Loading...
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 "workarounds.h"
24
25static inline u8 *ef4_tx_get_copy_buffer(struct ef4_tx_queue *tx_queue,
26 struct ef4_tx_buffer *buffer)
27{
28 unsigned int index = ef4_tx_queue_get_insert_index(tx_queue);
29 struct ef4_buffer *page_buf =
30 &tx_queue->cb_page[index >> (PAGE_SHIFT - EF4_TX_CB_ORDER)];
31 unsigned int offset =
32 ((index << EF4_TX_CB_ORDER) + NET_IP_ALIGN) & (PAGE_SIZE - 1);
33
34 if (unlikely(!page_buf->addr) &&
35 ef4_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
36 GFP_ATOMIC))
37 return NULL;
38 buffer->dma_addr = page_buf->dma_addr + offset;
39 buffer->unmap_len = 0;
40 return (u8 *)page_buf->addr + offset;
41}
42
43u8 *ef4_tx_get_copy_buffer_limited(struct ef4_tx_queue *tx_queue,
44 struct ef4_tx_buffer *buffer, size_t len)
45{
46 if (len > EF4_TX_CB_SIZE)
47 return NULL;
48 return ef4_tx_get_copy_buffer(tx_queue, buffer);
49}
50
51static void ef4_dequeue_buffer(struct ef4_tx_queue *tx_queue,
52 struct ef4_tx_buffer *buffer,
53 unsigned int *pkts_compl,
54 unsigned int *bytes_compl)
55{
56 if (buffer->unmap_len) {
57 struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
58 dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
59 if (buffer->flags & EF4_TX_BUF_MAP_SINGLE)
60 dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
61 DMA_TO_DEVICE);
62 else
63 dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
64 DMA_TO_DEVICE);
65 buffer->unmap_len = 0;
66 }
67
68 if (buffer->flags & EF4_TX_BUF_SKB) {
69 (*pkts_compl)++;
70 (*bytes_compl) += buffer->skb->len;
71 dev_consume_skb_any((struct sk_buff *)buffer->skb);
72 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
73 "TX queue %d transmission id %x complete\n",
74 tx_queue->queue, tx_queue->read_count);
75 }
76
77 buffer->len = 0;
78 buffer->flags = 0;
79}
80
81unsigned int ef4_tx_max_skb_descs(struct ef4_nic *efx)
82{
83 /* This is probably too much since we don't have any TSO support;
84 * it's a left-over from when we had Software TSO. But it's safer
85 * to leave it as-is than try to determine a new bound.
86 */
87 /* Header and payload descriptor for each output segment, plus
88 * one for every input fragment boundary within a segment
89 */
90 unsigned int max_descs = EF4_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
91
92 /* Possibly one more per segment for the alignment workaround,
93 * or for option descriptors
94 */
95 if (EF4_WORKAROUND_5391(efx))
96 max_descs += EF4_TSO_MAX_SEGS;
97
98 /* Possibly more for PCIe page boundaries within input fragments */
99 if (PAGE_SIZE > EF4_PAGE_SIZE)
100 max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
101 DIV_ROUND_UP(GSO_LEGACY_MAX_SIZE,
102 EF4_PAGE_SIZE));
103
104 return max_descs;
105}
106
107static void ef4_tx_maybe_stop_queue(struct ef4_tx_queue *txq1)
108{
109 /* We need to consider both queues that the net core sees as one */
110 struct ef4_tx_queue *txq2 = ef4_tx_queue_partner(txq1);
111 struct ef4_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 = READ_ONCE(txq1->read_count);
136 txq2->old_read_count = READ_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 EF4_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
148static int ef4_enqueue_skb_copy(struct ef4_tx_queue *tx_queue,
149 struct sk_buff *skb)
150{
151 unsigned int min_len = tx_queue->tx_min_size;
152 unsigned int copy_len = skb->len;
153 struct ef4_tx_buffer *buffer;
154 u8 *copy_buffer;
155 int rc;
156
157 EF4_BUG_ON_PARANOID(copy_len > EF4_TX_CB_SIZE);
158
159 buffer = ef4_tx_queue_get_insert_buffer(tx_queue);
160
161 copy_buffer = ef4_tx_get_copy_buffer(tx_queue, buffer);
162 if (unlikely(!copy_buffer))
163 return -ENOMEM;
164
165 rc = skb_copy_bits(skb, 0, copy_buffer, copy_len);
166 EF4_WARN_ON_PARANOID(rc);
167 if (unlikely(copy_len < min_len)) {
168 memset(copy_buffer + copy_len, 0, min_len - copy_len);
169 buffer->len = min_len;
170 } else {
171 buffer->len = copy_len;
172 }
173
174 buffer->skb = skb;
175 buffer->flags = EF4_TX_BUF_SKB;
176
177 ++tx_queue->insert_count;
178 return rc;
179}
180
181static struct ef4_tx_buffer *ef4_tx_map_chunk(struct ef4_tx_queue *tx_queue,
182 dma_addr_t dma_addr,
183 size_t len)
184{
185 const struct ef4_nic_type *nic_type = tx_queue->efx->type;
186 struct ef4_tx_buffer *buffer;
187 unsigned int dma_len;
188
189 /* Map the fragment taking account of NIC-dependent DMA limits. */
190 do {
191 buffer = ef4_tx_queue_get_insert_buffer(tx_queue);
192 dma_len = nic_type->tx_limit_len(tx_queue, dma_addr, len);
193
194 buffer->len = dma_len;
195 buffer->dma_addr = dma_addr;
196 buffer->flags = EF4_TX_BUF_CONT;
197 len -= dma_len;
198 dma_addr += dma_len;
199 ++tx_queue->insert_count;
200 } while (len);
201
202 return buffer;
203}
204
205/* Map all data from an SKB for DMA and create descriptors on the queue.
