xref: /linux/drivers/net/ethernet/sfc/tx.c (revision b889fcf63cb62e7fdb7816565e28f44dbe4a76a5)
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 static void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
26 			       struct efx_tx_buffer *buffer,
27 			       unsigned int *pkts_compl,
28 			       unsigned int *bytes_compl)
29 {
30 	if (buffer->unmap_len) {
31 		struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
32 		dma_addr_t unmap_addr = (buffer->dma_addr + buffer->len -
33 					 buffer->unmap_len);
34 		if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
35 			dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
36 					 DMA_TO_DEVICE);
37 		else
38 			dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
39 				       DMA_TO_DEVICE);
40 		buffer->unmap_len = 0;
41 	}
42 
43 	if (buffer->flags & EFX_TX_BUF_SKB) {
44 		(*pkts_compl)++;
45 		(*bytes_compl) += buffer->skb->len;
46 		dev_kfree_skb_any((struct sk_buff *) buffer->skb);
47 		netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
48 			   "TX queue %d transmission id %x complete\n",
49 			   tx_queue->queue, tx_queue->read_count);
50 	} else if (buffer->flags & EFX_TX_BUF_HEAP) {
51 		kfree(buffer->heap_buf);
52 	}
53 
54 	buffer->len = 0;
55 	buffer->flags = 0;
56 }
57 
58 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
59 			       struct sk_buff *skb);
60 
61 static inline unsigned
62 efx_max_tx_len(struct efx_nic *efx, dma_addr_t dma_addr)
63 {
64 	/* Depending on the NIC revision, we can use descriptor
65 	 * lengths up to 8K or 8K-1.  However, since PCI Express
66 	 * devices must split read requests at 4K boundaries, there is
67 	 * little benefit from using descriptors that cross those
68 	 * boundaries and we keep things simple by not doing so.
69 	 */
70 	unsigned len = (~dma_addr & (EFX_PAGE_SIZE - 1)) + 1;
71 
72 	/* Work around hardware bug for unaligned buffers. */
73 	if (EFX_WORKAROUND_5391(efx) && (dma_addr & 0xf))
74 		len = min_t(unsigned, len, 512 - (dma_addr & 0xf));
75 
76 	return len;
77 }
78 
79 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
80 {
81 	/* Header and payload descriptor for each output segment, plus
82 	 * one for every input fragment boundary within a segment
83 	 */
84 	unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
85 
86 	/* Possibly one more per segment for the alignment workaround */
87 	if (EFX_WORKAROUND_5391(efx))
88 		max_descs += EFX_TSO_MAX_SEGS;
89 
90 	/* Possibly more for PCIe page boundaries within input fragments */
91 	if (PAGE_SIZE > EFX_PAGE_SIZE)
92 		max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
93 				   DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE));
94 
95 	return max_descs;
96 }
97 
98 /* Get partner of a TX queue, seen as part of the same net core queue */
99 static struct efx_tx_queue *efx_tx_queue_partner(struct efx_tx_queue *tx_queue)
100 {
101 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
102 		return tx_queue - EFX_TXQ_TYPE_OFFLOAD;
103 	else
104 		return tx_queue + EFX_TXQ_TYPE_OFFLOAD;
105 }
106 
107 static void efx_tx_maybe_stop_queue(struct efx_tx_queue *txq1)
108 {
109 	/* We need to consider both queues that the net core sees as one */
110 	struct efx_tx_queue *txq2 = efx_tx_queue_partner(txq1);
111 	struct efx_nic *efx = txq1->efx;
112 	unsigned int fill_level;
113 
114 	fill_level = max(txq1->insert_count - txq1->old_read_count,
115 			 txq2->insert_count - txq2->old_read_count);
116 	if (likely(fill_level < efx->txq_stop_thresh))
117 		return;
118 
119 	/* We used the stale old_read_count above, which gives us a
120 	 * pessimistic estimate of the fill level (which may even
121 	 * validly be >= efx->txq_entries).  Now try again using
122 	 * read_count (more likely to be a cache miss).
123 	 *
124 	 * If we read read_count and then conditionally stop the
125 	 * queue, it is possible for the completion path to race with
126 	 * us and complete all outstanding descriptors in the middle,
127 	 * after which there will be no more completions to wake it.
128 	 * Therefore we stop the queue first, then read read_count
129 	 * (with a memory barrier to ensure the ordering), then
130 	 * restart the queue if the fill level turns out to be low
131 	 * enough.
132 	 */
133 	netif_tx_stop_queue(txq1->core_txq);
134 	smp_mb();
135 	txq1->old_read_count = ACCESS_ONCE(txq1->read_count);
136 	txq2->old_read_count = ACCESS_ONCE(txq2->read_count);
137 
138 	fill_level = max(txq1->insert_count - txq1->old_read_count,
139 			 txq2->insert_count - txq2->old_read_count);
140 	EFX_BUG_ON_PARANOID(fill_level >= efx->txq_entries);
141 	if (likely(fill_level < efx->txq_stop_thresh)) {
142 		smp_mb();
143 		if (likely(!efx->loopback_selftest))
144 			netif_tx_start_queue(txq1->core_txq);
145 	}
146 }
147 
148 /*
149  * Add a socket buffer to a TX queue
150  *
151  * This maps all fragments of a socket buffer for DMA and adds them to
152  * the TX queue.  The queue's insert pointer will be incremented by
153  * the number of fragments in the socket buffer.
