xref: /linux/drivers/net/ethernet/sfc/tx.c (revision d458cdf712e0c671e8e819abb16ecd6e44f9daec)
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->len -
69 					 buffer->unmap_len);
70 		if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
71 			dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
72 					 DMA_TO_DEVICE);
73 		else
74 			dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
75 				       DMA_TO_DEVICE);
76 		buffer->unmap_len = 0;
77 	}
78 
79 	if (buffer->flags & EFX_TX_BUF_SKB) {
80 		(*pkts_compl)++;
81 		(*bytes_compl) += buffer->skb->len;
82 		dev_kfree_skb_any((struct sk_buff *) buffer->skb);
83 		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
84 			   "TX queue %d transmission id %x complete\n",
85 			   tx_queue->queue, tx_queue->read_count);
86 	} else if (buffer->flags & EFX_TX_BUF_HEAP) {
87 		kfree(buffer->heap_buf);
88 	}
89 
90 	buffer->len = 0;
91 	buffer->flags = 0;
92 }
93 
94 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
95 			       struct sk_buff *skb);
96 
97 static inline unsigned
98 efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr)
99 {
100 	/* Depending on the NIC revision, we can use descriptor
101 	 * lengths up to 8K or 8K-1.  However, since PCI Express
102 	 * devices must split read requests at 4K boundaries, there is
103 	 * little benefit from using descriptors that cross those
104 	 * boundaries and we keep things simple by not doing so.
105 	 */
106 	unsigned len = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1;
107 
108 	/* Work around hardware bug for unaligned buffers. */
109 	if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf))
110 		len = min_t(unsigned, len, 512 - (dma_addr & 0xf));
111 
112 	return len;
113 }
114 
115 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
116 {
117 	/* Header and payload descriptor for each output segment, plus
118 	 * one for every input fragment boundary within a segment
119 	 */
120 	unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
121 
122 	/* Possibly one more per segment for the alignment workaround,
123 	 * or for option descriptors
124 	 */
125 	if (EFX_WORKAROUND_5391(efx) || efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
126 		max_descs += EFX_TSO_MAX_SEGS;
127 
128 	/* Possibly more for PCIe page boundaries within input fragments */
129 	if (PAGE_SIZE > EFX_PAGE_SIZE)
130 		max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
131 				   DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE));
132 
133 	return max_descs;
134 }
135 
136 /* Get partner of a TX queue, seen as part of the same net core queue */
137 static struct efx_tx_queue *efx_tx_queue_partner(struct efx_tx_queue *tx_queue)
138 {
139 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
140 		return tx_queue - EFX_TXQ_TYPE_OFFLOAD;
141 	else
142 		return tx_queue + EFX_TXQ_TYPE_OFFLOAD;
143 }
144 
145 static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
146 {
147 	/* We need to consider both queues that the net core sees as one */
148 	struct efx_tx_queue *txq2 = efx_tx_queue_partner(txq1);
149 	struct efx_nic *efx = txq1->efx;
150 	unsigned int fill_level;
151 
152 	fill_level = max(txq1->insert_count - txq1->old_read_count,
153 			 txq2->insert_count - txq2->old_read_count);
154 	if (likely(fill_level < efx->txq_stop_thresh))
155 		return;
156 
157 	/* We used the stale old_read_count above, which gives us a
158 	 * pessimistic estimate of the fill level (which may even
159 	 * validly be >= efx->txq_entries).  Now try again using
160 	 * read_count (more likely to be a cache miss).
161 	 *
162 	 * If we read read_count and then conditionally stop the
163 	 * queue, it is possible for the completion path to race with
164 	 * us and complete all outstanding descriptors in the middle,
165 	 * after which there will be no more completions to wake it.
166 	 * Therefore we stop the queue first, then read read_count
167 	 * (with a memory barrier to ensure the ordering), then
168 	 * restart the queue if the fill level turns out to be low
169 	 * enough.
170 	 */
171 	netif_tx_stop_queue(txq1->core_txq);
172 	smp_mb();
173 	txq1->old_read_count = ACCESS_ONCE(txq1->read_count);
174 	txq2->old_read_count = ACCESS_ONCE(txq2->read_count);
175 
176 	fill_level = max(txq1->insert_count - txq1->old_read_count,
177 			 txq2->insert_count - txq2->old_read_count);
178 	EFX_BUG_ON_PARANOID(fill_level >= efx->txq_entries);
179 	if (likely(fill_level < efx->txq_stop_thresh)) {
180 		smp_mb();
181 		if (likely(!efx->loopback_selftest))
182 			netif_tx_start_queue(txq1->core_txq);
183 	}
184 }
185 
186 #ifdef EFX_USE_PIO
187 
188 struct efx_short_copy_buffer {
189 	int used;
190 	u8 buf[L1_CACHE_BYTES];
191 };
192 
193 /* Copy to PIO, respecting that writes to PIO buffers must be dword aligned.
194  * Advances piobuf pointer. Leaves additional data in the copy buffer.
195  */
196 static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf,
197 				    u8 *data, int len,
198 				    struct efx_short_copy_buffer *copy_buf)
199 {
200 	int block_len = len & ~(sizeof(copy_buf->buf) - 1);
201 
202 	memcpy_toio(*piobuf, data, block_len);
203 	*piobuf += block_len;
204 	len -= block_len;
205 
206 	if (len) {
207 		data += block_len;
208 		BUG_ON(copy_buf->used);
209 		BUG_ON(len > sizeof(copy_buf->buf));
210 		memcpy(copy_buf->buf, data, len);
211 		copy_buf->used = len;
212 	}
213 }
214 
215 /* Copy to PIO, respecting dword alignment, popping data from copy buffer first.
