Linux Audio

Check our new training course

Loading...
v6.8
  1// SPDX-License-Identifier: GPL-2.0-only
  2/****************************************************************************
  3 * Driver for Solarflare network controllers and boards
  4 * Copyright 2005-2006 Fen Systems Ltd.
  5 * Copyright 2005-2013 Solarflare Communications Inc.
 
 
 
 
  6 */
  7
  8#include <linux/pci.h>
  9#include <linux/tcp.h>
 10#include <linux/ip.h>
 11#include <linux/in.h>
 12#include <linux/ipv6.h>
 13#include <linux/slab.h>
 14#include <net/ipv6.h>
 15#include <linux/if_ether.h>
 16#include <linux/highmem.h>
 17#include <linux/cache.h>
 18#include "net_driver.h"
 19#include "efx.h"
 20#include "io.h"
 21#include "nic.h"
 22#include "tx.h"
 23#include "tx_common.h"
 24#include "workarounds.h"
 25#include "ef10_regs.h"
 26
 27#ifdef EFX_USE_PIO
 28
 29#define EFX_PIOBUF_SIZE_DEF ALIGN(256, L1_CACHE_BYTES)
 30unsigned int efx_piobuf_size __read_mostly = EFX_PIOBUF_SIZE_DEF;
 31
 32#endif /* EFX_USE_PIO */
 
 
 
 
 
 
 33
 34static inline u8 *efx_tx_get_copy_buffer(struct efx_tx_queue *tx_queue,
 35					 struct efx_tx_buffer *buffer)
 
 
 36{
 37	unsigned int index = efx_tx_queue_get_insert_index(tx_queue);
 38	struct efx_buffer *page_buf =
 39		&tx_queue->cb_page[index >> (PAGE_SHIFT - EFX_TX_CB_ORDER)];
 40	unsigned int offset =
 41		((index << EFX_TX_CB_ORDER) + NET_IP_ALIGN) & (PAGE_SIZE - 1);
 
 
 
 
 
 
 
 
 42
 43	if (unlikely(!page_buf->addr) &&
 44	    efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
 45				 GFP_ATOMIC))
 46		return NULL;
 47	buffer->dma_addr = page_buf->dma_addr + offset;
 48	buffer->unmap_len = 0;
 49	return (u8 *)page_buf->addr + offset;
 
 
 50}
 51
 52u8 *efx_tx_get_copy_buffer_limited(struct efx_tx_queue *tx_queue,
 53				   struct efx_tx_buffer *buffer, size_t len)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 54{
 55	if (len > EFX_TX_CB_SIZE)
 56		return NULL;
 57	return efx_tx_get_copy_buffer(tx_queue, buffer);
 
 
 
 
 
 
 58}
 59
 60static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
 
 
 61{
 62	/* We need to consider all queues that the net core sees as one */
 63	struct efx_nic *efx = txq1->efx;
 64	struct efx_tx_queue *txq2;
 65	unsigned int fill_level;
 
 
 
 66
 67	fill_level = efx_channel_tx_old_fill_level(txq1->channel);
 68	if (likely(fill_level < efx->txq_stop_thresh))
 69		return;
 70
 71	/* We used the stale old_read_count above, which gives us a
 72	 * pessimistic estimate of the fill level (which may even
 73	 * validly be >= efx->txq_entries).  Now try again using
 74	 * read_count (more likely to be a cache miss).
 75	 *
 76	 * If we read read_count and then conditionally stop the
 77	 * queue, it is possible for the completion path to race with
 78	 * us and complete all outstanding descriptors in the middle,
 79	 * after which there will be no more completions to wake it.
 80	 * Therefore we stop the queue first, then read read_count
 81	 * (with a memory barrier to ensure the ordering), then
 82	 * restart the queue if the fill level turns out to be low
 83	 * enough.
 84	 */
 85	netif_tx_stop_queue(txq1->core_txq);
 86	smp_mb();
 87	efx_for_each_channel_tx_queue(txq2, txq1->channel)
 88		txq2->old_read_count = READ_ONCE(txq2->read_count);
 89
 90	fill_level = efx_channel_tx_old_fill_level(txq1->channel);
 91	EFX_WARN_ON_ONCE_PARANOID(fill_level >= efx->txq_entries);
 92	if (likely(fill_level < efx->txq_stop_thresh)) {
 93		smp_mb();
 94		if (likely(!efx->loopback_selftest))
 95			netif_tx_start_queue(txq1->core_txq);
 96	}
 
 
 
 97}
 98
 99static int efx_enqueue_skb_copy(struct efx_tx_queue *tx_queue,
100				struct sk_buff *skb)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
101{
102	unsigned int copy_len = skb->len;
 
103	struct efx_tx_buffer *buffer;
104	u8 *copy_buffer;
105	int rc;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
106
107	EFX_WARN_ON_ONCE_PARANOID(copy_len > EFX_TX_CB_SIZE);
 
108
109	buffer = efx_tx_queue_get_insert_buffer(tx_queue);
 
 
 
 
 
110
111	copy_buffer = efx_tx_get_copy_buffer(tx_queue, buffer);
112	if (unlikely(!copy_buffer))
113		return -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
114
115	rc = skb_copy_bits(skb, 0, copy_buffer, copy_len);
116	EFX_WARN_ON_PARANOID(rc);
117	buffer->len = copy_len;
 
 
 
 
 
 
 
 
118
 
119	buffer->skb = skb;
120	buffer->flags = EFX_TX_BUF_SKB;
121
122	++tx_queue->insert_count;
123	return rc;
124}
125
126#ifdef EFX_USE_PIO
 
127
128struct efx_short_copy_buffer {
129	int used;
130	u8 buf[L1_CACHE_BYTES];
131};
132
133/* Copy to PIO, respecting that writes to PIO buffers must be dword aligned.
134 * Advances piobuf pointer. Leaves additional data in the copy buffer.
135 */
136static void efx_memcpy_toio_aligned(struct efx_nic *efx, u8 __iomem **piobuf,
137				    u8 *data, int len,
138				    struct efx_short_copy_buffer *copy_buf)
139{
140	int block_len = len & ~(sizeof(copy_buf->buf) - 1);
141
142	__iowrite64_copy(*piobuf, data, block_len >> 3);
143	*piobuf += block_len;
144	len -= block_len;
 
 
 
 
 
 
 
145
146	if (len) {
147		data += block_len;
148		BUG_ON(copy_buf->used);
149		BUG_ON(len > sizeof(copy_buf->buf));
150		memcpy(copy_buf->buf, data, len);
151		copy_buf->used = len;
 
 
152	}
 
 
153}
154
155/* Copy to PIO, respecting dword alignment, popping data from copy buffer first.
156 * Advances piobuf pointer. Leaves additional data in the copy buffer.
 
 
157 */
158static void efx_memcpy_toio_aligned_cb(struct efx_nic *efx, u8 __iomem **piobuf,
159				       u8 *data, int len,
160				       struct efx_short_copy_buffer *copy_buf)
 
161{
162	if (copy_buf->used) {
163		/* if the copy buffer is partially full, fill it up and write */
164		int copy_to_buf =
165			min_t(int, sizeof(copy_buf->buf) - copy_buf->used, len);
166
167		memcpy(copy_buf->buf + copy_buf->used, data, copy_to_buf);
168		copy_buf->used += copy_to_buf;
169
170		/* if we didn't fill it up then we're done for now */
171		if (copy_buf->used < sizeof(copy_buf->buf))
 
 
 
 
 
172			return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
173
174		__iowrite64_copy(*piobuf, copy_buf->buf,
175				 sizeof(copy_buf->buf) >> 3);
176		*piobuf += sizeof(copy_buf->buf);
177		data += copy_to_buf;
178		len -= copy_to_buf;
179		copy_buf->used = 0;
180	}
 
181
182	efx_memcpy_toio_aligned(efx, piobuf, data, len, copy_buf);
183}
184
185static void efx_flush_copy_buffer(struct efx_nic *efx, u8 __iomem *piobuf,
186				  struct efx_short_copy_buffer *copy_buf)
187{
188	/* if there's anything in it, write the whole buffer, including junk */
189	if (copy_buf->used)
190		__iowrite64_copy(piobuf, copy_buf->buf,
191				 sizeof(copy_buf->buf) >> 3);
 
 
 
 
192}
193
194/* Traverse skb structure and copy fragments in to PIO buffer.
195 * Advances piobuf pointer.
196 */
197static void efx_skb_copy_bits_to_pio(struct efx_nic *efx, struct sk_buff *skb,
198				     u8 __iomem **piobuf,
199				     struct efx_short_copy_buffer *copy_buf)
200{
201	int i;
 
