobuf pointer. Leaves additional data in the copy buffer.
207  */
208 static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf,
209 				       u8 *data, int len,
210 				       struct efx_short_copy_buffer *copy_buf)
211 {
212 	if (copy_buf->used) {
213 		/* if the copy buffer is partially full, fill it up and write */
214 		int copy_to_buf =
215 			min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len);
216 
217 		memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf);
218 		copy_buf->used += copy_to_buf;
219 
220 		/* if we didn't fill it up then we're done for now */
221 		if (copy_buf->used < sizeof(copy_buf->buf))
222 			return;
223 
224 		__iowrite64_copy(*piobuf, copy_buf->buf,
225 				 sizeof(copy_buf->buf) >> 3);
226 		*piobuf += sizeof(copy_buf->buf);
227 		data += copy_to_buf;
228 		len -= copy_to_buf;
229 		copy_buf->used = 0;
230 	}
231 
232 	efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf);
233 }
234 
235 static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf,
236 				  struct efx_short_copy_buffer *copy_buf)
237 {
238 	/* if there's anything in it, write the whole buffer, including junk */
239 	if (copy_buf->used)
240 		__iowrite64_copy(piobuf, copy_buf->buf,
241 				 sizeof(copy_buf->buf) >> 3);
242 }
243 
244 /* Traverse skb structure and copy fragments in to PIO buffer.
245  * Advances piobuf pointer.
246  */
247 static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb,
248 				     u8 __iomem **piobuf,
249 				     struct efx_short_copy_buffer *copy_buf)
250 {
251 	int i;
252 
253 	efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb),
254 				copy_buf);
255 
256 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
257 		skb_frag_t *f = &skb_shinfo(skb)->frags[i];
258 		u8 *vaddr;
259 
260 		vaddr = kmap_atomic(skb_frag_page(f));
261 
262 		efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + f->page_offset,
263 					   skb_frag_size(f), copy_buf);
264 		kunmap_atomic(vaddr);
265 	}
266 
267 	EFX_BUG_ON_PARANOID(skb_shinfo(skb)->frag_list);
268 }
269 
270 static struct efx_tx_buffer *
271 efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
272 {
273 	struct efx_tx_buffer *buffer =
274 		efx_tx_queue_get_insert_buffer(tx_queue);
275 	u8 __iomem *piobuf = tx_queue->piobuf;
276 
277 	/* Copy to PIO buffer. Ensure the writes are padded to the end
278 	 * of a cache line, as this is required for write-combining to be
279 	 * effective on at least x86.
280 	 */
281 
282 	if (skb_shinfo(skb)->nr_frags) {
283 		/* The size of the copy buffer will ensure all writes
284 		 * are the size of a cache line.
285 		 */
286 		struct efx_short_copy_buffer copy_buf;
287 
288 		copy_buf.used = 0;
289 
290 		efx_skb_copy_bits_to_pio(tx_queue->efx, skb,
291 					 &piobuf, &copy_buf);
292 		efx_flush_copy_buffer(tx_queue->efx, piobuf, &copy_buf);
293 	} else {
294 		/* Pad the write to the size of a cache line.
295 		 * We can do this because we know the skb_shared_info sruct is
296 		 * after the source, and the destination buffer is big enough.
297 		 */
298 		BUILD_BUG_ON(L1_CACHE_BYTES >
299 			     SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
300 		__iowrite64_copy(tx_queue->piobuf, skb->data,
301 				 ALIGN(skb->len, L1_CACHE_BYTES) >> 3);
302 	}
303 
304 	EFX_POPULATE_QWORD_5(buffer->option,
305 			     ESF_DZ_TX_DESC_IS_OPT, 1,
306 			     ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO,
307 			     ESF_DZ_TX_PIO_CONT, 0,
308 			     ESF_DZ_TX_PIO_BYTE_CNT, skb->len,
309 			     ESF_DZ_TX_PIO_BUF_ADDR,
310 			     tx_queue->piobuf_offset);
311 	++tx_queue->pio_packets;
312 	++tx_queue->insert_count;
313 	return buffer;
314 }
315 #endif /* EFX_USE_PIO */
316 
317 /*
318  * Add a socket buffer to a TX queue
319  *
320  * This maps all fragments of a socket buffer for DMA and adds them to
321  * the TX queue.  The queue's insert pointer will be incremented by
322  * the number of fragments in the socket buffer.
323  *
324  * If any DMA mapping fails, any mapped fragments will be unmapped,
325  * the queue's insert pointer will be restored to its original value.
326  *
327  * This function is split out from efx_hard_start_xmit to allow the
328  * loopback test to direct packets via specific TX queues.
329  *
330  * Returns NETDEV_TX_OK.
331  * You must hold netif_tx_lock() to call this function.
