1 // SPDX-License-Identifier: GPL-2.0-only 2 /**************************************************************************** 3 * Driver for Solarflare network controllers and boards 4 * Copyright 2018 Solarflare Communications Inc. 5 * Copyright 2019-2020 Xilinx Inc. 6 * 7 * This program is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License version 2 as published 9 * by the Free Software Foundation, incorporated herein by reference. 10 */ 11 12 #include <net/ip6_checksum.h> 13 14 #include "net_driver.h" 15 #include "tx_common.h" 16 #include "nic_common.h" 17 #include "mcdi_functions.h" 18 #include "ef100_regs.h" 19 #include "io.h" 20 #include "ef100_tx.h" 21 #include "ef100_nic.h" 22 23 int ef100_tx_probe(struct efx_tx_queue *tx_queue) 24 { 25 /* Allocate an extra descriptor for the QMDA status completion entry */ 26 return efx_nic_alloc_buffer(tx_queue->efx, &tx_queue->txd.buf, 27 (tx_queue->ptr_mask + 2) * 28 sizeof(efx_oword_t), 29 GFP_KERNEL); 30 } 31 32 void ef100_tx_init(struct efx_tx_queue *tx_queue) 33 { 34 /* must be the inverse of lookup in efx_get_tx_channel */ 35 tx_queue->core_txq = 36 netdev_get_tx_queue(tx_queue->efx->net_dev, 37 tx_queue->channel->channel - 38 tx_queue->efx->tx_channel_offset); 39 40 /* This value is purely documentational; as EF100 never passes through 41 * the switch statement in tx.c:__efx_enqueue_skb(), that switch does 42 * not handle case 3. EF100's TSOv3 descriptors are generated by 43 * ef100_make_tso_desc(). 44 * Meanwhile, all efx_mcdi_tx_init() cares about is that it's not 2. 45 */ 46 tx_queue->tso_version = 3; 47 if (efx_mcdi_tx_init(tx_queue)) 48 netdev_WARN(tx_queue->efx->net_dev, 49 "failed to initialise TXQ %d\n", tx_queue->queue); 50 } 51 52 static bool ef100_tx_can_tso(struct efx_tx_queue *tx_queue, struct sk_buff *skb) 53 { 54 struct efx_nic *efx = tx_queue->efx; 55 struct ef100_nic_data *nic_data; 56 struct efx_tx_buffer *buffer; 57 size_t header_len; 58 u32 mss; 59 60 nic_data = efx->nic_data; 61 62 if (!skb_is_gso_tcp(skb)) 63 return false; 64 if (!(efx->net_dev->features & NETIF_F_TSO)) 65 return false; 66 67 mss = skb_shinfo(skb)->gso_size; 68 if (unlikely(mss < 4)) { 69 WARN_ONCE(1, "MSS of %u is too small for TSO\n", mss); 70 return false; 71 } 72 73 header_len = efx_tx_tso_header_length(skb); 74 if (header_len > nic_data->tso_max_hdr_len) 75 return false; 76 77 if (skb_shinfo(skb)->gso_segs > nic_data->tso_max_payload_num_segs) { 78 /* net_dev->gso_max_segs should've caught this */ 79 WARN_ON_ONCE(1); 80 return false; 81 } 82 83 if (skb->data_len / mss > nic_data->tso_max_frames) 84 return false; 85 86 /* net_dev->gso_max_size should've caught this */ 87 if (WARN_ON_ONCE(skb->data_len > nic_data->tso_max_payload_len)) 88 return false; 89 90 /* Reserve an empty buffer for the TSO V3 descriptor. 91 * Convey the length of the header since we already know it. 92 */ 93 buffer = efx_tx_queue_get_insert_buffer(tx_queue); 94 buffer->flags = EFX_TX_BUF_TSO_V3 | EFX_TX_BUF_CONT; 95 buffer->len = header_len; 96 buffer->unmap_len = 0; 97 buffer->skb = skb; 98 ++tx_queue->insert_count; 99 return true; 100 } 101 102 static efx_oword_t *ef100_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index) 103 { 104 if (likely(tx_queue->txd.buf.addr)) 105 return ((efx_oword_t *)tx_queue->txd.buf.addr) + index; 106 else 107 return NULL; 