xref: /openbmc/linux/drivers/net/ethernet/sfc/tx.c (revision 4f3db074)
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 		efx_nic_push_buffers(tx_queue);
436 
437 	tx_queue->tx_packets++;
438 
439 	return NETDEV_TX_OK;
440 
441  dma_err:
442 	netif_err(efx, tx_err, efx->net_dev,
443 		  " TX queue %d could not map skb with %d bytes %d "
444 		  "fragments for DMA\n", tx_queue->queue, skb->len,
445 		  skb_shinfo(skb)->nr_frags + 1);
446 
447 	/* Mark the packet as transmitted, and free the SKB ourselves */
448 	dev_kfree_skb_any(skb);
449 
450 	/* Work backwards until we hit the original insert pointer value */
451 	while (tx_queue->insert_count != old_insert_count) {
452 		unsigned int pkts_compl = 0, bytes_compl = 0;
453 		--tx_queue->insert_count;
454 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
455 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
456 	}
457 
458 	/* Free the fragment we were mid-way through pushing */
459 	if (unmap_len) {
460 		if (dma_flags & EFX_TX_BUF_MAP_SINGLE)
461 			dma_unmap_single(dma_dev, unmap_addr, unmap_len,
462 					 DMA_TO_DEVICE);
463 		else
464 			dma_unmap_page(dma_dev, unmap_addr, unmap_len,
465 				       DMA_TO_DEVICE);
466 	}
467 
468 	return NETDEV_TX_OK;
469 }
470 
471 /* Remove packets from the TX queue
472  *
473  * This removes packets from the TX queue, up to and including the
474  * specified index.
475  */
476 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
477 				unsigned int index,
478 				unsigned int *pkts_compl,
479 				unsigned int *bytes_compl)
480 {
481 	struct efx_nic *efx = tx_queue->efx;
482 	unsigned int stop_index, read_ptr;
483 
484 	stop_index = (index + 1) & tx_queue->ptr_mask;
485 	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
486 
487 	while (read_ptr != stop_index) {
488 		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
489 
490 		if (!(buffer->flags & EFX_TX_BUF_OPTION) &&
491 		    unlikely(buffer->len == 0)) {
492 			netif_err(efx, tx_err, efx->net_dev,
493 				  "TX queue %d spurious TX completion id %x\n",
494 				  tx_queue->queue, read_ptr);
495 			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
496 			return;
497 		}
498 
499 		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
500 
501 		++tx_queue->read_count;
502 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
503 	}
504 }
505 
506 /* Initiate a packet transmission.  We use one channel per CPU
507  * (sharing when we have more CPUs than channels).  On Falcon, the TX
508  * completion events will be directed back to the CPU that transmitted
509  * the packet, which should be cache-efficient.
510  *
511  * Context: non-blocking.
512  * Note that returning anything other than NETDEV_TX_OK will cause the
513  * OS to free the skb.