216  * Advances piobuf pointer. Leaves additional data in the copy buffer.
217  */
218 static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf,
219 				       u8 *data, int len,
220 				       struct efx_short_copy_buffer *copy_buf)
221 {
222 	if (copy_buf->used) {
223 		/* if the copy buffer is partially full, fill it up and write */
224 		int copy_to_buf =
225 			min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len);
226 
227 		memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf);
228 		copy_buf->used += copy_to_buf;
229 
230 		/* if we didn't fill it up then we're done for now */
231 		if (copy_buf->used < sizeof(copy_buf->buf))
232 			return;
233 
234 		memcpy_toio(*piobuf, copy_buf->buf, sizeof(copy_buf->buf));
235 		*piobuf += sizeof(copy_buf->buf);
236 		data += copy_to_buf;
237 		len -= copy_to_buf;
238 		copy_buf->used = 0;
239 	}
240 
241 	efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf);
242 }
243 
244 static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf,
245 				  struct efx_short_copy_buffer *copy_buf)
246 {
247 	/* if there's anything in it, write the whole buffer, including junk */
248 	if (copy_buf->used)
249 		memcpy_toio(piobuf, copy_buf->buf, sizeof(copy_buf->buf));
250 }
251 
252 /* Traverse skb structure and copy fragments in to PIO buffer.
253  * Advances piobuf pointer.
254  */
255 static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb,
256 				     u8 __iomem **piobuf,
257 				     struct efx_short_copy_buffer *copy_buf)
258 {
259 	int i;
260 
261 	efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb),
262 				copy_buf);
263 
264 	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
265 		skb_frag_t *f = &skb_shinfo(skb)->frags[i];
266 		u8 *vaddr;
267 
268 		vaddr = kmap_atomic(skb_frag_page(f));
269 
270 		efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + f->page_offset,
271 					   skb_frag_size(f), copy_buf);
272 		kunmap_atomic(vaddr);
273 	}
274 
275 	EFX_BUG_ON_PARANOID(skb_shinfo(skb)->frag_list);
276 }
277 
278 static struct efx_tx_buffer *
279 efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
280 {
281 	struct efx_tx_buffer *buffer =
282 		efx_tx_queue_get_insert_buffer(tx_queue);
283 	u8 __iomem *piobuf = tx_queue->piobuf;
284 
285 	/* Copy to PIO buffer. Ensure the writes are padded to the end
286 	 * of a cache line, as this is required for write-combining to be
287 	 * effective on at least x86.
288 	 */
289 
290 	if (skb_shinfo(skb)->nr_frags) {
291 		/* The size of the copy buffer will ensure all writes
292 		 * are the size of a cache line.
293 		 */
294 		struct efx_short_copy_buffer copy_buf;
295 
296 		copy_buf.used = 0;
297 
298 		efx_skb_copy_bits_to_pio(tx_queue->efx, skb,
299 					 &piobuf, &copy_buf);
300 		efx_flush_copy_buffer(tx_queue->efx, piobuf, &copy_buf);
301 	} else {
302 		/* Pad the write to the size of a cache line.
303 		 * We can do this because we know the skb_shared_info sruct is
304 		 * after the source, and the destination buffer is big enough.
305 		 */
306 		BUILD_BUG_ON(L1_CACHE_BYTES >
307 			     SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
308 		memcpy_toio(tx_queue->piobuf, skb->data,
309 			    ALIGN(skb->len, L1_CACHE_BYTES));
310 	}
311 
312 	EFX_POPULATE_QWORD_5(buffer->option,
313 			     ESF_DZ_TX_DESC_IS_OPT, 1,
314 			     ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO,
315 			     ESF_DZ_TX_PIO_CONT, 0,
316 			     ESF_DZ_TX_PIO_BYTE_CNT, skb->len,
317 			     ESF_DZ_TX_PIO_BUF_ADDR,
318 			     tx_queue->piobuf_offset);
319 	++tx_queue->pio_packets;
320 	++tx_queue->insert_count;
321 	return buffer;
322 }
323 #endif /* EFX_USE_PIO */
324 
325 /*
326  * Add a socket buffer to a TX queue
327  *
328  * This maps all fragments of a socket buffer for DMA and adds them to
329  * the TX queue.  The queue's insert pointer will be incremented by
330  * the number of fragments in the socket buffer.
331  *
332  * If any DMA mapping fails, any mapped fragments will be unmapped,
333  * the queue's insert pointer will be restored to its original value.
334  *
335  * This function is split out from efx_hard_start_xmit to allow the
336  * loopback test to direct packets via specific TX queues.
337  *
338  * Returns NETDEV_TX_OK.
339  * You must hold netif_tx_lock() to call this function.