154  *
155  * If any DMA mapping fails, any mapped fragments will be unmapped,
156  * the queue's insert pointer will be restored to its original value.
157  *
158  * This function is split out from efx_hard_start_xmit to allow the
159  * loopback test to direct packets via specific TX queues.
160  *
161  * Returns NETDEV_TX_OK.
162  * You must hold netif_tx_lock() to call this function.
163  */
164 netdev_tx_t efx_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
165 {
166 	struct efx_nic *efx = tx_queue->efx;
167 	struct device *dma_dev = &efx->pci_dev->dev;
168 	struct efx_tx_buffer *buffer;
169 	skb_frag_t *fragment;
170 	unsigned int len, unmap_len = 0, insert_ptr;
171 	dma_addr_t dma_addr, unmap_addr = 0;
172 	unsigned int dma_len;
173 	unsigned short dma_flags;
174 	int i = 0;
175 
176 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
177 
178 	if (skb_shinfo(skb)->gso_size)
179 		return efx_enqueue_skb_tso(tx_queue, skb);
180 
181 	/* Get size of the initial fragment */
182 	len = skb_headlen(skb);
183 
184 	/* Pad if necessary */
185 	if (EFX_WORKAROUND_15592(efx) && skb->len <= 32) {
186 		EFX_BUG_ON_PARANOID(skb->data_len);
187 		len = 32 + 1;
188 		if (skb_pad(skb, len - skb->len))
189 			return NETDEV_TX_OK;
190 	}
191 
192 	/* Map for DMA.  Use dma_map_single rather than dma_map_page
193 	 * since this is more efficient on machines with sparse
194 	 * memory.
195 	 */
196 	dma_flags = EFX_TX_BUF_MAP_SINGLE;
197 	dma_addr = dma_map_single(dma_dev, skb->data, len, PCI_DMA_TODEVICE);
198 
199 	/* Process all fragments */
200 	while (1) {
201 		if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
202 			goto dma_err;
203 
204 		/* Store fields for marking in the per-fragment final
205 		 * descriptor */
206 		unmap_len = len;
207 		unmap_addr = dma_addr;
208 
209 		/* Add to TX queue, splitting across DMA boundaries */
210 		do {
211 			insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
212 			buffer = &tx_queue->buffer[insert_ptr];
213 			EFX_BUG_ON_PARANOID(buffer->flags);
214 			EFX_BUG_ON_PARANOID(buffer->len);
215 			EFX_BUG_ON_PARANOID(buffer->unmap_len);
216 
217 			dma_len = efx_max_tx_len(efx, dma_addr);
218 			if (likely(dma_len >= len))
219 				dma_len = len;
220 
221 			/* Fill out per descriptor fields */
222 			buffer->len = dma_len;
223 			buffer->dma_addr = dma_addr;
224 			buffer->flags = EFX_TX_BUF_CONT;
225 			len -= dma_len;
226 			dma_addr += dma_len;
227 			++tx_queue->insert_count;
228 		} while (len);
229 
230 		/* Transfer ownership of the unmapping to the final buffer */
231 		buffer->flags = EFX_TX_BUF_CONT | dma_flags;
232 		buffer->unmap_len = unmap_len;
233 		unmap_len = 0;
234 
235 		/* Get address and size of next fragment */
236 		if (i >= skb_shinfo(skb)->nr_frags)
237 			break;
238 		fragment = &skb_shinfo(skb)->frags[i];
239 		len = skb_frag_size(fragment);
240 		i++;
241 		/* Map for DMA */
242 		dma_flags = 0;
243 		dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
244 					    DMA_TO_DEVICE);
245 	}
246 
247 	/* Transfer ownership of the skb to the final buffer */
248 	buffer->skb = skb;
249 	buffer->flags = EFX_TX_BUF_SKB | dma_flags;
250 
251 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
252 
253 	/* Pass off to hardware */
254 	efx_nic_push_buffers(tx_queue);
255 
256 	efx_tx_maybe_stop_queue(tx_queue);
257 
258 	return NETDEV_TX_OK;
259 
260  dma_err:
261 	netif_err(efx, tx_err, efx->net_dev,
262 		  " TX queue %d could not map skb with %d bytes %d "
263 		  "fragments for DMA\n", tx_queue->queue, skb->len,
264 		  skb_shinfo(skb)->nr_frags + 1);
265 
266 	/* Mark the packet as transmitted, and free the SKB ourselves */
267 	dev_kfree_skb_any(skb);
268 
269 	/* Work backwards until we hit the original insert pointer value */
270 	while (tx_queue->insert_count != tx_queue->write_count) {
271 		unsigned int pkts_compl = 0, bytes_compl = 0;
272 		--tx_queue->insert_count;
273 		insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
274 		buffer = &tx_queue->buffer[insert_ptr];
275 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
276 	}
277 
278 	/* Free the fragment we were mid-way through pushing */
279 	if (unmap_len) {
280 		if (dma_flags & EFX_TX_BUF_MAP_SINGLE)
281 			dma_unmap_single(dma_dev, unmap_addr, unmap_len,
282 					 DMA_TO_DEVICE);
283 		else
284 			dma_unmap_page(dma_dev, unmap_addr, unmap_len,
285 				       DMA_TO_DEVICE);
286 	}
287 
288 	return NETDEV_TX_OK;
289 }
290 
291 /* Remove packets from the TX queue
292  *
293  * This removes packets from the TX queue, up to and including the
294  * specified index.