514  */
515 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
516 				struct net_device *net_dev)
517 {
518 	struct efx_nic *efx = netdev_priv(net_dev);
519 	struct efx_tx_queue *tx_queue;
520 	unsigned index, type;
521 
522 	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
523 
524 	/* PTP "event" packet */
525 	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
526 	    unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
527 		return efx_ptp_tx(efx, skb);
528 	}
529 
530 	index = skb_get_queue_mapping(skb);
531 	type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
532 	if (index >= efx->n_tx_channels) {
533 		index -= efx->n_tx_channels;
534 		type |= EFX_TXQ_TYPE_HIGHPRI;
535 	}
536 	tx_queue = efx_get_tx_queue(efx, index, type);
537 
538 	return efx_enqueue_skb(tx_queue, skb);
539 }
540 
541 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
542 {
543 	struct efx_nic *efx = tx_queue->efx;
544 
545 	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
546 	tx_queue->core_txq =
547 		netdev_get_tx_queue(efx->net_dev,
548 				    tx_queue->queue / EFX_TXQ_TYPES +
549 				    ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
550 				     efx->n_tx_channels : 0));
551 }
552 
553 int efx_setup_tc(struct net_device *net_dev, u8 num_tc)
554 {
555 	struct efx_nic *efx = netdev_priv(net_dev);
556 	struct efx_channel *channel;
557 	struct efx_tx_queue *tx_queue;
558 	unsigned tc;
559 	int rc;
560 
561 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC)
562 		return -EINVAL;
563 
564 	if (num_tc == net_dev->num_tc)
565 		return 0;
566 
567 	for (tc = 0; tc < num_tc; tc++) {
568 		net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
569 		net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
570 	}
571 
572 	if (num_tc > net_dev->num_tc) {
573 		/* Initialise high-priority queues as necessary */
574 		efx_for_each_channel(channel, efx) {
575 			efx_for_each_possible_channel_tx_queue(tx_queue,
576 							       channel) {
577 				if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI))
578 					continue;
579 				if (!tx_queue->buffer) {
580 					rc = efx_probe_tx_queue(tx_queue);
581 					if (rc)
582 						return rc;
583 				}
584 				if (!tx_queue->initialised)
585 					efx_init_tx_queue(tx_queue);
586 				efx_init_tx_queue_core_txq(tx_queue);
587 			}
588 		}
589 	} else {
590 		/* Reduce number of classes before number of queues */
591 		net_dev->num_tc = num_tc;
592 	}
593 
594 	rc = netif_set_real_num_tx_queues(net_dev,
595 					  max_t(int, num_tc, 1) *
596 					  efx->n_tx_channels);
597 	if (rc)
598 		return rc;
599 
600 	/* Do not destroy high-priority queues when they become
601 	 * unused.  We would have to flush them first, and it is
602 	 * fairly difficult to flush a subset of TX queues.  Leave
603 	 * it to efx_fini_channels().
604 	 */
605 
606 	net_dev->num_tc = num_tc;
607 	return 0;
608 }
609 
610 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
611 {
612 	unsigned fill_level;
613 	struct efx_nic *efx = tx_queue->efx;
614 	struct efx_tx_queue *txq2;
615 	unsigned int pkts_compl = 0, bytes_compl = 0;
616 
617 	EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
618 
619 	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
620 	netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl);
621 
622 	if (pkts_compl > 1)
623 		++tx_queue->merge_events;
624 
625 	/* See if we need to restart the netif queue.  This memory
626 	 * barrier ensures that we write read_count (inside
627 	 * efx_dequeue_buffers()) before reading the queue status.
628 	 */
629 	smp_mb();
630 	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
631 	    likely(efx->port_enabled) &&
632 	    likely(netif_device_present(efx->net_dev))) {
633 		txq2 = efx_tx_queue_partner(tx_queue);
634 		fill_level = max(tx_queue->insert_count - tx_queue->read_count,
635 				 txq2->insert_count - txq2->read_count);
636 		if (fill_level <= efx->txq_wake_thresh)
637 			netif_tx_wake_queue(tx_queue->core_txq);
638 	}
639 
640 	/* Check whether the hardware queue is now empty */
641 	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
642 		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
643 		if (tx_queue->read_count == tx_queue->old_write_count) {
644 			smp_mb();
645 			tx_queue->empty_read_count =
646 				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
647 		}
648 	}
649 }
650 
651 /* Size of page-based TSO header buffers.  Larger blocks must be
652  * allocated from the heap.
653  */
654 #define TSOH_STD_SIZE	128
655 #define TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
656 
657 /* At most half the descriptors in the queue at any time will refer to
658  * a TSO header buffer, since they must always be followed by a
659  * payload descriptor referring to an skb.