340  */
341 netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
342 {
343 	struct efx_nic *efx = tx_queue->efx;
344 	struct device *dma_dev = &efx->pci_dev->dev;
345 	struct efx_tx_buffer *buffer;
346 	skb_frag_t *fragment;
347 	unsigned int len, unmap_len = 0;
348 	dma_addr_t dma_addr, unmap_addr = 0;
349 	unsigned int dma_len;
350 	unsigned short dma_flags;
351 	int i = 0;
352 
353 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
354 
355 	if (skb_shinfo(skb)->gso_size)
356 		return efx_enqueue_skb_tso(tx_queue, skb);
357 
358 	/* Get size of the initial fragment */
359 	len = skb_headlen(skb);
360 
361 	/* Pad if necessary */
362 	if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
363 		EFX_BUG_ON_PARANOID(skb->data_len);
364 		len = 32 + 1;
365 		if (skb_pad(skb, len - skb->len))
366 			return NETDEV_TX_OK;
367 	}
368 
369 	/* Consider using PIO for short packets */
370 #ifdef EFX_USE_PIO
371 	if (skb->len <= efx_piobuf_size && tx_queue->piobuf &&
372 	    efx_nic_tx_is_empty(tx_queue) &&
373 	    efx_nic_tx_is_empty(efx_tx_queue_partner(tx_queue))) {
374 		buffer = efx_enqueue_skb_pio(tx_queue, skb);
375 		dma_flags = EFX_TX_BUF_OPTION;
376 		goto finish_packet;
377 	}
378 #endif
379 
380 	/* Map for DMA.  Use dma_map_single rather than dma_map_page
381 	 * since this is more efficient on machines with sparse
382 	 * memory.
383 	 */
384 	dma_flags = EFX_TX_BUF_MAP_SINGLE;
385 	dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE);
386 
387 	/* Process all fragments */
388 	while (1) {
389 		if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
390 			goto dma_err;
391 
392 		/* Store fields for marking in the per-fragment final
393 		 * descriptor */
394 		unmap_len = len;
395 		unmap_addr = dma_addr;
396 
397 		/* Add to TX queue, splitting across DMA boundaries */
398 		do {
399 			buffer = efx_tx_queue_get_insert_buffer(tx_queue);
400 
401 			dma_len = efx_max_tx_len(efx, dma_addr);
402 			if (likely(dma_len >= len))
403 				dma_len = len;
404 
405 			/* Fill out per descriptor fields */
406 			buffer->len = dma_len;
407 			buffer->dma_addr = dma_addr;
408 			buffer->flags = EFX_TX_BUF_CONT;
409 			len -= dma_len;
410 			dma_addr += dma_len;
411 			++tx_queue->insert_count;
412 		} while (len);
413 
414 		/* Transfer ownership of the unmapping to the final buffer */
415 		buffer->flags = EFX_TX_BUF_CONT | dma_flags;
416 		buffer->unmap_len = unmap_len;
417 		unmap_len = 0;
418 
419 		/* Get address and size of next fragment */
420 		if (i >= skb_shinfo(skb)->nr_frags)
421 			break;
422 		fragment = &skb_shinfo(skb)->frags[i];
423 		len = skb_frag_size(fragment);
424 		i++;
425 		/* Map for DMA */
426 		dma_flags = 0;
427 		dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
428 					    DMA_TO_DEVICE);
429 	}
430 
431 	/* Transfer ownership of the skb to the final buffer */
432 finish_packet:
433 	buffer->skb = skb;
434 	buffer->flags = EFX_TX_BUF_SKB | dma_flags;
435 
436 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
437 
438 	/* Pass off to hardware */
439 	efx_nic_push_buffers(tx_queue);
440 
441 	efx_tx_maybe_stop_queue(tx_queue);
442 
443 	return NETDEV_TX_OK;
444 
445  dma_err:
446 	netif_err(efx, tx_err, efx->net_dev,
447 		  " TX queue %d could not map skb with %d bytes %d "
448 		  "fragments for DMA\n", tx_queue->queue, skb->len,
449 		  skb_shinfo(skb)->nr_frags + 1);
450 
451 	/* Mark the packet as transmitted, and free the SKB ourselves */
452 	dev_kfree_skb_any(skb);
453 
454 	/* Work backwards until we hit the original insert pointer value */
455 	while (tx_queue->insert_count != tx_queue->write_count) {
456 		unsigned int pkts_compl = 0, bytes_compl = 0;
457 		--tx_queue->insert_count;
458 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
459 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
460 	}
461 
462 	/* Free the fragment we were mid-way through pushing */
463 	if (unmap_len) {
464 		if (dma_flags & EFX_TX_BUF_MAP_SINGLE)
465 			dma_unmap_single(dma_dev, unmap_addr, unmap_len,
466 					 DMA_TO_DEVICE);
467 		else
468 			dma_unmap_page(dma_dev, unmap_addr, unmap_len,
469 				       DMA_TO_DEVICE);
470 	}
471 
472 	return NETDEV_TX_OK;
473 }
474 
475 /* Remove packets from the TX queue
476  *
477  * This removes packets from the TX queue, up to and including the
478  * specified index.
479  */
480 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
481 				unsigned int index,
482 				unsigned int *pkts_compl,
483 				unsigned int *bytes_compl)
484 {
485 	struct efx_nic *efx = tx_queue->efx;
486 	unsigned int stop_index, read_ptr;
487 
488 	stop_index = (index + 1) & tx_queue->ptr_mask;
489 	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
490 
491 	while (read_ptr != stop_index) {
492 		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
493 
494 		if (!(buffer->flags & EFX_TX_BUF_OPTION) &&
495 		    unlikely(buffer->len == 0)) {
496 			netif_err(efx, tx_err, efx->net_dev,
497 				  "TX queue %d spurious TX completion id %x\n",
498 				  tx_queue->queue, read_ptr);
499 			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
500 			return;
501 		}
502 
503 		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
504 
505 		++tx_queue->read_count;
506 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
507 	}
508 }
509 
510 /* Initiate a packet transmission.  We use one channel per CPU
511  * (sharing when we have more CPUs than channels).  On Falcon, the TX
512  * completion events will be directed back to the CPU that transmitted
513  * the packet, which should be cache-efficient.