 
 
 
202
203	efx_memcpy_toio_aligned(efx, piobuf, skb->data, skb_headlen(skb),
204				copy_buf);
205
206	for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) {
207		skb_frag_t *f = &skb_shinfo(skb)->frags[i];
208		u8 *vaddr;
209
210		vaddr = kmap_local_page(skb_frag_page(f));
 
 
 
211
212		efx_memcpy_toio_aligned_cb(efx, piobuf, vaddr + skb_frag_off(f),
213					   skb_frag_size(f), copy_buf);
214		kunmap_local(vaddr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
215	}
216
217	EFX_WARN_ON_ONCE_PARANOID(skb_shinfo(skb)->frag_list);
 
 
 
 
 
 
 
 
 
 
 
 
 
218}
219
220static int efx_enqueue_skb_pio(struct efx_tx_queue *tx_queue,
221			       struct sk_buff *skb)
222{
223	struct efx_tx_buffer *buffer =
224		efx_tx_queue_get_insert_buffer(tx_queue);
225	u8 __iomem *piobuf = tx_queue->piobuf;
226
227	/* Copy to PIO buffer. Ensure the writes are padded to the end
228	 * of a cache line, as this is required for write-combining to be
229	 * effective on at least x86.
230	 */
231
232	if (skb_shinfo(skb)->nr_frags) {
233		/* The size of the copy buffer will ensure all writes
234		 * are the size of a cache line.
235		 */
236		struct efx_short_copy_buffer copy_buf;
237
238		copy_buf.used = 0;
239
240		efx_skb_copy_bits_to_pio(tx_queue->efx, skb,
241					 &piobuf, &copy_buf);
242		efx_flush_copy_buffer(tx_queue->efx, piobuf, &copy_buf);
243	} else {
244		/* Pad the write to the size of a cache line.
245		 * We can do this because we know the skb_shared_info struct is
246		 * after the source, and the destination buffer is big enough.
247		 */
248		BUILD_BUG_ON(L1_CACHE_BYTES >
249			     SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
250		__iowrite64_copy(tx_queue->piobuf, skb->data,
251				 ALIGN(skb->len, L1_CACHE_BYTES) >> 3);
252	}
253
254	buffer->skb = skb;
255	buffer->flags = EFX_TX_BUF_SKB | EFX_TX_BUF_OPTION;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
256
257	EFX_POPULATE_QWORD_5(buffer->option,
258			     ESF_DZ_TX_DESC_IS_OPT, 1,
259			     ESF_DZ_TX_OPTION_TYPE, ESE_DZ_TX_OPTION_DESC_PIO,
260			     ESF_DZ_TX_PIO_CONT, 0,
261			     ESF_DZ_TX_PIO_BYTE_CNT, skb->len,
262			     ESF_DZ_TX_PIO_BUF_ADDR,
263			     tx_queue->piobuf_offset);
264	++tx_queue->insert_count;
265	return 0;
 
 
 
 
 
266}
267
268/* Decide whether we can use TX PIO, ie. write packet data directly into
269 * a buffer on the device.  This can reduce latency at the expense of
270 * throughput, so we only do this if both hardware and software TX rings
271 * are empty, including all queues for the channel.  This also ensures that
272 * only one packet at a time can be using the PIO buffer. If the xmit_more
273 * flag is set then we don't use this - there'll be another packet along
274 * shortly and we want to hold off the doorbell.
275 */
276static bool efx_tx_may_pio(struct efx_tx_queue *tx_queue)
277{
278	struct efx_channel *channel = tx_queue->channel;
279
280	if (!tx_queue->piobuf)
281		return false;
282
283	EFX_WARN_ON_ONCE_PARANOID(!channel->efx->type->option_descriptors);
 
 
 
 
 
284
285	efx_for_each_channel_tx_queue(tx_queue, channel)
286		if (!efx_nic_tx_is_empty(tx_queue, tx_queue->packet_write_count))
287			return false;
288
289	return true;
290}
291#endif /* EFX_USE_PIO */
292
293/* Send any pending traffic for a channel. xmit_more is shared across all
294 * queues for a channel, so we must check all of them.
295 */
296static void efx_tx_send_pending(struct efx_channel *channel)
297{
298	struct efx_tx_queue *q;
 
 
 
 
 
 
 
 
 
 
 
299
300	efx_for_each_channel_tx_queue(q, channel) {
301		if (q->xmit_pending)
302			efx_nic_push_buffers(q);
303	}
 
304}
305
306/*
307 * Add a socket buffer to a TX queue
308 *
309 * This maps all fragments of a socket buffer for DMA and adds them to
310 * the TX queue.  The queue's insert pointer will be incremented by
311 * the number of fragments in the socket buffer.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
312 *
313 * If any DMA mapping fails, any mapped fragments will be unmapped,
314 * the queue's insert pointer will be restored to its original value.
315 *
316 * This function is split out from efx_hard_start_xmit to allow the
317 * loopback test to direct packets via specific TX queues.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
318 *
319 * Returns NETDEV_TX_OK.
320 * You must hold netif_tx_lock() to call this function.
321 */
322netdev_tx_t __efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
323{
324	unsigned int old_insert_count = tx_queue->insert_count;
325	bool xmit_more = netdev_xmit_more();
326	bool data_mapped = false;
327	unsigned int segments;
328	unsigned int skb_len;
329	int rc;
330
331	skb_len = skb->len;
332	segments = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 0;
333	if (segments == 1)
334		segments = 0; /* Don't use TSO for a single segment. */
335
336	/* Handle TSO first - it's *possible* (although unlikely) that we might
337	 * be passed a packet to segment that's smaller than the copybreak/PIO
338	 * size limit.
339	 */
340	if (segments) {
341		switch (tx_queue->tso_version) {
342		case 1:
343			rc = efx_enqueue_skb_tso(tx_queue, skb, &data_mapped);
344			break;
345		case 2:
346			rc = efx_ef10_tx_tso_desc(tx_queue, skb, &data_mapped);
347			break;
348		case 0: /* No TSO on this queue, SW fallback needed */
349		default:
350			rc = -EINVAL;
351			break;
352		}
353		if (rc == -EINVAL) {
354			rc = efx_tx_tso_fallback(tx_queue, skb);
355			tx_queue->tso_fallbacks++;
356			if (rc == 0)
357				return 0;
358		}
359		if (rc)
360			goto err;
361#ifdef EFX_USE_PIO
362	} else if (skb_len <= efx_piobuf_size && !xmit_more &&
363		   efx_tx_may_pio(tx_queue)) {
364		/* Use PIO for short packets with an empty queue. */
365		if (efx_enqueue_skb_pio(tx_queue, skb))
366			goto err;
367		tx_queue->pio_packets++;
368		data_mapped = true;
369#endif
370	} else if (skb->data_len && skb_len <= EFX_TX_CB_SIZE) {
371		/* Pad short packets or coalesce short fragmented packets. */
372		if (efx_enqueue_skb_copy(tx_queue, skb))
373			goto err;
374		tx_queue->cb_packets++;
375		data_mapped = true;
376	}
377
378	/* Map for DMA and create descriptors if we haven't done so already. */
379	if (!data_mapped && (efx_tx_map_data(tx_queue, skb, segments)))
380		goto err;
 
 
 
 
 
 
381
382	efx_tx_maybe_stop_queue(tx_queue);
 
383
384	tx_queue->xmit_pending = true;
385
386	/* Pass off to hardware */
387	if (__netdev_tx_sent_queue(tx_queue->core_txq, skb_len, xmit_more))
388		efx_tx_send_pending(tx_queue->channel);
 
 
 
389
390	if (segments) {
391		tx_queue->tso_bursts++;
392		tx_queue->tso_packets += segments;
393		tx_queue->tx_packets  += segments;
394	} else {
395		tx_queue->tx_packets++;
 
 
 
 
396	}
397
398	return NETDEV_TX_OK;
 
399
 
 
 
 
 
 
400
401err:
402	efx_enqueue_unwind(tx_queue, old_insert_count);
403	dev_kfree_skb_any(skb);
404
405	/* If we're not expecting another transmit and we had something to push
406	 * on this queue or a partner queue then we need to push here to get the
407	 * previous packets out.
408	 */
409	if (!xmit_more)
410		efx_tx_send_pending(tx_queue->channel);
411
412	return NETDEV_TX_OK;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
413}
414
415/* Transmit a packet from an XDP buffer
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
416 *
417 * Returns number of packets sent on success, error code otherwise.
418 * Runs in NAPI context, either in our poll (for XDP TX) or a different NIC
419 * (for XDP redirect).
420 */
421int efx_xdp_tx_buffers(struct efx_nic *efx, int n, struct xdp_frame **xdpfs,
422		       bool flush)
 