295  */
296 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
297 				unsigned int index,
298 				unsigned int *pkts_compl,
299 				unsigned int *bytes_compl)
300 {
301 	struct efx_nic *efx = tx_queue->efx;
302 	unsigned int stop_index, read_ptr;
303 
304 	stop_index = (index + 1) & tx_queue->ptr_mask;
305 	read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
306 
307 	while (read_ptr != stop_index) {
308 		struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
309 		if (unlikely(buffer->len == 0)) {
310 			netif_err(efx, tx_err, efx->net_dev,
311 				  "TX queue %d spurious TX completion id %x\n",
312 				  tx_queue->queue, read_ptr);
313 			efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
314 			return;
315 		}
316 
317 		efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
318 
319 		++tx_queue->read_count;
320 		read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
321 	}
322 }
323 
324 /* Initiate a packet transmission.  We use one channel per CPU
325  * (sharing when we have more CPUs than channels).  On Falcon, the TX
326  * completion events will be directed back to the CPU that transmitted
327  * the packet, which should be cache-efficient.
328  *
329  * Context: non-blocking.
330  * Note that returning anything other than NETDEV_TX_OK will cause the
331  * OS to free the skb.
332  */
333 netdev_tx_t efx_hard_start_xmit(struct sk_buff *skb,
334 				struct net_device *net_dev)
335 {
336 	struct efx_nic *efx = netdev_priv(net_dev);
337 	struct efx_tx_queue *tx_queue;
338 	unsigned index, type;
339 
340 	EFX_WARN_ON_PARANOID(!netif_device_present(net_dev));
341 
342 	/* PTP "event" packet */
343 	if (unlikely(efx_xmit_with_hwtstamp(skb)) &&
344 	    unlikely(efx_ptp_is_ptp_tx(efx, skb))) {
345 		return efx_ptp_tx(efx, skb);
346 	}
347 
348 	index = skb_get_queue_mapping(skb);
349 	type = skb->ip_summed == CHECKSUM_PARTIAL ? EFX_TXQ_TYPE_OFFLOAD : 0;
350 	if (index >= efx->n_tx_channels) {
351 		index -= efx->n_tx_channels;
352 		type |= EFX_TXQ_TYPE_HIGHPRI;
353 	}
354 	tx_queue = efx_get_tx_queue(efx, index, type);
355 
356 	return efx_enqueue_skb(tx_queue, skb);
357 }
358 
359 void efx_init_tx_queue_core_txq(struct efx_tx_queue *tx_queue)
360 {
361 	struct efx_nic *efx = tx_queue->efx;
362 
363 	/* Must be inverse of queue lookup in efx_hard_start_xmit() */
364 	tx_queue->core_txq =
365 		netdev_get_tx_queue(efx->net_dev,
366 				    tx_queue->queue / EFX_TXQ_TYPES +
367 				    ((tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
368 				     efx->n_tx_channels : 0));
369 }
370 
371 int efx_setup_tc(struct net_device *net_dev, u8 num_tc)
372 {
373 	struct efx_nic *efx = netdev_priv(net_dev);
374 	struct efx_channel *channel;
375 	struct efx_tx_queue *tx_queue;
376 	unsigned tc;
377 	int rc;
378 
379 	if (efx_nic_rev(efx) < EFX_REV_FALCON_B0 || num_tc > EFX_MAX_TX_TC)
380 		return -EINVAL;
381 
382 	if (num_tc == net_dev->num_tc)
383 		return 0;
384 
385 	for (tc = 0; tc < num_tc; tc++) {
386 		net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
387 		net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
388 	}
389 
390 	if (num_tc > net_dev->num_tc) {
391 		/* Initialise high-priority queues as necessary */
392 		efx_for_each_channel(channel, efx) {
393 			efx_for_each_possible_channel_tx_queue(tx_queue,
394 							       channel) {
395 				if (!(tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI))
396 					continue;
397 				if (!tx_queue->buffer) {
398 					rc = efx_probe_tx_queue(tx_queue);
399 					if (rc)
400 						return rc;
401 				}
402 				if (!tx_queue->initialised)
403 					efx_init_tx_queue(tx_queue);
404 				efx_init_tx_queue_core_txq(tx_queue);
405 			}
406 		}
407 	} else {
408 		/* Reduce number of classes before number of queues */
409 		net_dev->num_tc = num_tc;
410 	}
411 
412 	rc = netif_set_real_num_tx_queues(net_dev,
413 					  max_t(int, num_tc, 1) *
414 					  efx->n_tx_channels);
415 	if (rc)
416 		return rc;
417 
418 	/* Do not destroy high-priority queues when they become
419 	 * unused.  We would have to flush them first, and it is
420 	 * fairly difficult to flush a subset of TX queues.  Leave
421 	 * it to efx_fini_channels().