514  *
515  * Context: non-blocking.
516  * Note that returning anything other than NETDEV_TX_OK will cause the
517  * OS to free the skb.
518  */
519 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
520 				struct net_device *net_dev)
521 {
522 	struct efx_nic *efx = netdev_priv(net_dev);
523 	struct efx_tx_queue *tx_queue;
524 	unsigned index, type;
525 
526 	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
527 
528 	/* PTP "event" packet */
529 	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
530 	    unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
531 		return efx_ptp_tx(efx, skb);
532 	}
533 
534 	index = skb_get_queue_mapping(skb);
535 	type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
536 	if (index >= efx->n_tx_channels) {
537 		index -= efx->n_tx_channels;
538 		type |= EFX_TXQ_TYPE_HIGHPRI;
539 	}
540 	tx_queue = efx_get_tx_queue(efx, index, type);
541 
542 	return efx_enqueue_skb(tx_queue, skb);
543 }
544 
545 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
546 {
547 	struct efx_nic *efx = tx_queue->efx;
548 
549 	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
550 	tx_queue->core_txq =
551 		netdev_get_tx_queue(efx->net_dev,
552 				    tx_queue->queue / EFX_TXQ_TYPES +
553 				    ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
554 				     efx->n_tx_channels : 0));
555 }
556 
557 int efx_setup_tc(struct net_device *net_dev, u8 num_tc)
558 {
559 	struct efx_nic *efx = netdev_priv(net_dev);
560 	struct efx_channel *channel;
561 	struct efx_tx_queue *tx_queue;
562 	unsigned tc;
563 	int rc;
564 
565 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC)
566 		return -EINVAL;
567 
568 	if (num_tc == net_dev->num_tc)
569 		return 0;
570 
571 	for (tc = 0; tc < num_tc; tc++) {
572 		net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
573 		net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
574 	}
575 
576 	if (num_tc > net_dev->num_tc) {
577 		/* Initialise high-priority queues as necessary */
578 		efx_for_each_channel(channel, efx) {
579 			efx_for_each_possible_channel_tx_queue(tx_queue,
580 							       channel) {
581 				if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI))
582 					continue;
583 				if (!tx_queue->buffer) {
584 					rc = efx_probe_tx_queue(tx_queue);
585 					if (rc)
586 						return rc;
587 				}
588 				if (!tx_queue->initialised)
589 					efx_init_tx_queue(tx_queue);
590 				efx_init_tx_queue_core_txq(tx_queue);
591 			}
592 		}
593 	} else {
594 		/* Reduce number of classes before number of queues */
595 		net_dev->num_tc = num_tc;
596 	}
597 
598 	rc = netif_set_real_num_tx_queues(net_dev,
599 					  max_t(int, num_tc, 1) *
600 					  efx->n_tx_channels);
601 	if (rc)
602 		return rc;
603 
604 	/* Do not destroy high-priority queues when they become
605 	 * unused.  We would have to flush them first, and it is
606 	 * fairly difficult to flush a subset of TX queues.  Leave
607 	 * it to efx_fini_channels().
608 	 */
609 
610 	net_dev->num_tc = num_tc;
611 	return 0;
612 }
613 
614 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
615 {
616 	unsigned fill_level;
617 	struct efx_nic *efx = tx_queue->efx;
618 	struct efx_tx_queue *txq2;
619 	unsigned int pkts_compl = 0, bytes_compl = 0;
620 
621 	EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
622 
623 	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
624 	netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl);
625 
626 	if (pkts_compl > 1)
627 		++tx_queue->merge_events;
628 
629 	/* See if we need to restart the netif queue.  This memory
630 	 * barrier ensures that we write read_count (inside
631 	 * efx_dequeue_buffers()) before reading the queue status.
632 	 */
633 	smp_mb();
634 	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
635 	    likely(efx->port_enabled) &&
636 	    likely(netif_device_present(efx->net_dev))) {
637 		txq2 = efx_tx_queue_partner(tx_queue);
638 		fill_level = max(tx_queue->insert_count - tx_queue->read_count,
639 				 txq2->insert_count - txq2->read_count);
640 		if (fill_level <= efx->txq_wake_thresh)
641 			netif_tx_wake_queue(tx_queue->core_txq);
642 	}
643 
644 	/* Check whether the hardware queue is now empty */
645 	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
646 		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
647 		if (tx_queue->read_count == tx_queue->old_write_count) {
648 			smp_mb();
649 			tx_queue->empty_read_count =
650 				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
651 		}
652 	}
653 }
654 
655 /* Size of page-based TSO header buffers.  Larger blocks must be
656  * allocated from the heap.
657  */
658 #define TSOH_STD_SIZE	128
659 #define TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
660 
661 /* At most half the descriptors in the queue at any time will refer to
662  * a TSO header buffer, since they must always be followed by a
663  * payload descriptor referring to an skb.