332  */
333 netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
334 {
335 	struct efx_nic *efx = tx_queue->efx;
336 	struct device *dma_dev = &efx->pci_dev->dev;
337 	struct efx_tx_buffer *buffer;
338 	unsigned int old_insert_count = tx_queue->insert_count;
339 	skb_frag_t *fragment;
340 	unsigned int len, unmap_len = 0;
341 	dma_addr_t dma_addr, unmap_addr = 0;
342 	unsigned int dma_len;
343 	unsigned short dma_flags;
344 	int i = 0;
345 
346 	if (skb_shinfo(skb)->gso_size)
347 		return efx_enqueue_skb_tso(tx_queue, skb);
348 
349 	/* Get size of the initial fragment */
350 	len = skb_headlen(skb);
351 
352 	/* Pad if necessary */
353 	if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
354 		EFX_BUG_ON_PARANOID(skb->data_len);
355 		len = 32 + 1;
356 		if (skb_pad(skb, len - skb->len))
357 			return NETDEV_TX_OK;
358 	}
359 
360 	/* Consider using PIO for short packets */
361 #ifdef EFX_USE_PIO
362 	if (skb->len <= efx_piobuf_size && !skb->xmit_more &&
363 	    efx_nic_may_tx_pio(tx_queue)) {
364 		buffer = efx_enqueue_skb_pio(tx_queue, skb);
365 		dma_flags = EFX_TX_BUF_OPTION;
366 		goto finish_packet;
367 	}
368 #endif
369 
370 	/* Map for DMA.  Use dma_map_single rather than dma_map_page
371 	 * since this is more efficient on machines with sparse
372 	 * memory.
373 	 */
374 	dma_flags = EFX_TX_BUF_MAP_SINGLE;
375 	dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE);
376 
377 	/* Process all fragments */
378 	while (1) {
379 		if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
380 			goto dma_err;
381 
382 		/* Store fields for marking in the per-fragment final
383 		 * descriptor */
384 		unmap_len = len;
385 		unmap_addr = dma_addr;
386 
387 		/* Add to TX queue, splitting across DMA boundaries */
388 		do {
389 			buffer = efx_tx_queue_get_insert_buffer(tx_queue);
390 
391 			dma_len = efx_max_tx_len(efx, dma_addr);
392 			if (likely(dma_len >= len))
393 				dma_len = len;
394 
395 			/* Fill out per descriptor fields */
396 			buffer->len = dma_len;
397 			buffer->dma_addr = dma_addr;
398 			buffer->flags = EFX_TX_BUF_CONT;
399 			len -= dma_len;
400 			dma_addr += dma_len;
401 			++tx_queue->insert_count;
402 		} while (len);
403 
404 		/* Transfer ownership of the unmapping to the final buffer */
405 		buffer->flags = EFX_TX_BUF_CONT | dma_flags;
406 		buffer->unmap_len = unmap_len;
407 		buffer->dma_offset = buffer->dma_addr - unmap_addr;
408 		unmap_len = 0;
409 
410 		/* Get address and size of next fragment */
411 		if (i >= skb_shinfo(skb)->nr_frags)
412 			break;
413 		fragment = &skb_shinfo(skb)->frags[i];
414 		len = skb_frag_size(fragment);
415 		i++;
416 		/* Map for DMA */
417 		dma_flags = 0;
418 		dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
419 					    DMA_TO_DEVICE);
420 	}
421 
422 	/* Transfer ownership of the skb to the final buffer */
423 #ifdef EFX_USE_PIO
424 finish_packet:
425 #endif
426 	buffer->skb = skb;
427 	buffer->flags = EFX_TX_BUF_SKB | dma_flags;
428 
429 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
430 
431 	efx_tx_maybe_stop_queue(tx_queue);
432 
433 	/* Pass off to hardware */
434 	if (!skb->xmit_more || netif_xmit_stopped(tx_queue->core_txq)) {
435 		struct efx_tx_queue *txq2 = efx_tx_queue_partner(tx_queue);
436 
437 		/* There could be packets left on the partner queue if those
438 		 * SKBs had skb->xmit_more set. If we do not push those they
439 		 * could be left for a long time and cause a netdev watchdog.
440 		 */
441 		if (txq2->xmit_more_available)
442 			efx_nic_push_buffers(txq2);
443 
444 		efx_nic_push_buffers(tx_queue);
445 	} else {
446 		tx_queue->xmit_more_available = skb->xmit_more;
447 	}
448 
449 	tx_queue->tx_packets++;
450 
451 	return NETDEV_TX_OK;
452 
453  dma_err:
454 	netif_err(efx, tx_err, efx->net_dev,
455 		  " TX queue %d could not map skb with %d bytes %d "
456 		  "fragments for DMA\n", tx_queue->queue, skb->len,
457 		  skb_shinfo(skb)->nr_frags + 1);
458 
459 	/* Mark the packet as transmitted, and free the SKB ourselves */
460 	dev_kfree_skb_any(skb);
461 
462 	/* Work backwards until we hit the original insert pointer value */
463 	while (tx_queue->insert_count != old_insert_count) {
464 		unsigned int pkts_compl = 0, bytes_compl = 0;
465 		--tx_queue->insert_count;
466 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
467 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
468 	}
469 
470 	/* Free the fragment we were mid-way through pushing */
471 	if (unmap_len) {
472 		if (dma_flags & EFX_TX_BUF_MAP_SINGLE)
473 			dma_unmap_single(dma_dev, unmap_addr, unmap_len,
474 					 DMA_TO_DEVICE);
475 		else
476 			dma_unmap_page(dma_dev, unmap_addr, unmap_len,
477 				       DMA_TO_DEVICE);
478 	}
479 
480 	return NETDEV_TX_OK;
481 }
482 
483 /* Remove packets from the TX queue
484  *
485  * This removes packets from the TX queue, up to and including the
486  * specified index.