660  */
661 static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue)
662 {
663 	return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE);
664 }
665 
666 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
667 {
668 	struct efx_nic *efx = tx_queue->efx;
669 	unsigned int entries;
670 	int rc;
671 
672 	/* Create the smallest power-of-two aligned ring */
673 	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
674 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
675 	tx_queue->ptr_mask = entries - 1;
676 
677 	netif_dbg(efx, probe, efx->net_dev,
678 		  "creating TX queue %d size %#x mask %#x\n",
679 		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
680 
681 	/* Allocate software ring */
682 	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
683 				   GFP_KERNEL);
684 	if (!tx_queue->buffer)
685 		return -ENOMEM;
686 
687 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) {
688 		tx_queue->tsoh_page =
689 			kcalloc(efx_tsoh_page_count(tx_queue),
690 				sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL);
691 		if (!tx_queue->tsoh_page) {
692 			rc = -ENOMEM;
693 			goto fail1;
694 		}
695 	}
696 
697 	/* Allocate hardware ring */
698 	rc = efx_nic_probe_tx(tx_queue);
699 	if (rc)
700 		goto fail2;
701 
702 	return 0;
703 
704 fail2:
705 	kfree(tx_queue->tsoh_page);
706 	tx_queue->tsoh_page = NULL;
707 fail1:
708 	kfree(tx_queue->buffer);
709 	tx_queue->buffer = NULL;
710 	return rc;
711 }
712 
713 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
714 {
715 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
716 		  "initialising TX queue %d\n", tx_queue->queue);
717 
718 	tx_queue->insert_count = 0;
719 	tx_queue->write_count = 0;
720 	tx_queue->old_write_count = 0;
721 	tx_queue->read_count = 0;
722 	tx_queue->old_read_count = 0;
723 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
724 
725 	/* Set up TX descriptor ring */
726 	efx_nic_init_tx(tx_queue);
727 
728 	tx_queue->initialised = true;
729 }
730 
731 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
732 {
733 	struct efx_tx_buffer *buffer;
734 
735 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
736 		  "shutting down TX queue %d\n", tx_queue->queue);
737 
738 	if (!tx_queue->buffer)
739 		return;
740 
741 	/* Free any buffers left in the ring */
742 	while (tx_queue->read_count != tx_queue->write_count) {
743 		unsigned int pkts_compl = 0, bytes_compl = 0;
744 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
745 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
746 
747 		++tx_queue->read_count;
748 	}
749 	netdev_tx_reset_queue(tx_queue->core_txq);
750 }
751 
752 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
753 {
754 	int i;
755 
756 	if (!tx_queue->buffer)
757 		return;
758 
759 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
760 		  "destroying TX queue %d\n", tx_queue->queue);
761 	efx_nic_remove_tx(tx_queue);
762 
763 	if (tx_queue->tsoh_page) {
764 		for (i = 0; i < efx_tsoh_page_count(tx_queue); i++)
765 			efx_nic_free_buffer(tx_queue->efx,
766 					    &tx_queue->tsoh_page[i]);
767 		kfree(tx_queue->tsoh_page);
768 		tx_queue->tsoh_page = NULL;
769 	}
770 
771 	kfree(tx_queue->buffer);
772 	tx_queue->buffer = NULL;
773 }
774 
775 
776 /* Efx TCP segmentation acceleration.
777  *
778  * Why?  Because by doing it here in the driver we can go significantly
779  * faster than the GSO.
780  *
781  * Requires TX checksum offload support.
782  */
783 
784 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
785 
786 /**
787  * struct tso_state - TSO state for an SKB
788  * @out_len: Remaining length in current segment
789  * @seqnum: Current sequence number
790  * @ipv4_id: Current IPv4 ID, host endian
791  * @packet_space: Remaining space in current packet
792  * @dma_addr: DMA address of current position
793  * @in_len: Remaining length in current SKB fragment
794  * @unmap_len: Length of SKB fragment
795  * @unmap_addr: DMA address of SKB fragment
796  * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0
797  * @protocol: Network protocol (after any VLAN header)
798  * @ip_off: Offset of IP header
799  * @tcp_off: Offset of TCP header
800  * @header_len: Number of bytes of header
801  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
802  * @header_dma_addr: Header DMA address, when using option descriptors
803  * @header_unmap_len: Header DMA mapped length, or 0 if not using option
804  *	descriptors
805  *
806  * The state used during segmentation.  It is put into this data structure
807  * just to make it easy to pass into inline functions.