422 	 */
423 
424 	net_dev->num_tc = num_tc;
425 	return 0;
426 }
427 
428 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
429 {
430 	unsigned fill_level;
431 	struct efx_nic *efx = tx_queue->efx;
432 	struct efx_tx_queue *txq2;
433 	unsigned int pkts_compl = 0, bytes_compl = 0;
434 
435 	EFX_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
436 
437 	efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
438 	netdev_tx_completed_queue(tx_queue->core_txq, pkts_compl, bytes_compl);
439 
440 	/* See if we need to restart the netif queue.  This memory
441 	 * barrier ensures that we write read_count (inside
442 	 * efx_dequeue_buffers()) before reading the queue status.
443 	 */
444 	smp_mb();
445 	if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
446 	    likely(efx->port_enabled) &&
447 	    likely(netif_device_present(efx->net_dev))) {
448 		txq2 = efx_tx_queue_partner(tx_queue);
449 		fill_level = max(tx_queue->insert_count - tx_queue->read_count,
450 				 txq2->insert_count - txq2->read_count);
451 		if (fill_level <= efx->txq_wake_thresh)
452 			netif_tx_wake_queue(tx_queue->core_txq);
453 	}
454 
455 	/* Check whether the hardware queue is now empty */
456 	if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
457 		tx_queue->old_write_count = ACCESS_ONCE(tx_queue->write_count);
458 		if (tx_queue->read_count == tx_queue->old_write_count) {
459 			smp_mb();
460 			tx_queue->empty_read_count =
461 				tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
462 		}
463 	}
464 }
465 
466 /* Size of page-based TSO header buffers.  Larger blocks must be
467  * allocated from the heap.
468  */
469 #define TSOH_STD_SIZE	128
470 #define TSOH_PER_PAGE	(PAGE_SIZE / TSOH_STD_SIZE)
471 
472 /* At most half the descriptors in the queue at any time will refer to
473  * a TSO header buffer, since they must always be followed by a
474  * payload descriptor referring to an skb.
475  */
476 static unsigned int efx_tsoh_page_count(struct efx_tx_queue *tx_queue)
477 {
478 	return DIV_ROUND_UP(tx_queue->ptr_mask + 1, 2 * TSOH_PER_PAGE);
479 }
480 
481 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
482 {
483 	struct efx_nic *efx = tx_queue->efx;
484 	unsigned int entries;
485 	int rc;
486 
487 	/* Create the smallest power-of-two aligned ring */
488 	entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
489 	EFX_BUG_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
490 	tx_queue->ptr_mask = entries - 1;
491 
492 	netif_dbg(efx, probe, efx->net_dev,
493 		  "creating TX queue %d size %#x mask %#x\n",
494 		  tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
495 
496 	/* Allocate software ring */
497 	tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
498 				   GFP_KERNEL);
499 	if (!tx_queue->buffer)
500 		return -ENOMEM;
501 
502 	if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD) {
503 		tx_queue->tsoh_page =
504 			kcalloc(efx_tsoh_page_count(tx_queue),
505 				sizeof(tx_queue->tsoh_page[0]), GFP_KERNEL);
506 		if (!tx_queue->tsoh_page) {
507 			rc = -ENOMEM;
508 			goto fail1;
509 		}
510 	}
511 
512 	/* Allocate hardware ring */
513 	rc = efx_nic_probe_tx(tx_queue);
514 	if (rc)
515 		goto fail2;
516 
517 	return 0;
518 
519 fail2:
520 	kfree(tx_queue->tsoh_page);
521 	tx_queue->tsoh_page = NULL;
522 fail1:
523 	kfree(tx_queue->buffer);
524 	tx_queue->buffer = NULL;
525 	return rc;
526 }
527 
528 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
529 {
530 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
531 		  "initialising TX queue %d\n", tx_queue->queue);
532 
533 	tx_queue->insert_count = 0;
534 	tx_queue->write_count = 0;
535 	tx_queue->old_write_count = 0;
536 	tx_queue->read_count = 0;
537 	tx_queue->old_read_count = 0;
538 	tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
539 
540 	/* Set up TX descriptor ring */
541 	efx_nic_init_tx(tx_queue);
542 
543 	tx_queue->initialised = true;
544 }
545 
546 void efx_release_tx_buffers(struct efx_tx_queue *tx_queue)
547 {
548 	struct efx_tx_buffer *buffer;
549 
550 	if (!tx_queue->buffer)
551 		return;
552 
553 	/* Free any buffers left in the ring */
554 	while (tx_queue->read_count != tx_queue->write_count) {
555 		unsigned int pkts_compl = 0, bytes_compl = 0;
556 		buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
557 		efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
558 
559 		++tx_queue->read_count;
560 	}
561 	netdev_tx_reset_queue(tx_queue->core_txq);
562 }
563 
564 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
565 {
566 	if (!tx_queue->initialised)
567 		return;
568 
569 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
570 		  "shutting down TX queue %d\n", tx_queue->queue);
571 
572 	tx_queue->initialised = false;
573 
574 	/* Flush TX queue, remove descriptor ring */
575 	efx_nic_fini_tx(tx_queue);
576 
577 	efx_release_tx_buffers(tx_queue);
578 }
579 
580 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
581 {
582 	int i;
583 
584 	if (!tx_queue->buffer)
585 		return;
586 
587 	netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
588 		  "destroying TX queue %d\n", tx_queue->queue);
589 	efx_nic_remove_tx(tx_queue);
590 
591 	if (tx_queue->tsoh_page) {
592 		for (i = 0; i < efx_tsoh_page_count(tx_queue); i++)
593 			efx_nic_free_buffer(tx_queue->efx,
594 					    &tx_queue->tsoh_page[i]);
595 		kfree(tx_queue->tsoh_page);
596 		tx_queue->tsoh_page = NULL;
597 	}
598 
599 	kfree(tx_queue->buffer);
600 	tx_queue->buffer = NULL;
601 }
602 
603 
604 /* Efx TCP segmentation acceleration.