664  */
665 static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue)
666 {
667 	return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE);
668 }
669 
670 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
671 {
672 	struct efx_nic *efx = tx_queue->efx;
673 	unsigned int entries;
674 	int rc;
675 
676 	/* Create the smallest power-of-two aligned ring */
677 	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
678 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
679 	tx_queue->ptr_mask = entries - 1;
680 
681 	netif_dbg(efx, probe, efx->net_dev,
682 		  "creating TX queue %d size %#x mask %#x\n",
683 		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
684 
685 	/* Allocate software ring */
686 	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
687 				   GFP_KERNEL);
688 	if (!tx_queue->buffer)
689 		return -ENOMEM;
690 
691 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) {
692 		tx_queue->tsoh_page =
693 			kcalloc(efx_tsoh_page_count(tx_queue),
694 				sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL);
695 		if (!tx_queue->tsoh_page) {
696 			rc = -ENOMEM;
697 			goto fail1;
698 		}
699 	}
700 
701 	/* Allocate hardware ring */
702 	rc = efx_nic_probe_tx(tx_queue);
703 	if (rc)
704 		goto fail2;
705 
706 	return 0;
707 
708 fail2:
709 	kfree(tx_queue->tsoh_page);
710 	tx_queue->tsoh_page = NULL;
711 fail1:
712 	kfree(tx_queue->buffer);
713 	tx_queue->buffer = NULL;
714 	return rc;
715 }
716 
717 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
718 {
719 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
720 		  "initialising TX queue %d\n", tx_queue->queue);
721 
722 	tx_queue->insert_count = 0;
723 	tx_queue->write_count = 0;
724 	tx_queue->old_write_count = 0;
725 	tx_queue->read_count = 0;
726 	tx_queue->old_read_count = 0;
727 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
728 
729 	/* Set up TX descriptor ring */
730 	efx_nic_init_tx(tx_queue);
731 
732 	tx_queue->initialised = true;
733 }
734 
735 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
736 {
737 	struct efx_tx_buffer *buffer;
738 
739 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
740 		  "shutting down TX queue %d\n", tx_queue->queue);
741 
742 	if (!tx_queue->buffer)
743 		return;
744 
745 	/* Free any buffers left in the ring */
746 	while (tx_queue->read_count != tx_queue->write_count) {
747 		unsigned int pkts_compl = 0, bytes_compl = 0;
748 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
749 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
750 
751 		++tx_queue->read_count;
752 	}
753 	netdev_tx_reset_queue(tx_queue->core_txq);
754 }
755 
756 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
757 {
758 	int i;
759 
760 	if (!tx_queue->buffer)
761 		return;
762 
763 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
764 		  "destroying TX queue %d\n", tx_queue->queue);
765 	efx_nic_remove_tx(tx_queue);
766 
767 	if (tx_queue->tsoh_page) {
768 		for (i = 0; i < efx_tsoh_page_count(tx_queue); i++)
769 			efx_nic_free_buffer(tx_queue->efx,
770 					    &tx_queue->tsoh_page[i]);
771 		kfree(tx_queue->tsoh_page);
772 		tx_queue->tsoh_page = NULL;
773 	}
774 
775 	kfree(tx_queue->buffer);
776 	tx_queue->buffer = NULL;
777 }
778 
779 
780 /* Efx TCP segmentation acceleration.
781  *
782  * Why?  Because by doing it here in the driver we can go significantly
783  * faster than the GSO.
784  *
785  * Requires TX checksum offload support.
786  */
787 
788 /* Number of bytes inserted at the start of a TSO header buffer,
789  * similar to NET_IP_ALIGN.
790  */
791 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
792 #define TSOH_OFFSET	0
793 #else
794 #define TSOH_OFFSET	NET_IP_ALIGN
795 #endif
796 
797 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
798 
799 /**
800  * struct tso_state - TSO state for an SKB
801  * @out_len: Remaining length in current segment
802  * @seqnum: Current sequence number
803  * @ipv4_id: Current IPv4 ID, host endian
804  * @packet_space: Remaining space in current packet
805  * @dma_addr: DMA address of current position
806  * @in_len: Remaining length in current SKB fragment
807  * @unmap_len: Length of SKB fragment
808  * @unmap_addr: DMA address of SKB fragment
809  * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0
810  * @protocol: Network protocol (after any VLAN header)
811  * @ip_off: Offset of IP header
812  * @tcp_off: Offset of TCP header
813  * @header_len: Number of bytes of header
814  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
815  * @header_dma_addr: Header DMA address, when using option descriptors
816  * @header_unmap_len: Header DMA mapped length, or 0 if not using option
817  *	descriptors
818  *
819  * The state used during segmentation.  It is put into this data structure
820  * just to make it easy to pass into inline functions.
821  */
822 struct tso_state {
823 	/* Output position */
824 	unsigned out_len;
825 	unsigned seqnum;
826 	u16 ipv4_id;
827 	unsigned packet_space;
828 
829 	/* Input position */
830 	dma_addr_t dma_addr;
831 	unsigned in_len;
832 	unsigned unmap_len;
833 	dma_addr_t unmap_addr;
834 	unsigned short dma_flags;
835 
836 	__be16 protocol;
837 	unsigned int ip_off;
838 	unsigned int tcp_off;
839 	unsigned header_len;
840 	unsigned int ip_base_len;
841 	dma_addr_t header_dma_addr;
842 	unsigned int header_unmap_len;
843 };
844 
845 
846 /*
847  * Verify that our various assumptions about sk_buffs and the conditions
848  * under which TSO will be attempted hold true.  Return the protocol number.
