xref: /openbmc/linux/drivers/net/ethernet/sfc/tx.c (revision 206204a1)
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_kfree_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 /* Get partner of a TX queue, seen as part of the same net core queue */
136 static struct efx_tx_queue *efx_tx_queue_partner(struct efx_tx_queue *tx_queue)
137 {
138 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
139 		return tx_queue - EFX_TXQ_TYPE_OFFLOAD;
140 	else
141 		return tx_queue + EFX_TXQ_TYPE_OFFLOAD;
142 }
143 
144 static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
145 {
146 	/* We need to consider both queues that the net core sees as one */
147 	struct efx_tx_queue *txq2 = efx_tx_queue_partner(txq1);
148 	struct efx_nic *efx = txq1->efx;
149 	unsigned int fill_level;
150 
151 	fill_level = max(txq1->insert_count - txq1->old_read_count,
152 			 txq2->insert_count - txq2->old_read_count);
153 	if (likely(fill_level < efx->txq_stop_thresh))
154 		return;
155 
156 	/* We used the stale old_read_count above, which gives us a
157 	 * pessimistic estimate of the fill level (which may even
158 	 * validly be >= efx->txq_entries).  Now try again using
159 	 * read_count (more likely to be a cache miss).
160 	 *
161 	 * If we read read_count and then conditionally stop the
162 	 * queue, it is possible for the completion path to race with
163 	 * us and complete all outstanding descriptors in the middle,
164 	 * after which there will be no more completions to wake it.
165 	 * Therefore we stop the queue first, then read read_count
166 	 * (with a memory barrier to ensure the ordering), then
167 	 * restart the queue if the fill level turns out to be low
168 	 * enough.
169 	 */
170 	netif_tx_stop_queue(txq1->core_txq);
171 	smp_mb();
172 	txq1->old_read_count = ACCESS_ONCE(txq1->read_count);
173 	txq2->old_read_count = ACCESS_ONCE(txq2->read_count);
174 
175 	fill_level = max(txq1->insert_count - txq1->old_read_count,
176 			 txq2->insert_count - txq2->old_read_count);
177 	EFX_BUG_ON_PARANOID(fill_level >= efx->txq_entries);
178 	if (likely(fill_level < efx->txq_stop_thresh)) {
179 		smp_mb();
180 		if (likely(!efx->loopback_selftest))
181 			netif_tx_start_queue(txq1->core_txq);
182 	}
183 }
184 
185 #ifdef EFX_USE_PIO
186 
187 struct efx_short_copy_buffer {
188 	int used;
189 	u8 buf[L1_CACHE_BYTES];
190 };
191 
192 /* Copy in explicit 64-bit writes. */
193 static void efx_memcpy_64(void __iomem *dest, void *src, size_t len)
194 {
195 	u64 *src64 = src;
196 	u64 __iomem *dest64 = dest;
197 	size_t l64 = len / 8;
198 	size_t i;
199 
200 	for (i = 0; i < l64; i++)
201 		writeq(src64[i], &dest64[i]);
202 }
203 
204 /* Copy to PIO, respecting that writes to PIO buffers must be dword aligned.
205  * Advances piobuf pointer. Leaves additional data in the copy buffer.
206  */
207 static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf,
208 				    u8 *data, int len,
209 				    struct efx_short_copy_buffer *copy_buf)
210 {
211 	int block_len = len & ~(sizeof(copy_buf->buf) - 1);
212 
213 	efx_memcpy_64(*piobuf, data, block_len);
214 	*piobuf += block_len;
215 	len -= block_len;
216 
217 	if (len) {
218 		data += block_len;
219 		BUG_ON(copy_buf->used);
220 		BUG_ON(len > sizeof(copy_buf->buf));
221 		memcpy(copy_buf->buf, data, len);
222 		copy_buf->used = len;
223 	}
224 }
225 
226 /* Copy to PIO, respecting dword alignment, popping data from copy buffer first.
227  * Advances piobuf pointer. Leaves additional data in the copy buffer.
228  */
229 static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf,
230 				       u8 *data, int len,
231 				       struct efx_short_copy_buffer *copy_buf)
232 {
233 	if (copy_buf->used) {
234 		/* if the copy buffer is partially full, fill it up and write */
235 		int copy_to_buf =
236 			min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len);
237 
238 		memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf);
239 		copy_buf->used += copy_to_buf;
240 
241 		/* if we didn't fill it up then we're done for now */
242 		if (copy_buf->used < sizeof(copy_buf->buf))
243 			return;
244 
245 		efx_memcpy_64(*piobuf, copy_buf->buf, sizeof(copy_buf->buf));
246 		*piobuf += sizeof(copy_buf->buf);
247 		data += copy_to_buf;
248 		len -= copy_to_buf;
249 		copy_buf->used = 0;
250 	}
251 
252 	efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf);
253 }
254 
255 static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf,
256 				  struct efx_short_copy_buffer *copy_buf)
257 {
258 	/* if there's anything in it, write the whole buffer, including junk */
259 	if (copy_buf->used)
260 		efx_memcpy_64(piobuf, copy_buf->buf, sizeof(copy_buf->buf));
261 }
262 
263 /* Traverse skb structure and copy fragments in to PIO buffer.
264  * Advances piobuf pointer.