487  */
488 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
489 				unsigned int index,
490 				unsigned int *pkts_compl,
491 				unsigned int *bytes_compl)
492 {
493 	struct efx_nic *efx = tx_queue->efx;
494 	unsigned int stop_index, read_ptr;
495 
496 	stop_index = (index + 1) & tx_queue->ptr_mask;
497 	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
498 
499 	while (read_ptr != stop_index) {
500 		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
501 
502 		if (!(buffer->flags & EFX_TX_BUF_OPTION) &&
503 		    unlikely(buffer->len == 0)) {
504 			netif_err(efx, tx_err, efx->net_dev,
505 				  "TX queue %d spurious TX completion id %x\n",
506 				  tx_queue->queue, read_ptr);
507 			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
508 			return;
509 		}
510 
511 		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
512 
513 		++tx_queue->read_count;
514 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
515 	}
516 }
517 
518 /* Initiate a packet transmission.  We use one channel per CPU
519  * (sharing when we have more CPUs than channels).  On Falcon, the TX
520  * completion events will be directed back to the CPU that transmitted
521  * the packet, which should be cache-efficient.
522  *
523  * Context: non-blocking.
524  * Note that returning anything other than NETDEV_TX_OK will cause the
525  * OS to free the skb.
526  */
527 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
528 				struct net_device *net_dev)
529 {
530 	struct efx_nic *efx = netdev_priv(net_dev);
531 	struct efx_tx_queue *tx_queue;
532 	unsigned index, type;
533 
534 	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
535 
536 	/* PTP "event" packet */
537 	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
538 	    unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
539 		return efx_ptp_tx(efx, skb);
540 	}
541 
542 	index = skb_get_queue_mapping(skb);
543 	type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
544 	if (index >= efx->n_tx_channels) {
545 		index -= efx->n_tx_channels;
546 		type |= EFX_TXQ_TYPE_HIGHPRI;
547 	}
548 	tx_queue = efx_get_tx_queue(efx, index, type);
549 
550 	return efx_enqueue_skb(tx_queue, skb);
551 }
552 
553 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
554 {
555 	struct efx_nic *efx = tx_queue->efx;
556 
557 	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
558 	tx_queue->core_txq =
559 		netdev_get_tx_queue(efx->net_dev,
560 				    tx_queue->queue / EFX_TXQ_TYPES +
561 				    ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
562 				     efx->n_tx_channels : 0));
563 }
564 
565 int efx_setup_tc(struct net_device *net_dev, u8 num_tc)
566 {
567 	struct efx_nic *efx = netdev_priv(net_dev);
568 	struct efx_channel *channel;
569 	struct efx_tx_queue *tx_queue;
570 	unsigned tc;
571 	int rc;
572 
573 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC)
574 		return -EINVAL;
575 
576 	if (num_tc == net_dev->num_tc)
577 		return 0;
578 
579 	for (tc = 0; tc < num_tc; tc++) {
580 		net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
581 		net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
582 	}
583 
584 	if (num_tc > net_dev->num_tc) {
585 		/* Initialise high-priority queues as necessary */
586 		efx_for_each_channel(channel, efx) {
587 			efx_for_each_possible_channel_tx_queue(tx_queue,
588 							       channel) {
589 				if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI))
590 					continue;
591 				if (!tx_queue->buffer) {
592 					rc = efx_probe_tx_queue(tx_queue);
593 					if (rc)
594 						return rc;
595 				}
596 				if (!tx_queue->initialised)
597 					efx_init_tx_queue(tx_queue);
598 				efx_init_tx_queue_core_txq(tx_queue);
599 			}
600 		}
601 	} else {
602 		/* Reduce number of classes before number of queues */
603 		net_dev->num_tc = num_tc;
604 	}
605 
606 	rc = netif_set_real_num_tx_queues(net_dev,
607 					  max_t(int, num_tc, 1) *
608 					  efx->n_tx_channels);
609 	if (rc)
610 		return rc;
611 
612 	/* Do not destroy high-priority queues when they become
613 	 * unused.  We would have to flush them first, and it is
614 	 * fairly difficult to flush a subset of TX queues.  Leave
615 	 * it to efx_fini_channels().