849  */
850 static __be16 efx_tso_check_protocol(struct sk_buff *skb)
851 {
852 	__be16 protocol = skb->protocol;
853 
854 	EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
855 			    protocol);
856 	if (protocol == htons(ETH_P_8021Q)) {
857 		struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
858 		protocol = veh->h_vlan_encapsulated_proto;
859 	}
860 
861 	if (protocol == htons(ETH_P_IP)) {
862 		EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
863 	} else {
864 		EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
865 		EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
866 	}
867 	EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
868 			     + (tcp_hdr(skb)->doff << 2u)) >
869 			    skb_headlen(skb));
870 
871 	return protocol;
872 }
873 
874 static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue,
875 			       struct efx_tx_buffer *buffer, unsigned int len)
876 {
877 	u8 *result;
878 
879 	EFX_BUG_ON_PARANOID(buffer->len);
880 	EFX_BUG_ON_PARANOID(buffer->flags);
881 	EFX_BUG_ON_PARANOID(buffer->unmap_len);
882 
883 	if (likely(len <= TSOH_STD_SIZE - TSOH_OFFSET)) {
884 		unsigned index =
885 			(tx_queue->insert_count & tx_queue->ptr_mask) / 2;
886 		struct efx_buffer *page_buf =
887 			&tx_queue->tsoh_page[index / TSOH_PER_PAGE];
888 		unsigned offset =
889 			TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + TSOH_OFFSET;
890 
891 		if (unlikely(!page_buf->addr) &&
892 		    efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
893 					 GFP_ATOMIC))
894 			return NULL;
895 
896 		result = (u8 *)page_buf->addr + offset;
897 		buffer->dma_addr = page_buf->dma_addr + offset;
898 		buffer->flags = EFX_TX_BUF_CONT;
899 	} else {
900 		tx_queue->tso_long_headers++;
901 
902 		buffer->heap_buf = kmalloc(TSOH_OFFSET + len, GFP_ATOMIC);
903 		if (unlikely(!buffer->heap_buf))
904 			return NULL;
905 		result = (u8 *)buffer->heap_buf + TSOH_OFFSET;
906 		buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP;
907 	}
908 
909 	buffer->len = len;
910 
911 	return result;
912 }
913 
914 /**
915  * efx_tx_queue_insert - push descriptors onto the TX queue
916  * @tx_queue:		Efx TX queue
917  * @dma_addr:		DMA address of fragment
918  * @len:		Length of fragment
919  * @final_buffer:	The final buffer inserted into the queue
920  *
921  * Push descriptors onto the TX queue.
922  */
923 static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
924 				dma_addr_t dma_addr, unsigned len,
925 				struct efx_tx_buffer **final_buffer)
926 {
927 	struct efx_tx_buffer *buffer;
928 	struct efx_nic *efx = tx_queue->efx;
929 	unsigned dma_len;
930 
931 	EFX_BUG_ON_PARANOID(len <= 0);
932 
933 	while (1) {
934 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
935 		++tx_queue->insert_count;
936 
937 		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
938 				    tx_queue->read_count >=
939 				    efx->txq_entries);
940 
941 		buffer->dma_addr = dma_addr;
942 
943 		dma_len = efx_max_tx_len(efx, dma_addr);
944 
945 		/* If there is enough space to send then do so */
946 		if (dma_len >= len)
947 			break;
948 
949 		buffer->len = dma_len;
950 		buffer->flags = EFX_TX_BUF_CONT;
951 		dma_addr += dma_len;
952 		len -= dma_len;
953 	}
954 
955 	EFX_BUG_ON_PARANOID(!len);
956 	buffer->len = len;
957 	*final_buffer = buffer;
958 }
959 
960 
961 /*
962  * Put a TSO header into the TX queue.
963  *
964  * This is special-cased because we know that it is small enough to fit in
965  * a single fragment, and we know it doesn't cross a page boundary.  It
966  * also allows us to not worry about end-of-packet etc.
967  */
968 static int efx_tso_put_header(struct efx_tx_queue *tx_queue,
969 			      struct efx_tx_buffer *buffer, u8 *header)
970 {
971 	if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) {
972 		buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev,
973 						  header, buffer->len,
974 						  DMA_TO_DEVICE);
975 		if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev,
976 					       buffer->dma_addr))) {
977 			kfree(buffer->heap_buf);
978 			buffer->len = 0;
979 			buffer->flags = 0;
980 			return -ENOMEM;
981 		}
982 		buffer->unmap_len = buffer->len;
983 		buffer->flags |= EFX_TX_BUF_MAP_SINGLE;
984 	}
985 
986 	++tx_queue->insert_count;
987 	return 0;
988 }
989 
990 
991 /* Remove buffers put into a tx_queue.  None of the buffers must have
992  * an skb attached.