605  *
606  * Why?  Because by doing it here in the driver we can go significantly
607  * faster than the GSO.
608  *
609  * Requires TX checksum offload support.
610  */
611 
612 /* Number of bytes inserted at the start of a TSO header buffer,
613  * similar to NET_IP_ALIGN.
614  */
615 #ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
616 #define TSOH_OFFSET	0
617 #else
618 #define TSOH_OFFSET	NET_IP_ALIGN
619 #endif
620 
621 #define PTR_DIFF(p1, p2)  ((u8 *)(p1) - (u8 *)(p2))
622 
623 /**
624  * struct tso_state - TSO state for an SKB
625  * @out_len: Remaining length in current segment
626  * @seqnum: Current sequence number
627  * @ipv4_id: Current IPv4 ID, host endian
628  * @packet_space: Remaining space in current packet
629  * @dma_addr: DMA address of current position
630  * @in_len: Remaining length in current SKB fragment
631  * @unmap_len: Length of SKB fragment
632  * @unmap_addr: DMA address of SKB fragment
633  * @dma_flags: TX buffer flags for DMA mapping - %EFX_TX_BUF_MAP_SINGLE or 0
634  * @protocol: Network protocol (after any VLAN header)
635  * @ip_off: Offset of IP header
636  * @tcp_off: Offset of TCP header
637  * @header_len: Number of bytes of header
638  * @ip_base_len: IPv4 tot_len or IPv6 payload_len, before TCP payload
639  *
640  * The state used during segmentation.  It is put into this data structure
641  * just to make it easy to pass into inline functions.
642  */
643 struct tso_state {
644 	/* Output position */
645 	unsigned out_len;
646 	unsigned seqnum;
647 	unsigned ipv4_id;
648 	unsigned packet_space;
649 
650 	/* Input position */
651 	dma_addr_t dma_addr;
652 	unsigned in_len;
653 	unsigned unmap_len;
654 	dma_addr_t unmap_addr;
655 	unsigned short dma_flags;
656 
657 	__be16 protocol;
658 	unsigned int ip_off;
659 	unsigned int tcp_off;
660 	unsigned header_len;
661 	unsigned int ip_base_len;
662 };
663 
664 
665 /*
666  * Verify that our various assumptions about sk_buffs and the conditions
667  * under which TSO will be attempted hold true.  Return the protocol number.
668  */
669 static __be16 efx_tso_check_protocol(struct sk_buff *skb)
670 {
671 	__be16 protocol = skb->protocol;
672 
673 	EFX_BUG_ON_PARANOID(((struct ethhdr *)skb->data)->h_proto !=
674 			    protocol);
675 	if (protocol == htons(ETH_P_8021Q)) {
676 		struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
677 		protocol = veh->h_vlan_encapsulated_proto;
678 	}
679 
680 	if (protocol == htons(ETH_P_IP)) {
681 		EFX_BUG_ON_PARANOID(ip_hdr(skb)->protocol != IPPROTO_TCP);
682 	} else {
683 		EFX_BUG_ON_PARANOID(protocol != htons(ETH_P_IPV6));
684 		EFX_BUG_ON_PARANOID(ipv6_hdr(skb)->nexthdr != NEXTHDR_TCP);
685 	}
686 	EFX_BUG_ON_PARANOID((PTR_DIFF(tcp_hdr(skb), skb->data)
687 			     + (tcp_hdr(skb)->doff << 2u)) >
688 			    skb_headlen(skb));
689 
690 	return protocol;
691 }
692 
693 static u8 *efx_tsoh_get_buffer(struct efx_tx_queue *tx_queue,
694 			       struct efx_tx_buffer *buffer, unsigned int len)
695 {
696 	u8 *result;
697 
698 	EFX_BUG_ON_PARANOID(buffer->len);
699 	EFX_BUG_ON_PARANOID(buffer->flags);
700 	EFX_BUG_ON_PARANOID(buffer->unmap_len);
701 
702 	if (likely(len <= TSOH_STD_SIZE - TSOH_OFFSET)) {
703 		unsigned index =
704 			(tx_queue->insert_count & tx_queue->ptr_mask) / 2;
705 		struct efx_buffer *page_buf =
706 			&tx_queue->tsoh_page[index / TSOH_PER_PAGE];
707 		unsigned offset =
708 			TSOH_STD_SIZE * (index % TSOH_PER_PAGE) + TSOH_OFFSET;
709 
710 		if (unlikely(!page_buf->addr) &&
711 		    efx_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE))
712 			return NULL;
713 
714 		result = (u8 *)page_buf->addr + offset;
715 		buffer->dma_addr = page_buf->dma_addr + offset;
716 		buffer->flags = EFX_TX_BUF_CONT;
717 	} else {
718 		tx_queue->tso_long_headers++;
719 
720 		buffer->heap_buf = kmalloc(TSOH_OFFSET + len, GFP_ATOMIC);
721 		if (unlikely(!buffer->heap_buf))
722 			return NULL;
723 		result = (u8 *)buffer->heap_buf + TSOH_OFFSET;
724 		buffer->flags = EFX_TX_BUF_CONT | EFX_TX_BUF_HEAP;
725 	}
726 
727 	buffer->len = len;
728 
729 	return result;
730 }
731 
732 /**
733  * efx_tx_queue_insert - push descriptors onto the TX queue
734  * @tx_queue:		Efx TX queue
735  * @dma_addr:		DMA address of fragment
736  * @len:		Length of fragment
737  * @final_buffer:	The final buffer inserted into the queue
738  *
739  * Push descriptors onto the TX queue.