808  */
809 struct tso_state {
810 	/* Output position */
811 	unsigned out_len;
812 	unsigned seqnum;
813 	u16 ipv4_id;
814 	unsigned packet_space;
815 
816 	/* Input position */
817 	dma_addr_t dma_addr;
818 	unsigned in_len;
819 	unsigned unmap_len;
820 	dma_addr_t unmap_addr;
821 	unsigned short dma_flags;
822 
823 	__be16 protocol;
824 	unsigned int ip_off;
825 	unsigned int tcp_off;
826 	unsigned header_len;
827 	unsigned int ip_base_len;
828 	dma_addr_t header_dma_addr;
829 	unsigned int header_unmap_len;
830 };
831 
832 
833 /*
834  * Verify that our various assumptions about sk_buffs and the conditions
835  * under which TSO will be attempted hold true.  Return the protocol number.
836  */
837 static __be16 efx_tso_check_protocol(struct sk_buff *skb)
838 {
839 	__be16 protocol = skb->protocol;
840 
841 	EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
842 			    protocol);
843 	if (protocol == htons(ETH_P_8021Q)) {
844 		struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
845 		protocol = veh->h_vlan_encapsulated_proto;
846 	}
847 
848 	if (protocol == htons(ETH_P_IP)) {
849 		EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
850 	} else {
851 		EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
852 		EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
853 	}
854 	EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
855 			     + (tcp_hdr(skb)->doff << 2u)) >
856 			    skb_headlen(skb));
857 
858 	return protocol;
859 }
860 
861 static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue,
862 			       struct efx_tx_buffer *buffer, unsigned int len)
863 {
864 	u8 *result;
865 
866 	EFX_BUG_ON_PARANOID(buffer->len);
867 	EFX_BUG_ON_PARANOID(buffer->flags);
868 	EFX_BUG_ON_PARANOID(buffer->unmap_len);
869 
870 	if (likely(len <= TSOH_STD_SIZE - NET_IP_ALIGN)) {
871 		unsigned index =
872 			(tx_queue->insert_count & tx_queue->ptr_mask) / 2;
873 		struct efx_buffer *page_buf =
874 			&tx_queue->tsoh_page[index / TSOH_PER_PAGE];
875 		unsigned offset =
876 			TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + NET_IP_ALIGN;
877 
878 		if (unlikely(!page_buf->addr) &&
879 		    efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
880 					 GFP_ATOMIC))
881 			return NULL;
882 
883 		result = (u8 *)page_buf->addr + offset;
884 		buffer->dma_addr = page_buf->dma_addr + offset;
885 		buffer->flags = EFX_TX_BUF_CONT;
886 	} else {
887 		tx_queue->tso_long_headers++;
888 
889 		buffer->heap_buf = kmalloc(NET_IP_ALIGN + len, GFP_ATOMIC);
890 		if (unlikely(!buffer->heap_buf))
891 			return NULL;
892 		result = (u8 *)buffer->heap_buf + NET_IP_ALIGN;
893 		buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP;
894 	}
895 
896 	buffer->len = len;
897 
898 	return result;
899 }
900 
901 /**
902  * efx_tx_queue_insert - push descriptors onto the TX queue
903  * @tx_queue:		Efx TX queue
904  * @dma_addr:		DMA address of fragment
905  * @len:		Length of fragment
906  * @final_buffer:	The final buffer inserted into the queue
907  *
908  * Push descriptors onto the TX queue.