206 */
207static int ef4_tx_map_data(struct ef4_tx_queue *tx_queue, struct sk_buff *skb)
208{
209 struct ef4_nic *efx = tx_queue->efx;
210 struct device *dma_dev = &efx->pci_dev->dev;
211 unsigned int frag_index, nr_frags;
212 dma_addr_t dma_addr, unmap_addr;
213 unsigned short dma_flags;
214 size_t len, unmap_len;
215
216 nr_frags = skb_shinfo(skb)->nr_frags;
217 frag_index = 0;
218
219 /* Map header data. */
220 len = skb_headlen(skb);
221 dma_addr = dma_map_single(dma_dev, skb->data, len, DMA_TO_DEVICE);
222 dma_flags = EF4_TX_BUF_MAP_SINGLE;
223 unmap_len = len;
224 unmap_addr = dma_addr;
225
226 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
227 return -EIO;
228
229 /* Add descriptors for each fragment. */
230 do {
231 struct ef4_tx_buffer *buffer;
232 skb_frag_t *fragment;
233
234 buffer = ef4_tx_map_chunk(tx_queue, dma_addr, len);
235
236 /* The final descriptor for a fragment is responsible for
237 * unmapping the whole fragment.
238 */
239 buffer->flags = EF4_TX_BUF_CONT | dma_flags;
240 buffer->unmap_len = unmap_len;
241 buffer->dma_offset = buffer->dma_addr - unmap_addr;
242
243 if (frag_index >= nr_frags) {
244 /* Store SKB details with the final buffer for
245 * the completion.
246 */
247 buffer->skb = skb;
248 buffer->flags = EF4_TX_BUF_SKB | dma_flags;
249 return 0;
250 }
251
252 /* Move on to the next fragment. */
253 fragment = &skb_shinfo(skb)->frags[frag_index++];
254 len = skb_frag_size(fragment);
255 dma_addr = skb_frag_dma_map(dma_dev, fragment,
256 0, len, DMA_TO_DEVICE);
257 dma_flags = 0;
258 unmap_len = len;
259 unmap_addr = dma_addr;
260
261 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
262 return -EIO;
263 } while (1);
264}
265
266/* Remove buffers put into a tx_queue. None of the buffers must have
267 * an skb attached.
268 */
269static void ef4_enqueue_unwind(struct ef4_tx_queue *tx_queue)
270{
271 struct ef4_tx_buffer *buffer;
272
273 /* Work backwards until we hit the original insert pointer value */
274 while (tx_queue->insert_count != tx_queue->write_count) {
275 --tx_queue->insert_count;
276 buffer = __ef4_tx_queue_get_insert_buffer(tx_queue);
277 ef4_dequeue_buffer(tx_queue, buffer, NULL, NULL);
278 }
279}
280
281/*
282 * Add a socket buffer to a TX queue
283 *
284 * This maps all fragments of a socket buffer for DMA and adds them to
285 * the TX queue. The queue's insert pointer will be incremented by
286 * the number of fragments in the socket buffer.
287 *
288 * If any DMA mapping fails, any mapped fragments will be unmapped,
289 * the queue's insert pointer will be restored to its original value.
290 *
291 * This function is split out from ef4_hard_start_xmit to allow the
292 * loopback test to direct packets via specific TX queues.
293 *
294 * Returns NETDEV_TX_OK.
295 * You must hold netif_tx_lock() to call this function.
296 */
297netdev_tx_t ef4_enqueue_skb(struct ef4_tx_queue *tx_queue, struct sk_buff *skb)
298{
299 bool data_mapped = false;
300 unsigned int skb_len;
301
302 skb_len = skb->len;
303 EF4_WARN_ON_PARANOID(skb_is_gso(skb));
304
305 if (skb_len < tx_queue->tx_min_size ||
306 (skb->data_len && skb_len <= EF4_TX_CB_SIZE)) {
307 /* Pad short packets or coalesce short fragmented packets. */
308 if (ef4_enqueue_skb_copy(tx_queue, skb))
309 goto err;
310 tx_queue->cb_packets++;
311 data_mapped = true;
312 }
313
314 /* Map for DMA and create descriptors if we haven't done so already. */
315 if (!data_mapped && (ef4_tx_map_data(tx_queue, skb)))
316 goto err;
317
318 /* Update BQL */
319 netdev_tx_sent_queue(tx_queue->core_txq, skb_len);
320
321 /* Pass off to hardware */
322 if (!netdev_xmit_more() || netif_xmit_stopped(tx_queue->core_txq)) {
323 struct ef4_tx_queue *txq2 = ef4_tx_queue_partner(tx_queue);
324
325 /* There could be packets left on the partner queue if those
326 * SKBs had skb->xmit_more set. If we do not push those they
327 * could be left for a long time and cause a netdev watchdog.
328 */
329 if (txq2->xmit_more_available)
330 ef4_nic_push_buffers(txq2);
331
332 ef4_nic_push_buffers(tx_queue);
333 } else {
334 tx_queue->xmit_more_available = netdev_xmit_more();
335 }
336
337 tx_queue->tx_packets++;
338
339 ef4_tx_maybe_stop_queue(tx_queue);
340
341 return NETDEV_TX_OK;
342
343
344err:
345 ef4_enqueue_unwind(tx_queue);
346 dev_kfree_skb_any(skb);
347 return NETDEV_TX_OK;
348}
349
350/* Remove packets from the TX queue
351 *
352 * This removes packets from the TX queue, up to and including the
353 * specified index.