740  */
741 static void efx_tx_queue_insert(struct efx_tx_queue *tx_queue,
742 				dma_addr_t dma_addr, unsigned len,
743 				struct efx_tx_buffer **final_buffer)
744 {
745 	struct efx_tx_buffer *buffer;
746 	struct efx_nic *efx = tx_queue->efx;
747 	unsigned dma_len, insert_ptr;
748 
749 	EFX_BUG_ON_PARANOID(len <= 0);
750 
751 	while (1) {
752 		insert_ptr = tx_queue->insert_count & tx_queue->ptr_mask;
753 		buffer = &tx_queue->buffer[insert_ptr];
754 		++tx_queue->insert_count;
755 
756 		EFX_BUG_ON_PARANOID(tx_queue->insert_count -
757 				    tx_queue->read_count >=
758 				    efx->txq_entries);
759 
760 		EFX_BUG_ON_PARANOID(buffer->len);
761 		EFX_BUG_ON_PARANOID(buffer->unmap_len);
762 		EFX_BUG_ON_PARANOID(buffer->flags);
763 
764 		buffer->dma_addr = dma_addr;
765 
766 		dma_len = efx_max_tx_len(efx, dma_addr);
767 
768 		/* If there is enough space to send then do so */
769 		if (dma_len >= len)
770 			break;
771 
772 		buffer->len = dma_len;
773 		buffer->flags = EFX_TX_BUF_CONT;
774 		dma_addr += dma_len;
775 		len -= dma_len;
776 	}
777 
778 	EFX_BUG_ON_PARANOID(!len);
779 	buffer->len = len;
780 	*final_buffer = buffer;
781 }
782 
783 
784 /*
785  * Put a TSO header into the TX queue.
786  *
787  * This is special-cased because we know that it is small enough to fit in
788  * a single fragment, and we know it doesn't cross a page boundary.  It
789  * also allows us to not worry about end-of-packet etc.
790  */
791 static int efx_tso_put_header(struct efx_tx_queue *tx_queue,
792 			      struct efx_tx_buffer *buffer, u8 *header)
793 {
794 	if (unlikely(buffer->flags & EFX_TX_BUF_HEAP)) {
795 		buffer->dma_addr = dma_map_single(&tx_queue->efx->pci_dev->dev,
796 						  header, buffer->len,
797 						  DMA_TO_DEVICE);
798 		if (unlikely(dma_mapping_error(&tx_queue->efx->pci_dev->dev,
799 					       buffer->dma_addr))) {
800 			kfree(buffer->heap_buf);
801 			buffer->len = 0;
802 			buffer->flags = 0;
803 			return -ENOMEM;
804 		}
805 		buffer->unmap_len = buffer->len;
806 		buffer->flags |= EFX_TX_BUF_MAP_SINGLE;
807 	}
808 
809 	++tx_queue->insert_count;
810 	return 0;
811 }
812 
813 
814 /* Remove buffers put into a tx_queue.  None of the buffers must have
815  * an skb attached.