616 	 */
617 
618 	net_dev->num_tc = num_tc;
619 	return 0;
620 }
621 
622 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
623 {
624 	unsigned fill_level;
625 	struct efx_nic *efx = tx_queue->efx;
626 	struct efx_tx_queue *txq2;
627 	unsigned int pkts_compl = 0, bytes_compl = 0;
628 
629 	EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
630 
631 	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
632 	tx_queue->pkts_compl += pkts_compl;
633 	tx_queue->bytes_compl += bytes_compl;
634 
635 	if (pkts_compl > 1)
636 		++tx_queue->merge_events;
637 
638 	/* See if we need to restart the netif queue.  This memory
639 	 * barrier ensures that we write read_count (inside
640 	 * efx_dequeue_buffers()) before reading the queue status.
641 	 */
642 	smp_mb();
643 	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
644 	    likely(efx->port_enabled) &&
645 	    likely(netif_device_present(efx->net_dev))) {
646 		txq2 = efx_tx_queue_partner(tx_queue);
647 		fill_level = max(tx_queue->insert_count - tx_queue->read_count,
648 				 txq2->insert_count - txq2->read_count);
649 		if (fill_level <= efx->txq_wake_thresh)
650 			netif_tx_wake_queue(tx_queue->core_txq);
651 	}
652 
653 	/* Check whether the hardware queue is now empty */
654 	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
655 		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
656 		if (tx_queue->read_count == tx_queue->old_write_count) {
657 			smp_mb();
658 			tx_queue->empty_read_count =
659 				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
660 		}
661 	}
662 }
663 
664 /* Size of page-based TSO header buffers.  Larger blocks must be
665  * allocated from the heap.
666  */
667 #define TSOH_STD_SIZE	128
668 #define TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
669 
670 /* At most half the descriptors in the queue at any time will refer to
671  * a TSO header buffer, since they must always be followed by a
672  * payload descriptor referring to an skb.
673  */
674 static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue)
675 {
676 	return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE);
677 }
678 
679 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
680 {
681 	struct efx_nic *efx = tx_queue->efx;
682 	unsigned int entries;
683 	int rc;
684 
685 	/* Create the smallest power-of-two aligned ring */
686 	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
687 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
688 	tx_queue->ptr_mask = entries - 1;
689 
690 	netif_dbg(efx, probe, efx->net_dev,
691 		  "creating TX queue %d size %#x mask %#x\n",
692 		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
693 
694 	/* Allocate software ring */
695 	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
696 				   GFP_KERNEL);
697 	if (!tx_queue->buffer)
698 		return -ENOMEM;
699 
700 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) {
701 		tx_queue->tsoh_page =
702 			kcalloc(efx_tsoh_page_count(tx_queue),
703 				sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL);
704 		if (!tx_queue->tsoh_page) {
705 			rc = -ENOMEM;
706 			goto fail1;
707 		}
708 	}
709 
710 	/* Allocate hardware ring */
711 	rc = efx_nic_probe_tx(tx_queue);
712 	if (rc)
713 		goto fail2;
714 
715 	return 0;
716 
717 fail2:
718 	kfree(tx_queue->tsoh_page);
719 	tx_queue->tsoh_page = NULL;
720 fail1:
721 	kfree(tx_queue->buffer);
722 	tx_queue->buffer = NULL;
723 	return rc;
724 }
725 
726 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
727 {
728 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
729 		  "initialising TX queue %d\n", tx_queue->queue);
730 
731 	tx_queue->insert_count = 0;
732 	tx_queue->write_count = 0;
733 	tx_queue->old_write_count = 0;
734 	tx_queue->read_count = 0;
735 	tx_queue->old_read_count = 0;
736 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
737 	tx_queue->xmit_more_available = false;
738 
739 	/* Set up TX descriptor ring */
740 	efx_nic_init_tx(tx_queue);
741 
742 	tx_queue->initialised = true;
743 }
744 
745 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
746 {
747 	struct efx_tx_buffer *buffer;
748 
749 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
750 		  "shutting down TX queue %d\n", tx_queue->queue);
751 
752 	if (!tx_queue->buffer)
753 		return;
754 
755 	/* Free any buffers left in the ring */
756 	while (tx_queue->read_count != tx_queue->write_count) {
757 		unsigned int pkts_compl = 0, bytes_compl = 0;
758 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
759 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
760 
761 		++tx_queue->read_count;
762 	}
763 	tx_queue->xmit_more_available = false;
764 	netdev_tx_reset_queue(tx_queue->core_txq);
765 }
766 
767 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
768 {
769 	int i;
770 
771 	if (!tx_queue->buffer)
772 		return;
773 
774 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
775 		  "destroying TX queue %d\n", tx_queue->queue);
776 	efx_nic_remove_tx(tx_queue);
777 
778 	if (tx_queue->tsoh_page) {
779 		for (i = 0; i < efx_tsoh_page_count(tx_queue); i++)
780 			efx_nic_free_buffer(tx_queue->efx,
781 					    &tx_queue->tsoh_page[i]);
782 		kfree(tx_queue->tsoh_page);
783 		tx_queue->tsoh_page = NULL;
784 	}
785 
786 	kfree(tx_queue->buffer);
787 	tx_queue->buffer = NULL;
788 }
789 
790 
791 /* Efx TCP segmentation acceleration.
792  *
793  * Why?  Because by doing it here in the driver we can go significantly
794  * faster than the GSO.