108 } 109 110 static void ef100_notify_tx_desc(struct efx_tx_queue *tx_queue) 111 { 112 unsigned int write_ptr; 113 efx_dword_t reg; 114 115 tx_queue->xmit_pending = false; 116 117 if (unlikely(tx_queue->notify_count == tx_queue->write_count)) 118 return; 119 120 write_ptr = tx_queue->write_count & tx_queue->ptr_mask; 121 /* The write pointer goes into the high word */ 122 EFX_POPULATE_DWORD_1(reg, ERF_GZ_TX_RING_PIDX, write_ptr); 123 efx_writed_page(tx_queue->efx, ®, 124 ER_GZ_TX_RING_DOORBELL, tx_queue->queue); 125 tx_queue->notify_count = tx_queue->write_count; 126 } 127 128 static void ef100_tx_push_buffers(struct efx_tx_queue *tx_queue) 129 { 130 ef100_notify_tx_desc(tx_queue); 131 ++tx_queue->pushes; 132 } 133 134 static void ef100_set_tx_csum_partial(const struct sk_buff *skb, 135 struct efx_tx_buffer *buffer, efx_oword_t *txd) 136 { 137 efx_oword_t csum; 138 int csum_start; 139 140 if (!skb || skb->ip_summed != CHECKSUM_PARTIAL) 141 return; 142 143 /* skb->csum_start has the offset from head, but we need the offset 144 * from data. 145 */ 146 csum_start = skb_checksum_start_offset(skb); 147 EFX_POPULATE_OWORD_3(csum, 148 ESF_GZ_TX_SEND_CSO_PARTIAL_EN, 1, 149 ESF_GZ_TX_SEND_CSO_PARTIAL_START_W, 150 csum_start >> 1, 151 ESF_GZ_TX_SEND_CSO_PARTIAL_CSUM_W, 152 skb->csum_offset >> 1); 153 EFX_OR_OWORD(*txd, *txd, csum); 154 } 155 156 static void ef100_set_tx_hw_vlan(const struct sk_buff *skb, efx_oword_t *txd) 157 { 158 u16 vlan_tci = skb_vlan_tag_get(skb); 159 efx_oword_t vlan; 160 161 EFX_POPULATE_OWORD_2(vlan, 162 ESF_GZ_TX_SEND_VLAN_INSERT_EN, 1, 163 ESF_GZ_TX_SEND_VLAN_INSERT_TCI, vlan_tci); 164 EFX_OR_OWORD(*txd, *txd, vlan); 165 } 166 167 static void ef100_make_send_desc(struct efx_nic *efx, 168 const struct sk_buff *skb, 169 struct efx_tx_buffer *buffer, efx_oword_t *txd, 170 unsigned int segment_count) 171 { 172 /* TX send descriptor */ 173 EFX_POPULATE_OWORD_3(*txd, 174 ESF_GZ_TX_SEND_NUM_SEGS, segment_count, 175 ESF_GZ_TX_SEND_LEN, buffer->len, 176 ESF_GZ_TX_SEND_ADDR, buffer->dma_addr); 177 178 if (likely(efx->net_dev->features & NETIF_F_HW_CSUM)) 179 ef100_set_tx_csum_partial(skb, buffer, txd); 180 if (efx->net_dev->features & NETIF_F_HW_VLAN_CTAG_TX && 181 skb && skb_vlan_tag_present(skb)) 182 ef100_set_tx_hw_vlan(skb, txd); 183 } 184 185 static void ef100_make_tso_desc(struct efx_nic *efx, 186 const struct sk_buff *skb, 187 struct efx_tx_buffer *buffer, efx_oword_t *txd, 188 unsigned int segment_count) 189 { 190 bool gso_partial = skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL; 191 unsigned int len, ip_offset, tcp_offset, payload_segs; 192 u32 mangleid = ESE_GZ_TX_DESC_IP4_ID_INC_MOD16; 193 unsigned int outer_ip_offset, outer_l4_offset; 194 u16 vlan_tci = skb_vlan_tag_get(skb); 195 u32 mss = skb_shinfo(skb)->gso_size; 196 bool encap = skb->encapsulation; 197 bool udp_encap = false; 198 u16 vlan_enable = 0; 199 struct tcphdr *tcp; 200 bool outer_csum; 201 u32 paylen; 202 203 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_FIXEDID) 204 mangleid = ESE_GZ_TX_DESC_IP4_ID_NO_OP; 205 if (efx->net_dev->features & NETIF_F_HW_VLAN_CTAG_TX) 206 vlan_enable = skb_vlan_tag_present(skb); 207 208 len = skb->len - buffer->len; 209 /* We use 1 for the TSO descriptor and 1 for the header */ 210 payload_segs = segment_count - 2; 211 if (encap) { 212 outer_ip_offset = skb_network_offset(skb); 213 outer_l4_offset = skb_transport_offset(skb); 214 ip_offset = skb_inner_network_offset(skb); 215 tcp_offset = skb_inner_transport_offset(skb); 216 if (skb_shinfo(skb)->gso_type & 217 (SKB_GSO_UDP_TUNNEL | SKB_GSO_UDP_TUNNEL_CSUM)) 218 udp_encap = true; 219 } else { 220 ip_offset = skb_network_offset(skb); 221 tcp_offset = skb_transport_offset(skb); 222 outer_ip_offset = outer_l4_offset = 0; 223 } 224 outer_csum = skb_shinfo(skb)->gso_type & SKB_GSO_UDP_TUNNEL_CSUM; 225 226 /* subtract TCP payload length from inner checksum */ 227 tcp = (void *)skb->data + tcp_offset; 228 paylen = skb->len - tcp_offset; 229 csum_replace_by_diff(&tcp->check, (__force __wsum)htonl(paylen)); 230 231 EFX_POPULATE_OWORD_19(*txd, 232 ESF_GZ_TX_DESC_TYPE, ESE_GZ_TX_DESC_TYPE_TSO, 233 ESF_GZ_TX_TSO_MSS, mss, 234 ESF_GZ_TX_TSO_HDR_NUM_SEGS, 1, 235 ESF_GZ_TX_TSO_PAYLOAD_NUM_SEGS, payload_segs, 236 ESF_GZ_TX_TSO_HDR_LEN_W, buffer->len >> 1, 237 ESF_GZ_TX_TSO_PAYLOAD_LEN, len, 238 ESF_GZ_TX_TSO_CSO_OUTER_L4, outer_csum, 239 ESF_GZ_TX_TSO_CSO_INNER_L4, 1, 240 ESF_GZ_TX_TSO_INNER_L3_OFF_W, ip_offset >> 1, 241 ESF_GZ_TX_TSO_INNER_L4_OFF_W, tcp_offset >> 1, 242 ESF_GZ_TX_TSO_ED_INNER_IP4_ID, mangleid, 243 ESF_GZ_TX_TSO_ED_INNER_IP_LEN, 1, 244 ESF_GZ_TX_TSO_OUTER_L3_OFF_W, outer_ip_offset >> 1, 245 ESF_GZ_TX_TSO_OUTER_L4_OFF_W, outer_l4_offset >> 1, 246 ESF_GZ_TX_TSO_ED_OUTER_UDP_LEN, udp_encap && !gso_partial, 247 ESF_GZ_TX_TSO_ED_OUTER_IP_LEN, encap && !gso_partial, 248 ESF_GZ_TX_TSO_ED_OUTER_IP4_ID, encap ? mangleid : 249 ESE_GZ_TX_DESC_IP4_ID_NO_OP, 250 ESF_GZ_TX_TSO_VLAN_INSERT_EN, vlan_enable, 251 ESF_GZ_TX_TSO_VLAN_INSERT_TCI, vlan_tci 252 ); 253 } 254 255 static void ef100_tx_make_descriptors(struct efx_tx_queue *tx_queue, 256 const struct sk_buff *skb, 257 unsigned int segment_count, 258 struct efx_rep *efv) 259 { 260 unsigned int old_write_count = tx_queue->write_count; 261 unsigned int new_write_count = old_write_count; 262 struct efx_tx_buffer *buffer; 263 unsigned int next_desc_type; 264 unsigned int write_ptr; 265 efx_oword_t *txd; 266 unsigned int nr_descs = tx_queue->insert_count - old_write_count; 267 268 if (unlikely(nr_descs == 0)) 269 return; 270 271 if (segment_count) 272 next_desc_type = ESE_GZ_TX_DESC_TYPE_TSO; 273 else 274 next_desc_type = ESE_GZ_TX_DESC_TYPE_SEND; 275 276 if (unlikely(efv)) { 277 /* Create TX override descriptor */ 278 write_ptr = new_write_count & tx_queue->ptr_mask; 279 txd = ef100_tx_desc(tx_queue, write_ptr); 280 ++new_write_count; 281 282 tx_queue->packet_write_count = new_write_count; 283 EFX_POPULATE_OWORD_3(*txd, 284 ESF_GZ_TX_DESC_TYPE, ESE_GZ_TX_DESC_TYPE_PREFIX, 285 ESF_GZ_TX_PREFIX_EGRESS_MPORT, efv->mport, 286 ESF_GZ_TX_PREFIX_EGRESS_MPORT_EN, 1); 287 nr_descs--; 288 } 289 290 /* if it's a raw write (such as XDP) then always SEND single frames */ 291 if (!skb) 292 nr_descs = 1; 293 294 do { 295 write_ptr = new_write_count & tx_queue->ptr_mask; 296 buffer = &tx_queue->buffer[write_ptr]; 297 txd = ef100_tx_desc(tx_queue, write_ptr); 298 ++new_write_count; 299 300 /* Create TX descriptor ring entry */ 301 tx_queue->packet_write_count = new_write_count; 302 303 switch (next_desc_type) { 304 case ESE_GZ_TX_DESC_TYPE_SEND: 305 ef100_make_send_desc(tx_queue->efx, skb, 306 buffer, txd, nr_descs); 307 break; 308 case ESE_GZ_TX_DESC_TYPE_TSO: 309 /* TX TSO descriptor */ 310 WARN_ON_ONCE(!