drivers/net/ethernet/sfc/tx_common.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /****************************************************************************
   3  * Driver for Solarflare network controllers and boards
   4  * Copyright 2018 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 "net_driver.h"
  12 #include "efx.h"
  13 #include "nic_common.h"
  14 #include "tx_common.h"
  15 #include <net/gso.h>
  16
  17 static unsigned int efx_tx_cb_page_count(struct efx_tx_queue *tx_queue)
  18 {
  19         return DIV_ROUND_UP(tx_queue->ptr_mask + 1,
  20                             PAGE_SIZE >> EFX_TX_CB_ORDER);
  21 }
  22
  23 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
  24 {
  25         struct efx_nic *efx = tx_queue->efx;
  26         unsigned int entries;
  27         int rc;
  28
  29         /* Create the smallest power-of-two aligned ring */
  30         entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
  31         EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
  32         tx_queue->ptr_mask = entries - 1;
  33
  34         netif_dbg(efx, probe, efx->net_dev,
  35                   "creating TX queue %d size %#x mask %#x\n",
  36                   tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
  37
  38         /* Allocate software ring */
  39         tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
  40                                    GFP_KERNEL);
  41         if (!tx_queue->buffer)
  42                 return -ENOMEM;
  43
  44         tx_queue->cb_page = kcalloc(efx_tx_cb_page_count(tx_queue),
  45                                     sizeof(tx_queue->cb_page[0]), GFP_KERNEL);
  46         if (!tx_queue->cb_page) {
  47                 rc = -ENOMEM;
  48                 goto fail1;
  49         }
  50
  51         /* Allocate hardware ring, determine TXQ type */
  52         rc = efx_nic_probe_tx(tx_queue);
  53         if (rc)
  54                 goto fail2;
  55
  56         tx_queue->channel->tx_queue_by_type[tx_queue->type] = tx_queue;
  57         return 0;
  58
  59 fail2:
  60         kfree(tx_queue->cb_page);
  61         tx_queue->cb_page = NULL;
  62 fail1:
  63         kfree(tx_queue->buffer);
  64         tx_queue->buffer = NULL;
  65         return rc;
  66 }
  67
  68 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
  69 {
  70         struct efx_nic *efx = tx_queue->efx;
  71
  72         netif_dbg(efx, drv, efx->net_dev,
  73                   "initialising TX queue %d\n", tx_queue->queue);
  74
  75         tx_queue->insert_count = 0;
  76         tx_queue->notify_count = 0;
  77         tx_queue->write_count = 0;
  78         tx_queue->packet_write_count = 0;
  79         tx_queue->old_write_count = 0;
  80         tx_queue->read_count = 0;
  81         tx_queue->old_read_count = 0;
  82         tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
  83         tx_queue->xmit_pending = false;
  84         tx_queue->timestamping = (efx_ptp_use_mac_tx_timestamps(efx) &&
  85                                   tx_queue->channel == efx_ptp_channel(efx));
  86         tx_queue->completed_timestamp_major = 0;
  87         tx_queue->completed_timestamp_minor = 0;
  88
  89         tx_queue->old_complete_packets = tx_queue->complete_packets;
  90         tx_queue->old_complete_bytes = tx_queue->complete_bytes;
  91         tx_queue->old_tso_bursts = tx_queue->tso_bursts;
  92         tx_queue->old_tso_packets = tx_queue->tso_packets;
  93
  94         tx_queue->xdp_tx = efx_channel_is_xdp_tx(tx_queue->channel);
  95         tx_queue->tso_version = 0;
  96
  97         /* Set up TX descriptor ring */
  98         efx_nic_init_tx(tx_queue);
  99
 100         tx_queue->initialised = true;
 101 }
 102
 103 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 104 {
 105         struct efx_tx_buffer *buffer;
 106
 107         netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 108                   "shutting down TX queue %d\n", tx_queue->queue);
 109
 110         tx_queue->initialised = false;
 111
 112         if (!tx_queue->buffer)
 113                 return;
 114
 115         /* Free any buffers left in the ring */
 116         while (tx_queue->read_count != tx_queue->write_count) {
 117                 unsigned int xdp_pkts_compl = 0, xdp_bytes_compl = 0;
 118                 unsigned int pkts_compl = 0, bytes_compl = 0;