265  */
266 static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb,
267 				     u8 __iomem **piobuf,
268 				     struct efx_short_copy_buffer *copy_buf)
269 {
270 	int i;
271 
272 	efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb),
273 				copy_buf);
274 
275 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
276 		skb_frag_t *f = &skb_shinfo(skb)->frags[i];
277 		u8 *vaddr;
278 
279 		vaddr = kmap_atomic(skb_frag_page(f));
280 
281 		efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + f->page_offset,
282 					   skb_frag_size(f), copy_buf);
283 		kunmap_atomic(vaddr);
284 	}
285 
286 	EFX_BUG_ON_PARANOID(skb_shinfo(skb)->frag_list);
287 }
288 
289 static struct efx_tx_buffer *
290 efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
291 {
292 	struct efx_tx_buffer *buffer =
293 		efx_tx_queue_get_insert_buffer(tx_queue);
294 	u8 __iomem *piobuf = tx_queue->piobuf;
295 
296 	/* Copy to PIO buffer. Ensure the writes are padded to the end
297 	 * of a cache line, as this is required for write-combining to be
298 	 * effective on at least x86.
299 	 */
300 
301 	if (skb_shinfo(skb)->nr_frags) {
302 		/* The size of the copy buffer will ensure all writes
303 		 * are the size of a cache line.
304 		 */
305 		struct efx_short_copy_buffer copy_buf;
306 
307 		copy_buf.used = 0;
308 
309 		efx_skb_copy_bits_to_pio(tx_queue->efx, skb,
310 					 &piobuf, &copy_buf);
311 		efx_flush_copy_buffer(tx_queue->efx, piobuf, &copy_buf);
312 	} else {
313 		/* Pad the write to the size of a cache line.
314 		 * We can do this because we know the skb_shared_info sruct is
315 		 * after the source, and the destination buffer is big enough.
316 		 */
317 		BUILD_BUG_ON(L1_CACHE_BYTES >
318 			     SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
319 		efx_memcpy_64(tx_queue->piobuf, skb->data,
320 			      ALIGN(skb->len, L1_CACHE_BYTES));
321 	}
322 
323 	EFX_POPULATE_QWORD_5(buffer->option,
324 			     ESF_DZ_TX_DESC_IS_OPT, 1,
325 			     ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO,
326 			     ESF_DZ_TX_PIO_CONT, 0,
327 			     ESF_DZ_TX_PIO_BYTE_CNT, skb->len,
328 			     ESF_DZ_TX_PIO_BUF_ADDR,
329 			     tx_queue->piobuf_offset);
330 	++tx_queue->pio_packets;
331 	++tx_queue->insert_count;
332 	return buffer;
333 }
334 #endif /* EFX_USE_PIO */
335 
336 /*
337  * Add a socket buffer to a TX queue
338  *
339  * This maps all fragments of a socket buffer for DMA and adds them to
340  * the TX queue.  The queue's insert pointer will be incremented by
341  * the number of fragments in the socket buffer.
342  *
343  * If any DMA mapping fails, any mapped fragments will be unmapped,
344  * the queue's insert pointer will be restored to its original value.
345  *
346  * This function is split out from efx_hard_start_xmit to allow the
347  * loopback test to direct packets via specific TX queues.
348  *
349  * Returns NETDEV_TX_OK.
350  * You must hold netif_tx_lock() to call this function.
351  */
352 netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
353 {
354 	struct efx_nic *efx = tx_queue->efx;
355 	struct device *dma_dev = &efx->pci_dev->dev;
356 	struct efx_tx_buffer *buffer;
357 	skb_frag_t *fragment;
358 	unsigned int len, unmap_len = 0;
359 	dma_addr_t dma_addr, unmap_addr = 0;
360 	unsigned int dma_len;
361 	unsigned short dma_flags;
362 	int i = 0;
363 
364 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
365 
366 	if (skb_shinfo(skb)->gso_size)
367 		return efx_enqueue_skb_tso(tx_queue, skb);
368 
369 	/* Get size of the initial fragment */
370 	len = skb_headlen(skb);
371 
372 	/* Pad if necessary */
373 	if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
374 		EFX_BUG_ON_PARANOID(skb->data_len);
375 		len = 32 + 1;
376 		if (skb_pad(skb, len - skb->len))
377 			return NETDEV_TX_OK;
378 	}
379 
380 	/* Consider using PIO for short packets */
381 #ifdef EFX_USE_PIO
382 	if (skb->len <= efx_piobuf_size && tx_queue->piobuf &&
383 	    efx_nic_tx_is_empty(tx_queue) &&
384 	    efx_nic_tx_is_empty(efx_tx_queue_partner(tx_queue))) {
385 		buffer = efx_enqueue_skb_pio(tx_queue, skb);
386 		dma_flags = EFX_TX_BUF_OPTION;
387 		goto finish_packet;
388 	}
389 #endif
390 
391 	/* Map for DMA.  Use dma_map_single rather than dma_map_page
392 	 * since this is more efficient on machines with sparse
393 	 * memory.