354 */
355static void ef4_dequeue_buffers(struct ef4_tx_queue *tx_queue,
356 unsigned int index,
357 unsigned int *pkts_compl,
358 unsigned int *bytes_compl)
359{
360 struct ef4_nic *efx = tx_queue->efx;
361 unsigned int stop_index, read_ptr;
362
363 stop_index = (index + 1) & tx_queue->ptr_mask;
364 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
365
366 while (read_ptr != stop_index) {
367 struct ef4_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
368
369 if (!(buffer->flags & EF4_TX_BUF_OPTION) &&
370 unlikely(buffer->len == 0)) {
371 netif_err(efx, tx_err, efx->net_dev,
372 "TX queue %d spurious TX completion id %x\n",
373 tx_queue->queue, read_ptr);
374 ef4_schedule_reset(efx, RESET_TYPE_TX_SKIP);
375 return;
376 }
377
378 ef4_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
379
380 ++tx_queue->read_count;
381 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
382 }
383}
384
385/* Initiate a packet transmission. We use one channel per CPU
386 * (sharing when we have more CPUs than channels). On Falcon, the TX
387 * completion events will be directed back to the CPU that transmitted
388 * the packet, which should be cache-efficient.
389 *
390 * Context: non-blocking.
391 * Note that returning anything other than NETDEV_TX_OK will cause the
392 * OS to free the skb.
393 */
394netdev_tx_t ef4_hard_start_xmit(struct sk_buff *skb,
395 struct net_device *net_dev)
396{
397 struct ef4_nic *efx = netdev_priv(net_dev);
398 struct ef4_tx_queue *tx_queue;
399 unsigned index, type;
400
401 EF4_WARN_ON_PARANOID(!netif_device_present(net_dev));
402
403 index = skb_get_queue_mapping(skb);
404 type = skb->ip_summed == CHECKSUM_PARTIAL ? EF4_TXQ_TYPE_OFFLOAD : 0;
405 if (index >= efx->n_tx_channels) {
406 index -= efx->n_tx_channels;
407 type |= EF4_TXQ_TYPE_HIGHPRI;
408 }
409 tx_queue = ef4_get_tx_queue(efx, index, type);
410
411 return ef4_enqueue_skb(tx_queue, skb);
412}
413
414void ef4_init_tx_queue_core_txq(struct ef4_tx_queue *tx_queue)
415{
416 struct ef4_nic *efx = tx_queue->efx;
417
418 /* Must be inverse of queue lookup in ef4_hard_start_xmit() */
419 tx_queue->core_txq =
420 netdev_get_tx_queue(efx->net_dev,
421 tx_queue->queue / EF4_TXQ_TYPES +
422 ((tx_queue->queue & EF4_TXQ_TYPE_HIGHPRI) ?
423 efx->n_tx_channels : 0));
424}
425
426int ef4_setup_tc(struct net_device *net_dev, enum tc_setup_type type,
427 void *type_data)
428{
429 struct ef4_nic *efx = netdev_priv(net_dev);
430 struct tc_mqprio_qopt *mqprio = type_data;
431 struct ef4_channel *channel;
432 struct ef4_tx_queue *tx_queue;
433 unsigned tc, num_tc;
434 int rc;
435
436 if (type != TC_SETUP_QDISC_MQPRIO)
437 return -EOPNOTSUPP;
438
439 num_tc = mqprio->num_tc;
440
441 if (ef4_nic_rev(efx) < EF4_REV_FALCON_B0 || num_tc > EF4_MAX_TX_TC)
442 return -EINVAL;
443
444 mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
445
446 if (num_tc == net_dev->num_tc)
447 return 0;
448
449 for (tc = 0; tc < num_tc; tc++) {
450 net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
451 net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
452 }
453
454 if (num_tc > net_dev->num_tc) {
455 /* Initialise high-priority queues as necessary */
456 ef4_for_each_channel(channel, efx) {
457 ef4_for_each_possible_channel_tx_queue(tx_queue,
458 channel) {
459 if (!(tx_queue->queue & EF4_TXQ_TYPE_HIGHPRI))
460 continue;
461 if (!tx_queue->buffer) {
462 rc = ef4_probe_tx_queue(tx_queue);
463 if (rc)
464 return rc;
465 }
466 if (!tx_queue->initialised)
467 ef4_init_tx_queue(tx_queue);
468 ef4_init_tx_queue_core_txq(tx_queue);
469 }
470 }
471 } else {
472 /* Reduce number of classes before number of queues */
473 net_dev->num_tc = num_tc;
474 }
475
476 rc = netif_set_real_num_tx_queues(net_dev,
477 max_t(int, num_tc, 1) *
478 efx->n_tx_channels);
479 if (rc)
480 return rc;
481
482 /* Do not destroy high-priority queues when they become
483 * unused. We would have to flush them first, and it is
484 * fairly difficult to flush a subset of TX queues. Leave
485 * it to ef4_fini_channels().
486 */
487
488 net_dev->num_tc = num_tc;
489 return 0;
490}
491
492void ef4_xmit_done(struct ef4_tx_queue *tx_queue, unsigned int index)
493{
494 unsigned fill_level;
495 struct ef4_nic *efx = tx_queue->efx;
496 struct ef4_tx_queue *txq2;
497 unsigned int pkts_compl = 0, bytes_compl = 0;
498
499 EF4_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
500
501 ef4_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
502 tx_queue->pkts_compl += pkts_compl;
503 tx_queue->bytes_compl += bytes_compl;
504
505 if (pkts_compl > 1)
506 ++tx_queue->merge_events;
507
508 /* See if we need to restart the netif queue. This memory
509 * barrier ensures that we write read_count (inside
510 * ef4_dequeue_buffers()) before reading the queue status.