795  *
796  * Requires TX checksum offload support.
797  */
798 
799 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
800 
801 /**
802  * struct tso_state - TSO state for an SKB
803  * @out_len: Remaining length in current segment
804  * @seqnum: Current sequence number
805  * @ipv4_id: Current IPv4 ID, host endian
806  * @packet_space: Remaining space in current packet
807  * @dma_addr: DMA address of current position
808  * @in_len: Remaining length in current SKB fragment
809  * @unmap_len: Length of SKB fragment
810  * @unmap_addr: DMA address of SKB fragment
811  * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0
812  * @protocol: Network protocol (after any VLAN header)
813  * @ip_off: Offset of IP header
814  * @tcp_off: Offset of TCP header
815  * @header_len: Number of bytes of header
816  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
817  * @header_dma_addr: Header DMA address, when using option descriptors
818  * @header_unmap_len: Header DMA mapped length, or 0 if not using option
819  *	descriptors
820  *
821  * The state used during segmentation.  It is put into this data structure
822  * just to make it easy to pass into inline functions.
823  */
824 struct tso_state {
825 	/* Output position */
826 	unsigned out_len;
827 	unsigned seqnum;
828 	u16 ipv4_id;
829 	unsigned packet_space;
830 
831 	/* Input position */
832 	dma_addr_t dma_addr;
833 	unsigned in_len;
834 	unsigned unmap_len;
835 	dma_addr_t unmap_addr;
836 	unsigned short dma_flags;
837 
838 	__be16 protocol;
839 	unsigned int ip_off;
840 	unsigned int tcp_off;
841 	unsigned header_len;
842 	unsigned int ip_base_len;
843 	dma_addr_t header_dma_addr;
844 	unsigned int header_unmap_len;
845 };
846 
847 
848 /*
849  * Verify that our various assumptions about sk_buffs and the conditions
850  * under which TSO will be attempted hold true.  Return the protocol number.
851  */
852 static __be16 efx_tso_check_protocol(struct sk_buff *skb)
853 {
854 	__be16 protocol = skb->protocol;
855 
856 	EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
857 			    protocol);
858 	if (protocol == htons(ETH_P_8021Q)) {
859 		struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
860 		protocol = veh->h_vlan_encapsulated_proto;
861 	}
862 
863 	if (protocol == htons(ETH_P_IP)) {
864 		EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
865 	} else {
866 		EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
867 		EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
868 	}
869 	EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
870 			     + (tcp_hdr(skb)->doff << 2u)) >
871 			    skb_headlen(skb));
872 
873 	return protocol;
874 }
875 
876 static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue,
877 			       struct efx_tx_buffer *buffer, unsigned int len)
878 {
879 	u8 *result;
880 
881 	EFX_BUG_ON_PARANOID(buffer->len);
882 	EFX_BUG_ON_PARANOID(buffer->flags);
883 	EFX_BUG_ON_PARANOID(buffer->unmap_len);
884 
885 	if (likely(len <= TSOH_STD_SIZE - NET_IP_ALIGN)) {
886 		unsigned index =
887 			(tx_queue->insert_count & tx_queue->ptr_mask) / 2;
888 		struct efx_buffer *page_buf =
889 			&tx_queue->tsoh_page[index / TSOH_PER_PAGE];
890 		unsigned offset =
891 			TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + NET_IP_ALIGN;
892 
893 		if (unlikely(!page_buf->addr) &&
894 		    efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
895 					 GFP_ATOMIC))
896 			return NULL;
897 
898 		result = (u8 *)page_buf->addr + offset;
899 		buffer->dma_addr = page_buf->dma_addr + offset;
900 		buffer->flags = EFX_TX_BUF_CONT;
901 	} else {
902 		tx_queue->tso_long_headers++;
903 
904 		buffer->heap_buf = kmalloc(NET_IP_ALIGN + len, GFP_ATOMIC);
905 		if (unlikely(!buffer->heap_buf))
906 			return NULL;
907 		result = (u8 *)buffer->heap_buf + NET_IP_ALIGN;
908 		buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP;
909 	}
910 
911 	buffer->len = len;
912 
913 	return result;
914 }
915 
916 /**
917  * efx_tx_queue_insert - push descriptors onto the TX queue
918  * @tx_queue:		Efx TX queue
919  * @dma_addr:		DMA address of fragment
920  * @len:		Length of fragment
921  * @final_buffer:	The final buffer inserted into the queue
922  *
923  * Push descriptors onto the TX queue.
924  */
925 static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
926 				dma_addr_t dma_addr, unsigned len,
927 				struct efx_tx_buffer **final_buffer)
928 {
929 	struct efx_tx_buffer *buffer;
930 	struct efx_nic *efx = tx_queue->efx;
931 	unsigned dma_len;
932 
933 	EFX_BUG_ON_PARANOID(len <= 0);
934 
935 	while (1) {
936 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
937 		++tx_queue->insert_count;
938 
939 		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
940 				    tx_queue->read_count >=
941 				    efx->txq_entries);
942 
943 		buffer->dma_addr = dma_addr;
944 
945 		dma_len = efx_max_tx_len(efx, dma_addr);
946 
947 		/* If there is enough space to send then do so */
948 		if (dma_len >= len)
949 			break;
950 
951 		buffer->len = dma_len;
952 		buffer->flags = EFX_TX_BUF_CONT;
953 		dma_addr += dma_len;
954 		len -= dma_len;
955 	}
956 
957 	EFX_BUG_ON_PARANOID(!len);
958 	buffer->len = len;
959 	*final_buffer = buffer;
960 }
961 
962 
963 /*
964  * Put a TSO header into the TX queue.