394 	 */
395 	dma_flags = EFX_TX_BUF_MAP_SINGLE;
396 	dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE);
397 
398 	/* Process all fragments */
399 	while (1) {
400 		if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
401 			goto dma_err;
402 
403 		/* Store fields for marking in the per-fragment final
404 		 * descriptor */
405 		unmap_len = len;
406 		unmap_addr = dma_addr;
407 
408 		/* Add to TX queue, splitting across DMA boundaries */
409 		do {
410 			buffer = efx_tx_queue_get_insert_buffer(tx_queue);
411 
412 			dma_len = efx_max_tx_len(efx, dma_addr);
413 			if (likely(dma_len >= len))
414 				dma_len = len;
415 
416 			/* Fill out per descriptor fields */
417 			buffer->len = dma_len;
418 			buffer->dma_addr = dma_addr;
419 			buffer->flags = EFX_TX_BUF_CONT;
420 			len -= dma_len;
421 			dma_addr += dma_len;
422 			++tx_queue->insert_count;
423 		} while (len);
424 
425 		/* Transfer ownership of the unmapping to the final buffer */
426 		buffer->flags = EFX_TX_BUF_CONT | dma_flags;
427 		buffer->unmap_len = unmap_len;
428 		buffer->dma_offset = buffer->dma_addr - unmap_addr;
429 		unmap_len = 0;
430 
431 		/* Get address and size of next fragment */
432 		if (i >= skb_shinfo(skb)->nr_frags)
433 			break;
434 		fragment = &skb_shinfo(skb)->frags[i];
435 		len = skb_frag_size(fragment);
436 		i++;
437 		/* Map for DMA */
438 		dma_flags = 0;
439 		dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
440 					    DMA_TO_DEVICE);
441 	}
442 
443 	/* Transfer ownership of the skb to the final buffer */
444 #ifdef EFX_USE_PIO
445 finish_packet:
446 #endif
447 	buffer->skb = skb;
448 	buffer->flags = EFX_TX_BUF_SKB | dma_flags;
449 
450 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
451 
452 	/* Pass off to hardware */
453 	efx_nic_push_buffers(tx_queue);
454 
455 	efx_tx_maybe_stop_queue(tx_queue);
456 
457 	return NETDEV_TX_OK;
458 
459  dma_err:
460 	netif_err(efx, tx_err, efx->net_dev,
461 		  " TX queue %d could not map skb with %d bytes %d "
462 		  "fragments for DMA\n", tx_queue->queue, skb->len,
463 		  skb_shinfo(skb)->nr_frags + 1);
464 
465 	/* Mark the packet as transmitted, and free the SKB ourselves */
466 	dev_kfree_skb_any(skb);
467 
468 	/* Work backwards until we hit the original insert pointer value */
469 	while (tx_queue->insert_count != tx_queue->write_count) {
470 		unsigned int pkts_compl = 0, bytes_compl = 0;
471 		--tx_queue->insert_count;
472 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
473 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
474 	}
475 
476 	/* Free the fragment we were mid-way through pushing */
477 	if (unmap_len) {
478 		if (dma_flags & EFX_TX_BUF_MAP_SINGLE)
479 			dma_unmap_single(dma_dev, unmap_addr, unmap_len,
480 					 DMA_TO_DEVICE);
481 		else
482 			dma_unmap_page(dma_dev, unmap_addr, unmap_len,
483 				       DMA_TO_DEVICE);
484 	}
485 
486 	return NETDEV_TX_OK;
487 }
488 
489 /* Remove packets from the TX queue
490  *
491  * This removes packets from the TX queue, up to and including the
492  * specified index.
493  */
494 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
495 				unsigned int index,
496 				unsigned int *pkts_compl,
497 				unsigned int *bytes_compl)
498 {
499 	struct efx_nic *efx = tx_queue->efx;
500 	unsigned int stop_index, read_ptr;
501 
502 	stop_index = (index + 1) & tx_queue->ptr_mask;
503 	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
504 
505 	while (read_ptr != stop_index) {
506 		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
507 
508 		if (!(buffer->flags & EFX_TX_BUF_OPTION) &&
509 		    unlikely(buffer->len == 0)) {
510 			netif_err(efx, tx_err, efx->net_dev,
511 				  "TX queue %d spurious TX completion id %x\n",
512 				  tx_queue->queue, read_ptr);
513 			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
514 			return;
515 		}
516 
517 		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
518 
519 		++tx_queue->read_count;
520 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
521 	}
522 }
523 
524 /* Initiate a packet transmission.  We use one channel per CPU
525  * (sharing when we have more CPUs than channels).  On Falcon, the TX
526  * completion events will be directed back to the CPU that transmitted
527  * the packet, which should be cache-efficient.
528  *
529  * Context: non-blocking.
530  * Note that returning anything other than NETDEV_TX_OK will cause the
531  * OS to free the skb.