993  */
994 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
995 {
996 	struct efx_tx_buffer *buffer;
997 
998 	/* Work backwards until we hit the original insert pointer value */
999 	while (tx_queue->insert_count != tx_queue->write_count) {
1000 		--tx_queue->insert_count;
1001 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
1002 		efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
1003 	}
1004 }
1005 
1006 
1007 /* Parse the SKB header and initialise state. */
1008 static int tso_start(struct tso_state *st, struct efx_nic *efx,
1009 		     const struct sk_buff *skb)
1010 {
1011 	bool use_options = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1012 	struct device *dma_dev = &efx->pci_dev->dev;
1013 	unsigned int header_len, in_len;
1014 	dma_addr_t dma_addr;
1015 
1016 	st->ip_off = skb_network_header(skb) - skb->data;
1017 	st->tcp_off = skb_transport_header(skb) - skb->data;
1018 	header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
1019 	in_len = skb_headlen(skb) - header_len;
1020 	st->header_len = header_len;
1021 	st->in_len = in_len;
1022 	if (st->protocol == htons(ETH_P_IP)) {
1023 		st->ip_base_len = st->header_len - st->ip_off;
1024 		st->ipv4_id = ntohs(ip_hdr(skb)->id);
1025 	} else {
1026 		st->ip_base_len = st->header_len - st->tcp_off;
1027 		st->ipv4_id = 0;
1028 	}
1029 	st->seqnum = ntohl(tcp_hdr(skb)->seq);
1030 
1031 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
1032 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
1033 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
1034 
1035 	st->out_len = skb->len - header_len;
1036 
1037 	if (!use_options) {
1038 		st->header_unmap_len = 0;
1039 
1040 		if (likely(in_len == 0)) {
1041 			st->dma_flags = 0;
1042 			st->unmap_len = 0;
1043 			return 0;
1044 		}
1045 
1046 		dma_addr = dma_map_single(dma_dev, skb->data + header_len,
1047 					  in_len, DMA_TO_DEVICE);
1048 		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
1049 		st->dma_addr = dma_addr;
1050 		st->unmap_addr = dma_addr;
1051 		st->unmap_len = in_len;
1052 	} else {
1053 		dma_addr = dma_map_single(dma_dev, skb->data,
1054 					  skb_headlen(skb), DMA_TO_DEVICE);
1055 		st->header_dma_addr = dma_addr;
1056 		st->header_unmap_len = skb_headlen(skb);
1057 		st->dma_flags = 0;
1058 		st->dma_addr = dma_addr + header_len;
1059 		st->unmap_len = 0;
1060 	}
1061 
1062 	return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0;
1063 }
1064 
1065 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
1066 			    skb_frag_t *frag)
1067 {
1068 	st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0,
1069 					  skb_frag_size(frag), DMA_TO_DEVICE);
1070 	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
1071 		st->dma_flags = 0;
1072 		st->unmap_len = skb_frag_size(frag);
1073 		st->in_len = skb_frag_size(frag);
1074 		st->dma_addr = st->unmap_addr;
1075 		return 0;
1076 	}
1077 	return -ENOMEM;
1078 }
1079 
1080 
1081 /**
1082  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1083  * @tx_queue:		Efx TX queue
1084  * @skb:		Socket buffer
1085  * @st:			TSO state
1086  *
1087  * Form descriptors for the current fragment, until we reach the end
1088  * of fragment or end-of-packet.
1089  */
1090 static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
1091 					  const struct sk_buff *skb,
1092 					  struct tso_state *st)
1093 {
1094 	struct efx_tx_buffer *buffer;
1095 	int n;
1096 
1097 	if (st->in_len == 0)
1098 		return;
1099 	if (st->packet_space == 0)
1100 		return;
1101 
1102 	EFX_BUG_ON_PARANOID(st->in_len <= 0);
1103 	EFX_BUG_ON_PARANOID(st->packet_space <= 0);
1104 
1105 	n = min(st->in_len, st->packet_space);
1106 
1107 	st->packet_space -= n;
1108 	st->out_len -= n;
1109 	st->in_len -= n;
1110 
1111 	efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
1112 
1113 	if (st->out_len == 0) {
1114 		/* Transfer ownership of the skb */
1115 		buffer->skb = skb;
1116 		buffer->flags = EFX_TX_BUF_SKB;
1117 	} else if (st->packet_space != 0) {
1118 		buffer->flags = EFX_TX_BUF_CONT;
1119 	}
1120 
1121 	if (st->in_len == 0) {
1122 		/* Transfer ownership of the DMA mapping */
1123 		buffer->unmap_len = st->unmap_len;
1124 		buffer->flags |= st->dma_flags;
1125 		st->unmap_len = 0;
1126 	}
1127 
1128 	st->dma_addr += n;
1129 }
1130 
1131 
1132 /**
1133  * tso_start_new_packet - generate a new header and prepare for the new packet
1134  * @tx_queue:		Efx TX queue
1135  * @skb:		Socket buffer
1136  * @st:			TSO state
1137  *
1138  * Generate a new header and prepare for the new packet.  Return 0 on
1139  * success, or -%ENOMEM if failed to alloc header.
1140  */
1141 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
1142 				const struct sk_buff *skb,
1143 				struct tso_state *st)
1144 {
1145 	struct efx_tx_buffer *buffer =
1146 		efx_tx_queue_get_insert_buffer(tx_queue);
1147 	bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
1148 	u8 tcp_flags_clear;
1149 
1150 	if (!is_last) {
1151 		st->packet_space = skb_shinfo(skb)->gso_size;
1152 		tcp_flags_clear = 0x09; /* mask out FIN and PSH */
1153 	} else {
1154 		st->packet_space = st->out_len;
1155 		tcp_flags_clear = 0x00;
1156 	}
1157 
1158 	if (!st->header_unmap_len) {
1159 		/* Allocate and insert a DMA-mapped header buffer. */
1160 		struct tcphdr *tsoh_th;
1161 		unsigned ip_length;
1162 		u8 *header;
1163 		int rc;
1164 
1165 		header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
1166 		if (!header)
1167 			return -ENOMEM;
1168 
1169 		tsoh_th = (struct tcphdr *)(header + st->tcp_off);
1170 
1171 		/* Copy and update the headers. */
1172 		memcpy(header, skb->data, st->header_len);
1173 
1174 		tsoh_th->seq = htonl(st->seqnum);
1175 		((u8 *)tsoh_th)[13] &= ~tcp_flags_clear;
1176 
1177 		ip_length = st->ip_base_len + st->packet_space;
1178 
1179 		if (st->protocol == htons(ETH_P_IP)) {
1180 			struct iphdr *tsoh_iph =
1181 				(struct iphdr *)(header + st->ip_off);
1182 
1183 			tsoh_iph->tot_len = htons(ip_length);
1184 			tsoh_iph->id = htons(st->ipv4_id);
1185 		} else {
1186 			struct ipv6hdr *tsoh_iph =
1187 				(struct ipv6hdr *)(header + st->ip_off);
1188 
1189 			tsoh_iph->payload_len = htons(ip_length);
1190 		}
1191 
1192 		rc = efx_tso_put_header(tx_queue, buffer, header);
1193 		if (unlikely(rc))
1194 			return rc;
1195 	} else {
1196 		/* Send the original headers with a TSO option descriptor
1197 		 * in front
1198 		 */
1199 		u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear;
1200 
1201 		buffer->flags = EFX_TX_BUF_OPTION;
1202 		buffer->len = 0;
1203 		buffer->unmap_len = 0;
1204 		EFX_POPULATE_QWORD_5(buffer->option,
1205 				     ESF_DZ_TX_DESC_IS_OPT, 1,
1206 				     ESF_DZ_TX_OPTION_TYPE,
1207 				     ESE_DZ_TX_OPTION_DESC_TSO,
1208 				     ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags,
1209 				     ESF_DZ_TX_TSO_IP_ID, st->ipv4_id,
1210 				     ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum);
1211 		++tx_queue->insert_count;
1212 
1213 		/* We mapped the headers in tso_start().  Unmap them
1214 		 * when the last segment is completed.