423{
424	struct efx_tx_buffer *tx_buffer;
425	struct efx_tx_queue *tx_queue;
426	struct xdp_frame *xdpf;
427	dma_addr_t dma_addr;
428	unsigned int len;
429	int space;
430	int cpu;
431	int i = 0;
432
433	if (unlikely(n && !xdpfs))
434		return -EINVAL;
435	if (unlikely(!n))
436		return 0;
437
438	cpu = raw_smp_processor_id();
439	if (unlikely(cpu >= efx->xdp_tx_queue_count))
440		return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
441
442	tx_queue = efx->xdp_tx_queues[cpu];
443	if (unlikely(!tx_queue))
444		return -EINVAL;
 
 
 
 
 
 
 
 
 
 
 
445
446	if (!tx_queue->initialised)
447		return -EINVAL;
448
449	if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
450		HARD_TX_LOCK(efx->net_dev, tx_queue->core_txq, cpu);
451
452	/* If we're borrowing net stack queues we have to handle stop-restart
453	 * or we might block the queue and it will be considered as frozen
454	 */
455	if (efx->xdp_txq_queues_mode == EFX_XDP_TX_QUEUES_BORROWED) {
456		if (netif_tx_queue_stopped(tx_queue->core_txq))
457			goto unlock;
458		efx_tx_maybe_stop_queue(tx_queue);
459	}
460
461	/* Check for available space. We should never need multiple
462	 * descriptors per frame.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
463	 */
464	space = efx->txq_entries +
465		tx_queue->read_count - tx_queue->insert_count;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
466
467	for (i = 0; i < n; i++) {
468		xdpf = xdpfs[i];
 
 
 
 
 
 
 
 
 
 
 
 
469
470		if (i >= space)
471			break;
 
 
 
472
473		/* We'll want a descriptor for this tx. */
474		prefetchw(__efx_tx_queue_get_insert_buffer(tx_queue));
 
 
 
 
 
 
 
 
 
475
476		len = xdpf->len;
477
478		/* Map for DMA. */
479		dma_addr = dma_map_single(&efx->pci_dev->dev,
480					  xdpf->data, len,
481					  DMA_TO_DEVICE);
482		if (dma_mapping_error(&efx->pci_dev->dev, dma_addr))
483			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
484
485		/*  Create descriptor and set up for unmapping DMA. */
486		tx_buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
487		tx_buffer->xdpf = xdpf;
488		tx_buffer->flags = EFX_TX_BUF_XDP |
489				   EFX_TX_BUF_MAP_SINGLE;
490		tx_buffer->dma_offset = 0;
491		tx_buffer->unmap_len = len;
492		tx_queue->tx_packets++;
493	}
494
495	/* Pass mapped frames to hardware. */
496	if (flush && i > 0)
497		efx_nic_push_buffers(tx_queue);
498
499unlock:
500	if (efx->xdp_txq_queues_mode != EFX_XDP_TX_QUEUES_DEDICATED)
501		HARD_TX_UNLOCK(efx->net_dev, tx_queue->core_txq);
502
503	return i == 0 ? -EIO : i;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
504}
505
506/* Initiate a packet transmission.  We use one channel per CPU
507 * (sharing when we have more CPUs than channels).
 
 
 
 
508 *
509 * Context: non-blocking.
510 * Should always return NETDEV_TX_OK and consume the skb.
511 */
512netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
513				struct net_device *net_dev)
 
514{
515	struct efx_nic *efx = efx_netdev_priv(net_dev);
516	struct efx_tx_queue *tx_queue;
517	unsigned index, type;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
518
519	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
520	index = skb_get_queue_mapping(skb);
521	type = efx_tx_csum_type_skb(skb);
 
 
522
523	/* PTP "event" packet */
524	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
525	    ((efx_ptp_use_mac_tx_timestamps(efx) && efx->ptp_data) ||
526	    unlikely(efx_ptp_is_ptp_tx(efx, skb)))) {
527		/* There may be existing transmits on the channel that are
528		 * waiting for this packet to trigger the doorbell write.
529		 * We need to send the packets at this point.
530		 */
531		efx_tx_send_pending(efx_get_tx_channel(efx, index));
532		return efx_ptp_tx(efx, skb);
 
 
533	}
534
535	tx_queue = efx_get_tx_queue(efx, index, type);
536	if (WARN_ON_ONCE(!tx_queue)) {
537		/* We don't have a TXQ of the right type.
538		 * This should never happen, as we don't advertise offload
539		 * features unless we can support them.
540		 */
541		dev_kfree_skb_any(skb);
542		/* If we're not expecting another transmit and we had something to push
543		 * on this queue or a partner queue then we need to push here to get the
544		 * previous packets out.
545		 */
546		if (!netdev_xmit_more())
547			efx_tx_send_pending(efx_get_tx_channel(efx, index));
548		return NETDEV_TX_OK;
549	}
550
551	return __efx_enqueue_skb(tx_queue, skb);
 
 
 
 
 
 
552}
553
554void efx_xmit_done_single(struct efx_tx_queue *tx_queue)
 
 
 
 
 
 
 
 
 
 
 
 
 
555{
556	unsigned int pkts_compl = 0, bytes_compl = 0;
557	unsigned int efv_pkts_compl = 0;
558	unsigned int read_ptr;
559	bool finished = false;
560
561	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 
562
563	while (!finished) {
564		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
565
566		if (!efx_tx_buffer_in_use(buffer)) {
567			struct efx_nic *efx = tx_queue->efx;
568
569			netif_err(efx, hw, efx->net_dev,
570				  "TX queue %d spurious single TX completion\n",
571				  tx_queue->queue);
572			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
573			return;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
574		}
575
576		/* Need to check the flag before dequeueing. */
577		if (buffer->flags & EFX_TX_BUF_SKB)
578			finished = true;
579		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
580				   &efv_pkts_compl);
 
 
 
 
 
581
582		++tx_queue->read_count;
583		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 
 
584	}
585
586	tx_queue->pkts_compl += pkts_compl;
587	tx_queue->bytes_compl += bytes_compl;
 
 
588
589	EFX_WARN_ON_PARANOID(pkts_compl + efv_pkts_compl != 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
590
591	efx_xmit_done_check_empty(tx_queue);
 
592}
593
594void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
 
 
 
 
 
 
595{
596	struct efx_nic *efx = tx_queue->efx;
597
598	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
599	tx_queue->core_txq =
600		netdev_get_tx_queue(efx->net_dev,
601				    tx_queue->channel->channel);
 
 
 
 
602}
v3.5.6
 
   1/****************************************************************************
   2 * Driver for Solarflare Solarstorm network controllers and boards
   3 * Copyright 2005-2006 Fen Systems Ltd.
   4 * Copyright 2005-2010 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 "net_driver.h"
  21#include "efx.h"
 
  22#include "nic.h"
 
 
  23#include "workarounds.h"
 
 
 
 
 