511 */
512 smp_mb();
513 if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
514 likely(efx->port_enabled) &&
515 likely(netif_device_present(efx->net_dev))) {
516 txq2 = ef4_tx_queue_partner(tx_queue);
517 fill_level = max(tx_queue->insert_count - tx_queue->read_count,
518 txq2->insert_count - txq2->read_count);
519 if (fill_level <= efx->txq_wake_thresh)
520 netif_tx_wake_queue(tx_queue->core_txq);
521 }
522
523 /* Check whether the hardware queue is now empty */
524 if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
525 tx_queue->old_write_count = READ_ONCE(tx_queue->write_count);
526 if (tx_queue->read_count == tx_queue->old_write_count) {
527 smp_mb();
528 tx_queue->empty_read_count =
529 tx_queue->read_count | EF4_EMPTY_COUNT_VALID;
530 }
531 }
532}
533
534static unsigned int ef4_tx_cb_page_count(struct ef4_tx_queue *tx_queue)
535{
536 return DIV_ROUND_UP(tx_queue->ptr_mask + 1, PAGE_SIZE >> EF4_TX_CB_ORDER);
537}
538
539int ef4_probe_tx_queue(struct ef4_tx_queue *tx_queue)
540{
541 struct ef4_nic *efx = tx_queue->efx;
542 unsigned int entries;
543 int rc;
544
545 /* Create the smallest power-of-two aligned ring */
546 entries = max(roundup_pow_of_two(efx->txq_entries), EF4_MIN_DMAQ_SIZE);
547 EF4_BUG_ON_PARANOID(entries > EF4_MAX_DMAQ_SIZE);
548 tx_queue->ptr_mask = entries - 1;
549
550 netif_dbg(efx, probe, efx->net_dev,
551 "creating TX queue %d size %#x mask %#x\n",
552 tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
553
554 /* Allocate software ring */
555 tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
556 GFP_KERNEL);
557 if (!tx_queue->buffer)
558 return -ENOMEM;
559
560 tx_queue->cb_page = kcalloc(ef4_tx_cb_page_count(tx_queue),
561 sizeof(tx_queue->cb_page[0]), GFP_KERNEL);
562 if (!tx_queue->cb_page) {
563 rc = -ENOMEM;
564 goto fail1;
565 }
566
567 /* Allocate hardware ring */
568 rc = ef4_nic_probe_tx(tx_queue);
569 if (rc)
570 goto fail2;
571
572 return 0;
573
574fail2:
575 kfree(tx_queue->cb_page);
576 tx_queue->cb_page = NULL;
577fail1:
578 kfree(tx_queue->buffer);
579 tx_queue->buffer = NULL;
580 return rc;
581}
582
583void ef4_init_tx_queue(struct ef4_tx_queue *tx_queue)
584{
585 struct ef4_nic *efx = tx_queue->efx;
586
587 netif_dbg(efx, drv, efx->net_dev,
588 "initialising TX queue %d\n", tx_queue->queue);
589
590 tx_queue->insert_count = 0;
591 tx_queue->write_count = 0;
592 tx_queue->old_write_count = 0;
593 tx_queue->read_count = 0;
594 tx_queue->old_read_count = 0;
595 tx_queue->empty_read_count = 0 | EF4_EMPTY_COUNT_VALID;
596 tx_queue->xmit_more_available = false;
597
598 /* Some older hardware requires Tx writes larger than 32. */
599 tx_queue->tx_min_size = EF4_WORKAROUND_15592(efx) ? 33 : 0;
600
601 /* Set up TX descriptor ring */
602 ef4_nic_init_tx(tx_queue);
603
604 tx_queue->initialised = true;
605}
606
607void ef4_fini_tx_queue(struct ef4_tx_queue *tx_queue)
608{
609 struct ef4_tx_buffer *buffer;
610
611 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
612 "shutting down TX queue %d\n", tx_queue->queue);
613
614 if (!tx_queue->buffer)
615 return;
616
617 /* Free any buffers left in the ring */
618 while (tx_queue->read_count != tx_queue->write_count) {
619 unsigned int pkts_compl = 0, bytes_compl = 0;
620 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
621 ef4_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
622
623 ++tx_queue->read_count;
624 }
625 tx_queue->xmit_more_available = false;
626 netdev_tx_reset_queue(tx_queue->core_txq);
627}
628
629void ef4_remove_tx_queue(struct ef4_tx_queue *tx_queue)
630{
631 int i;
632
633 if (!tx_queue->buffer)
634 return;
635
636 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
637 "destroying TX queue %d\n", tx_queue->queue);
638 ef4_nic_remove_tx(tx_queue);
639
640 if (tx_queue->cb_page) {
641 for (i = 0; i < ef4_tx_cb_page_count(tx_queue); i++)
642 ef4_nic_free_buffer(tx_queue->efx,
643 &tx_queue->cb_page[i]);
644 kfree(tx_queue->cb_page);
645 tx_queue->cb_page = NULL;
646 }
647
648 kfree(tx_queue->buffer);
649 tx_queue->buffer = NULL;
650}
1/****************************************************************************
2 * Driver for Solarflare network controllers and boards
3 * Copyright 2005-2006 Fen Systems Ltd.
4 * Copyright 2005-2013 Solarflare Communications Inc.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published
8 * by the Free Software Foundation, incorporated herein by reference.
