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.h"
  14 #include "tx_common.h"
  15
  16 static unsigned int efx_tx_cb_page_count(struct efx_tx_queue *tx_queue)
  17 {
  18         return DIV_ROUND_UP(tx_queue->ptr_mask + 1,
  19                             PAGE_SIZE >> EFX_TX_CB_ORDER);
  20 }
  21
  22 int efx_probe_tx_queue(struct efx_tx_queue *tx_queue)
  23 {
  24         struct efx_nic *efx = tx_queue->efx;
  25         unsigned int entries;
  26         int rc;
  27
  28         /* Create the smallest power-of-two aligned ring */
  29         entries = max(roundup_pow_of_two(efx->txq_entries), EFX_MIN_DMAQ_SIZE);
  30         EFX_WARN_ON_PARANOID(entries > EFX_MAX_DMAQ_SIZE);
  31         tx_queue->ptr_mask = entries - 1;
  32
  33         netif_dbg(efx, probe, efx->net_dev,
  34                   "creating TX queue %d size %#x mask %#x\n",
  35                   tx_queue->queue, efx->txq_entries, tx_queue->ptr_mask);
  36
  37         /* Allocate software ring */
  38         tx_queue->buffer = kcalloc(entries, sizeof(*tx_queue->buffer),
  39                                    GFP_KERNEL);
  40         if (!tx_queue->buffer)
  41                 return -ENOMEM;
  42
  43         tx_queue->cb_page = kcalloc(efx_tx_cb_page_count(tx_queue),
  44                                     sizeof(tx_queue->cb_page[0]), GFP_KERNEL);
  45         if (!tx_queue->cb_page) {
  46                 rc = -ENOMEM;
  47                 goto fail1;
  48         }
  49
  50         /* Allocate hardware ring */
  51         rc = efx_nic_probe_tx(tx_queue);
  52         if (rc)
  53                 goto fail2;
  54
  55         return 0;
  56
  57 fail2:
  58         kfree(tx_queue->cb_page);
  59         tx_queue->cb_page = NULL;
  60 fail1:
  61         kfree(tx_queue->buffer);
  62         tx_queue->buffer = NULL;
  63         return rc;
  64 }
  65
  66 void efx_init_tx_queue(struct efx_tx_queue *tx_queue)
  67 {
  68         struct efx_nic *efx = tx_queue->efx;
  69
  70         netif_dbg(efx, drv, efx->net_dev,
  71                   "initialising TX queue %d\n", tx_queue->queue);
  72
  73         tx_queue->insert_count = 0;
  74         tx_queue->write_count = 0;
  75         tx_queue->packet_write_count = 0;
  76         tx_queue->old_write_count = 0;
  77         tx_queue->read_count = 0;
  78         tx_queue->old_read_count = 0;
  79         tx_queue->empty_read_count = 0 | EFX_EMPTY_COUNT_VALID;
  80         tx_queue->xmit_more_available = false;
  81         tx_queue->timestamping = (efx_ptp_use_mac_tx_timestamps(efx) &&
  82                                   tx_queue->channel == efx_ptp_channel(efx));
  83         tx_queue->completed_timestamp_major = 0;
  84         tx_queue->completed_timestamp_minor = 0;
  85
  86         tx_queue->xdp_tx = efx_channel_is_xdp_tx(tx_queue->channel);
  87
  88         /* Set up default function pointers. These may get replaced by
  89          * efx_nic_init_tx() based off NIC/queue capabilities.