 
  24
  25/*
  26 * TX descriptor ring full threshold
  27 *
  28 * The tx_queue descriptor ring fill-level must fall below this value
  29 * before we restart the netif queue
  30 */
  31#define EFX_TXQ_THRESHOLD(_efx) ((_efx)->txq_entries / 2u)
  32
  33static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
  34			       struct efx_tx_buffer *buffer,
  35			       unsigned int *pkts_compl,
  36			       unsigned int *bytes_compl)
  37{
  38	if (buffer->unmap_len) {
  39		struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
  40		dma_addr_t unmap_addr = (buffer->dma_addr + buffer->len -
  41					 buffer->unmap_len);
  42		if (buffer->unmap_single)
  43			pci_unmap_single(pci_dev, unmap_addr, buffer->unmap_len,
  44					 PCI_DMA_TODEVICE);
  45		else
  46			pci_unmap_page(pci_dev, unmap_addr, buffer->unmap_len,
  47				       PCI_DMA_TODEVICE);
  48		buffer->unmap_len = 0;
  49		buffer->unmap_single = false;
  50	}
  51
  52	if (buffer->skb) {
  53		(*pkts_compl)++;
  54		(*bytes_compl) += buffer->skb->len;
  55		dev_kfree_skb_any((struct sk_buff *) buffer->skb);
  56		buffer->skb = NULL;
  57		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
  58			   "TX queue %d transmission id %x complete\n",
  59			   tx_queue->queue, tx_queue->read_count);
  60	}
  61}
  62
  63/**
  64 * struct efx_tso_header - a DMA mapped buffer for packet headers
  65 * @next: Linked list of free ones.
  66 *	The list is protected by the TX queue lock.
  67 * @dma_unmap_len: Length to unmap for an oversize buffer, or 0.
  68 * @dma_addr: The DMA address of the header below.
  69 *
  70 * This controls the memory used for a TSO header.  Use TSOH_DATA()
  71 * to find the packet header data.  Use TSOH_SIZE() to calculate the
  72 * total size required for a given packet header length.  TSO headers
  73 * in the free list are exactly %TSOH_STD_SIZE bytes in size.
  74 */
  75struct efx_tso_header {
  76	union {
  77		struct efx_tso_header *next;
  78		size_t unmap_len;
  79	};
  80	dma_addr_t dma_addr;
  81};
  82
  83static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
  84			       struct sk_buff *skb);
  85static void efx_fini_tso(struct efx_tx_queue *tx_queue);
  86static void efx_tsoh_heap_free(struct efx_tx_queue *tx_queue,
  87			       struct efx_tso_header *tsoh);
  88
  89static void efx_tsoh_free(struct efx_tx_queue *tx_queue,
  90			  struct efx_tx_buffer *buffer)
  91{
  92	if (buffer->tsoh) {
  93		if (likely(!buffer->tsoh->unmap_len)) {
  94			buffer->tsoh->next = tx_queue->tso_headers_free;
  95			tx_queue->tso_headers_free = buffer->tsoh;
  96		} else {
  97			efx_tsoh_heap_free(tx_queue, buffer->tsoh);
  98		}
  99		buffer->tsoh = NULL;
 100	}
 101}
 102
 103
 104static inline unsigned
 105efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr)
 106{
 107	/* Depending on the NIC revision, we can use descriptor
 108	 * lengths up to 8K or 8K-1.  However, since PCI Express
 109	 * devices must split read requests at 4K boundaries, there is
 110	 * little benefit from using descriptors that cross those
 111	 * boundaries and we keep things simple by not doing so.
 112	 */
 113	unsigned len = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1;
 114
 115	/* Work around hardware bug for unaligned buffers. */
 116	if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf))
 117		len = min_t(unsigned, len, 512 - (dma_addr & 0xf));
 118
 119	return len;
 120}
 121
 122unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
 123{
 124	/* Header and payload descriptor for each output segment, plus
 125	 * one for every input fragment boundary within a segment
 
 
 
 
 
 
 126	 */
 127	unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
 
 
 
 128
 129	/* Possibly one more per segment for the alignment workaround */
 130	if (EFX_WORKAROUND_5391(efx))
 131		max_descs += EFX_TSO_MAX_SEGS;
 132
 133	/* Possibly more for PCIe page boundaries within input fragments */
 134	if (PAGE_SIZE > EFX_PAGE_SIZE)
 135		max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
 136				   DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE));
 137
 138	return max_descs;
 139}
 140
 141/*
 142 * Add a socket buffer to a TX queue
 143 *
 144 * This maps all fragments of a socket buffer for DMA and adds them to
 145 * the TX queue.  The queue's insert pointer will be incremented by
 146 * the number of fragments in the socket buffer.
 147 *
 148 * If any DMA mapping fails, any mapped fragments will be unmapped,
 149 * the queue's insert pointer will be restored to its original value.
 150 *
 151 * This function is split out from efx_hard_start_xmit to allow the
 152 * loopback test to direct packets via specific TX queues.
 153 *
 154 * Returns NETDEV_TX_OK or NETDEV_TX_BUSY
 155 * You must hold netif_tx_lock() to call this function.
 156 */
 157netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 158{
 159	struct efx_nic *efx = tx_queue->efx;
 160	struct pci_dev *pci_dev = efx->pci_dev;
 161	struct efx_tx_buffer *buffer;
 162	skb_frag_t *fragment;
 163	unsigned int len, unmap_len = 0, fill_level, insert_ptr;
 164	dma_addr_t dma_addr, unmap_addr = 0;
 165	unsigned int dma_len;
 166	bool unmap_single;
 167	int q_space, i = 0;
 168	netdev_tx_t rc = NETDEV_TX_OK;
 169
 170	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
 171
 172	if (skb_shinfo(skb)->gso_size)
 173		return efx_enqueue_skb_tso(tx_queue, skb);
 174
 175	/* Get size of the initial fragment */
 176	len = skb_headlen(skb);
 177
 178	/* Pad if necessary */
 179	if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
 180		EFX_BUG_ON_PARANOID(skb->data_len);
 181		len = 32 + 1;
 182		if (skb_pad(skb, len - skb->len))
 183			return NETDEV_TX_OK;
 184	}
 185
 186	fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 187	q_space = efx->txq_entries - 1 - fill_level;
 188
 189	/* Map for DMA.  Use pci_map_single rather than pci_map_page
 190	 * since this is more efficient on machines with sparse
 191	 * memory.
 192	 */
 193	unmap_single = true;
 194	dma_addr = pci_map_single(pci_dev, skb->data, len, PCI_DMA_TODEVICE);
 195
 196	/* Process all fragments */
 197	while (1) {
 198		if (unlikely(pci_dma_mapping_error(pci_dev, dma_addr)))
 199			goto pci_err;
 200
 201		/* Store fields for marking in the per-fragment final
 202		 * descriptor */
 203		unmap_len = len;
 204		unmap_addr = dma_addr;
 205
 206		/* Add to TX queue, splitting across DMA boundaries */
 207		do {
 208			if (unlikely(q_space-- <= 0)) {
 209				/* It might be that completions have
 210				 * happened since the xmit path last
 211				 * checked.  Update the xmit path's
 212				 * copy of read_count.
 213				 */
 214				netif_tx_stop_queue(tx_queue->core_txq);
 215				/* This memory barrier protects the
 216				 * change of queue state from the access
 217				 * of read_count. */
 218				smp_mb();
 219				tx_queue->old_read_count =
 220					ACCESS_ONCE(tx_queue->read_count);
 221				fill_level = (tx_queue->insert_count
 222					      - tx_queue->old_read_count);
 223				q_space = efx->txq_entries - 1 - fill_level;
 224				if (unlikely(q_space-- <= 0)) {
 225					rc = NETDEV_TX_BUSY;
 226					goto unwind;
 227				}
 228				smp_mb();
 229				if (likely(!efx->loopback_selftest))
 230					netif_tx_start_queue(
 231						tx_queue->core_txq);
 232			}
 233
 234			insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
 235			buffer = &tx_queue->buffer[insert_ptr];
 236			efx_tsoh_free(tx_queue, buffer);
 237			EFX_BUG_ON_PARANOID(buffer->tsoh);
 238			EFX_BUG_ON_PARANOID(buffer->skb);
 239			EFX_BUG_ON_PARANOID(buffer->len);
 240			EFX_BUG_ON_PARANOID(!buffer->continuation);
 241			EFX_BUG_ON_PARANOID(buffer->unmap_len);
 242
 243			dma_len = efx_max_tx_len(efx, dma_addr);
 244			if (likely(dma_len >= len))
 245				dma_len = len;
 246
 247			/* Fill out per descriptor fields */
 248			buffer->len = dma_len;
 249			buffer->dma_addr = dma_addr;
 250			len -= dma_len;
 251			dma_addr += dma_len;
 252			++tx_queue->insert_count;
 253		} while (len);
 254
 255		/* Transfer ownership of the unmapping to the final buffer */
 256		buffer->unmap_single = unmap_single;
 257		buffer->unmap_len = unmap_len;
 258		unmap_len = 0;
 259
 260		/* Get address and size of next fragment */
 261		if (i >= skb_shinfo(skb)->nr_frags)
 262			break;
 263		fragment = &skb_shinfo(skb)->frags[i];
 264		len = skb_frag_size(fragment);
 265		i++;
 266		/* Map for DMA */
 267		unmap_single = false;
 268		dma_addr = skb_frag_dma_map(&pci_dev->dev, fragment, 0, len,
 269					    DMA_TO_DEVICE);
 270	}
 271
 272	/* Transfer ownership of the skb to the final buffer */
 273	buffer->skb = skb;
 274	buffer->continuation = false;
 275
 276	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
 
 
 277
 278	/* Pass off to hardware */
 279	efx_nic_push_buffers(tx_queue);
 280
 281	return NETDEV_TX_OK;
 
 
 