532  */
533 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
534 				struct net_device *net_dev)
535 {
536 	struct efx_nic *efx = netdev_priv(net_dev);
537 	struct efx_tx_queue *tx_queue;
538 	unsigned index, type;
539 
540 	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
541 
542 	/* PTP "event" packet */
543 	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
544 	    unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
545 		return efx_ptp_tx(efx, skb);
546 	}
547 
548 	index = skb_get_queue_mapping(skb);
549 	type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
550 	if (index >= efx->n_tx_channels) {
551 		index -= efx->n_tx_channels;
552 		type |= EFX_TXQ_TYPE_HIGHPRI;
553 	}
554 	tx_queue = efx_get_tx_queue(efx, index, type);
555 
556 	return efx_enqueue_skb(tx_queue, skb);
557 }
558 
559 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
560 {
561 	struct efx_nic *efx = tx_queue->efx;
562 
563 	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
564 	tx_queue->core_txq =
565 		netdev_get_tx_queue(efx->net_dev,
566 				    tx_queue->queue / EFX_TXQ_TYPES +
567 				    ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
568 				     efx->n_tx_channels : 0));
569 }
570 
571 int efx_setup_tc(struct net_device *net_dev, u8 num_tc)
572 {
573 	struct efx_nic *efx = netdev_priv(net_dev);
574 	struct efx_channel *channel;
575 	struct efx_tx_queue *tx_queue;
576 	unsigned tc;
577 	int rc;
578 
579 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC)
580 		return -EINVAL;
581 
582 	if (num_tc == net_dev->num_tc)
583 		return 0;
584 
585 	for (tc = 0; tc < num_tc; tc++) {
586 		net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
587 		net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
588 	}
589 
590 	if (num_tc > net_dev->num_tc) {
591 		/* Initialise high-priority queues as necessary */
592 		efx_for_each_channel(channel, efx) {
593 			efx_for_each_possible_channel_tx_queue(tx_queue,
594 							       channel) {
595 				if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI))
596 					continue;
597 				if (!tx_queue->buffer) {
598 					rc = efx_probe_tx_queue(tx_queue);
599 					if (rc)
600 						return rc;
601 				}
602 				if (!tx_queue->initialised)
603 					efx_init_tx_queue(tx_queue);
604 				efx_init_tx_queue_core_txq(tx_queue);
605 			}
606 		}
607 	} else {
608 		/* Reduce number of classes before number of queues */
609 		net_dev->num_tc = num_tc;
610 	}
611 
612 	rc = netif_set_real_num_tx_queues(net_dev,
613 					  max_t(int, num_tc, 1) *
614 					  efx->n_tx_channels);
615 	if (rc)
616 		return rc;
617 
618 	/* Do not destroy high-priority queues when they become
619 	 * unused.  We would have to flush them first, and it is
620 	 * fairly difficult to flush a subset of TX queues.  Leave
621 	 * it to efx_fini_channels().
622 	 */
623 
624 	net_dev->num_tc = num_tc;
625 	return 0;
626 }
627 
628 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
629 {
630 	unsigned fill_level;
631 	struct efx_nic *efx = tx_queue->efx;
632 	struct efx_tx_queue *txq2;
633 	unsigned int pkts_compl = 0, bytes_compl = 0;
634 
635 	EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
636 
637 	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
638 	netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl);
639 
640 	if (pkts_compl > 1)
641 		++tx_queue->merge_events;
642 
643 	/* See if we need to restart the netif queue.  This memory
644 	 * barrier ensures that we write read_count (inside
645 	 * efx_dequeue_buffers()) before reading the queue status.
646 	 */
647 	smp_mb();
648 	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
649 	    likely(efx->port_enabled) &&
650 	    likely(netif_device_present(efx->net_dev))) {
651 		txq2 = efx_tx_queue_partner(tx_queue);
652 		fill_level = max(tx_queue->insert_count - tx_queue->read_count,
653 				 txq2->insert_count - txq2->read_count);
654 		if (fill_level <= efx->txq_wake_thresh)
655 			netif_tx_wake_queue(tx_queue->core_txq);
656 	}
657 
658 	/* Check whether the hardware queue is now empty */
659 	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
660 		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
661 		if (tx_queue->read_count == tx_queue->old_write_count) {
662 			smp_mb();
663 			tx_queue->empty_read_count =
664 				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
665 		}
666 	}
667 }
668 
669 /* Size of page-based TSO header buffers.  Larger blocks must be
670  * allocated from the heap.
671  */
672 #define TSOH_STD_SIZE	128
673 #define TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
674 
675 /* At most half the descriptors in the queue at any time will refer to
676  * a TSO header buffer, since they must always be followed by a
677  * payload descriptor referring to an skb.