909  */
910 static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
911 				dma_addr_t dma_addr, unsigned len,
912 				struct efx_tx_buffer **final_buffer)
913 {
914 	struct efx_tx_buffer *buffer;
915 	struct efx_nic *efx = tx_queue->efx;
916 	unsigned dma_len;
917 
918 	EFX_BUG_ON_PARANOID(len <= 0);
919 
920 	while (1) {
921 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
922 		++tx_queue->insert_count;
923 
924 		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
925 				    tx_queue->read_count >=
926 				    efx->txq_entries);
927 
928 		buffer->dma_addr = dma_addr;
929 
930 		dma_len = efx_max_tx_len(efx, dma_addr);
931 
932 		/* If there is enough space to send then do so */
933 		if (dma_len >= len)
934 			break;
935 
936 		buffer->len = dma_len;
937 		buffer->flags = EFX_TX_BUF_CONT;
938 		dma_addr += dma_len;
939 		len -= dma_len;
940 	}
941 
942 	EFX_BUG_ON_PARANOID(!len);
943 	buffer->len = len;
944 	*final_buffer = buffer;
945 }
946 
947 
948 /*
949  * Put a TSO header into the TX queue.
950  *
951  * This is special-cased because we know that it is small enough to fit in
952  * a single fragment, and we know it doesn't cross a page boundary.  It
953  * also allows us to not worry about end-of-packet etc.
954  */
955 static int efx_tso_put_header(struct efx_tx_queue *tx_queue,
956 			      struct efx_tx_buffer *buffer, u8 *header)
957 {
958 	if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) {
959 		buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev,
960 						  header, buffer->len,
961 						  DMA_TO_DEVICE);
962 		if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev,
963 					       buffer->dma_addr))) {
964 			kfree(buffer->heap_buf);
965 			buffer->len = 0;
966 			buffer->flags = 0;
967 			return -ENOMEM;
968 		}
969 		buffer->unmap_len = buffer->len;
970 		buffer->dma_offset = 0;
971 		buffer->flags |= EFX_TX_BUF_MAP_SINGLE;
972 	}
973 
974 	++tx_queue->insert_count;
975 	return 0;
976 }
977 
978 
979 /* Remove buffers put into a tx_queue.  None of the buffers must have
980  * an skb attached.
981  */
982 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
983 			       unsigned int insert_count)
984 {
985 	struct efx_tx_buffer *buffer;
986 
987 	/* Work backwards until we hit the original insert pointer value */
988 	while (tx_queue->insert_count != insert_count) {
989 		--tx_queue->insert_count;
990 		buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
991 		efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
992 	}
993 }
994 
995 
996 /* Parse the SKB header and initialise state. */
997 static int tso_start(struct tso_state *st, struct efx_nic *efx,
998 		     const struct sk_buff *skb)
999 {
1000 	bool use_opt_desc = efx_nic_rev(efx) >= EFX_REV_HUNT_A0;
1001 	struct device *dma_dev = &efx->pci_dev->dev;
1002 	unsigned int header_len, in_len;
1003 	dma_addr_t dma_addr;
1004 
1005 	st->ip_off = skb_network_header(skb) - skb->data;
1006 	st->tcp_off = skb_transport_header(skb) - skb->data;
1007 	header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
1008 	in_len = skb_headlen(skb) - header_len;
1009 	st->header_len = header_len;
1010 	st->in_len = in_len;
1011 	if (st->protocol == htons(ETH_P_IP)) {
1012 		st->ip_base_len = st->header_len - st->ip_off;
1013 		st->ipv4_id = ntohs(ip_hdr(skb)->id);
1014 	} else {
1015 		st->ip_base_len = st->header_len - st->tcp_off;