816  */
817 static void efx_enqueue_unwind(struct efx_tx_queue *tx_queue)
818 {
819 	struct efx_tx_buffer *buffer;
820 
821 	/* Work backwards until we hit the original insert pointer value */
822 	while (tx_queue->insert_count != tx_queue->write_count) {
823 		--tx_queue->insert_count;
824 		buffer = &tx_queue->buffer[tx_queue->insert_count &
825 					   tx_queue->ptr_mask];
826 		efx_dequeue_buffer(tx_queue, buffer, NULL, NULL);
827 	}
828 }
829 
830 
831 /* Parse the SKB header and initialise state. */
832 static void tso_start(struct tso_state *st, const struct sk_buff *skb)
833 {
834 	st->ip_off = skb_network_header(skb) - skb->data;
835 	st->tcp_off = skb_transport_header(skb) - skb->data;
836 	st->header_len = st->tcp_off + (tcp_hdr(skb)->doff << 2u);
837 	if (st->protocol == htons(ETH_P_IP)) {
838 		st->ip_base_len = st->header_len - st->ip_off;
839 		st->ipv4_id = ntohs(ip_hdr(skb)->id);
840 	} else {
841 		st->ip_base_len = st->header_len - st->tcp_off;
842 		st->ipv4_id = 0;
843 	}
844 	st->seqnum = ntohl(tcp_hdr(skb)->seq);
845 
846 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->urg);
847 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->syn);
848 	EFX_BUG_ON_PARANOID(tcp_hdr(skb)->rst);
849 
850 	st->out_len = skb->len - st->header_len;
851 	st->unmap_len = 0;
852 	st->dma_flags = 0;
853 }
854 
855 static int tso_get_fragment(struct tso_state *st, struct efx_nic *efx,
856 			    skb_frag_t *frag)
857 {
858 	st->unmap_addr = skb_frag_dma_map(&efx->pci_dev->dev, frag, 0,
859 					  skb_frag_size(frag), DMA_TO_DEVICE);
860 	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
861 		st->dma_flags = 0;
862 		st->unmap_len = skb_frag_size(frag);
863 		st->in_len = skb_frag_size(frag);
864 		st->dma_addr = st->unmap_addr;
865 		return 0;
866 	}
867 	return -ENOMEM;
868 }
869 
870 static int tso_get_head_fragment(struct tso_state *st, struct efx_nic *efx,
871 				 const struct sk_buff *skb)
872 {
873 	int hl = st->header_len;
874 	int len = skb_headlen(skb) - hl;
875 
876 	st->unmap_addr = dma_map_single(&efx->pci_dev->dev, skb->data + hl,
877 					len, DMA_TO_DEVICE);
878 	if (likely(!dma_mapping_error(&efx->pci_dev->dev, st->unmap_addr))) {
879 		st->dma_flags = EFX_TX_BUF_MAP_SINGLE;
880 		st->unmap_len = len;
881 		st->in_len = len;
882 		st->dma_addr = st->unmap_addr;
883 		return 0;
884 	}
885 	return -ENOMEM;
886 }
887 
888 
889 /**
890  * tso_fill_packet_with_fragment - form descriptors for the current fragment
891  * @tx_queue:		Efx TX queue
892  * @skb:		Socket buffer
893  * @st:			TSO state
894  *
895  * Form descriptors for the current fragment, until we reach the end
896  * of fragment or end-of-packet.
897  */
898 static void tso_fill_packet_with_fragment(struct efx_tx_queue *tx_queue,
899 					  const struct sk_buff *skb,
900 					  struct tso_state *st)
901 {
902 	struct efx_tx_buffer *buffer;
903 	int n;
904 
905 	if (st->in_len == 0)
906 		return;
907 	if (st->packet_space == 0)
908 		return;
909 
910 	EFX_BUG_ON_PARANOID(st->in_len <= 0);
911 	EFX_BUG_ON_PARANOID(st->packet_space <= 0);
912 
913 	n = min(st->in_len, st->packet_space);
914 
915 	st->packet_space -= n;
916 	st->out_len -= n;
917 	st->in_len -= n;
918 
919 	efx_tx_queue_insert(tx_queue, st->dma_addr, n, &buffer);
920 
921 	if (st->out_len == 0) {
922 		/* Transfer ownership of the skb */
923 		buffer->skb = skb;
924 		buffer->flags = EFX_TX_BUF_SKB;
925 	} else if (st->packet_space != 0) {
926 		buffer->flags = EFX_TX_BUF_CONT;
927 	}
928 
929 	if (st->in_len == 0) {
930 		/* Transfer ownership of the DMA mapping */
931 		buffer->unmap_len = st->unmap_len;
932 		buffer->flags |= st->dma_flags;
933 		st->unmap_len = 0;
934 	}
935 
936 	st->dma_addr += n;
937 }
938 
939 
940 /**
941  * tso_start_new_packet - generate a new header and prepare for the new packet
942  * @tx_queue:		Efx TX queue
943  * @skb:		Socket buffer
944  * @st:			TSO state
945  *
946  * Generate a new header and prepare for the new packet.  Return 0 on
947  * success, or -%ENOMEM if failed to alloc header.