678  */
679 static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue)
680 {
681 	return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE);
682 }
683 
684 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
685 {
686 	struct efx_nic *efx = tx_queue->efx;
687 	unsigned int entries;
688 	int rc;
689 
690 	/* Create the smallest power-of-two aligned ring */
691 	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
692 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
693 	tx_queue->ptr_mask = entries - 1;
694 
695 	netif_dbg(efx, probe, efx->net_dev,
696 		  "creating TX queue %d size %#x mask %#x\n",
697 		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
698 
699 	/* Allocate software ring */
700 	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
701 				   GFP_KERNEL);
702 	if (!tx_queue->buffer)
703 		return -ENOMEM;
704 
705 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) {
706 		tx_queue->tsoh_page =
707 			kcalloc(efx_tsoh_page_count(tx_queue),
708 				sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL);
709 		if (!tx_queue->tsoh_page) {
710 			rc = -ENOMEM;
711 			goto fail1;
712 		}
713 	}
714 
715 	/* Allocate hardware ring */
716 	rc = efx_nic_probe_tx(tx_queue);
717 	if (rc)
718 		goto fail2;
719 
720 	return 0;
721 
722 fail2:
723 	kfree(tx_queue->tsoh_page);
724 	tx_queue->tsoh_page = NULL;
725 fail1:
726 	kfree(tx_queue->buffer);
727 	tx_queue->buffer = NULL;
728 	return rc;
729 }
730 
731 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
732 {
733 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
734 		  "initialising TX queue %d\n", tx_queue->queue);
735 
736 	tx_queue->insert_count = 0;
737 	tx_queue->write_count = 0;
738 	tx_queue->old_write_count = 0;
739 	tx_queue->read_count = 0;
740 	tx_queue->old_read_count = 0;
741 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
742 
743 	/* Set up TX descriptor ring */
744 	efx_nic_init_tx(tx_queue);
745 
746 	tx_queue->initialised = true;
747 }
748 
749 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
750 {
751 	struct efx_tx_buffer *buffer;
752 
753 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
754 		  "shutting down TX queue %d\n", tx_queue->queue);
755 
756 	if (!tx_queue->buffer)
757 		return;
758 
759 	/* Free any buffers left in the ring */
760 	while (tx_queue->read_count != tx_queue->write_count) {
761 		unsigned int pkts_compl = 0, bytes_compl = 0;
762 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
763 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
764 
765 		++tx_queue->read_count;
766 	}
767 	netdev_tx_reset_queue(tx_queue->core_txq);
768 }
769 
770 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
771 {
772 	int i;
773 
774 	if (!tx_queue->buffer)
775 		return;
776 
777 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
778 		  "destroying TX queue %d\n", tx_queue->queue);
779 	efx_nic_remove_tx(tx_queue);
780 
781 	if (tx_queue->tsoh_page) {
782 		for (i = 0; i < efx_tsoh_page_count(tx_queue); i++)
783 			efx_nic_free_buffer(tx_queue->efx,
784 					    &tx_queue->tsoh_page[i]);
785 		kfree(tx_queue->tsoh_page);
786 		tx_queue->tsoh_page = NULL;
787 	}
788 
789 	kfree(tx_queue->buffer);
790 	tx_queue->buffer = NULL;
791 }
792 
793 
794 /* Efx TCP segmentation acceleration.
795  *
796  * Why?  Because by doing it here in the driver we can go significantly
797  * faster than the GSO.
798  *
799  * Requires TX checksum offload support.
800  */
801 
802 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
803 
804 /**
805  * struct tso_state - TSO state for an SKB
806  * @out_len: Remaining length in current segment
807  * @seqnum: Current sequence number
808  * @ipv4_id: Current IPv4 ID, host endian
809  * @packet_space: Remaining space in current packet
810  * @dma_addr: DMA address of current position
811  * @in_len: Remaining length in current SKB fragment
812  * @unmap_len: Length of SKB fragment
813  * @unmap_addr: DMA address of SKB fragment
814  * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0
815  * @protocol: Network protocol (after any VLAN header)
816  * @ip_off: Offset of IP header
817  * @tcp_off: Offset of TCP header
818  * @header_len: Number of bytes of header
819  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
820  * @header_dma_addr: Header DMA address, when using option descriptors
821  * @header_unmap_len: Header DMA mapped length, or 0 if not using option
822  *	descriptors
823  *
824  * The state used during segmentation.  It is put into this data structure
825  * just to make it easy to pass into inline functions.
826  */
827 struct tso_state {
828 	/* Output position */
829 	unsigned out_len;
830 	unsigned seqnum;
831 	u16 ipv4_id;
832 	unsigned packet_space;
833 
834 	/* Input position */
835 	dma_addr_t dma_addr;
836 	unsigned in_len;
837 	unsigned unmap_len;
838 	dma_addr_t unmap_addr;
839 	unsigned short dma_flags;
840 
841 	__be16 protocol;
842 	unsigned int ip_off;
843 	unsigned int tcp_off;
844 	unsigned header_len;
845 	unsigned int ip_base_len;
846 	dma_addr_t header_dma_addr;
847 	unsigned int header_unmap_len;
848 };
849 
850 
851 /*
852  * Verify that our various assumptions about sk_buffs and the conditions
853  * under which TSO will be attempted hold true.  Return the protocol number.
854  */
855 static __be16 efx_tso_check_protocol(struct sk_buff *skb)
856 {
857 	__be16 protocol = skb->protocol;
858 
859 	EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
860 			    protocol);
861 	if (protocol == htons(ETH_P_8021Q)) {
862 		struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
863 		protocol = veh->h_vlan_encapsulated_proto;
864 	}
865 
866 	if (protocol == htons(ETH_P_IP)) {
867 		EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
868 	} else {
869 		EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
870 		EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
871 	}
872 	EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
873 			     + (tcp_hdr(skb)->doff << 2u)) >
874 			    skb_headlen(skb));
875 
876 	return protocol;
877 }
878 
879 static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue,
880 			       struct efx_tx_buffer *buffer, unsigned int len)
881 {
882 	u8 *result;
883 
884 	EFX_BUG_ON_PARANOID(buffer->len);
885 	EFX_BUG_ON_PARANOID(buffer->flags);
886 	EFX_BUG_ON_PARANOID(buffer->unmap_len);
887 
888 	if (likely(len <= TSOH_STD_SIZE - NET_IP_ALIGN)) {
889 		unsigned index =
890 			(tx_queue->insert_count & tx_queue->ptr_mask) / 2;
891 		struct efx_buffer *page_buf =
892 			&tx_queue->tsoh_page[index / TSOH_PER_PAGE];
893 		unsigned offset =
894 			TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + NET_IP_ALIGN;
895 
896 		if (unlikely(!page_buf->addr) &&
897 		    efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
898 					 GFP_ATOMIC))
899 			return NULL;
900 
901 		result = (u8 *)page_buf->addr + offset;
902 		buffer->dma_addr = page_buf->dma_addr + offset;
903 		buffer->flags = EFX_TX_BUF_CONT;
904 	} else {
905 		tx_queue->tso_long_headers++;
906 
907 		buffer->heap_buf = kmalloc(NET_IP_ALIGN + len, GFP_ATOMIC);
908 		if (unlikely(!buffer->heap_buf))
909 			return NULL;
910 		result = (u8 *)buffer->heap_buf + NET_IP_ALIGN;
911 		buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP;
912 	}
913 
914 	buffer->len = len;
915 
916 	return result;
917 }
918 
919 /**
920  * efx_tx_queue_insert - push descriptors onto the TX queue
921  * @tx_queue:		Efx TX queue
922  * @dma_addr:		DMA address of fragment
923  * @len:		Length of fragment
924  * @final_buffer:	The final buffer inserted into the queue
925  *
926  * Push descriptors onto the TX queue.