 119                 unsigned int efv_pkts_compl = 0;
 120
 121                 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
 122                 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
 123                                    &efv_pkts_compl, &xdp_pkts_compl,
 124                                    &xdp_bytes_compl);
 125
 126                 ++tx_queue->read_count;
 127         }
 128         tx_queue->xmit_pending = false;
 129         netdev_tx_reset_queue(tx_queue->core_txq);
 130 }
 131
 132 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 133 {
 134         int i;
 135
 136         if (!tx_queue->buffer)
 137                 return;
 138
 139         netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 140                   "destroying TX queue %d\n", tx_queue->queue);
 141         efx_nic_remove_tx(tx_queue);
 142
 143         if (tx_queue->cb_page) {
 144                 for (i = 0; i < efx_tx_cb_page_count(tx_queue); i++)
 145                         efx_nic_free_buffer(tx_queue->efx,
 146                                             &tx_queue->cb_page[i]);
 147                 kfree(tx_queue->cb_page);
 148                 tx_queue->cb_page = NULL;
 149         }
 150
 151         kfree(tx_queue->buffer);
 152         tx_queue->buffer = NULL;
 153         tx_queue->channel->tx_queue_by_type[tx_queue->type] = NULL;
 154 }
 155
 156 void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 157                         struct efx_tx_buffer *buffer,
 158                         unsigned int *pkts_compl,
 159                         unsigned int *bytes_compl,
 160                         unsigned int *efv_pkts_compl,
 161                         unsigned int *xdp_pkts,
 162                         unsigned int *xdp_bytes)
 163 {
 164         if (buffer->unmap_len) {
 165                 struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
 166                 dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
 167
 168                 if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
 169                         dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
 170                                          DMA_TO_DEVICE);
 171                 else
 172                         dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
 173                                        DMA_TO_DEVICE);
 174                 buffer->unmap_len = 0;
 175         }
 176
 177         if (buffer->flags & EFX_TX_BUF_SKB) {
 178                 struct sk_buff *skb = (struct sk_buff *)buffer->skb;
 179
 180                 if (unlikely(buffer->flags & EFX_TX_BUF_EFV)) {
 181                         EFX_WARN_ON_PARANOID(!efv_pkts_compl);
 182                         (*efv_pkts_compl)++;
 183                 } else {
 184                         EFX_WARN_ON_PARANOID(!pkts_compl || !bytes_compl);
 185                         (*pkts_compl)++;
 186                         (*bytes_compl) += skb->len;
 187                 }
 188
 189                 if (tx_queue->timestamping &&
 190                     (tx_queue->completed_timestamp_major ||
 191                      tx_queue->completed_timestamp_minor)) {
 192                         struct skb_shared_hwtstamps hwtstamp;
 193
 194                         hwtstamp.hwtstamp =
 195                                 efx_ptp_nic_to_kernel_time(tx_queue);
 196                         skb_tstamp_tx(skb, &hwtstamp);
 197
 198                         tx_queue->completed_timestamp_major = 0;
 199                         tx_queue->completed_timestamp_minor = 0;
 200                 }
 201                 dev_consume_skb_any((struct sk_buff *)buffer->skb);
 202                 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
 203                            "TX queue %d transmission id %x complete\n",
 204                            tx_queue->queue, tx_queue->read_count);
 205         } else if (buffer->flags & EFX_TX_BUF_XDP) {
 206                 xdp_return_frame_rx_napi(buffer->xdpf);
 207                 if (xdp_pkts)
 208                         (*xdp_pkts)++;
 209                 if (xdp_bytes)
 210                         (*xdp_bytes) += buffer->xdpf->len;
 211         }
 212
 213         buffer->len = 0;
 214         buffer->flags = 0;
 215 }
 216
 217 /* Remove packets from the TX queue
 218  *
 219  * This removes packets from the TX queue, up to and including the
 220  * specified index.