1016 		st->ipv4_id = 0;
1017 	}
1018 	st->seqnum = ntohl(tcp_hdr(skb)->seq);
1019 
1020 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
1021 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
1022 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
1023 
1024 	st->out_len = skb->len - header_len;
1025 
1026 	if (!use_opt_desc) {
1027 		st->header_unmap_len = 0;
1028 
1029 		if (likely(in_len == 0)) {
1030 			st->dma_flags = 0;
1031 			st->unmap_len = 0;
1032 			return 0;
1033 		}
1034 
1035 		dma_addr = dma_map_single(dma_dev, skb->data + header_len,
1036 					  in_len, DMA_TO_DEVICE);
1037 		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
1038 		st->dma_addr = dma_addr;
1039 		st->unmap_addr = dma_addr;
1040 		st->unmap_len = in_len;
1041 	} else {
1042 		dma_addr = dma_map_single(dma_dev, skb->data,
1043 					  skb_headlen(skb), DMA_TO_DEVICE);
1044 		st->header_dma_addr = dma_addr;
1045 		st->header_unmap_len = skb_headlen(skb);
1046 		st->dma_flags = 0;
1047 		st->dma_addr = dma_addr + header_len;
1048 		st->unmap_len = 0;
1049 	}
1050 
1051 	return unlikely(dma_mapping_error(dma_dev, dma_addr)) ? -ENOMEM : 0;
1052 }
1053 
1054 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
1055 			    skb_frag_t *frag)
1056 {
1057 	st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0,
1058 					  skb_frag_size(frag), DMA_TO_DEVICE);
1059 	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
1060 		st->dma_flags = 0;
1061 		st->unmap_len = skb_frag_size(frag);
1062 		st->in_len = skb_frag_size(frag);
1063 		st->dma_addr = st->unmap_addr;
1064 		return 0;
1065 	}
1066 	return -ENOMEM;
1067 }
1068 
1069 
1070 /**
1071  * tso_fill_packet_with_fragment - form descriptors for the current fragment
1072  * @tx_queue:		Efx TX queue
1073  * @skb:		Socket buffer
1074  * @st:			TSO state
1075  *
1076  * Form descriptors for the current fragment, until we reach the end
1077  * of fragment or end-of-packet.
1078  */
1079 static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
1080 					  const struct sk_buff *skb,
1081 					  struct tso_state *st)
1082 {
1083 	struct efx_tx_buffer *buffer;
1084 	int n;
1085 
1086 	if (st->in_len == 0)
1087 		return;
1088 	if (st->packet_space == 0)
1089 		return;
1090 
1091 	EFX_BUG_ON_PARANOID(st->in_len <= 0);
1092 	EFX_BUG_ON_PARANOID(st->packet_space <= 0);
1093 
1094 	n = min(st->in_len, st->packet_space);
1095 
1096 	st->packet_space -= n;
1097 	st->out_len -= n;
1098 	st->in_len -= n;
1099 
1100 	efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
1101 
1102 	if (st->out_len == 0) {
1103 		/* Transfer ownership of the skb */
1104 		buffer->skb = skb;
1105 		buffer->flags = EFX_TX_BUF_SKB;
1106 	} else if (st->packet_space != 0) {
1107 		buffer->flags = EFX_TX_BUF_CONT;
1108 	}
1109 
1110 	if (st->in_len == 0) {
1111 		/* Transfer ownership of the DMA mapping */
1112 		buffer->unmap_len = st->unmap_len;
1113 		buffer->dma_offset = buffer->unmap_len - buffer->len;
1114 		buffer->flags |= st->dma_flags;
1115 		st->unmap_len = 0;
1116 	}
1117 
1118 	st->dma_addr += n;
1119 }
1120 
1121 
1122 /**
1123  * tso_start_new_packet - generate a new header and prepare for the new packet
1124  * @tx_queue:		Efx TX queue
1125  * @skb:		Socket buffer
1126  * @st:			TSO state
1127  *
1128  * Generate a new header and prepare for the new packet.  Return 0 on
1129  * success, or -%ENOMEM if failed to alloc header.