  90          */
  91         tx_queue->handle_tso = efx_enqueue_skb_tso;
  92
  93         /* Set up TX descriptor ring */
  94         efx_nic_init_tx(tx_queue);
  95
  96         tx_queue->initialised = true;
  97 }
  98
  99 void efx_fini_tx_queue(struct efx_tx_queue *tx_queue)
 100 {
 101         struct efx_tx_buffer *buffer;
 102
 103         netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 104                   "shutting down TX queue %d\n", tx_queue->queue);
 105
 106         if (!tx_queue->buffer)
 107                 return;
 108
 109         /* Free any buffers left in the ring */
 110         while (tx_queue->read_count != tx_queue->write_count) {
 111                 unsigned int pkts_compl = 0, bytes_compl = 0;
 112
 113                 buffer = &tx_queue->buffer[tx_queue->read_count & tx_queue->ptr_mask];
 114                 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
 115
 116                 ++tx_queue->read_count;
 117         }
 118         tx_queue->xmit_more_available = false;
 119         netdev_tx_reset_queue(tx_queue->core_txq);
 120 }
 121
 122 void efx_remove_tx_queue(struct efx_tx_queue *tx_queue)
 123 {
 124         int i;
 125
 126         if (!tx_queue->buffer)
 127                 return;
 128
 129         netif_dbg(tx_queue->efx, drv, tx_queue->efx->net_dev,
 130                   "destroying TX queue %d\n", tx_queue->queue);
 131         efx_nic_remove_tx(tx_queue);
 132
 133         if (tx_queue->cb_page) {
 134                 for (i = 0; i < efx_tx_cb_page_count(tx_queue); i++)
 135                         efx_nic_free_buffer(tx_queue->efx,
 136                                             &tx_queue->cb_page[i]);
 137                 kfree(tx_queue->cb_page);
 138                 tx_queue->cb_page = NULL;
 139         }
 140
 141         kfree(tx_queue->buffer);
 142         tx_queue->buffer = NULL;
 143 }
 144
 145 void efx_dequeue_buffer(struct efx_tx_queue *tx_queue,
 146                         struct efx_tx_buffer *buffer,
 147                         unsigned int *pkts_compl,
 148                         unsigned int *bytes_compl)
 149 {
 150         if (buffer->unmap_len) {
 151                 struct device *dma_dev = &tx_queue->efx->pci_dev->dev;
 152                 dma_addr_t unmap_addr = buffer->dma_addr - buffer->dma_offset;
 153
 154                 if (buffer->flags & EFX_TX_BUF_MAP_SINGLE)
 155                         dma_unmap_single(dma_dev, unmap_addr, buffer->unmap_len,
 156                                          DMA_TO_DEVICE);
 157                 else
 158                         dma_unmap_page(dma_dev, unmap_addr, buffer->unmap_len,
 159                                        DMA_TO_DEVICE);
 160                 buffer->unmap_len = 0;
 161         }
 162
 163         if (buffer->flags & EFX_TX_BUF_SKB) {
 164                 struct sk_buff *skb = (struct sk_buff *)buffer->skb;
 165
 166                 EFX_WARN_ON_PARANOID(!pkts_compl || !bytes_compl);
 167                 (*pkts_compl)++;
 168                 (*bytes_compl) += skb->len;
 169                 if (tx_queue->timestamping &&
 170                     (tx_queue->completed_timestamp_major ||
 171                      tx_queue->completed_timestamp_minor)) {
 172                         struct skb_shared_hwtstamps hwtstamp;
 173
 174                         hwtstamp.hwtstamp =
 175                                 efx_ptp_nic_to_kernel_time(tx_queue);
 176                         skb_tstamp_tx(skb, &hwtstamp);
 177
 178                         tx_queue->completed_timestamp_major = 0;
 179                         tx_queue->completed_timestamp_minor = 0;
 180                 }
 181                 dev_consume_skb_any((struct sk_buff *)buffer->skb);
 182                 netif_vdbg(tx_queue->efx, tx_done, tx_queue->efx->net_dev,
 183                            "TX queue %d transmission id %x complete\n",
 184                            tx_queue->queue, tx_queue->read_count);
 185         } else if (buffer->flags & EFX_TX_BUF_XDP) {
 186                 xdp_return_frame_rx_napi(buffer->xdpf);
 187         }
 188
 189         buffer->len = 0;
 190         buffer->flags = 0;
 191 }
 192
 193 /* Remove packets from the TX queue
 194  *
 195  * This removes packets from the TX queue, up to and including the
 196  * specified index.