948  */
949 static int tso_start_new_packet(struct efx_tx_queue *tx_queue,
950 				const struct sk_buff *skb,
951 				struct tso_state *st)
952 {
953 	struct efx_tx_buffer *buffer =
954 		&tx_queue->buffer[tx_queue->insert_count & tx_queue->ptr_mask];
955 	struct tcphdr *tsoh_th;
956 	unsigned ip_length;
957 	u8 *header;
958 	int rc;
959 
960 	/* Allocate and insert a DMA-mapped header buffer. */
961 	header = efx_tsoh_get_buffer(tx_queue, buffer, st->header_len);
962 	if (!header)
963 		return -ENOMEM;
964 
965 	tsoh_th = (struct tcphdr *)(header + st->tcp_off);
966 
967 	/* Copy and update the headers. */
968 	memcpy(header, skb->data, st->header_len);
969 
970 	tsoh_th->seq = htonl(st->seqnum);
971 	st->seqnum += skb_shinfo(skb)->gso_size;
972 	if (st->out_len > skb_shinfo(skb)->gso_size) {
973 		/* This packet will not finish the TSO burst. */
974 		st->packet_space = skb_shinfo(skb)->gso_size;
975 		tsoh_th->fin = 0;
976 		tsoh_th->psh = 0;
977 	} else {
978 		/* This packet will be the last in the TSO burst. */
979 		st->packet_space = st->out_len;
980 		tsoh_th->fin = tcp_hdr(skb)->fin;
981 		tsoh_th->psh = tcp_hdr(skb)->psh;
982 	}
983 	ip_length = st->ip_base_len + st->packet_space;
984 
985 	if (st->protocol == htons(ETH_P_IP)) {
986 		struct iphdr *tsoh_iph = (struct iphdr *)(header + st->ip_off);
987 
988 		tsoh_iph->tot_len = htons(ip_length);
989 
990 		/* Linux leaves suitable gaps in the IP ID space for us to fill. */
991 		tsoh_iph->id = htons(st->ipv4_id);
992 		st->ipv4_id++;
993 	} else {
994 		struct ipv6hdr *tsoh_iph =
995 			(struct ipv6hdr *)(header + st->ip_off);
996 
997 		tsoh_iph->payload_len = htons(ip_length);
998 	}
999 
1000 	rc = efx_tso_put_header(tx_queue, buffer, header);
1001 	if (unlikely(rc))
1002 		return rc;
1003 
1004 	++tx_queue->tso_packets;
1005 
1006 	return 0;
1007 }
1008 
1009 
1010 /**
1011  * efx_enqueue_skb_tso - segment and transmit a TSO socket buffer
1012  * @tx_queue:		Efx TX queue
1013  * @skb:		Socket buffer
1014  *
1015  * Context: You must hold netif_tx_lock() to call this function.
1016  *
1017  * Add socket buffer @skb to @tx_queue, doing TSO or return != 0 if
1018  * @skb was not enqueued.  In all cases @skb is consumed.  Return
1019  * %NETDEV_TX_OK.
1020  */
1021 static int efx_enqueue_skb_tso(struct efx_tx_queue *tx_queue,
1022 			       struct sk_buff *skb)
1023 {
1024 	struct efx_nic *efx = tx_queue->efx;
1025 	int frag_i, rc;
1026 	struct tso_state state;
1027 
1028 	/* Find the packet protocol and sanity-check it */
1029 	state.protocol = efx_tso_check_protocol(skb);
1030 
1031 	EFX_BUG_ON_PARANOID(tx_queue->write_count != tx_queue->insert_count);
1032 
1033 	tso_start(&state, skb);
1034 
1035 	/* Assume that skb header area contains exactly the headers, and
1036 	 * all payload is in the frag list.
1037 	 */
1038 	if (skb_headlen(skb) == state.header_len) {
1039 		/* Grab the first payload fragment. */
1040 		EFX_BUG_ON_PARANOID(skb_shinfo(skb)->nr_frags < 1);
1041 		frag_i = 0;
1042 		rc = tso_get_fragment(&state, efx,
1043 				      skb_shinfo(skb)->frags + frag_i);
1044 		if (rc)
1045 			goto mem_err;
1046 	} else {
1047 		rc = tso_get_head_fragment(&state, efx, skb);
1048 		if (rc)
1049 			goto mem_err;
1050 		frag_i = -1;
1051 	}
1052 
1053 	if (tso_start_new_packet(tx_queue, skb, &state) < 0)
1054 		goto mem_err;
1055 
1056 	while (1) {
1057 		tso_fill_packet_with_fragment(tx_queue, skb, &state);
1058 
1059 		/* Move onto the next fragment? */
1060 		if (state.in_len == 0) {
1061 			if (++frag_i >= skb_shinfo(skb)->nr_frags)
1062 				/* End of payload reached. */
1063 				break;
1064 			rc = tso_get_fragment(&state, efx,
1065 					      skb_shinfo(skb)->frags + frag_i);
1066 			if (rc)
1067 				goto mem_err;
1068 		}
1069 
1070 		/* Start at new packet? */
1071 		if (state.packet_space == 0 &&
1072 		    tso_start_new_packet(tx_queue, skb, &state) < 0)
1073 			goto mem_err;
1074 	}
1075 
1076 	netdev_tx_sent_queue(tx_queue->core_txq, skb->len);
1077 
1078 	/* Pass off to hardware */
1079 	efx_nic_push_buffers(tx_queue);
1080 
1081 	efx_tx_maybe_stop_queue(tx_queue);
1082 
1083 	tx_queue->tso_bursts++;
1084 	return NETDEV_TX_OK;
1085 
1086  mem_err:
1087 	netif_err(efx, tx_err, efx->net_dev,
1088 		  "Out of memory for TSO headers, or DMA mapping error\n");
1089 	dev_kfree_skb_any(skb);
1090 
1091 	/* Free the DMA mapping we were in the process of writing out */
1092 	if (state.unmap_len) {
1093 		if (state.dma_flags & EFX_TX_BUF_MAP_SINGLE)
1094 			dma_unmap_single(&efx->pci_dev->dev, state.unmap_addr,
1095 					 state.unmap_len, DMA_TO_DEVICE);
1096 		else
1097 			dma_unmap_page(&efx->pci_dev->dev, state.unmap_addr,
1098 				       state.unmap_len, DMA_TO_DEVICE);
1099 	}
1100 
1101 	efx_enqueue_unwind(tx_queue);
1102 	return NETDEV_TX_OK;
1103 }
1104