9 */
10
11#include <linux/pci.h>
12#include <linux/tcp.h>
13#include <linux/ip.h>
14#include <linux/in.h>
15#include <linux/ipv6.h>
16#include <linux/slab.h>
17#include <net/ipv6.h>
18#include <linux/if_ether.h>
19#include <linux/highmem.h>
20#include <linux/cache.h>
21#include "net_driver.h"
22#include "efx.h"
23#include "io.h"
24#include "nic.h"
25#include "tx.h"
26#include "workarounds.h"
27
28static inline u8 *ef4_tx_get_copy_buffer(struct ef4_tx_queue *tx_queue,
29 struct ef4_tx_buffer *buffer)
30{
31 unsigned int index = ef4_tx_queue_get_insert_index(tx_queue);
32 struct ef4_buffer *page_buf =
33 &tx_queue->cb_page[index >> (PAGE_SHIFT - EF4_TX_CB_ORDER)];
34 unsigned int offset =
35 ((index << EF4_TX_CB_ORDER) + NET_IP_ALIGN) & (PAGE_SIZE - 1);
36
37 if (unlikely(!page_buf->addr) &&
38 ef4_nic_alloc_buffer(tx_queue->efx, page_buf, PAGE_SIZE,
39 GFP_ATOMIC))
40 return NULL;
41 buffer->dma_addr = page_buf->dma_addr + offset;
42 buffer->unmap_len = 0;
43 return (u8 *)page_buf->addr + offset;
44}
45
46u8 *ef4_tx_get_copy_buffer_limited(struct ef4_tx_queue *tx_queue,
47 struct ef4_tx_buffer *buffer, size_t len)
48{
49 if (len > EF4_TX_CB_SIZE)
50 return NULL;
51 return ef4_tx_get_copy_buffer(tx_queue, buffer);
52}
53
54static void ef4_dequeue_buffer(struct ef4_tx_queue *tx_queue,
55 struct ef4_tx_buffer *buffer,
56 unsigned int *pkts_compl,
57 unsigned int *bytes_compl)
58{
59 if (buffer->unmap_len) {
60 struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
61 dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
62 if (buffer->flags & EF4_TX_BUF_MAP_SINGLE)
63 dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
64 DMA_TO_DEVICE);
65 else
66 dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
67 DMA_TO_DEVICE);
68 buffer->unmap_len = 0;
69 }
70
71 if (buffer->flags & EF4_TX_BUF_SKB) {
72 (*pkts_compl)++;
73 (*bytes_compl) += buffer->skb->len;
74 dev_consume_skb_any((struct sk_buff *)buffer->skb);
75 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
76 "TX queue %d transmission id %x complete\n",
77 tx_queue->queue, tx_queue->read_count);
78 }
79
80 buffer->len = 0;
81 buffer->flags = 0;
82}
83
84unsigned int ef4_tx_max_skb_descs(struct ef4_nic *efx)
85{
86 /* This is probably too much since we don't have any TSO support;
87 * it's a left-over from when we had Software TSO. But it's safer
88 * to leave it as-is than try to determine a new bound.
89 */
90 /* Header and payload descriptor for each output segment, plus
91 * one for every input fragment boundary within a segment
92 */
93 unsigned int max_descs = EF4_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
94
95 /* Possibly one more per segment for the alignment workaround,
96 * or for option descriptors
97 */
98 if (EF4_WORKAROUND_5391(efx))
99 max_descs += EF4_TSO_MAX_SEGS;
100
101 /* Possibly more for PCIe page boundaries within input fragments */
102 if (PAGE_SIZE > EF4_PAGE_SIZE)
103 max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
104 DIV_ROUND_UP(GSO_MAX_SIZE, EF4_PAGE_SIZE));
105
106 return max_descs;
107}
108
109static void ef4_tx_maybe_stop_queue(struct ef4_tx_queue *txq1)
110{
111 /* We need to consider both queues that the net core sees as one */
112 struct ef4_tx_queue *txq2 = ef4_tx_queue_partner(txq1);
113 struct ef4_nic *efx = txq1->efx;
114 unsigned int fill_level;
115
116 fill_level = max(txq1->insert_count - txq1->old_read_count,
117 txq2->insert_count - txq2->old_read_count);
118 if (likely(fill_level < efx->txq_stop_thresh))
119 return;
120
121 /* We used the stale old_read_count above, which gives us a
122 * pessimistic estimate of the fill level (which may even
123 * validly be >= efx->txq_entries). Now try again using
124 * read_count (more likely to be a cache miss).
125 *
126 * If we read read_count and then conditionally stop the
127 * queue, it is possible for the completion path to race with
128 * us and complete all outstanding descriptors in the middle,
129 * after which there will be no more completions to wake it.
130 * Therefore we stop the queue first, then read read_count
131 * (with a memory barrier to ensure the ordering), then
132 * restart the queue if the fill level turns out to be low
133 * enough.
134 */
135 netif_tx_stop_queue(txq1->core_txq);
136 smp_mb();
137 txq1->old_read_count = READ_ONCE(txq1->read_count);
138 txq2->old_read_count = READ_ONCE(txq2->read_count);
139
140 fill_level = max(txq1->insert_count - txq1->old_read_count,
141 txq2->insert_count - txq2->old_read_count);
142 EF4_BUG_ON_PARANOID(fill_level >= efx->txq_entries);
143 if (likely(fill_level < efx->txq_stop_thresh)) {
144 smp_mb();
145 if (likely(!efx->loopback_selftest))
146 netif_tx_start_queue(txq1->core_txq);
147 }
148}
149
150static int ef4_enqueue_skb_copy(struct ef4_tx_queue *tx_queue,
151 struct sk_buff *skb)
152{
153 unsigned int min_len = tx_queue->tx_min_size;
154 unsigned int copy_len = skb->len;
155 struct ef4_tx_buffer *buffer;
156 u8 *copy_buffer;
157 int rc;
158
159 EF4_BUG_ON_PARANOID(copy_len > EF4_TX_CB_SIZE);
160
161 buffer = ef4_tx_queue_get_insert_buffer(tx_queue);
162
163 copy_buffer = ef4_tx_get_copy_buffer(tx_queue, buffer);
164 if (unlikely(!copy_buffer))
165 return -ENOMEM;
166
167 rc = skb_copy_bits(skb, 0, copy_buffer, copy_len);
168 EF4_WARN_ON_PARANOID(rc);
169 if (unlikely(copy_len < min_len)) {
170 memset(copy_buffer + copy_len, 0, min_len - copy_len);
171 buffer->len = min_len;
172 } else {
173 buffer->len = copy_len;
174 }
175
176 buffer->skb = skb;
177 buffer->flags = EF4_TX_BUF_SKB;
178
179 ++tx_queue->insert_count;
180 return rc;
181}
182
183static struct ef4_tx_buffer *ef4_tx_map_chunk(struct ef4_tx_queue *tx_queue,
184 dma_addr_t dma_addr,
185 size_t len)
186{
187 const struct ef4_nic_type *nic_type = tx_queue->efx->type;
188 struct ef4_tx_buffer *buffer;
189 unsigned int dma_len;
190
191 /* Map the fragment taking account of NIC-dependent DMA limits. */
192 do {
193 buffer = ef4_tx_queue_get_insert_buffer(tx_queue);
194 dma_len = nic_type->tx_limit_len(tx_queue, dma_addr, len);
195
196 buffer->len = dma_len;
197 buffer->dma_addr = dma_addr;
198 buffer->flags = EF4_TX_BUF_CONT;
199 len -= dma_len;
200 dma_addr += dma_len;
201 ++tx_queue->insert_count;
202 } while (len);
203
204 return buffer;
205}
206
207/* Map all data from an SKB for DMA and create descriptors on the queue.