1215 		 */
1216 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
1217 		buffer->dma_addr = st->header_dma_addr;
1218 		buffer->len = st->header_len;
1219 		if (is_last) {
1220 			buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE;
1221 			buffer->unmap_len = st->header_unmap_len;
1222 			/* Ensure we only unmap them once in case of a
1223 			 * later DMA mapping error and rollback
1224 			 */
1225 			st->header_unmap_len = 0;
1226 		} else {
1227 			buffer->flags = EFX_TX_BUF_CONT;
1228 			buffer->unmap_len = 0;
1229 		}
1230 		++tx_queue->insert_count;
1231 	}
1232 
1233 	st->seqnum += skb_shinfo(skb)->gso_size;
1234 
1235 	/* Linux leaves suitable gaps in the IP ID space for us to fill. */
1236 	++st->ipv4_id;
1237 
1238 	++tx_queue->tso_packets;
1239 
1240 	return 0;
1241 }
1242 
1243 
1244 /**
1245  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1246  * @tx_queue:		Efx TX queue
1247  * @skb:		Socket buffer
1248  *
1249  * Context: You must hold netif_tx_lock() to call this function.
1250  *
1251  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1252  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1253  * %NETDEV_TX_OK.
1254  */
1255 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1256 			       struct sk_buff *skb)
1257 {
1258 	struct efx_nic *efx = tx_queue->efx;
1259 	int frag_i, rc;
1260 	struct tso_state state;
1261 
1262 	/* Find the packet protocol and sanity-check it */
1263 	state.protocol = efx_tso_check_protocol(skb);
1264 
1265 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1266 
1267 	rc = tso_start(&state, efx, skb);
1268 	if (rc)
1269 		goto mem_err;
1270 
1271 	if (likely(state.in_len == 0)) {
1272 		/* Grab the first payload fragment. */
1273 		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1274 		frag_i = 0;
1275 		rc = tso_get_fragment(&state, efx,
1276 				      skb_shinfo(skb)->frags + frag_i);
1277 		if (rc)
1278 			goto mem_err;
1279 	} else {
1280 		/* Payload starts in the header area. */
1281 		frag_i = -1;
1282 	}
1283 
1284 	if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1285 		goto mem_err;
1286 
1287 	while (1) {
1288 		tso_fill_packet_with_fragment(tx_queue, skb, &state);
1289 
1290 		/* Move onto the next fragment? */
1291 		if (state.in_len == 0) {
1292 			if (++frag_i >= skb_shinfo(skb)->nr_frags)
1293 				/* End of payload reached. */
1294 				break;
1295 			rc = tso_get_fragment(&state, efx,
1296 					      skb_shinfo(skb)->frags + frag_i);
1297 			if (rc)
1298 				goto mem_err;
1299 		}
1300 
1301 		/* Start at new packet? */
1302 		if (state.packet_space == 0 &&
1303 		    tso_start_new_packet(tx_queue, skb, &state) < 0)
1304 			goto mem_err;
1305 	}
1306 
1307 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
1308 
1309 	/* Pass off to hardware */
1310 	efx_nic_push_buffers(tx_queue);
1311 
1312 	efx_tx_maybe_stop_queue(tx_queue);
1313 
1314 	tx_queue->tso_bursts++;
1315 	return NETDEV_TX_OK;
1316 
1317  mem_err:
1318 	netif_err(efx, tx_err, efx->net_dev,
1319 		  "Out of memory for TSO headers, or DMA mapping error\n");
1320 	dev_kfree_skb_any(skb);
1321 
1322 	/* Free the DMA mapping we were in the process of writing out */
1323 	if (state.unmap_len) {
1324 		if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE)
1325 			dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr,
1326 					 state.unmap_len, DMA_TO_DEVICE);
1327 		else
1328 			dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr,
1329 				       state.unmap_len, DMA_TO_DEVICE);
1330 	}
1331 
1332 	/* Free the header DMA mapping, if using option descriptors */
1333 	if (state.header_unmap_len)
1334 		dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr,
1335 				 state.header_unmap_len, DMA_TO_DEVICE);
1336 
1337 	efx_enqueue_unwind(tx_queue);
1338 	return NETDEV_TX_OK;
1339 }
1340