drivers/net/ethernet/sfc/nic.c

   1 /****************************************************************************
   2  * Driver for Solarflare Solarstorm network controllers and boards
   3  * Copyright 2005-2006 Fen Systems Ltd.
   4  * Copyright 2006-2011 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/bitops.h>
  12 #include <linux/delay.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/pci.h>
  15 #include <linux/module.h>
  16 #include <linux/seq_file.h>
  17 #include "net_driver.h"
  18 #include "bitfield.h"
  19 #include "efx.h"
  20 #include "nic.h"
  21 #include "regs.h"
  22 #include "io.h"
  23 #include "workarounds.h"
  24
  25 /**************************************************************************
  26  *
  27  * Configurable values
  28  *
  29  **************************************************************************
  30  */
  31
  32 /* This is set to 16 for a good reason.  In summary, if larger than
  33  * 16, the descriptor cache holds more than a default socket
  34  * buffer's worth of packets (for UDP we can only have at most one
  35  * socket buffer's worth outstanding).  This combined with the fact
  36  * that we only get 1 TX event per descriptor cache means the NIC
  37  * goes idle.
  38  */
  39 #define TX_DC_ENTRIES 16
  40 #define TX_DC_ENTRIES_ORDER 1
  41
  42 #define RX_DC_ENTRIES 64
  43 #define RX_DC_ENTRIES_ORDER 3
  44
  45 /* If EFX_MAX_INT_ERRORS internal errors occur within
  46  * EFX_INT_ERROR_EXPIRE seconds, we consider the NIC broken and
  47  * disable it.
  48  */
  49 #define EFX_INT_ERROR_EXPIRE 3600
  50 #define EFX_MAX_INT_ERRORS 5
  51
  52 /* We poll for events every FLUSH_INTERVAL ms, and check FLUSH_POLL_COUNT times
  53  */
  54 #define EFX_FLUSH_INTERVAL 10
  55 #define EFX_FLUSH_POLL_COUNT 100
  56
  57 /* Size and alignment of special buffers (4KB) */
  58 #define EFX_BUF_SIZE 4096
  59
  60 /* Depth of RX flush request fifo */
  61 #define EFX_RX_FLUSH_COUNT 4
  62
  63 /* Generated event code for efx_generate_test_event() */
  64 #define EFX_CHANNEL_MAGIC_TEST(_channel)        \
  65         (0x00010100 + (_channel)->channel)
  66
  67 /* Generated event code for efx_generate_fill_event() */
  68 #define EFX_CHANNEL_MAGIC_FILL(_channel)        \
  69         (0x00010200 + (_channel)->channel)
  70
  71 /**************************************************************************
  72  *
  73  * Solarstorm hardware access
  74  *
  75  **************************************************************************/
  76
  77 static inline void efx_write_buf_tbl(struct efx_nic *efx, efx_qword_t *value,
  78                                      unsigned int index)
  79 {
  80         efx_sram_writeq(efx, efx->membase + efx->type->buf_tbl_base,
  81                         value, index);
  82 }
  83
  84 /* Read the current event from the event queue */
  85 static inline efx_qword_t *efx_event(struct efx_channel *channel,
  86                                      unsigned int index)
  87 {
  88         return ((efx_qword_t *) (channel->eventq.addr)) +
  89                 (index & channel->eventq_mask);
  90 }
  91
  92 /* See if an event is present
  93  *
  94  * We check both the high and low dword of the event for all ones.  We
  95  * wrote all ones when we cleared the event, and no valid event can
  96  * have all ones in either its high or low dwords.  This approach is
  97  * robust against reordering.
  98  *
  99  * Note that using a single 64-bit comparison is incorrect; even
 100  * though the CPU read will be atomic, the DMA write may not be.