927  */
928 static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
929 				dma_addr_t dma_addr, unsigned len,
930 				struct efx_tx_buffer **final_buffer)
931 {
932 	struct efx_tx_buffer *buffer;
933 	struct efx_nic *efx = tx_queue->efx;
934 	unsigned dma_len;
935 
936 	EFX_BUG_ON_PARANOID(len <= 0);
937 
938 	while (1) {
939 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
940 		++tx_queue->insert_count;
941 
942 		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
943 				    tx_queue->read_count >=
944 				    efx->txq_entries);
945 
946 		buffer->dma_addr = dma_addr;
947 
948 		dma_len = efx_max_tx_len(efx, dma_addr);
949 
950 		/* If there is enough space to send then do so */
951 		if (dma_len >= len)
952 			break;
953 
954 		buffer->len = dma_len;
955 		buffer->flags = EFX_TX_BUF_CONT;
956 		dma_addr += dma_len;
957 		len -= dma_len;
958 	}
959 
960 	EFX_BUG_ON_PARANOID(!len);
961 	buffer->len = len;
962 	*final_buffer = buffer;
963 }
964 
965 
966 /*
967  * Put a TSO header into the TX queue.
968  *
969  * This is special-cased because we know that it is small enough to fit in
970  * a single fragment, and we know it doesn't cross a page boundary.  It
971  * also allows us to not worry about end-of-packet etc.
972  */
973 static int efx_tso_put_header(struct efx_tx_queue *tx_queue,
974 			      struct efx_tx_buffer *buffer, u8 *header)
975 {
976 	if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) {
977 		buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev,
978 						  header, buffer->len,
979 						  DMA_TO_DEVICE);
980 		if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev,
981 					       buffer->dma_addr))) {
982 			kfree(buffer->heap_buf);
983 			buffer->len = 0;
984 			buffer->flags = 0;
985 			return -ENOMEM;
986 		}
987 		buffer->unmap_len = buffer->len;
988 		buffer->dma_offset = 0;
989 		buffer->flags |= EFX_TX_BUF_MAP_SINGLE;
990 	}
991 
992 	++tx_queue->insert_count;
993 	return 0;
994 }
995 
996 
997 /* Remove buffers put into a tx_queue.  None of the buffers must have
998  * an skb attached.
999  */
1000 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
1001 {
1002 	struct efx_tx_buffer *buffer;
1003 
1004 	/* Work backwards until we hit the original insert pointer value */
1005 	while (tx_queue->insert_count != tx_queue->write_count) {
1006 		--tx_queue->insert_count;
1007 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
1008 		efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
1009 	}
1010 }
1011 
1012 
1013 /* Parse the SKB header and initialise state. */
1014 static int tso_start(struct tso_state *st, struct efx_nic *efx,
1015 		     const struct sk_buff *skb)
1016 {
1017 	bool use_opt_desc = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1018 	struct device *dma_dev = &efx->pci_dev->dev;
1019 	unsigned int header_len, in_len;
1020 	dma_addr_t dma_addr;
1021 
1022 	st->ip_off = skb_network_header(skb) - skb->data;
1023 	st->tcp_off = skb_transport_header(skb) - skb->data;
1024 	header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
1025 	in_len = skb_headlen(skb) - header_len;
1026 	st->header_len = header_len;
1027 	st->in_len = in_len;
1028 	if (st->protocol == htons(ETH_P_IP)) {
1029 		st->ip_base_len = st->header_len - st->ip_off;
1030 		st->ipv4_id = ntohs(ip_hdr(skb)->id);
1031 	} else {
1032 		st->ip_base_len = st->header_len - st->tcp_off;
1033 		st->ipv4_id = 0;
1034 	}
1035 	st->seqnum = ntohl(tcp_hdr(skb)->seq);
1036 
1037 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
1038 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
1039 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
1040 
1041 	st->out_len = skb->len - header_len;
1042 
1043 	if (!use_opt_desc) {
1044 		st->header_unmap_len = 0;
1045 
1046 		if (likely(in_len == 0)) {
1047 			st->dma_flags = 0;
1048 			st->unmap_len = 0;
1049 			return 0;
1050 		}
1051 
1052 		dma_addr = dma_map_single(dma_dev, skb->data + header_len,
1053 					  in_len, DMA_TO_DEVICE);
1054 		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
1055 		st->dma_addr = dma_addr;
1056 		st->unmap_addr = dma_addr;
1057 		st->unmap_len = in_len;
1058 	} else {
1059 		dma_addr = dma_map_single(dma_dev, skb->data,
1060 					  skb_headlen(skb), DMA_TO_DEVICE);
1061 		st->header_dma_addr = dma_addr;
1062 		st->header_unmap_len = skb_headlen(skb);
1063 		st->dma_flags = 0;
1064 		st->dma_addr = dma_addr + header_len;
1065 		st->unmap_len = 0;
1066 	}
1067 
1068 	return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0;
1069 }
1070 
1071 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
1072 			    skb_frag_t *frag)
1073 {
1074 	st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0,
1075 					  skb_frag_size(frag), DMA_TO_DEVICE);
1076 	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
1077 		st->dma_flags = 0;
1078 		st->unmap_len = skb_frag_size(frag);
1079 		st->in_len = skb_frag_size(frag);
1080 		st->dma_addr = st->unmap_addr;
1081 		return 0;
1082 	}
1083 	return -ENOMEM;
1084 }
1085 
1086 
1087 /**
1088  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1089  * @tx_queue:		Efx TX queue
1090  * @skb:		Socket buffer
1091  * @st:			TSO state
1092  *
1093  * Form descriptors for the current fragment, until we reach the end
1094  * of fragment or end-of-packet.