965  *
966  * This is special-cased because we know that it is small enough to fit in
967  * a single fragment, and we know it doesn't cross a page boundary.  It
968  * also allows us to not worry about end-of-packet etc.
969  */
970 static int efx_tso_put_header(struct efx_tx_queue *tx_queue,
971 			      struct efx_tx_buffer *buffer, u8 *header)
972 {
973 	if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) {
974 		buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev,
975 						  header, buffer->len,
976 						  DMA_TO_DEVICE);
977 		if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev,
978 					       buffer->dma_addr))) {
979 			kfree(buffer->heap_buf);
980 			buffer->len = 0;
981 			buffer->flags = 0;
982 			return -ENOMEM;
983 		}
984 		buffer->unmap_len = buffer->len;
985 		buffer->dma_offset = 0;
986 		buffer->flags |= EFX_TX_BUF_MAP_SINGLE;
987 	}
988 
989 	++tx_queue->insert_count;
990 	return 0;
991 }
992 
993 
994 /* Remove buffers put into a tx_queue.  None of the buffers must have
995  * an skb attached.
996  */
997 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
998 			       unsigned int insert_count)
999 {
1000 	struct efx_tx_buffer *buffer;
1001 
1002 	/* Work backwards until we hit the original insert pointer value */
1003 	while (tx_queue->insert_count != insert_count) {
1004 		--tx_queue->insert_count;
1005 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
1006 		efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
1007 	}
1008 }
1009 
1010 
1011 /* Parse the SKB header and initialise state. */
1012 static int tso_start(struct tso_state *st, struct efx_nic *efx,
1013 		     struct efx_tx_queue *tx_queue,
1014 		     const struct sk_buff *skb)
1015 {
1016 	struct device *dma_dev = &efx->pci_dev->dev;
1017 	unsigned int header_len, in_len;
1018 	bool use_opt_desc = false;
1019 	dma_addr_t dma_addr;
1020 
1021 	if (tx_queue->tso_version == 1)
1022 		use_opt_desc = true;
1023 
1024 	st->ip_off = skb_network_header(skb) - skb->data;
1025 	st->tcp_off = skb_transport_header(skb) - skb->data;
1026 	header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
1027 	in_len = skb_headlen(skb) - header_len;
1028 	st->header_len = header_len;
1029 	st->in_len = in_len;
1030 	if (st->protocol == htons(ETH_P_IP)) {
1031 		st->ip_base_len = st->header_len - st->ip_off;
1032 		st->ipv4_id = ntohs(ip_hdr(skb)->id);
1033 	} else {
1034 		st->ip_base_len = st->header_len - st->tcp_off;
1035 		st->ipv4_id = 0;
1036 	}
1037 	st->seqnum = ntohl(tcp_hdr(skb)->seq);
1038 
1039 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
1040 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
1041 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
1042 
1043 	st->out_len = skb->len - header_len;
1044 
1045 	if (!use_opt_desc) {
1046 		st->header_unmap_len = 0;
1047 
1048 		if (likely(in_len == 0)) {
1049 			st->dma_flags = 0;
1050 			st->unmap_len = 0;
1051 			return 0;
1052 		}
1053 
1054 		dma_addr = dma_map_single(dma_dev, skb->data + header_len,
1055 					  in_len, DMA_TO_DEVICE);
1056 		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
1057 		st->dma_addr = dma_addr;
1058 		st->unmap_addr = dma_addr;
1059 		st->unmap_len = in_len;
1060 	} else {
1061 		dma_addr = dma_map_single(dma_dev, skb->data,
1062 					  skb_headlen(skb), DMA_TO_DEVICE);
1063 		st->header_dma_addr = dma_addr;
1064 		st->header_unmap_len = skb_headlen(skb);
1065 		st->dma_flags = 0;
1066 		st->dma_addr = dma_addr + header_len;
1067 		st->unmap_len = 0;
1068 	}
1069 
1070 	return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0;
1071 }
1072 
1073 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
1074 			    skb_frag_t *frag)
1075 {
1076 	st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0,
1077 					  skb_frag_size(frag), DMA_TO_DEVICE);
1078 	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
1079 		st->dma_flags = 0;
1080 		st->unmap_len = skb_frag_size(frag);
1081 		st->in_len = skb_frag_size(frag);
1082 		st->dma_addr = st->unmap_addr;
1083 		return 0;
1084 	}
1085 	return -ENOMEM;
1086 }
1087 
1088 
1089 /**
1090  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1091  * @tx_queue:		Efx TX queue
1092  * @skb:		Socket buffer
1093  * @st:			TSO state
1094  *
1095  * Form descriptors for the current fragment, until we reach the end
1096  * of fragment or end-of-packet.