208 */
209static int ef4_tx_map_data(struct ef4_tx_queue *tx_queue, struct sk_buff *skb)
210{
211 struct ef4_nic *efx = tx_queue->efx;
212 struct device *dma_dev = &efx->pci_dev->dev;
213 unsigned int frag_index, nr_frags;
214 dma_addr_t dma_addr, unmap_addr;
215 unsigned short dma_flags;
216 size_t len, unmap_len;
217
218 nr_frags = skb_shinfo(skb)->nr_frags;
219 frag_index = 0;
220
221 /* Map header data. */
222 len = skb_headlen(skb);
223 dma_addr = dma_map_single(dma_dev, skb->data, len, DMA_TO_DEVICE);
224 dma_flags = EF4_TX_BUF_MAP_SINGLE;
225 unmap_len = len;
226 unmap_addr = dma_addr;
227
228 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
229 return -EIO;
230
231 /* Add descriptors for each fragment. */
232 do {
233 struct ef4_tx_buffer *buffer;
234 skb_frag_t *fragment;
235
236 buffer = ef4_tx_map_chunk(tx_queue, dma_addr, len);
237
238 /* The final descriptor for a fragment is responsible for
239 * unmapping the whole fragment.
240 */
241 buffer->flags = EF4_TX_BUF_CONT | dma_flags;
242 buffer->unmap_len = unmap_len;
243 buffer->dma_offset = buffer->dma_addr - unmap_addr;
244
245 if (frag_index >= nr_frags) {
246 /* Store SKB details with the final buffer for
247 * the completion.
248 */
249 buffer->skb = skb;
250 buffer->flags = EF4_TX_BUF_SKB | dma_flags;
251 return 0;
252 }
253
254 /* Move on to the next fragment. */
255 fragment = &skb_shinfo(skb)->frags[frag_index++];
256 len = skb_frag_size(fragment);
257 dma_addr = skb_frag_dma_map(dma_dev, fragment,
258 0, len, DMA_TO_DEVICE);
259 dma_flags = 0;
260 unmap_len = len;
261 unmap_addr = dma_addr;
262
263 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
264 return -EIO;
265 } while (1);
266}
267
268/* Remove buffers put into a tx_queue. None of the buffers must have
269 * an skb attached.
270 */
271static void ef4_enqueue_unwind(struct ef4_tx_queue *tx_queue)
272{
273 struct ef4_tx_buffer *buffer;
274
275 /* Work backwards until we hit the original insert pointer value */
276 while (tx_queue->insert_count != tx_queue->write_count) {
277 --tx_queue->insert_count;
278 buffer = __ef4_tx_queue_get_insert_buffer(tx_queue);
279 ef4_dequeue_buffer(tx_queue, buffer, NULL, NULL);
280 }
281}
282
283/*
284 * Add a socket buffer to a TX queue
285 *
286 * This maps all fragments of a socket buffer for DMA and adds them to
287 * the TX queue. The queue's insert pointer will be incremented by
288 * the number of fragments in the socket buffer.
289 *
290 * If any DMA mapping fails, any mapped fragments will be unmapped,
291 * the queue's insert pointer will be restored to its original value.
292 *
293 * This function is split out from ef4_hard_start_xmit to allow the
294 * loopback test to direct packets via specific TX queues.
295 *
296 * Returns NETDEV_TX_OK.
297 * You must hold netif_tx_lock() to call this function.
298 */
299netdev_tx_t ef4_enqueue_skb(struct ef4_tx_queue *tx_queue, struct sk_buff *skb)
300{
301 bool data_mapped = false;
302 unsigned int skb_len;
303
304 skb_len = skb->len;
305 EF4_WARN_ON_PARANOID(skb_is_gso(skb));
306
307 if (skb_len < tx_queue->tx_min_size ||
308 (skb->data_len && skb_len <= EF4_TX_CB_SIZE)) {
309 /* Pad short packets or coalesce short fragmented packets. */
310 if (ef4_enqueue_skb_copy(tx_queue, skb))
311 goto err;
312 tx_queue->cb_packets++;
313 data_mapped = true;
314 }
315
316 /* Map for DMA and create descriptors if we haven't done so already. */
317 if (!data_mapped && (ef4_tx_map_data(tx_queue, skb)))
318 goto err;
319
320 /* Update BQL */
321 netdev_tx_sent_queue(tx_queue->core_txq, skb_len);
322
323 /* Pass off to hardware */
324 if (!skb->xmit_more || netif_xmit_stopped(tx_queue->core_txq)) {
325 struct ef4_tx_queue *txq2 = ef4_tx_queue_partner(tx_queue);
326
327 /* There could be packets left on the partner queue if those
328 * SKBs had skb->xmit_more set. If we do not push those they
329 * could be left for a long time and cause a netdev watchdog.