1095  */
1096 static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
1097 					  const struct sk_buff *skb,
1098 					  struct tso_state *st)
1099 {
1100 	struct efx_tx_buffer *buffer;
1101 	int n;
1102 
1103 	if (st->in_len == 0)
1104 		return;
1105 	if (st->packet_space == 0)
1106 		return;
1107 
1108 	EFX_BUG_ON_PARANOID(st->in_len <= 0);
1109 	EFX_BUG_ON_PARANOID(st->packet_space <= 0);
1110 
1111 	n = min(st->in_len, st->packet_space);
1112 
1113 	st->packet_space -= n;
1114 	st->out_len -= n;
1115 	st->in_len -= n;
1116 
1117 	efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
1118 
1119 	if (st->out_len == 0) {
1120 		/* Transfer ownership of the skb */
1121 		buffer->skb = skb;
1122 		buffer->flags = EFX_TX_BUF_SKB;
1123 	} else if (st->packet_space != 0) {
1124 		buffer->flags = EFX_TX_BUF_CONT;
1125 	}
1126 
1127 	if (st->in_len == 0) {
1128 		/* Transfer ownership of the DMA mapping */
1129 		buffer->unmap_len = st->unmap_len;
1130 		buffer->dma_offset = buffer->unmap_len - buffer->len;
1131 		buffer->flags |= st->dma_flags;
1132 		st->unmap_len = 0;
1133 	}
1134 
1135 	st->dma_addr += n;
1136 }
1137 
1138 
1139 /**
1140  * tso_start_new_packet - generate a new header and prepare for the new packet
1141  * @tx_queue:		Efx TX queue
1142  * @skb:		Socket buffer
1143  * @st:			TSO state
1144  *
1145  * Generate a new header and prepare for the new packet.  Return 0 on
1146  * success, or -%ENOMEM if failed to alloc header.
1147  */
1148 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
1149 				const struct sk_buff *skb,
1150 				struct tso_state *st)
1151 {
1152 	struct efx_tx_buffer *buffer =
1153 		efx_tx_queue_get_insert_buffer(tx_queue);
1154 	bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
1155 	u8 tcp_flags_clear;
1156 
1157 	if (!is_last) {
1158 		st->packet_space = skb_shinfo(skb)->gso_size;
1159 		tcp_flags_clear = 0x09; /* mask out FIN and PSH */
1160 	} else {
1161 		st->packet_space = st->out_len;
1162 		tcp_flags_clear = 0x00;
1163 	}
1164 
1165 	if (!st->header_unmap_len) {
1166 		/* Allocate and insert a DMA-mapped header buffer. */
1167 		struct tcphdr *tsoh_th;
1168 		unsigned ip_length;
1169 		u8 *header;
1170 		int rc;
1171 
1172 		header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
1173 		if (!header)
1174 			return -ENOMEM;
1175 
1176 		tsoh_th = (struct tcphdr *)(header + st->tcp_off);
1177 
1178 		/* Copy and update the headers. */
1179 		memcpy(header, skb->data, st->header_len);
1180 
1181 		tsoh_th->seq = htonl(st->seqnum);
1182 		((u8 *)tsoh_th)[13] &= ~tcp_flags_clear;
1183 
1184 		ip_length = st->ip_base_len + st->packet_space;
1185 
1186 		if (st->protocol == htons(ETH_P_IP)) {
1187 			struct iphdr *tsoh_iph =
1188 				(struct iphdr *)(header + st->ip_off);
1189 
1190 			tsoh_iph->tot_len = htons(ip_length);
1191 			tsoh_iph->id = htons(st->ipv4_id);
1192 		} else {
1193 			struct ipv6hdr *tsoh_iph =
1194 				(struct ipv6hdr *)(header + st->ip_off);
1195 
1196 			tsoh_iph->payload_len = htons(ip_length);
1197 		}
1198 
1199 		rc = efx_tso_put_header(tx_queue, buffer, header);
1200 		if (unlikely(rc))
1201 			return rc;
1202 	} else {
1203 		/* Send the original headers with a TSO option descriptor
1204 		 * in front
1205 		 */
1206 		u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear;
1207 
1208 		buffer->flags = EFX_TX_BUF_OPTION;
1209 		buffer->len = 0;
1210 		buffer->unmap_len = 0;
1211 		EFX_POPULATE_QWORD_5(buffer->option,
1212 				     ESF_DZ_TX_DESC_IS_OPT, 1,
1213 				     ESF_DZ_TX_OPTION_TYPE,
1214 				     ESE_DZ_TX_OPTION_DESC_TSO,
1215 				     ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags,
1216 				     ESF_DZ_TX_TSO_IP_ID, st->ipv4_id,
1217 				     ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum);
1218 		++tx_queue->insert_count;
1219 
1220 		/* We mapped the headers in tso_start().  Unmap them
1221 		 * when the last segment is completed.