 101  */
 102 static inline int efx_event_present(efx_qword_t *event)
 103 {
 104         return !(EFX_DWORD_IS_ALL_ONES(event->dword[0]) |
 105                   EFX_DWORD_IS_ALL_ONES(event->dword[1]));
 106 }
 107
 108 static bool efx_masked_compare_oword(const efx_oword_t *a, const efx_oword_t *b,
 109                                      const efx_oword_t *mask)
 110 {
 111         return ((a->u64[0] ^ b->u64[0]) & mask->u64[0]) ||
 112                 ((a->u64[1] ^ b->u64[1]) & mask->u64[1]);
 113 }
 114
 115 int efx_nic_test_registers(struct efx_nic *efx,
 116                            const struct efx_nic_register_test *regs,
 117                            size_t n_regs)
 118 {
 119         unsigned address = 0, i, j;
 120         efx_oword_t mask, imask, original, reg, buf;
 121
 122         /* Falcon should be in loopback to isolate the XMAC from the PHY */
 123         WARN_ON(!LOOPBACK_INTERNAL(efx));
 124
 125         for (i = 0; i < n_regs; ++i) {
 126                 address = regs[i].address;
 127                 mask = imask = regs[i].mask;
 128                 EFX_INVERT_OWORD(imask);
 129
 130                 efx_reado(efx, &original, address);
 131
 132                 /* bit sweep on and off */
 133                 for (j = 0; j < 128; j++) {
 134                         if (!EFX_EXTRACT_OWORD32(mask, j, j))
 135                                 continue;
 136
 137                         /* Test this testable bit can be set in isolation */
 138                         EFX_AND_OWORD(reg, original, mask);
 139                         EFX_SET_OWORD32(reg, j, j, 1);
 140
 141                         efx_writeo(efx, &reg, address);
 142                         efx_reado(efx, &buf, address);
 143
 144                         if (efx_masked_compare_oword(&reg, &buf, &mask))
 145                                 goto fail;
 146
 147                         /* Test this testable bit can be cleared in isolation */
 148                         EFX_OR_OWORD(reg, original, mask);
 149                         EFX_SET_OWORD32(reg, j, j, 0);
 150
 151                         efx_writeo(efx, &reg, address);
 152                         efx_reado(efx, &buf, address);
 153
 154                         if (efx_masked_compare_oword(&reg, &buf, &mask))
 155                                 goto fail;
 156                 }
 157
 158                 efx_writeo(efx, &original, address);
 159         }
 160
 161         return 0;
 162
 163 fail:
 164         netif_err(efx, hw, efx->net_dev,
 165                   "wrote "EFX_OWORD_FMT" read "EFX_OWORD_FMT
 166                   " at address 0x%x mask "EFX_OWORD_FMT"\n", EFX_OWORD_VAL(reg),
 167                   EFX_OWORD_VAL(buf), address, EFX_OWORD_VAL(mask));
 168         return -EIO;
 169 }
 170
 171 /**************************************************************************
 172  *
 173  * Special buffer handling
 174  * Special buffers are used for event queues and the TX and RX
 175  * descriptor rings.
 176  *
 177  *************************************************************************/
 178
 179 /*
 180  * Initialise a special buffer
 181  *
 182  * This will define a buffer (previously allocated via
 183  * efx_alloc_special_buffer()) in the buffer table, allowing
 184  * it to be used for event queues, descriptor rings etc.
 185  */
 186 static void
 187 efx_init_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
 188 {
 189         efx_qword_t buf_desc;
 190         int index;
 191         dma_addr_t dma_addr;
 192         int i;
 193
 194         EFX_BUG_ON_PARANOID(!buffer->addr);
 195
 196         /* Write buffer descriptors to NIC */
 197         for (i = 0; i < buffer->entries; i++) {
 198                 index = buffer->index + i;
 199                 dma_addr = buffer->dma_addr + (i * 4096);
 200                 netif_dbg(efx, probe, efx->net_dev,
 201                           "mapping special buffer %d at %llx\n",
 202                           index, (unsigned long long)dma_addr);
 203                 EFX_POPULATE_QWORD_3(buf_desc,
 204                                      FRF_AZ_BUF_ADR_REGION, 0,
 205                                      FRF_AZ_BUF_ADR_FBUF, dma_addr >> 12,
 206                                      FRF_AZ_BUF_OWNER_ID_FBUF, 0);
 207                 efx_write_buf_tbl(efx, &buf_desc, index);
 208         }
 209 }
 210
 211 /* Unmaps a buffer and clears the buffer table entries */
 212 static void
 213 efx_fini_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
 214 {
 215         efx_oword_t buf_tbl_upd;
 216         unsigned int start = buffer->index;
 217         unsigned int end = (buffer->index + buffer->entries - 1);
 218
 219         if (!buffer->entries)
 220                 return;
 221
 222         netif_dbg(efx, hw, efx->net_dev, "unmapping special buffers %d-%d\n",
 223                   buffer->index, buffer->index + buffer->entries - 1);
 224
 225         EFX_POPULATE_OWORD_4(buf_tbl_upd,
 226                              FRF_AZ_BUF_UPD_CMD, 0,
 227                              FRF_AZ_BUF_CLR_CMD, 1,
 228                              FRF_AZ_BUF_CLR_END_ID, end,
 229                              FRF_AZ_BUF_CLR_START_ID, start);
 230         efx_writeo(efx, &buf_tbl_upd, FR_AZ_BUF_TBL_UPD);
 231 }
 232
 233 /*
 234  * Allocate a new special buffer
 235  *
 236  * This allocates memory for a new buffer, clears it and allocates a
 237  * new buffer ID range.  It does not write into the buffer table.