 282
 283 pci_err:
 284	netif_err(efx, tx_err, efx->net_dev,
 285		  " TX queue %d could not map skb with %d bytes %d "
 286		  "fragments for DMA\n", tx_queue->queue, skb->len,
 287		  skb_shinfo(skb)->nr_frags + 1);
 288
 289	/* Mark the packet as transmitted, and free the SKB ourselves */
 290	dev_kfree_skb_any(skb);
 291
 292 unwind:
 293	/* Work backwards until we hit the original insert pointer value */
 294	while (tx_queue->insert_count != tx_queue->write_count) {
 295		unsigned int pkts_compl = 0, bytes_compl = 0;
 296		--tx_queue->insert_count;
 297		insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
 298		buffer = &tx_queue->buffer[insert_ptr];
 299		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
 300		buffer->len = 0;
 301	}
 302
 303	/* Free the fragment we were mid-way through pushing */
 304	if (unmap_len) {
 305		if (unmap_single)
 306			pci_unmap_single(pci_dev, unmap_addr, unmap_len,
 307					 PCI_DMA_TODEVICE);
 308		else
 309			pci_unmap_page(pci_dev, unmap_addr, unmap_len,
 310				       PCI_DMA_TODEVICE);
 311	}
 312
 313	return rc;
 314}
 315
 316/* Remove packets from the TX queue
 317 *
 318 * This removes packets from the TX queue, up to and including the
 319 * specified index.
 320 */
 321static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 322				unsigned int index,
 323				unsigned int *pkts_compl,
 324				unsigned int *bytes_compl)
 325{
 326	struct efx_nic *efx = tx_queue->efx;
 327	unsigned int stop_index, read_ptr;
 
 
 328
 329	stop_index = (index + 1) & tx_queue->ptr_mask;
 330	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 331
 332	while (read_ptr != stop_index) {
 333		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
 334		if (unlikely(buffer->len == 0)) {
 335			netif_err(efx, tx_err, efx->net_dev,
 336				  "TX queue %d spurious TX completion id %x\n",
 337				  tx_queue->queue, read_ptr);
 338			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
 339			return;
 340		}
 341
 342		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
 343		buffer->continuation = true;
 344		buffer->len = 0;
 345
 346		++tx_queue->read_count;
 347		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 348	}
 349}
 350
 351/* Initiate a packet transmission.  We use one channel per CPU
 352 * (sharing when we have more CPUs than channels).  On Falcon, the TX
 353 * completion events will be directed back to the CPU that transmitted
 354 * the packet, which should be cache-efficient.
 355 *
 356 * Context: non-blocking.
 357 * Note that returning anything other than NETDEV_TX_OK will cause the
 358 * OS to free the skb.
 359 */
 360netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
 361				struct net_device *net_dev)
 362{
 363	struct efx_nic *efx = netdev_priv(net_dev);
 364	struct efx_tx_queue *tx_queue;
 365	unsigned index, type;
 366
 367	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
 368
 369	index = skb_get_queue_mapping(skb);
 370	type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
 371	if (index >= efx->n_tx_channels) {
 372		index -= efx->n_tx_channels;
 373		type |= EFX_TXQ_TYPE_HIGHPRI;
 
 374	}
 375	tx_queue = efx_get_tx_queue(efx, index, type);
 376
 377	return efx_enqueue_skb(tx_queue, skb);
 378}
 379
 380void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
 
 381{
 382	struct efx_nic *efx = tx_queue->efx;
 383
 384	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
 385	tx_queue->core_txq =
 386		netdev_get_tx_queue(efx->net_dev,
 387				    tx_queue->queue / EFX_TXQ_TYPES +
 388				    ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
 389				     efx->n_tx_channels : 0));
 390}
 391
 392int efx_setup_tc(struct net_device *net_dev, u8 num_tc)
 
 
 
 
 
 393{
 394	struct efx_nic *efx = netdev_priv(net_dev);
 395	struct efx_channel *channel;
 396	struct efx_tx_queue *tx_queue;
 397	unsigned tc;
 398	int rc;
 399
 400	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC)
 401		return -EINVAL;
 402
 403	if (num_tc == net_dev->num_tc)
 404		return 0;
 
 405
 406	for (tc = 0; tc < num_tc; tc++) {
 407		net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
 408		net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
 409	}
 410
 411	if (num_tc > net_dev->num_tc) {
 412		/* Initialise high-priority queues as necessary */
 413		efx_for_each_channel(channel, efx) {
 414			efx_for_each_possible_channel_tx_queue(tx_queue,
 415							       channel) {
 416				if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI))
 417					continue;
 418				if (!tx_queue->buffer) {
 419					rc = efx_probe_tx_queue(tx_queue);
 420					if (rc)
 421						return rc;
 422				}
 423				if (!tx_queue->initialised)
 424					efx_init_tx_queue(tx_queue);
 425				efx_init_tx_queue_core_txq(tx_queue);
 426			}
 427		}
 428	} else {
 429		/* Reduce number of classes before number of queues */
 430		net_dev->num_tc = num_tc;
 431	}
 432
 433	rc = netif_set_real_num_tx_queues(net_dev,
 434					  max_t(int, num_tc, 1) *
 435					  efx->n_tx_channels);
 436	if (rc)
 437		return rc;
 438
 439	/* Do not destroy high-priority queues when they become
 440	 * unused.  We would have to flush them first, and it is
 441	 * fairly difficult to flush a subset of TX queues.  Leave
 442	 * it to efx_fini_channels().
 443	 */
 444
 445	net_dev->num_tc = num_tc;
 446	return 0;
 447}
 448
 449void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 
 450{
 451	unsigned fill_level;
 452	struct efx_nic *efx = tx_queue->efx;
 453	unsigned int pkts_compl = 0, bytes_compl = 0;
 
 
 
 
 
 454
 455	EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
 456
 457	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
 458	netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl);
 459
 460	/* See if we need to restart the netif queue.  This barrier
 461	 * separates the update of read_count from the test of the
 462	 * queue state. */
 463	smp_mb();
 464	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
 465	    likely(efx->port_enabled) &&
 466	    likely(netif_device_present(efx->net_dev))) {
 467		fill_level = tx_queue->insert_count - tx_queue->read_count;
 468		if (fill_level < EFX_TXQ_THRESHOLD(efx))
 469			netif_tx_wake_queue(tx_queue->core_txq);
 
 
 
 
 
 470	}
 471
 472	/* Check whether the hardware queue is now empty */
 473	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
 474		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
 475		if (tx_queue->read_count == tx_queue->old_write_count) {
 476			smp_mb();
 477			tx_queue->empty_read_count =
 478				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
 479		}
 480	}
 481}
 482
 483int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
 484{
 485	struct efx_nic *efx = tx_queue->efx;
 486	unsigned int entries;
 487	int i, rc;
 488
 489	/* Create the smallest power-of-two aligned ring */
 490	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
 491	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
 492	tx_queue->ptr_mask = entries - 1;
 493
 494	netif_dbg(efx, probe, efx->net_dev,
 495		  "creating TX queue %d size %#x mask %#x\n",
 496		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
 497
 498	/* Allocate software ring */
 499	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
 500				   GFP_KERNEL);
 501	if (!tx_queue->buffer)
 502		return -ENOMEM;
 503	for (i = 0; i <= tx_queue->ptr_mask; ++i)
 504		tx_queue->buffer[i].continuation = true;
 505
 506	/* Allocate hardware ring */
 507	rc = efx_nic_probe_tx(tx_queue);
 508	if (rc)
 509		goto fail;
 510
 
 
 
 
 
 
 
 
 511	return 0;
 512
 513 fail:
 514	kfree(tx_queue->buffer);
 515	tx_queue->buffer = NULL;
 516	return rc;
 517}
 518
 519void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
 
 
 
 
 
 
 
 
 520{
 521	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 522		  "initialising TX queue %d\n", tx_queue->queue);
 
 
 523
 524	tx_queue->insert_count = 0;
 525	tx_queue->write_count = 0;
 526	tx_queue->old_write_count = 0;
 527	tx_queue->read_count = 0;
 528	tx_queue->old_read_count = 0;
 529	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
 530
 531	/* Set up TX descriptor ring */
 532	efx_nic_init_tx(tx_queue);
 
 533
 534	tx_queue->initialised = true;
 535}
 
 536
 537void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
 
 
 
 538{
 539	struct efx_tx_buffer *buffer;
 540
 541	if (!tx_queue->buffer)
 542		return;
 543
 544	/* Free any buffers left in the ring */
 545	while (tx_queue->read_count != tx_queue->write_count) {
 546		unsigned int pkts_compl = 0, bytes_compl = 0;
 547		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
 548		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
 549		buffer->continuation = true;
 550		buffer->len = 0;
 551
 552		++tx_queue->read_count;
 