1097  */
1098 static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
1099 					  const struct sk_buff *skb,
1100 					  struct tso_state *st)
1101 {
1102 	struct efx_tx_buffer *buffer;
1103 	int n;
1104 
1105 	if (st->in_len == 0)
1106 		return;
1107 	if (st->packet_space == 0)
1108 		return;
1109 
1110 	EFX_BUG_ON_PARANOID(st->in_len <= 0);
1111 	EFX_BUG_ON_PARANOID(st->packet_space <= 0);
1112 
1113 	n = min(st->in_len, st->packet_space);
1114 
1115 	st->packet_space -= n;
1116 	st->out_len -= n;
1117 	st->in_len -= n;
1118 
1119 	efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
1120 
1121 	if (st->out_len == 0) {
1122 		/* Transfer ownership of the skb */
1123 		buffer->skb = skb;
1124 		buffer->flags = EFX_TX_BUF_SKB;
1125 	} else if (st->packet_space != 0) {
1126 		buffer->flags = EFX_TX_BUF_CONT;
1127 	}
1128 
1129 	if (st->in_len == 0) {
1130 		/* Transfer ownership of the DMA mapping */
1131 		buffer->unmap_len = st->unmap_len;
1132 		buffer->dma_offset = buffer->unmap_len - buffer->len;
1133 		buffer->flags |= st->dma_flags;
1134 		st->unmap_len = 0;
1135 	}
1136 
1137 	st->dma_addr += n;
1138 }
1139 
1140 
1141 /**
1142  * tso_start_new_packet - generate a new header and prepare for the new packet
1143  * @tx_queue:		Efx TX queue
1144  * @skb:		Socket buffer
1145  * @st:			TSO state
1146  *
1147  * Generate a new header and prepare for the new packet.  Return 0 on
1148  * success, or -%ENOMEM if failed to alloc header.
1149  */
1150 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
1151 				const struct sk_buff *skb,
1152 				struct tso_state *st)
1153 {
1154 	struct efx_tx_buffer *buffer =
1155 		efx_tx_queue_get_insert_buffer(tx_queue);
1156 	bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
1157 	u8 tcp_flags_clear;
1158 
1159 	if (!is_last) {
1160 		st->packet_space = skb_shinfo(skb)->gso_size;
1161 		tcp_flags_clear = 0x09; /* mask out FIN and PSH */
1162 	} else {
1163 		st->packet_space = st->out_len;
1164 		tcp_flags_clear = 0x00;
1165 	}
1166 
1167 	if (!st->header_unmap_len) {
1168 		/* Allocate and insert a DMA-mapped header buffer. */
1169 		struct tcphdr *tsoh_th;
1170 		unsigned ip_length;
1171 		u8 *header;
1172 		int rc;
1173 
1174 		header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
1175 		if (!header)
1176 			return -ENOMEM;
1177 
1178 		tsoh_th = (struct tcphdr *)(header + st->tcp_off);
1179 
1180 		/* Copy and update the headers. */
1181 		memcpy(header, skb->data, st->header_len);
1182 
1183 		tsoh_th->seq = htonl(st->seqnum);
1184 		((u8 *)tsoh_th)[13] &= ~tcp_flags_clear;
1185 
1186 		ip_length = st->ip_base_len + st->packet_space;
1187 
1188 		if (st->protocol == htons(ETH_P_IP)) {
1189 			struct iphdr *tsoh_iph =
1190 				(struct iphdr *)(header + st->ip_off);
1191 
1192 			tsoh_iph->tot_len = htons(ip_length);
1193 			tsoh_iph->id = htons(st->ipv4_id);
1194 		} else {
1195 			struct ipv6hdr *tsoh_iph =
1196 				(struct ipv6hdr *)(header + st->ip_off);
1197 
1198 			tsoh_iph->payload_len = htons(ip_length);
1199 		}
1200 
1201 		rc = efx_tso_put_header(tx_queue, buffer, header);
1202 		if (unlikely(rc))
1203 			return rc;
1204 	} else {
1205 		/* Send the original headers with a TSO option descriptor
1206 		 * in front
1207 		 */
1208 		u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear;
1209 
1210 		buffer->flags = EFX_TX_BUF_OPTION;
1211 		buffer->len = 0;
1212 		buffer->unmap_len = 0;
1213 		EFX_POPULATE_QWORD_5(buffer->option,
1214 				     ESF_DZ_TX_DESC_IS_OPT, 1,
1215 				     ESF_DZ_TX_OPTION_TYPE,
1216 				     ESE_DZ_TX_OPTION_DESC_TSO,
1217 				     ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags,
1218 				     ESF_DZ_TX_TSO_IP_ID, st->ipv4_id,
1219 				     ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum);
1220 		++tx_queue->insert_count;
1221 
1222 		/* We mapped the headers in tso_start().  Unmap them
1223 		 * when the last segment is completed.