 221  */
 222 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 223                                 unsigned int index,
 224                                 unsigned int *pkts_compl,
 225                                 unsigned int *bytes_compl,
 226                                 unsigned int *efv_pkts_compl,
 227                                 unsigned int *xdp_pkts,
 228                                 unsigned int *xdp_bytes)
 229 {
 230         struct efx_nic *efx = tx_queue->efx;
 231         unsigned int stop_index, read_ptr;
 232
 233         stop_index = (index + 1) & tx_queue->ptr_mask;
 234         read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 235
 236         while (read_ptr != stop_index) {
 237                 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
 238
 239                 if (!efx_tx_buffer_in_use(buffer)) {
 240                         netif_err(efx, tx_err, efx->net_dev,
 241                                   "TX queue %d spurious TX completion id %d\n",
 242                                   tx_queue->queue, read_ptr);
 243                         efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
 244                         return;
 245                 }
 246
 247                 efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl,
 248                                    efv_pkts_compl, xdp_pkts, xdp_bytes);
 249
 250                 ++tx_queue->read_count;
 251                 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 252         }
 253 }
 254
 255 void efx_xmit_done_check_empty(struct efx_tx_queue *tx_queue)
 256 {
 257         if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
 258                 tx_queue->old_write_count = READ_ONCE(tx_queue->write_count);
 259                 if (tx_queue->read_count == tx_queue->old_write_count) {
 260                         /* Ensure that read_count is flushed. */
 261                         smp_mb();
 262                         tx_queue->empty_read_count =
 263                                 tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
 264                 }
 265         }
 266 }
 267
 268 int efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 269 {
 270         unsigned int fill_level, pkts_compl = 0, bytes_compl = 0;
 271         unsigned int xdp_pkts_compl = 0, xdp_bytes_compl = 0;
 272         unsigned int efv_pkts_compl = 0;
 273         struct efx_nic *efx = tx_queue->efx;
 274
 275         EFX_WARN_ON_ONCE_PARANOID(index > tx_queue->ptr_mask);
 276
 277         efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl,
 278                             &efv_pkts_compl, &xdp_pkts_compl, &xdp_bytes_compl);
 279         tx_queue->pkts_compl += pkts_compl;
 280         tx_queue->bytes_compl += bytes_compl;
 281         tx_queue->complete_xdp_packets += xdp_pkts_compl;
 282         tx_queue->complete_xdp_bytes += xdp_bytes_compl;
 283
 284         if (pkts_compl + efv_pkts_compl > 1)
 285                 ++tx_queue->merge_events;
 286
 287         /* See if we need to restart the netif queue.  This memory
 288          * barrier ensures that we write read_count (inside
 289          * efx_dequeue_buffers()) before reading the queue status.
 290          */
 291         smp_mb();
 292         if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
 293             likely(efx->port_enabled) &&
 294             likely(netif_device_present(efx->net_dev))) {
 295                 fill_level = efx_channel_tx_fill_level(tx_queue->channel);
 296                 if (fill_level <= efx->txq_wake_thresh)
 297                         netif_tx_wake_queue(tx_queue->core_txq);
 298         }
 299
 300         efx_xmit_done_check_empty(tx_queue);
 301
 302         return pkts_compl + efv_pkts_compl;
 303 }
 304
 305 /* Remove buffers put into a tx_queue for the current packet.
 306  * None of the buffers must have an skb attached.
 307  */
 308 void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
 309                         unsigned int insert_count)
 310 {
 311         unsigned int xdp_bytes_compl = 0;
 312         unsigned int xdp_pkts_compl = 0;
 313         unsigned int efv_pkts_compl = 0;
 314         struct efx_tx_buffer *buffer;
 315         unsigned int bytes_compl = 0;
 316         unsigned int pkts_compl = 0;
 317
 318         /* Work backwards until we hit the original insert pointer value */
 319         while (tx_queue->insert_count != insert_count) {
 320                 --tx_queue->insert_count;
 321                 buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
 322                 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl,
 323                                    &efv_pkts_compl, &xdp_pkts_compl,
 324                                    &xdp_bytes_compl);
 325         }
 326 }
 327
 328 struct efx_tx_buffer *efx_tx_map_chunk(struct efx_tx_queue *tx_queue,
 329                                        dma_addr_t dma_addr, size_t len)
 330 {
 331         const struct efx_nic_type *nic_type = tx_queue->efx->type;
 332         struct efx_tx_buffer *buffer;
 333         unsigned int dma_len;
 334
 335         /* Map the fragment taking account of NIC-dependent DMA limits. */
 336         do {
 337                 buffer = efx_tx_queue_get_insert_buffer(tx_queue);
 338
 339                 if (nic_type->tx_limit_len)
 340                         dma_len = nic_type->tx_limit_len(tx_queue, dma_addr, len);
 341                 else
 342                         dma_len = len;
 343
 344                 buffer->len = dma_len;
 345                 buffer->dma_addr = dma_addr;
 346                 buffer->flags = EFX_TX_BUF_CONT;
 347                 len -= dma_len;
 348                 dma_addr += dma_len;
 349                 ++tx_queue->insert_count;
 350         } while (len);
 351
 352         return buffer;
 353 }
 354
 355 int efx_tx_tso_header_length(struct sk_buff *skb)
 356 {
 357         size_t header_len;
 358
 359         if (skb->encapsulation)
 360                 header_len = skb_inner_transport_offset(skb) +
 361                                 (inner_tcp_hdr(skb)->doff << 2u);
 362         else
 363                 header_len = skb_transport_offset(skb) +
 364                                 (tcp_hdr(skb)->doff << 2u);
 365         return header_len;
 366 }
 367
 368 /* Map all data from an SKB for DMA and create descriptors on the queue. */
 369 int efx_tx_map_data(struct efx_tx_queue *tx_queue, struct sk_buff *skb,
 370                     unsigned int segment_count)
 371 {
 372         struct efx_nic *efx = tx_queue->efx;
 373         struct device *dma_dev = &efx->pci_dev->dev;
 374         unsigned int frag_index, nr_frags;
 375         dma_addr_t dma_addr, unmap_addr;
 376         unsigned short dma_flags;
 377         size_t len, unmap_len;
 378
 379         nr_frags = skb_shinfo(skb)->nr_frags;
 380         frag_index = 0;
 381
 382         /* Map header data. */
 383         len = skb_headlen(skb);
 384         dma_addr = dma_map_single(dma_dev, skb->data, len, DMA_TO_DEVICE);
 385         dma_flags = EFX_TX_BUF_MAP_SINGLE;
 386         unmap_len = len;
 387         unmap_addr = dma_addr;
 388
 389         if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
 390                 return -EIO;
 391
 392         if (segment_count) {
 393                 /* For TSO we need to put the header in to a separate
 394                  * descriptor. Map this separately if necessary.