(buffer->flags & EFX_TX_BUF_TSO_V3)); 311 ef100_make_tso_desc(tx_queue->efx, skb, 312 buffer, txd, nr_descs); 313 break; 314 default: 315 /* TX segment descriptor */ 316 EFX_POPULATE_OWORD_3(*txd, 317 ESF_GZ_TX_DESC_TYPE, ESE_GZ_TX_DESC_TYPE_SEG, 318 ESF_GZ_TX_SEG_LEN, buffer->len, 319 ESF_GZ_TX_SEG_ADDR, buffer->dma_addr); 320 } 321 /* if it's a raw write (such as XDP) then always SEND */ 322 next_desc_type = skb ? ESE_GZ_TX_DESC_TYPE_SEG : 323 ESE_GZ_TX_DESC_TYPE_SEND; 324 /* mark as an EFV buffer if applicable */ 325 if (unlikely(efv)) 326 buffer->flags |= EFX_TX_BUF_EFV; 327 328 } while (new_write_count != tx_queue->insert_count); 329 330 wmb(); /* Ensure descriptors are written before they are fetched */ 331 332 tx_queue->write_count = new_write_count; 333 334 /* The write_count above must be updated before reading 335 * channel->holdoff_doorbell to avoid a race with the 336 * completion path, so ensure these operations are not 337 * re-ordered. This also flushes the update of write_count 338 * back into the cache. 339 */ 340 smp_mb(); 341 } 342 343 void ef100_tx_write(struct efx_tx_queue *tx_queue) 344 { 345 ef100_tx_make_descriptors(tx_queue, NULL, 0, NULL); 346 ef100_tx_push_buffers(tx_queue); 347 } 348 349 void ef100_ev_tx(struct efx_channel *channel, const efx_qword_t *p_event) 350 { 351 unsigned int tx_done = 352 EFX_QWORD_FIELD(*p_event, ESF_GZ_EV_TXCMPL_NUM_DESC); 353 unsigned int qlabel = 354 EFX_QWORD_FIELD(*p_event, ESF_GZ_EV_TXCMPL_Q_LABEL); 355 struct efx_tx_queue *tx_queue = 356 efx_channel_get_tx_queue(channel, qlabel); 357 unsigned int tx_index = (tx_queue->read_count + tx_done - 1) & 358 tx_queue->ptr_mask; 359 360 efx_xmit_done(tx_queue, tx_index); 361 } 362 363 /* Add a socket buffer to a TX queue 364 * 365 * You must hold netif_tx_lock() to call this function. 366 * 367 * Returns 0 on success, error code otherwise. In case of an error this 368 * function will free the SKB. 369 */ 370 int ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb) 371 { 372 return __ef100_enqueue_skb(tx_queue, skb, NULL); 373 } 374 375 int __ef100_enqueue_skb(struct efx_tx_queue *tx_queue, struct sk_buff *skb, 376 struct efx_rep *efv) 377 { 378 unsigned int old_insert_count = tx_queue->insert_count; 379 struct efx_nic *efx = tx_queue->efx; 380 bool xmit_more = netdev_xmit_more(); 381 unsigned int fill_level; 382 unsigned int segments; 383 int rc; 384 385 if (!tx_queue->buffer || !tx_queue->ptr_mask) { 386 netif_stop_queue(efx->net_dev); 387 dev_kfree_skb_any(skb); 388 return -ENODEV; 389 } 390 391 segments = skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 0; 392 if (segments == 1) 393 segments = 0; /* Don't use TSO/GSO for a single segment. */ 394 if (segments && !ef100_tx_can_tso(tx_queue, skb)) { 395 rc = efx_tx_tso_fallback(tx_queue, skb); 396 tx_queue->tso_fallbacks++; 397 if (rc) 398 goto err; 399 else 400 return 0; 401 } 402 403 if (unlikely(efv)) { 404 struct efx_tx_buffer *buffer = __efx_tx_queue_get_insert_buffer(tx_queue); 405 406 /* Drop representor packets if the queue is stopped. 407 * We currently don't assert backoff to representors so this is 408 * to make sure representor traffic can't starve the main 409 * net device. 