330 */
331 if (txq2->xmit_more_available)
332 ef4_nic_push_buffers(txq2);
333
334 ef4_nic_push_buffers(tx_queue);
335 } else {
336 tx_queue->xmit_more_available = skb->xmit_more;
337 }
338
339 tx_queue->tx_packets++;
340
341 ef4_tx_maybe_stop_queue(tx_queue);
342
343 return NETDEV_TX_OK;
344
345
346err:
347 ef4_enqueue_unwind(tx_queue);
348 dev_kfree_skb_any(skb);
349 return NETDEV_TX_OK;
350}
351
352/* Remove packets from the TX queue
353 *
354 * This removes packets from the TX queue, up to and including the
355 * specified index.
356 */
357static void ef4_dequeue_buffers(struct ef4_tx_queue *tx_queue,
358 unsigned int index,
359 unsigned int *pkts_compl,
360 unsigned int *bytes_compl)
361{
362 struct ef4_nic *efx = tx_queue->efx;
363 unsigned int stop_index, read_ptr;
364
365 stop_index = (index + 1) & tx_queue->ptr_mask;
366 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
367
368 while (read_ptr != stop_index) {
369 struct ef4_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
370
371 if (!(buffer->flags & EF4_TX_BUF_OPTION) &&
372 unlikely(buffer->len == 0)) {
373 netif_err(efx, tx_err, efx->net_dev,
374 "TX queue %d spurious TX completion id %x\n",
375 tx_queue->queue, read_ptr);
376 ef4_schedule_reset(efx, RESET_TYPE_TX_SKIP);
377 return;
378 }
379
380 ef4_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
381
382 ++tx_queue->read_count;
383 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
384 }
385}
386
387/* Initiate a packet transmission. We use one channel per CPU
388 * (sharing when we have more CPUs than channels). On Falcon, the TX
389 * completion events will be directed back to the CPU that transmitted
390 * the packet, which should be cache-efficient.
391 *
392 * Context: non-blocking.
393 * Note that returning anything other than NETDEV_TX_OK will cause the
394 * OS to free the skb.
395 */
396netdev_tx_t ef4_hard_start_xmit(struct sk_buff *skb,
397 struct net_device *net_dev)
398{
399 struct ef4_nic *efx = netdev_priv(net_dev);
400 struct ef4_tx_queue *tx_queue;
401 unsigned index, type;
402
403 EF4_WARN_ON_PARANOID(!netif_device_present(net_dev));
404
405 index = skb_get_queue_mapping(skb);
406 type = skb->ip_summed == CHECKSUM_PARTIAL ? EF4_TXQ_TYPE_OFFLOAD : 0;
407 if (index >= efx->n_tx_channels) {
408 index -= efx->n_tx_channels;
409 type |= EF4_TXQ_TYPE_HIGHPRI;
410 }
411 tx_queue = ef4_get_tx_queue(efx, index, type);
412
413 return ef4_enqueue_skb(tx_queue, skb);
414}
415
416void ef4_init_tx_queue_core_txq(struct ef4_tx_queue *tx_queue)
417{
418 struct ef4_nic *efx = tx_queue->efx;
419
420 /* Must be inverse of queue lookup in ef4_hard_start_xmit() */
421 tx_queue->core_txq =
422 netdev_get_tx_queue(efx->net_dev,
423 tx_queue->queue / EF4_TXQ_TYPES +
424 ((tx_queue->queue & EF4_TXQ_TYPE_HIGHPRI) ?
425 efx->n_tx_channels : 0));
426}
427
428int ef4_setup_tc(struct net_device *net_dev, enum tc_setup_type type,
429 void *type_data)
430{
431 struct ef4_nic *efx = netdev_priv(net_dev);
432 struct tc_mqprio_qopt *mqprio = type_data;
433 struct ef4_channel *channel;
434 struct ef4_tx_queue *tx_queue;
435 unsigned tc, num_tc;
436 int rc;
437
438 if (type != TC_SETUP_QDISC_MQPRIO)
439 return -EOPNOTSUPP;
440
441 num_tc = mqprio->num_tc;
442
443 if (ef4_nic_rev(efx) < EF4_REV_FALCON_B0 || num_tc > EF4_MAX_TX_TC)
444 return -EINVAL;
445
446 mqprio->hw = TC_MQPRIO_HW_OFFLOAD_TCS;
447
448 if (num_tc == net_dev->num_tc)
449 return 0;
450
451 for (tc = 0; tc < num_tc; tc++) {
452 net_dev->tc_to_txq[tc].offset = tc * efx->n_tx_channels;
453 net_dev->tc_to_txq[tc].count = efx->n_tx_channels;
454 }
455
456 if (num_tc > net_dev->num_tc) {
457 /* Initialise high-priority queues as necessary */
458 ef4_for_each_channel(channel, efx) {
459 ef4_for_each_possible_channel_tx_queue(tx_queue,
460 channel) {
461 if (!(tx_queue->queue & EF4_TXQ_TYPE_HIGHPRI))
462 continue;
463 if (!tx_queue->buffer) {
464 rc = ef4_probe_tx_queue(tx_queue);
465 if (rc)
466 return rc;
467 }
468 if (!tx_queue->initialised)
469 ef4_init_tx_queue(tx_queue);
470 ef4_init_tx_queue_core_txq(tx_queue);
471 }
472 }
473 } else {
474 /* Reduce number of classes before number of queues */
475 net_dev->num_tc = num_tc;
476 }
477
478 rc = netif_set_real_num_tx_queues(net_dev,
479 max_t(int, num_tc, 1) *
480 efx->n_tx_channels);
481 if (rc)
482 return rc;
483
484 /* Do not destroy high-priority queues when they become
485 * unused. We would have to flush them first, and it is
486 * fairly difficult to flush a subset of TX queues. Leave
487 * it to ef4_fini_channels().