 197  */
 198 static void efx_dequeue_buffers(struct efx_tx_queue *tx_queue,
 199                                 unsigned int index,
 200                                 unsigned int *pkts_compl,
 201                                 unsigned int *bytes_compl)
 202 {
 203         struct efx_nic *efx = tx_queue->efx;
 204         unsigned int stop_index, read_ptr;
 205
 206         stop_index = (index + 1) & tx_queue->ptr_mask;
 207         read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 208
 209         while (read_ptr != stop_index) {
 210                 struct efx_tx_buffer *buffer = &tx_queue->buffer[read_ptr];
 211
 212                 if (!efx_tx_buffer_in_use(buffer)) {
 213                         netif_err(efx, tx_err, efx->net_dev,
 214                                   "TX queue %d spurious TX completion id %d\n",
 215                                   tx_queue->queue, read_ptr);
 216                         efx_schedule_reset(efx, RESET_TYPE_TX_SKIP);
 217                         return;
 218                 }
 219
 220                 efx_dequeue_buffer(tx_queue, buffer, pkts_compl, bytes_compl);
 221
 222                 ++tx_queue->read_count;
 223                 read_ptr = tx_queue->read_count & tx_queue->ptr_mask;
 224         }
 225 }
 226
 227 void efx_xmit_done_check_empty(struct efx_tx_queue *tx_queue)
 228 {
 229         if ((int)(tx_queue->read_count - tx_queue->old_write_count) >= 0) {
 230                 tx_queue->old_write_count = READ_ONCE(tx_queue->write_count);
 231                 if (tx_queue->read_count == tx_queue->old_write_count) {
 232                         /* Ensure that read_count is flushed. */
 233                         smp_mb();
 234                         tx_queue->empty_read_count =
 235                                 tx_queue->read_count | EFX_EMPTY_COUNT_VALID;
 236                 }
 237         }
 238 }
 239
 240 void efx_xmit_done(struct efx_tx_queue *tx_queue, unsigned int index)
 241 {
 242         unsigned int fill_level, pkts_compl = 0, bytes_compl = 0;
 243         struct efx_nic *efx = tx_queue->efx;
 244         struct efx_tx_queue *txq2;
 245
 246         EFX_WARN_ON_ONCE_PARANOID(index > tx_queue->ptr_mask);
 247
 248         efx_dequeue_buffers(tx_queue, index, &pkts_compl, &bytes_compl);
 249         tx_queue->pkts_compl += pkts_compl;
 250         tx_queue->bytes_compl += bytes_compl;
 251
 252         if (pkts_compl > 1)
 253                 ++tx_queue->merge_events;
 254
 255         /* See if we need to restart the netif queue.  This memory
 256          * barrier ensures that we write read_count (inside
 257          * efx_dequeue_buffers()) before reading the queue status.
 258          */
 259         smp_mb();
 260         if (unlikely(netif_tx_queue_stopped(tx_queue->core_txq)) &&
 261             likely(efx->port_enabled) &&
 262             likely(netif_device_present(efx->net_dev))) {
 263                 txq2 = efx_tx_queue_partner(tx_queue);
 264                 fill_level = max(tx_queue->insert_count - tx_queue->read_count,
 265                                  txq2->insert_count - txq2->read_count);
 266                 if (fill_level <= efx->txq_wake_thresh)
 267                         netif_tx_wake_queue(tx_queue->core_txq);
 268         }
 269
 270         efx_xmit_done_check_empty(tx_queue);
 271 }
 272
 273 /* Remove buffers put into a tx_queue for the current packet.
 274  * None of the buffers must have an skb attached.
 275  */
 276 void efx_enqueue_unwind(struct efx_tx_queue *tx_queue,
 277                         unsigned int insert_count)
 278 {
 279         struct efx_tx_buffer *buffer;
 280         unsigned int bytes_compl = 0;
 281         unsigned int pkts_compl = 0;
 282
 283         /* Work backwards until we hit the original insert pointer value */
 284         while (tx_queue->insert_count != insert_count) {
 285                 --tx_queue->insert_count;
 286                 buffer = __efx_tx_queue_get_insert_buffer(tx_queue);
 287                 efx_dequeue_buffer(tx_queue, buffer, &pkts_compl, &bytes_compl);
 288         }
 289 }
 290
 291 struct efx_tx_buffer *efx_tx_map_chunk(struct efx_tx_queue *tx_queue,
 292                                        dma_addr_t dma_addr, size_t len)
 293 {
 294         const struct efx_nic_type *nic_type = tx_queue->efx->type;
 295         struct efx_tx_buffer *buffer;
 296         unsigned int dma_len;
 297
 298         /* Map the fragment taking account of NIC-dependent DMA limits. */
 299         do {