1130  */
1131 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
1132 				const struct sk_buff *skb,
1133 				struct tso_state *st)
1134 {
1135 	struct efx_tx_buffer *buffer =
1136 		efx_tx_queue_get_insert_buffer(tx_queue);
1137 	bool is_last = st->out_len <= skb_shinfo(skb)->gso_size;
1138 	u8 tcp_flags_clear;
1139 
1140 	if (!is_last) {
1141 		st->packet_space = skb_shinfo(skb)->gso_size;
1142 		tcp_flags_clear = 0x09; /* mask out FIN and PSH */
1143 	} else {
1144 		st->packet_space = st->out_len;
1145 		tcp_flags_clear = 0x00;
1146 	}
1147 
1148 	if (!st->header_unmap_len) {
1149 		/* Allocate and insert a DMA-mapped header buffer. */
1150 		struct tcphdr *tsoh_th;
1151 		unsigned ip_length;
1152 		u8 *header;
1153 		int rc;
1154 
1155 		header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
1156 		if (!header)
1157 			return -ENOMEM;
1158 
1159 		tsoh_th = (struct tcphdr *)(header + st->tcp_off);
1160 
1161 		/* Copy and update the headers. */
1162 		memcpy(header, skb->data, st->header_len);
1163 
1164 		tsoh_th->seq = htonl(st->seqnum);
1165 		((u8 *)tsoh_th)[13] &= ~tcp_flags_clear;
1166 
1167 		ip_length = st->ip_base_len + st->packet_space;
1168 
1169 		if (st->protocol == htons(ETH_P_IP)) {
1170 			struct iphdr *tsoh_iph =
1171 				(struct iphdr *)(header + st->ip_off);
1172 
1173 			tsoh_iph->tot_len = htons(ip_length);
1174 			tsoh_iph->id = htons(st->ipv4_id);
1175 		} else {
1176 			struct ipv6hdr *tsoh_iph =
1177 				(struct ipv6hdr *)(header + st->ip_off);
1178 
1179 			tsoh_iph->payload_len = htons(ip_length);
1180 		}
1181 
1182 		rc = efx_tso_put_header(tx_queue, buffer, header);
1183 		if (unlikely(rc))
1184 			return rc;
1185 	} else {
1186 		/* Send the original headers with a TSO option descriptor
1187 		 * in front
1188 		 */
1189 		u8 tcp_flags = ((u8 *)tcp_hdr(skb))[13] & ~tcp_flags_clear;
1190 
1191 		buffer->flags = EFX_TX_BUF_OPTION;
1192 		buffer->len = 0;
1193 		buffer->unmap_len = 0;
1194 		EFX_POPULATE_QWORD_5(buffer->option,
1195 				     ESF_DZ_TX_DESC_IS_OPT, 1,
1196 				     ESF_DZ_TX_OPTION_TYPE,
1197 				     ESE_DZ_TX_OPTION_DESC_TSO,
1198 				     ESF_DZ_TX_TSO_TCP_FLAGS, tcp_flags,
1199 				     ESF_DZ_TX_TSO_IP_ID, st->ipv4_id,
1200 				     ESF_DZ_TX_TSO_TCP_SEQNO, st->seqnum);
1201 		++tx_queue->insert_count;
1202 
1203 		/* We mapped the headers in tso_start().  Unmap them
1204 		 * when the last segment is completed.
1205 		 */
1206 		buffer = efx_tx_queue_get_insert_buffer(tx_queue);
1207 		buffer->dma_addr = st->header_dma_addr;
1208 		buffer->len = st->header_len;
1209 		if (is_last) {
1210 			buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_MAP_SINGLE;
1211 			buffer->unmap_len = st->header_unmap_len;
1212 			buffer->dma_offset = 0;
1213 			/* Ensure we only unmap them once in case of a
1214 			 * later DMA mapping error and rollback
1215 			 */
1216 			st->header_unmap_len = 0;
1217 		} else {
1218 			buffer->flags = EFX_TX_BUF_CONT;
1219 			buffer->unmap_len = 0;
1220 		}
1221 		++tx_queue->insert_count;
1222 	}
1223 
1224 	st->seqnum += skb_shinfo(skb)->gso_size;
1225 
1226 	/* Linux leaves suitable gaps in the IP ID space for us to fill. */
1227 	++st->ipv4_id;
1228 
1229 	++tx_queue->tso_packets;
1230 
1231 	++tx_queue->tx_packets;
1232 
1233 	return 0;
1234 }
1235 
1236 
1237 /**
1238  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1239  * @tx_queue:		Efx TX queue
1240  * @skb:		Socket buffer
1241  *
1242  * Context: You must hold netif_tx_lock() to call this function.