 395                  */
 396                 size_t header_len = efx_tx_tso_header_length(skb);
 397
 398                 if (header_len != len) {
 399                         tx_queue->tso_long_headers++;
 400                         efx_tx_map_chunk(tx_queue, dma_addr, header_len);
 401                         len -= header_len;
 402                         dma_addr += header_len;
 403                 }
 404         }
 405
 406         /* Add descriptors for each fragment. */
 407         do {
 408                 struct efx_tx_buffer *buffer;
 409                 skb_frag_t *fragment;
 410
 411                 buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
 412
 413                 /* The final descriptor for a fragment is responsible for
 414                  * unmapping the whole fragment.
 415                  */
 416                 buffer->flags = EFX_TX_BUF_CONT | dma_flags;
 417                 buffer->unmap_len = unmap_len;
 418                 buffer->dma_offset = buffer->dma_addr - unmap_addr;
 419
 420                 if (frag_index >= nr_frags) {
 421                         /* Store SKB details with the final buffer for
 422                          * the completion.
 423                          */
 424                         buffer->skb = skb;
 425                         buffer->flags = EFX_TX_BUF_SKB | dma_flags;
 426                         return 0;
 427                 }
 428
 429                 /* Move on to the next fragment. */
 430                 fragment = &skb_shinfo(skb)->frags[frag_index++];
 431                 len = skb_frag_size(fragment);
 432                 dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
 433                                             DMA_TO_DEVICE);
 434                 dma_flags = 0;
 435                 unmap_len = len;
 436                 unmap_addr = dma_addr;
 437
 438                 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
 439                         return -EIO;
 440         } while (1);
 441 }
 442
 443 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
 444 {
 445         /* Header and payload descriptor for each output segment, plus
 446          * one for every input fragment boundary within a segment
 447          */
 448         unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
 449
 450         /* Possibly one more per segment for option descriptors */
 451         if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
 452                 max_descs += EFX_TSO_MAX_SEGS;
 453
 454         /* Possibly more for PCIe page boundaries within input fragments */
 455         if (PAGE_SIZE > EFX_PAGE_SIZE)
 456                 max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
 457                                    DIV_ROUND_UP(GSO_LEGACY_MAX_SIZE,
 458                                                 EFX_PAGE_SIZE));
 459
 460         return max_descs;
 461 }
 462
 463 /*
 464  * Fallback to software TSO.
 465  *
 466  * This is used if we are unable to send a GSO packet through hardware TSO.
 467  * This should only ever happen due to per-queue restrictions - unsupported
 468  * packets should first be filtered by the feature flags.
 469  *
 470  * Returns 0 on success, error code otherwise.
 471  */
 472 int efx_tx_tso_fallback(struct efx_tx_queue *tx_queue, struct sk_buff *skb)
 473 {
 474         struct sk_buff *segments, *next;
 475
 476         segments = skb_gso_segment(skb, 0);
 477         if (IS_ERR(segments))
 478                 return PTR_ERR(segments);
 479
 480         dev_consume_skb_any(skb);
 481
 482         skb_list_walk_safe(segments, skb, next) {
 483                 skb_mark_not_on_list(skb);
 484                 efx_enqueue_skb(tx_queue, skb);
 485         }
 486
 487         return 0;
 488 }