1222 		 */
1223 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
1224 		buffer->dma_addr = st->header_dma_addr;
1225 		buffer->len = st->header_len;
1226 		if (is_last) {
1227 			buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE;
1228 			buffer->unmap_len = st->header_unmap_len;
1229 			buffer->dma_offset = 0;
1230 			/* Ensure we only unmap them once in case of a
1231 			 * later DMA mapping error and rollback
1232 			 */
1233 			st->header_unmap_len = 0;
1234 		} else {
1235 			buffer->flags = EFX_TX_BUF_CONT;
1236 			buffer->unmap_len = 0;
1237 		}
1238 		++tx_queue->insert_count;
1239 	}
1240 
1241 	st->seqnum += skb_shinfo(skb)->gso_size;
1242 
1243 	/* Linux leaves suitable gaps in the IP ID space for us to fill. */
1244 	++st->ipv4_id;
1245 
1246 	++tx_queue->tso_packets;
1247 
1248 	return 0;
1249 }
1250 
1251 
1252 /**
1253  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1254  * @tx_queue:		Efx TX queue
1255  * @skb:		Socket buffer
1256  *
1257  * Context: You must hold netif_tx_lock() to call this function.
1258  *
1259  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1260  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1261  * %NETDEV_TX_OK.
1262  */
1263 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1264 			       struct sk_buff *skb)
1265 {
1266 	struct efx_nic *efx = tx_queue->efx;
1267 	int frag_i, rc;
1268 	struct tso_state state;
1269 
1270 	/* Find the packet protocol and sanity-check it */
1271 	state.protocol = efx_tso_check_protocol(skb);
1272 
1273 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1274 
1275 	rc = tso_start(&state, efx, skb);
1276 	if (rc)
1277 		goto mem_err;
1278 
1279 	if (likely(state.in_len == 0)) {
1280 		/* Grab the first payload fragment. */
1281 		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1282 		frag_i = 0;
1283 		rc = tso_get_fragment(&state, efx,
1284 				      skb_shinfo(skb)->frags + frag_i);
1285 		if (rc)
1286 			goto mem_err;
1287 	} else {
1288 		/* Payload starts in the header area. */
1289 		frag_i = -1;
1290 	}
1291 
1292 	if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1293 		goto mem_err;
1294 
1295 	while (1) {
1296 		tso_fill_packet_with_fragment(tx_queue, skb, &state);
1297 
1298 		/* Move onto the next fragment? */
1299 		if (state.in_len == 0) {
1300 			if (++frag_i >= skb_shinfo(skb)->nr_frags)
1301 				/* End of payload reached. */
1302 				break;
1303 			rc = tso_get_fragment(&state, efx,
1304 					      skb_shinfo(skb)->frags + frag_i);
1305 			if (rc)
1306 				goto mem_err;
1307 		}
1308 
1309 		/* Start at new packet? */
1310 		if (state.packet_space == 0 &&
1311 		    tso_start_new_packet(tx_queue, skb, &state) < 0)
1312 			goto mem_err;
1313 	}
1314 
1315 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
1316 
1317 	/* Pass off to hardware */
1318 	efx_nic_push_buffers(tx_queue);
1319 
1320 	efx_tx_maybe_stop_queue(tx_queue);
1321 
1322 	tx_queue->tso_bursts++;
1323 	return NETDEV_TX_OK;
1324 
1325  mem_err:
1326 	netif_err(efx, tx_err, efx->net_dev,
1327 		  "Out of memory for TSO headers, or DMA mapping error\n");
1328 	dev_kfree_skb_any(skb);
1329 
1330 	/* Free the DMA mapping we were in the process of writing out */
1331 	if (state.unmap_len) {
1332 		if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE)
1333 			dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr,
1334 					 state.unmap_len, DMA_TO_DEVICE);
1335 		else
1336 			dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr,
1337 				       state.unmap_len, DMA_TO_DEVICE);
1338 	}
1339 
1340 	/* Free the header DMA mapping, if using option descriptors */
1341 	if (state.header_unmap_len)
1342 		dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr,
1343 				 state.header_unmap_len, DMA_TO_DEVICE);
1344 
1345 	efx_enqueue_unwind(tx_queue);
1346 	return NETDEV_TX_OK;
1347 }
1348