410 * And, of course, if there are no TX descriptors left. 411 */ 412 if (netif_tx_queue_stopped(tx_queue->core_txq) || 413 unlikely(efx_tx_buffer_in_use(buffer))) { 414 atomic64_inc(&efv->stats.tx_errors); 415 rc = -ENOSPC; 416 goto err; 417 } 418 419 /* Also drop representor traffic if it could cause us to 420 * stop the queue. If we assert backoff and we haven't 421 * received traffic on the main net device recently then the 422 * TX watchdog can go off erroneously. 423 */ 424 fill_level = efx_channel_tx_old_fill_level(tx_queue->channel); 425 fill_level += efx_tx_max_skb_descs(efx); 426 if (fill_level > efx->txq_stop_thresh) { 427 struct efx_tx_queue *txq2; 428 429 /* Refresh cached fill level and re-check */ 430 efx_for_each_channel_tx_queue(txq2, tx_queue->channel) 431 txq2->old_read_count = READ_ONCE(txq2->read_count); 432 433 fill_level = efx_channel_tx_old_fill_level(tx_queue->channel); 434 fill_level += efx_tx_max_skb_descs(efx); 435 if (fill_level > efx->txq_stop_thresh) { 436 atomic64_inc(&efv->stats.tx_errors); 437 rc = -ENOSPC; 438 goto err; 439 } 440 } 441 442 buffer->flags = EFX_TX_BUF_OPTION | EFX_TX_BUF_EFV; 443 tx_queue->insert_count++; 444 } 445 446 /* Map for DMA and create descriptors */ 447 rc = efx_tx_map_data(tx_queue, skb, segments); 448 if (rc) 449 goto err; 450 ef100_tx_make_descriptors(tx_queue, skb, segments, efv); 451 452 fill_level = efx_channel_tx_old_fill_level(tx_queue->channel); 453 if (fill_level > efx->txq_stop_thresh) { 454 struct efx_tx_queue *txq2; 455 456 /* Because of checks above, representor traffic should 457 * not be able to stop the queue. 458 */ 459 WARN_ON(efv); 460 461 netif_tx_stop_queue(tx_queue->core_txq); 462 /* Re-read after a memory barrier in case we've raced with 463 * the completion path. Otherwise there's a danger we'll never 464 * restart the queue if all completions have just happened. 465 */ 466 smp_mb(); 467 efx_for_each_channel_tx_queue(txq2, tx_queue->channel) 468 txq2->old_read_count = READ_ONCE(txq2->read_count); 469 fill_level = efx_channel_tx_old_fill_level(tx_queue->channel); 470 if (fill_level < efx->txq_stop_thresh) 471 netif_tx_start_queue(tx_queue->core_txq); 472 } 473 474 tx_queue->xmit_pending = true; 475 476 /* If xmit_more then we don't need to push the doorbell, unless there 477 * are 256 descriptors already queued in which case we have to push to 478 * ensure we never push more than 256 at once. 479 * 480 * Always push for representor traffic, and don't account it to parent 481 * PF netdevice's BQL. 482 */ 483 if (unlikely(efv) || 484 __netdev_tx_sent_queue(tx_queue->core_txq, skb->len, xmit_more) || 485 tx_queue->write_count - tx_queue->notify_count > 255) 486 ef100_tx_push_buffers(tx_queue); 487 488 if (segments) { 489 tx_queue->tso_bursts++; 490 tx_queue->tso_packets += segments; 491 tx_queue->tx_packets += segments; 492 } else { 493 tx_queue->tx_packets++; 494 } 495 return 0; 496 497 err: 498 efx_enqueue_unwind(tx_queue, old_insert_count); 499 if (!IS_ERR_OR_NULL(skb)) 500 dev_kfree_skb_any(skb); 501 502 /* If we're not expecting another transmit and we had something to push 503 * on this queue then we need to push here to get the previous packets 504 * out. We only enter this branch from before the xmit_more handling 505 * above, so xmit_pending still refers to the old state. 506 */ 507 if (tx_queue->xmit_pending && !xmit_more) 508 ef100_tx_push_buffers(tx_queue); 509 return rc; 510 } 511