1224 		 */
1225 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
1226 		buffer->dma_addr = st->header_dma_addr;
1227 		buffer->len = st->header_len;
1228 		if (is_last) {
1229 			buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE;
1230 			buffer->unmap_len = st->header_unmap_len;
1231 			buffer->dma_offset = 0;
1232 			/* Ensure we only unmap them once in case of a
1233 			 * later DMA mapping error and rollback
1234 			 */
1235 			st->header_unmap_len = 0;
1236 		} else {
1237 			buffer->flags = EFX_TX_BUF_CONT;
1238 			buffer->unmap_len = 0;
1239 		}
1240 		++tx_queue->insert_count;
1241 	}
1242 
1243 	st->seqnum += skb_shinfo(skb)->gso_size;
1244 
1245 	/* Linux leaves suitable gaps in the IP ID space for us to fill. */
1246 	++st->ipv4_id;
1247 
1248 	++tx_queue->tso_packets;
1249 
1250 	++tx_queue->tx_packets;
1251 
1252 	return 0;
1253 }
1254 
1255 
1256 /**
1257  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1258  * @tx_queue:		Efx TX queue
1259  * @skb:		Socket buffer
1260  *
1261  * Context: You must hold netif_tx_lock() to call this function.
1262  *
1263  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1264  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1265  * %NETDEV_TX_OK.
1266  */
1267 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1268 			       struct sk_buff *skb)
1269 {
1270 	struct efx_nic *efx = tx_queue->efx;
1271 	unsigned int old_insert_count = tx_queue->insert_count;
1272 	int frag_i, rc;
1273 	struct tso_state state;
1274 
1275 	/* Find the packet protocol and sanity-check it */
1276 	state.protocol = efx_tso_check_protocol(skb);
1277 
1278 	rc = tso_start(&state, efx, tx_queue, skb);
1279 	if (rc)
1280 		goto mem_err;
1281 
1282 	if (likely(state.in_len == 0)) {
1283 		/* Grab the first payload fragment. */
1284 		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1285 		frag_i = 0;
1286 		rc = tso_get_fragment(&state, efx,
1287 				      skb_shinfo(skb)->frags + frag_i);
1288 		if (rc)
1289 			goto mem_err;
1290 	} else {
1291 		/* Payload starts in the header area. */
1292 		frag_i = -1;
1293 	}
1294 
1295 	if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1296 		goto mem_err;
1297 
1298 	while (1) {
1299 		tso_fill_packet_with_fragment(tx_queue, skb, &state);
1300 
1301 		/* Move onto the next fragment? */
1302 		if (state.in_len == 0) {
1303 			if (++frag_i >= skb_shinfo(skb)->nr_frags)
1304 				/* End of payload reached. */
1305 				break;
1306 			rc = tso_get_fragment(&state, efx,
1307 					      skb_shinfo(skb)->frags + frag_i);
1308 			if (rc)
1309 				goto mem_err;
1310 		}
1311 
1312 		/* Start at new packet? */
1313 		if (state.packet_space == 0 &&
1314 		    tso_start_new_packet(tx_queue, skb, &state) < 0)
1315 			goto mem_err;
1316 	}
1317 
1318 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
1319 
1320 	efx_tx_maybe_stop_queue(tx_queue);
1321 
1322 	/* Pass off to hardware */
1323 	if (!skb->xmit_more || netif_xmit_stopped(tx_queue->core_txq)) {
1324 		struct efx_tx_queue *txq2 = efx_tx_queue_partner(tx_queue);
1325 
1326 		/* There could be packets left on the partner queue if those
1327 		 * SKBs had skb->xmit_more set. If we do not push those they
1328 		 * could be left for a long time and cause a netdev watchdog.
1329 		 */
1330 		if (txq2->xmit_more_available)
1331 			efx_nic_push_buffers(txq2);
1332 
1333 		efx_nic_push_buffers(tx_queue);
1334 	} else {
1335 		tx_queue->xmit_more_available = skb->xmit_more;
1336 	}
1337 
1338 	tx_queue->tso_bursts++;
1339 	return NETDEV_TX_OK;
1340 
1341  mem_err:
1342 	netif_err(efx, tx_err, efx->net_dev,
1343 		  "Out of memory for TSO headers, or DMA mapping error\n");
1344 	dev_kfree_skb_any(skb);
1345 
1346 	/* Free the DMA mapping we were in the process of writing out */
1347 	if (state.unmap_len) {
1348 		if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE)
1349 			dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr,
1350 					 state.unmap_len, DMA_TO_DEVICE);
1351 		else
1352 			dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr,
1353 				       state.unmap_len, DMA_TO_DEVICE);
1354 	}
1355 
1356 	/* Free the header DMA mapping, if using option descriptors */
1357 	if (state.header_unmap_len)
1358 		dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr,
1359 				 state.header_unmap_len, DMA_TO_DEVICE);
1360 
1361 	efx_enqueue_unwind(tx_queue, old_insert_count);
1362 	return NETDEV_TX_OK;
1363 }
1364