 
 553	}
 554	netdev_tx_reset_queue(tx_queue->core_txq);
 555}
 556
 557void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 558{
 559	if (!tx_queue->initialised)
 560		return;
 561
 562	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 563		  "shutting down TX queue %d\n", tx_queue->queue);
 564
 565	tx_queue->initialised = false;
 566
 567	/* Flush TX queue, remove descriptor ring */
 568	efx_nic_fini_tx(tx_queue);
 569
 570	efx_release_tx_buffers(tx_queue);
 571
 572	/* Free up TSO header cache */
 573	efx_fini_tso(tx_queue);
 574}
 575
 576void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 577{
 578	if (!tx_queue->buffer)
 579		return;
 580
 581	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 582		  "destroying TX queue %d\n", tx_queue->queue);
 583	efx_nic_remove_tx(tx_queue);
 584
 585	kfree(tx_queue->buffer);
 586	tx_queue->buffer = NULL;
 587}
 588
 589
 590/* Efx TCP segmentation acceleration.
 591 *
 592 * Why?  Because by doing it here in the driver we can go significantly
 593 * faster than the GSO.
 594 *
 595 * Requires TX checksum offload support.
 596 */
 597
 598/* Number of bytes inserted at the start of a TSO header buffer,
 599 * similar to NET_IP_ALIGN.
 600 */
 601#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
 602#define TSOH_OFFSET	0
 603#else
 604#define TSOH_OFFSET	NET_IP_ALIGN
 605#endif
 606
 607#define TSOH_BUFFER(tsoh)	((u8 *)(tsoh + 1) + TSOH_OFFSET)
 608
 609/* Total size of struct efx_tso_header, buffer and padding */
 610#define TSOH_SIZE(hdr_len)					\
 611	(sizeof(struct efx_tso_header) + TSOH_OFFSET + hdr_len)
 612
 613/* Size of blocks on free list.  Larger blocks must be allocated from
 614 * the heap.
 615 */
 616#define TSOH_STD_SIZE		128
 617
 618#define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
 619#define ETH_HDR_LEN(skb)  (skb_network_header(skb) - (skb)->data)
 620#define SKB_TCP_OFF(skb)  PTR_DIFF(tcp_hdr(skb), (skb)->data)
 621#define SKB_IPV4_OFF(skb) PTR_DIFF(ip_hdr(skb), (skb)->data)
 622#define SKB_IPV6_OFF(skb) PTR_DIFF(ipv6_hdr(skb), (skb)->data)
 623
 624/**
 625 * struct tso_state - TSO state for an SKB
 626 * @out_len: Remaining length in current segment
 627 * @seqnum: Current sequence number
 628 * @ipv4_id: Current IPv4 ID, host endian
 629 * @packet_space: Remaining space in current packet
 630 * @dma_addr: DMA address of current position
 631 * @in_len: Remaining length in current SKB fragment
 632 * @unmap_len: Length of SKB fragment
 633 * @unmap_addr: DMA address of SKB fragment
 634 * @unmap_single: DMA single vs page mapping flag
 635 * @protocol: Network protocol (after any VLAN header)
 636 * @header_len: Number of bytes of header
 637 * @full_packet_size: Number of bytes to put in each outgoing segment
 638 *
 639 * The state used during segmentation.  It is put into this data structure
 640 * just to make it easy to pass into inline functions.
 641 */
 642struct tso_state {
 643	/* Output position */
 644	unsigned out_len;
 645	unsigned seqnum;
 646	unsigned ipv4_id;
 647	unsigned packet_space;
 648
 649	/* Input position */
 650	dma_addr_t dma_addr;
 651	unsigned in_len;
 652	unsigned unmap_len;
 653	dma_addr_t unmap_addr;
 654	bool unmap_single;
 655
 656	__be16 protocol;
 657	unsigned header_len;
 658	int full_packet_size;
 659};
 660
 661
 662/*
 663 * Verify that our various assumptions about sk_buffs and the conditions
 664 * under which TSO will be attempted hold true.  Return the protocol number.
 665 */
 666static __be16 efx_tso_check_protocol(struct sk_buff *skb)
 667{
 668	__be16 protocol = skb->protocol;
 
 
 
 
 
 669
 670	EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
 671			    protocol);
 672	if (protocol == htons(ETH_P_8021Q)) {
 673		/* Find the encapsulated protocol; reset network header
 674		 * and transport header based on that. */
 675		struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
 676		protocol = veh->h_vlan_encapsulated_proto;
 677		skb_set_network_header(skb, sizeof(*veh));
 678		if (protocol == htons(ETH_P_IP))
 679			skb_set_transport_header(skb, sizeof(*veh) +
 680						 4 * ip_hdr(skb)->ihl);
 681		else if (protocol == htons(ETH_P_IPV6))
 682			skb_set_transport_header(skb, sizeof(*veh) +
 683						 sizeof(struct ipv6hdr));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 684	}
 685
 686	if (protocol == htons(ETH_P_IP)) {
 687		EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
 688	} else {
 689		EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
 690		EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
 691	}
 692	EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
 693			     + (tcp_hdr(skb)->doff << 2u)) >
 694			    skb_headlen(skb));
 695
 696	return protocol;
 697}
 698
 
 699
 700/*
 701 * Allocate a page worth of efx_tso_header structures, and string them
 702 * into the tx_queue->tso_headers_free linked list. Return 0 or -ENOMEM.
 703 */
 704static int efx_tsoh_block_alloc(struct efx_tx_queue *tx_queue)
 705{
 706
 707	struct pci_dev *pci_dev = tx_queue->efx->pci_dev;
 708	struct efx_tso_header *tsoh;
 709	dma_addr_t dma_addr;
 710	u8 *base_kva, *kva;
 711
 712	base_kva = pci_alloc_consistent(pci_dev, PAGE_SIZE, &dma_addr);
 713	if (base_kva == NULL) {
 714		netif_err(tx_queue->efx, tx_err, tx_queue->efx->net_dev,
 715			  "Unable to allocate page for TSO headers\n");
 716		return -ENOMEM;
 717	}
 718
 719	/* pci_alloc_consistent() allocates pages. */
 720	EFX_BUG_ON_PARANOID(dma_addr & (PAGE_SIZE - 1u));
 721
 722	for (kva = base_kva; kva < base_kva + PAGE_SIZE; kva += TSOH_STD_SIZE) {
 723		tsoh = (struct efx_tso_header *)kva;
 724		tsoh->dma_addr = dma_addr + (TSOH_BUFFER(tsoh) - base_kva);
 725		tsoh->next = tx_queue->tso_headers_free;
 726		tx_queue->tso_headers_free = tsoh;
 727	}
 728
 729	return 0;
 730}
 
 731
 
 
 
 
 
 
 732
 733/* Free up a TSO header, and all others in the same page. */
 734static void efx_tsoh_block_free(struct efx_tx_queue *tx_queue,
 735				struct efx_tso_header *tsoh,
 736				struct pci_dev *pci_dev)
 737{
 738	struct efx_tso_header **p;
 739	unsigned long base_kva;
 740	dma_addr_t base_dma;
 741
 742	base_kva = (unsigned long)tsoh & PAGE_MASK;
 743	base_dma = tsoh->dma_addr & PAGE_MASK;
 744
 745	p = &tx_queue->tso_headers_free;
 746	while (*p != NULL) {
 747		if (((unsigned long)*p & PAGE_MASK) == base_kva)
 748			*p = (*p)->next;
 749		else
 750			p = &(*p)->next;
 751	}
 752
 753	pci_free_consistent(pci_dev, PAGE_SIZE, (void *)base_kva, base_dma);
 754}
 755
 756static struct efx_tso_header *
 757efx_tsoh_heap_alloc(struct efx_tx_queue *tx_queue, size_t header_len)
 758{
 759	struct efx_tso_header *tsoh;
 760
 761	tsoh = kmalloc(TSOH_SIZE(header_len), GFP_ATOMIC | GFP_DMA);
 762	if (unlikely(!tsoh))
 763		return NULL;
 764
 765	tsoh->dma_addr = pci_map_single(tx_queue->efx->pci_dev,
 766					TSOH_BUFFER(tsoh), header_len,
 767					PCI_DMA_TODEVICE);
 768	if (unlikely(pci_dma_mapping_error(tx_queue->efx->pci_dev,
 769					   tsoh->dma_addr))) {
 770		kfree(tsoh);
 771		return NULL;
 772	}
 773
 774	tsoh->unmap_len = header_len;
 775	return tsoh;
 776}
 777
 778static void
 779efx_tsoh_heap_free(struct efx_tx_queue *tx_queue, struct efx_tso_header *tsoh)
 780{
 781	pci_unmap_single(tx_queue->efx->pci_dev,
 782			 tsoh->dma_addr, tsoh->unmap_len,
 783			 PCI_DMA_TODEVICE);
 784	kfree(tsoh);
 785}
 786
 787/**
 788 * efx_tx_queue_insert - push descriptors onto the TX queue
 789 * @tx_queue:		Efx TX queue
 790 * @dma_addr:		DMA address of fragment
 791 * @len:		Length of fragment
 792 * @final_buffer:	The final buffer inserted into the queue
 793 *
 794 * Push descriptors onto the TX queue.  Return 0 on success or 1 if
 795 * @tx_queue full.
 