488 */
489
490 net_dev->num_tc = num_tc;
491 return 0;
492}
493
494void ef4_xmit_done(struct ef4_tx_queue *tx_queue, unsigned int index)
495{
496 unsigned fill_level;
497 struct ef4_nic *efx = tx_queue->efx;
498 struct ef4_tx_queue *txq2;
499 unsigned int pkts_compl = 0, bytes_compl = 0;
500
501 EF4_BUG_ON_PARANOID(index > tx_queue->ptr_mask);
502
503 ef4_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
504 tx_queue->pkts_compl += pkts_compl;
505 tx_queue->bytes_compl += bytes_compl;
506
507 if (pkts_compl > 1)
508 ++tx_queue->merge_events;
509
510 /* See if we need to restart the netif queue. This memory
511 * barrier ensures that we write read_count (inside
512 * ef4_dequeue_buffers()) before reading the queue status.
513 */
514 smp_mb();
515 if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
516 likely(efx->port_enabled) &&
517 likely(netif_device_present(efx->net_dev))) {
518 txq2 = ef4_tx_queue_partner(tx_queue);
519 fill_level = max(tx_queue->insert_count - tx_queue->read_count,
520 txq2->insert_count - txq2->read_count);
521 if (fill_level <= efx->txq_wake_thresh)
522 netif_tx_wake_queue(tx_queue->core_txq);
523 }
524
525 /* Check whether the hardware queue is now empty */
526 if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
527 tx_queue->old_write_count = READ_ONCE(tx_queue->write_count);
528 if (tx_queue->read_count == tx_queue->old_write_count) {
529 smp_mb();
530 tx_queue->empty_read_count =
531 tx_queue->read_count | EF4_EMPTY_COUNT_VALID;
532 }
533 }
534}
535
536static unsigned int ef4_tx_cb_page_count(struct ef4_tx_queue *tx_queue)
537{
538 return DIV_ROUND_UP(tx_queue->ptr_mask + 1, PAGE_SIZE >> EF4_TX_CB_ORDER);
539}
540
541int ef4_probe_tx_queue(struct ef4_tx_queue *tx_queue)
542{
543 struct ef4_nic *efx = tx_queue->efx;
544 unsigned int entries;
545 int rc;
546
547 /* Create the smallest power-of-two aligned ring */
548 entries = max(roundup_pow_of_two(efx->txq_entries), EF4_MIN_DMAQ_SIZE);
549 EF4_BUG_ON_PARANOID(entries > EF4_MAX_DMAQ_SIZE);
550 tx_queue->ptr_mask = entries - 1;
551
552 netif_dbg(efx, probe, efx->net_dev,
553 "creating TX queue %d size %#x mask %#x\n",
554 tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
555
556 /* Allocate software ring */
557 tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
558 GFP_KERNEL);
559 if (!tx_queue->buffer)
560 return -ENOMEM;
561
562 tx_queue->cb_page = kcalloc(ef4_tx_cb_page_count(tx_queue),
563 sizeof(tx_queue->cb_page[0]), GFP_KERNEL);
564 if (!tx_queue->cb_page) {
565 rc = -ENOMEM;
566 goto fail1;
567 }
568
569 /* Allocate hardware ring */
570 rc = ef4_nic_probe_tx(tx_queue);
571 if (rc)
572 goto fail2;
573
574 return 0;
575
576fail2:
577 kfree(tx_queue->cb_page);
578 tx_queue->cb_page = NULL;
579fail1:
580 kfree(tx_queue->buffer);
581 tx_queue->buffer = NULL;
582 return rc;
583}
584
585void ef4_init_tx_queue(struct ef4_tx_queue *tx_queue)
586{
587 struct ef4_nic *efx = tx_queue->efx;
588
589 netif_dbg(efx, drv, efx->net_dev,
590 "initialising TX queue %d\n", tx_queue->queue);
591
592 tx_queue->insert_count = 0;
593 tx_queue->write_count = 0;
594 tx_queue->old_write_count = 0;
595 tx_queue->read_count = 0;
596 tx_queue->old_read_count = 0;
597 tx_queue->empty_read_count = 0 | EF4_EMPTY_COUNT_VALID;
598 tx_queue->xmit_more_available = false;
599
600 /* Some older hardware requires Tx writes larger than 32. */
601 tx_queue->tx_min_size = EF4_WORKAROUND_15592(efx) ? 33 : 0;
602
603 /* Set up TX descriptor ring */
604 ef4_nic_init_tx(tx_queue);
605
606 tx_queue->initialised = true;
607}
608
609void ef4_fini_tx_queue(struct ef4_tx_queue *tx_queue)
610{
611 struct ef4_tx_buffer *buffer;
612
613 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
614 "shutting down TX queue %d\n", tx_queue->queue);
615
616 if (!tx_queue->buffer)
617 return;
618
619 /* Free any buffers left in the ring */
620 while (tx_queue->read_count != tx_queue->write_count) {
621 unsigned int pkts_compl = 0, bytes_compl = 0;
622 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
623 ef4_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
624
625 ++tx_queue->read_count;
626 }
627 tx_queue->xmit_more_available = false;
628 netdev_tx_reset_queue(tx_queue->core_txq);
629}
630
631void ef4_remove_tx_queue(struct ef4_tx_queue *tx_queue)
632{
633 int i;
634
635 if (!tx_queue->buffer)
636 return;
637
638 netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
639 "destroying TX queue %d\n", tx_queue->queue);
640 ef4_nic_remove_tx(tx_queue);
641
642 if (tx_queue->cb_page) {
643 for (i = 0; i < ef4_tx_cb_page_count(tx_queue); i++)
644 ef4_nic_free_buffer(tx_queue->efx,
645 &tx_queue->cb_page[i]);
646 kfree(tx_queue->cb_page);
647 tx_queue->cb_page = NULL;
648 }
649
650 kfree(tx_queue->buffer);
651 tx_queue->buffer = NULL;
652}