1243  *
1244  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1245  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1246  * %NETDEV_TX_OK.
1247  */
1248 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1249 			       struct sk_buff *skb)
1250 {
1251 	struct efx_nic *efx = tx_queue->efx;
1252 	unsigned int old_insert_count = tx_queue->insert_count;
1253 	int frag_i, rc;
1254 	struct tso_state state;
1255 
1256 	/* Find the packet protocol and sanity-check it */
1257 	state.protocol = efx_tso_check_protocol(skb);
1258 
1259 	rc = tso_start(&state, efx, skb);
1260 	if (rc)
1261 		goto mem_err;
1262 
1263 	if (likely(state.in_len == 0)) {
1264 		/* Grab the first payload fragment. */
1265 		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1266 		frag_i = 0;
1267 		rc = tso_get_fragment(&state, efx,
1268 				      skb_shinfo(skb)->frags + frag_i);
1269 		if (rc)
1270 			goto mem_err;
1271 	} else {
1272 		/* Payload starts in the header area. */
1273 		frag_i = -1;
1274 	}
1275 
1276 	if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1277 		goto mem_err;
1278 
1279 	while (1) {
1280 		tso_fill_packet_with_fragment(tx_queue, skb, &state);
1281 
1282 		/* Move onto the next fragment? */
1283 		if (state.in_len == 0) {
1284 			if (++frag_i >= skb_shinfo(skb)->nr_frags)
1285 				/* End of payload reached. */
1286 				break;
1287 			rc = tso_get_fragment(&state, efx,
1288 					      skb_shinfo(skb)->frags + frag_i);
1289 			if (rc)
1290 				goto mem_err;
1291 		}
1292 
1293 		/* Start at new packet? */
1294 		if (state.packet_space == 0 &&
1295 		    tso_start_new_packet(tx_queue, skb, &state) < 0)
1296 			goto mem_err;
1297 	}
1298 
1299 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
1300 
1301 	efx_tx_maybe_stop_queue(tx_queue);
1302 
1303 	/* Pass off to hardware */
1304 	if (!skb->xmit_more || netif_xmit_stopped(tx_queue->core_txq))
1305 		efx_nic_push_buffers(tx_queue);
1306 
1307 	tx_queue->tso_bursts++;
1308 	return NETDEV_TX_OK;
1309 
1310  mem_err:
1311 	netif_err(efx, tx_err, efx->net_dev,
1312 		  "Out of memory for TSO headers, or DMA mapping error\n");
1313 	dev_kfree_skb_any(skb);
1314 
1315 	/* Free the DMA mapping we were in the process of writing out */
1316 	if (state.unmap_len) {
1317 		if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE)
1318 			dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr,
1319 					 state.unmap_len, DMA_TO_DEVICE);
1320 		else
1321 			dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr,
1322 				       state.unmap_len, DMA_TO_DEVICE);
1323 	}
1324 
1325 	/* Free the header DMA mapping, if using option descriptors */
1326 	if (state.header_unmap_len)
1327 		dma_unmap_single(&efx->pci_dev->dev, state.header_dma_addr,
1328 				 state.header_unmap_len, DMA_TO_DEVICE);
1329 
1330 	efx_enqueue_unwind(tx_queue, old_insert_count);
1331 	return NETDEV_TX_OK;
1332 }
1333