 796 */
 797static int efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
 798			       dma_addr_t dma_addr, unsigned len,
 799			       struct efx_tx_buffer **final_buffer)
 800{
 801	struct efx_tx_buffer *buffer;
 802	struct efx_nic *efx = tx_queue->efx;
 803	unsigned dma_len, fill_level, insert_ptr;
 804	int q_space;
 
 
 
 
 805
 806	EFX_BUG_ON_PARANOID(len <= 0);
 
 
 
 807
 808	fill_level = tx_queue->insert_count - tx_queue->old_read_count;
 809	/* -1 as there is no way to represent all descriptors used */
 810	q_space = efx->txq_entries - 1 - fill_level;
 811
 812	while (1) {
 813		if (unlikely(q_space-- <= 0)) {
 814			/* It might be that completions have happened
 815			 * since the xmit path last checked.  Update
 816			 * the xmit path's copy of read_count.
 817			 */
 818			netif_tx_stop_queue(tx_queue->core_txq);
 819			/* This memory barrier protects the change of
 820			 * queue state from the access of read_count. */
 821			smp_mb();
 822			tx_queue->old_read_count =
 823				ACCESS_ONCE(tx_queue->read_count);
 824			fill_level = (tx_queue->insert_count
 825				      - tx_queue->old_read_count);
 826			q_space = efx->txq_entries - 1 - fill_level;
 827			if (unlikely(q_space-- <= 0)) {
 828				*final_buffer = NULL;
 829				return 1;
 830			}
 831			smp_mb();
 832			netif_tx_start_queue(tx_queue->core_txq);
 833		}
 834
 835		insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
 836		buffer = &tx_queue->buffer[insert_ptr];
 837		++tx_queue->insert_count;
 838
 839		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
 840				    tx_queue->read_count >=
 841				    efx->txq_entries);
 842
 843		efx_tsoh_free(tx_queue, buffer);
 844		EFX_BUG_ON_PARANOID(buffer->len);
 845		EFX_BUG_ON_PARANOID(buffer->unmap_len);
 846		EFX_BUG_ON_PARANOID(buffer->skb);
 847		EFX_BUG_ON_PARANOID(!buffer->continuation);
 848		EFX_BUG_ON_PARANOID(buffer->tsoh);
 849
 850		buffer->dma_addr = dma_addr;
 
 851
 852		dma_len = efx_max_tx_len(efx, dma_addr);
 
 853
 854		/* If there is enough space to send then do so */
 855		if (dma_len >= len)
 856			break;
 857
 858		buffer->len = dma_len; /* Don't set the other members */
 859		dma_addr += dma_len;
 860		len -= dma_len;
 861	}
 862
 863	EFX_BUG_ON_PARANOID(!len);
 864	buffer->len = len;
 865	*final_buffer = buffer;
 866	return 0;
 867}
 868
 869
 870/*
 871 * Put a TSO header into the TX queue.
 872 *
 873 * This is special-cased because we know that it is small enough to fit in
 874 * a single fragment, and we know it doesn't cross a page boundary.  It
 875 * also allows us to not worry about end-of-packet etc.
 876 */
 877static void efx_tso_put_header(struct efx_tx_queue *tx_queue,
 878			       struct efx_tso_header *tsoh, unsigned len)
 879{
 880	struct efx_tx_buffer *buffer;
 881
 882	buffer = &tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask];
 883	efx_tsoh_free(tx_queue, buffer);
 884	EFX_BUG_ON_PARANOID(buffer->len);
 885	EFX_BUG_ON_PARANOID(buffer->unmap_len);
 886	EFX_BUG_ON_PARANOID(buffer->skb);
 887	EFX_BUG_ON_PARANOID(!buffer->continuation);
 888	EFX_BUG_ON_PARANOID(buffer->tsoh);
 889	buffer->len = len;
 890	buffer->dma_addr = tsoh->dma_addr;
 891	buffer->tsoh = tsoh;
 892
 893	++tx_queue->insert_count;
 894}
 895
 896
 897/* Remove descriptors put into a tx_queue. */
 898static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
 899{
 900	struct efx_tx_buffer *buffer;
 901	dma_addr_t unmap_addr;
 902
 903	/* Work backwards until we hit the original insert pointer value */
 904	while (tx_queue->insert_count != tx_queue->write_count) {
 905		--tx_queue->insert_count;
 906		buffer = &tx_queue->buffer[tx_queue->insert_count &
 907					   tx_queue->ptr_mask];
 908		efx_tsoh_free(tx_queue, buffer);
 909		EFX_BUG_ON_PARANOID(buffer->skb);
 910		if (buffer->unmap_len) {
 911			unmap_addr = (buffer->dma_addr + buffer->len -
 912				      buffer->unmap_len);
 913			if (buffer->unmap_single)
 914				pci_unmap_single(tx_queue->efx->pci_dev,
 915						 unmap_addr, buffer->unmap_len,
 916						 PCI_DMA_TODEVICE);
 917			else
 918				pci_unmap_page(tx_queue->efx->pci_dev,
 919					       unmap_addr, buffer->unmap_len,
 920					       PCI_DMA_TODEVICE);
 921			buffer->unmap_len = 0;
 922		}
 923		buffer->len = 0;
 924		buffer->continuation = true;
 925	}
 926}
 927
 928
 929/* Parse the SKB header and initialise state. */
 930static void tso_start(struct tso_state *st, const struct sk_buff *skb)
 931{
 932	/* All ethernet/IP/TCP headers combined size is TCP header size
 933	 * plus offset of TCP header relative to start of packet.
 934	 */
 935	st->header_len = ((tcp_hdr(skb)->doff << 2u)
 936			  + PTR_DIFF(tcp_hdr(skb), skb->data));
 937	st->full_packet_size = st->header_len + skb_shinfo(skb)->gso_size;
 938
 939	if (st->protocol == htons(ETH_P_IP))
 940		st->ipv4_id = ntohs(ip_hdr(skb)->id);
 941	else
 942		st->ipv4_id = 0;
 943	st->seqnum = ntohl(tcp_hdr(skb)->seq);
 944
 945	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
 946	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
 947	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
 948
 949	st->packet_space = st->full_packet_size;
 950	st->out_len = skb->len - st->header_len;
 951	st->unmap_len = 0;
 952	st->unmap_single = false;
 953}
 954
 955static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
 956			    skb_frag_t *frag)
 957{
 958	st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0,
 959					  skb_frag_size(frag), DMA_TO_DEVICE);
 960	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
 961		st->unmap_single = false;
 962		st->unmap_len = skb_frag_size(frag);
 963		st->in_len = skb_frag_size(frag);
 964		st->dma_addr = st->unmap_addr;
 965		return 0;
 966	}
 967	return -ENOMEM;
 968}
 969
 970static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
 971				 const struct sk_buff *skb)
 972{
 973	int hl = st->header_len;
 974	int len = skb_headlen(skb) - hl;
 975
 976	st->unmap_addr = pci_map_single(efx->pci_dev, skb->data + hl,
 977					len, PCI_DMA_TODEVICE);
 978	if (likely(!pci_dma_mapping_error(efx->pci_dev, st->unmap_addr))) {
 979		st->unmap_single = true;
 980		st->unmap_len = len;
 981		st->in_len = len;
 982		st->dma_addr = st->unmap_addr;
 983		return 0;
 984	}
 985	return -ENOMEM;
 986}
 987
 