 300                 buffer = efx_tx_queue_get_insert_buffer(tx_queue);
 301                 dma_len = nic_type->tx_limit_len(tx_queue, dma_addr, len);
 302
 303                 buffer->len = dma_len;
 304                 buffer->dma_addr = dma_addr;
 305                 buffer->flags = EFX_TX_BUF_CONT;
 306                 len -= dma_len;
 307                 dma_addr += dma_len;
 308                 ++tx_queue->insert_count;
 309         } while (len);
 310
 311         return buffer;
 312 }
 313
 314 /* Map all data from an SKB for DMA and create descriptors on the queue. */
 315 int efx_tx_map_data(struct efx_tx_queue *tx_queue, struct sk_buff *skb,
 316                     unsigned int segment_count)
 317 {
 318         struct efx_nic *efx = tx_queue->efx;
 319         struct device *dma_dev = &efx->pci_dev->dev;
 320         unsigned int frag_index, nr_frags;
 321         dma_addr_t dma_addr, unmap_addr;
 322         unsigned short dma_flags;
 323         size_t len, unmap_len;
 324
 325         nr_frags = skb_shinfo(skb)->nr_frags;
 326         frag_index = 0;
 327
 328         /* Map header data. */
 329         len = skb_headlen(skb);
 330         dma_addr = dma_map_single(dma_dev, skb->data, len, DMA_TO_DEVICE);
 331         dma_flags = EFX_TX_BUF_MAP_SINGLE;
 332         unmap_len = len;
 333         unmap_addr = dma_addr;
 334
 335         if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
 336                 return -EIO;
 337
 338         if (segment_count) {
 339                 /* For TSO we need to put the header in to a separate
 340                  * descriptor. Map this separately if necessary.
 341                  */
 342                 size_t header_len = skb_transport_header(skb) - skb->data +
 343                                 (tcp_hdr(skb)->doff << 2u);
 344
 345                 if (header_len != len) {
 346                         tx_queue->tso_long_headers++;
 347                         efx_tx_map_chunk(tx_queue, dma_addr, header_len);
 348                         len -= header_len;
 349                         dma_addr += header_len;
 350                 }
 351         }
 352
 353         /* Add descriptors for each fragment. */
 354         do {
 355                 struct efx_tx_buffer *buffer;
 356                 skb_frag_t *fragment;
 357
 358                 buffer = efx_tx_map_chunk(tx_queue, dma_addr, len);
 359
 360                 /* The final descriptor for a fragment is responsible for
 361                  * unmapping the whole fragment.
 362                  */
 363                 buffer->flags = EFX_TX_BUF_CONT | dma_flags;
 364                 buffer->unmap_len = unmap_len;
 365                 buffer->dma_offset = buffer->dma_addr - unmap_addr;
 366
 367                 if (frag_index >= nr_frags) {
 368                         /* Store SKB details with the final buffer for
 369                          * the completion.
 370                          */
 371                         buffer->skb = skb;
 372                         buffer->flags = EFX_TX_BUF_SKB | dma_flags;
 373                         return 0;
 374                 }
 375
 376                 /* Move on to the next fragment. */
 377                 fragment = &skb_shinfo(skb)->frags[frag_index++];
 378                 len = skb_frag_size(fragment);
 379                 dma_addr = skb_frag_dma_map(dma_dev, fragment, 0, len,
 380                                             DMA_TO_DEVICE);
 381                 dma_flags = 0;
 382                 unmap_len = len;
 383                 unmap_addr = dma_addr;
 384
 385                 if (unlikely(dma_mapping_error(dma_dev, dma_addr)))
 386                         return -EIO;
 387         } while (1);
 388 }
 389
 390 unsigned int efx_tx_max_skb_descs(struct efx_nic *efx)
 391 {
 392         /* Header and payload descriptor for each output segment, plus
 393          * one for every input fragment boundary within a segment
 394          */
 395         unsigned int max_descs = EFX_TSO_MAX_SEGS * 2 + MAX_SKB_FRAGS;
 396
 397         /* Possibly one more per segment for option descriptors */
 398         if (efx_nic_rev(efx) >= EFX_REV_HUNT_A0)
 399                 max_descs += EFX_TSO_MAX_SEGS;
 400
 401         /* Possibly more for PCIe page boundaries within input fragments */
 402         if (PAGE_SIZE > EFX_PAGE_SIZE)
 403                 max_descs += max_t(unsigned int, MAX_SKB_FRAGS,
 404                                    DIV_ROUND_UP(GSO_MAX_SIZE, EFX_PAGE_SIZE));
 405
 406         return max_descs;
 407 }