 988
 989/**
 990 * tso_fill_packet_with_fragment - form descriptors for the current fragment
 991 * @tx_queue:		Efx TX queue
 992 * @skb:		Socket buffer
 993 * @st:			TSO state
 994 *
 995 * Form descriptors for the current fragment, until we reach the end
 996 * of fragment or end-of-packet.  Return 0 on success, 1 if not enough
 997 * space in @tx_queue.
 998 */
 999static int tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
1000					 const struct sk_buff *skb,
1001					 struct tso_state *st)
1002{
1003	struct efx_tx_buffer *buffer;
1004	int n, end_of_packet, rc;
1005
1006	if (st->in_len == 0)
1007		return 0;
1008	if (st->packet_space == 0)
1009		return 0;
1010
1011	EFX_BUG_ON_PARANOID(st->in_len <= 0);
1012	EFX_BUG_ON_PARANOID(st->packet_space <= 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1013
1014	n = min(st->in_len, st->packet_space);
1015
1016	st->packet_space -= n;
1017	st->out_len -= n;
1018	st->in_len -= n;
1019
1020	rc = efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
1021	if (likely(rc == 0)) {
1022		if (st->out_len == 0)
1023			/* Transfer ownership of the skb */
1024			buffer->skb = skb;
1025
1026		end_of_packet = st->out_len == 0 || st->packet_space == 0;
1027		buffer->continuation = !end_of_packet;
1028
1029		if (st->in_len == 0) {
1030			/* Transfer ownership of the pci mapping */
1031			buffer->unmap_len = st->unmap_len;
1032			buffer->unmap_single = st->unmap_single;
1033			st->unmap_len = 0;
1034		}
1035	}
1036
1037	st->dma_addr += n;
1038	return rc;
1039}
1040
1041
1042/**
1043 * tso_start_new_packet - generate a new header and prepare for the new packet
1044 * @tx_queue:		Efx TX queue
1045 * @skb:		Socket buffer
1046 * @st:			TSO state
1047 *
1048 * Generate a new header and prepare for the new packet.  Return 0 on
1049 * success, or -1 if failed to alloc header.
1050 */
1051static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
1052				const struct sk_buff *skb,
1053				struct tso_state *st)
1054{
1055	struct efx_tso_header *tsoh;
1056	struct tcphdr *tsoh_th;
1057	unsigned ip_length;
1058	u8 *header;
1059
1060	/* Allocate a DMA-mapped header buffer. */
1061	if (likely(TSOH_SIZE(st->header_len) <= TSOH_STD_SIZE)) {
1062		if (tx_queue->tso_headers_free == NULL) {
1063			if (efx_tsoh_block_alloc(tx_queue))
1064				return -1;
1065		}
1066		EFX_BUG_ON_PARANOID(!tx_queue->tso_headers_free);
1067		tsoh = tx_queue->tso_headers_free;
1068		tx_queue->tso_headers_free = tsoh->next;
1069		tsoh->unmap_len = 0;
1070	} else {
1071		tx_queue->tso_long_headers++;
1072		tsoh = efx_tsoh_heap_alloc(tx_queue, st->header_len);
1073		if (unlikely(!tsoh))
1074			return -1;
1075	}
1076
1077	header = TSOH_BUFFER(tsoh);
1078	tsoh_th = (struct tcphdr *)(header + SKB_TCP_OFF(skb));
1079
1080	/* Copy and update the headers. */
1081	memcpy(header, skb->data, st->header_len);
1082
1083	tsoh_th->seq = htonl(st->seqnum);
1084	st->seqnum += skb_shinfo(skb)->gso_size;
1085	if (st->out_len > skb_shinfo(skb)->gso_size) {
1086		/* This packet will not finish the TSO burst. */
1087		ip_length = st->full_packet_size - ETH_HDR_LEN(skb);
1088		tsoh_th->fin = 0;
1089		tsoh_th->psh = 0;
1090	} else {
1091		/* This packet will be the last in the TSO burst. */
1092		ip_length = st->header_len - ETH_HDR_LEN(skb) + st->out_len;
1093		tsoh_th->fin = tcp_hdr(skb)->fin;
1094		tsoh_th->psh = tcp_hdr(skb)->psh;
1095	}
1096
1097	if (st->protocol == htons(ETH_P_IP)) {
1098		struct iphdr *tsoh_iph =
1099			(struct iphdr *)(header + SKB_IPV4_OFF(skb));
1100
1101		tsoh_iph->tot_len = htons(ip_length);
1102
1103		/* Linux leaves suitable gaps in the IP ID space for us to fill. */
1104		tsoh_iph->id = htons(st->ipv4_id);
1105		st->ipv4_id++;
1106	} else {
1107		struct ipv6hdr *tsoh_iph =
1108			(struct ipv6hdr *)(header + SKB_IPV6_OFF(skb));
1109
1110		tsoh_iph->payload_len = htons(ip_length - sizeof(*tsoh_iph));
1111	}
1112
1113	st->packet_space = skb_shinfo(skb)->gso_size;
1114	++tx_queue->tso_packets;
1115
1116	/* Form a descriptor for this header. */
1117	efx_tso_put_header(tx_queue, tsoh, st->header_len);
1118
1119	return 0;
1120}
1121
1122
1123/**
1124 * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1125 * @tx_queue:		Efx TX queue
1126 * @skb:		Socket buffer
1127 *
1128 * Context: You must hold netif_tx_lock() to call this function.
1129 *
1130 * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1131 * @skb was not enqueued.  In all cases @skb is consumed.  Return
1132 * %NETDEV_TX_OK or %NETDEV_TX_BUSY.
1133 */
1134static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1135			       struct sk_buff *skb)
1136{
1137	struct efx_nic *efx = tx_queue->efx;
1138	int frag_i, rc, rc2 = NETDEV_TX_OK;
1139	struct tso_state state;
 
1140
1141	/* Find the packet protocol and sanity-check it */
1142	state.protocol = efx_tso_check_protocol(skb);
1143
1144	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
 
1145
1146	tso_start(&state, skb);
 
1147
1148	/* Assume that skb header area contains exactly the headers, and
1149	 * all payload is in the frag list.
1150	 */
1151	if (skb_headlen(skb) == state.header_len) {
1152		/* Grab the first payload fragment. */
1153		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1154		frag_i = 0;
1155		rc = tso_get_fragment(&state, efx,
1156				      skb_shinfo(skb)->frags + frag_i);
1157		if (rc)
1158			goto mem_err;
1159	} else {
1160		rc = tso_get_head_fragment(&state, efx, skb);
1161		if (rc)
1162			goto mem_err;
1163		frag_i = -1;
1164	}
1165
1166	if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1167		goto mem_err;
1168
1169	while (1) {
1170		rc = tso_fill_packet_with_fragment(tx_queue, skb, &state);
1171		if (unlikely(rc)) {
1172			rc2 = NETDEV_TX_BUSY;
1173			goto unwind;
1174		}
1175
1176		/* Move onto the next fragment? */
1177		if (state.in_len == 0) {
1178			if (++frag_i >= skb_shinfo(skb)->nr_frags)
1179				/* End of payload reached. */
1180				break;
1181			rc = tso_get_fragment(&state, efx,
1182					      skb_shinfo(skb)->frags + frag_i);
1183			if (rc)
1184				goto mem_err;
1185		}
1186
1187		/* Start at new packet? */
1188		if (state.packet_space == 0 &&
1189		    tso_start_new_packet(tx_queue, skb, &state) < 0)
1190			goto mem_err;
1191	}
1192
1193	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
1194
1195	/* Pass off to hardware */
1196	efx_nic_push_buffers(tx_queue);
1197
1198	tx_queue->tso_bursts++;
1199	return NETDEV_TX_OK;
1200
1201 mem_err:
1202	netif_err(efx, tx_err, efx->net_dev,
1203		  "Out of memory for TSO headers, or PCI mapping error\n");
1204	dev_kfree_skb_any(skb);
1205
1206 unwind:
1207	/* Free the DMA mapping we were in the process of writing out */
1208	if (state.unmap_len) {
1209		if (state.unmap_single)
1210			pci_unmap_single(efx->pci_dev, state.unmap_addr,
1211					 state.unmap_len, PCI_DMA_TODEVICE);
1212		else
1213			pci_unmap_page(efx->pci_dev, state.unmap_addr,
1214				       state.unmap_len, PCI_DMA_TODEVICE);
1215	}
1216
1217	efx_enqueue_unwind(tx_queue);
1218	return rc2;
1219}
1220
1221
1222/*
1223 * Free up all TSO datastructures associated with tx_queue. This
1224 * routine should be called only once the tx_queue is both empty and
1225 * will no longer be used.
1226 */
1227static void efx_fini_tso(struct efx_tx_queue *tx_queue)
1228{
1229	unsigned i;
1230
1231	if (tx_queue->buffer) {
1232		for (i = 0; i <= tx_queue->ptr_mask; ++i)
1233			efx_tsoh_free(tx_queue, &tx_queue->buffer[i]);
1234	}
1235
1236	while (tx_queue->tso_headers_free != NULL)
1237		efx_tsoh_block_free(tx_queue, tx_queue->tso_headers_free,
1238				    tx_queue->efx->pci_dev);
1239}