 238  *
 239  * This call will allocate 4KB buffers, since 8KB buffers can't be
 240  * used for event queues and descriptor rings.
 241  */
 242 static int efx_alloc_special_buffer(struct efx_nic *efx,
 243                                     struct efx_special_buffer *buffer,
 244                                     unsigned int len)
 245 {
 246         len = ALIGN(len, EFX_BUF_SIZE);
 247
 248         buffer->addr = dma_alloc_coherent(&efx->pci_dev->dev, len,
 249                                           &buffer->dma_addr, GFP_KERNEL);
 250         if (!buffer->addr)
 251                 return -ENOMEM;
 252         buffer->len = len;
 253         buffer->entries = len / EFX_BUF_SIZE;
 254         BUG_ON(buffer->dma_addr & (EFX_BUF_SIZE - 1));
 255
 256         /* All zeros is a potentially valid event so memset to 0xff */
 257         memset(buffer->addr, 0xff, len);
 258
 259         /* Select new buffer ID */
 260         buffer->index = efx->next_buffer_table;
 261         efx->next_buffer_table += buffer->entries;
 262
 263         netif_dbg(efx, probe, efx->net_dev,
 264                   "allocating special buffers %d-%d at %llx+%x "
 265                   "(virt %p phys %llx)\n", buffer->index,
 266                   buffer->index + buffer->entries - 1,
 267                   (u64)buffer->dma_addr, len,
 268                   buffer->addr, (u64)virt_to_phys(buffer->addr));
 269
 270         return 0;
 271 }
 272
 273 static void
 274 efx_free_special_buffer(struct efx_nic *efx, struct efx_special_buffer *buffer)
 275 {
 276         if (!buffer->addr)
 277                 return;
 278
 279         netif_dbg(efx, hw, efx->net_dev,
 280                   "deallocating special buffers %d-%d at %llx+%x "
 281                   "(virt %p phys %llx)\n", buffer->index,
 282                   buffer->index + buffer->entries - 1,
 283                   (u64)buffer->dma_addr, buffer->len,
 284                   buffer->addr, (u64)virt_to_phys(buffer->addr));
 285
 286         dma_free_coherent(&efx->pci_dev->dev, buffer->len, buffer->addr,
 287                           buffer->dma_addr);
 288         buffer->addr = NULL;
 289         buffer->entries = 0;
 290 }
 291
 292 /**************************************************************************
 293  *
 294  * Generic buffer handling
 295  * These buffers are used for interrupt status and MAC stats
 296  *
 297  **************************************************************************/
 298
 299 int efx_nic_alloc_buffer(struct efx_nic *efx, struct efx_buffer *buffer,
 300                          unsigned int len)
 301 {
 302         buffer->addr = pci_alloc_consistent(efx->pci_dev, len,
 303                                             &buffer->dma_addr);
 304         if (!buffer->addr)
 305                 return -ENOMEM;
 306         buffer->len = len;
 307         memset(buffer->addr, 0, len);
 308         return 0;
 309 }
 310
 311 void efx_nic_free_buffer(struct efx_nic *efx, struct efx_buffer *buffer)
 312 {
 313         if (buffer->addr) {
 314                 pci_free_consistent(efx->pci_dev, buffer->len,
 315                                     buffer->addr, buffer->dma_addr);
 316                 buffer->addr = NULL;
 317         }
 318 }
 319
 320 /**************************************************************************
 321  *
 322  * TX path
 323  *
 324  **************************************************************************/
 325
 326 /* Returns a pointer to the specified transmit descriptor in the TX
 327  * descriptor queue belonging to the specified channel.
 328  */
 329 static inline efx_qword_t *
 330 efx_tx_desc(struct efx_tx_queue *tx_queue, unsigned int index)
 331 {
 332         return ((efx_qword_t *) (tx_queue->txd.addr)) + index;
 333 }
 334
 335 /* This writes to the TX_DESC_WPTR; write pointer for TX descriptor ring */
 336 static inline void efx_notify_tx_desc(struct efx_tx_queue *tx_queue)
 337 {
 338         unsigned write_ptr;
 339         efx_dword_t reg;
 340
 341         write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
 342         EFX_POPULATE_DWORD_1(reg, FRF_AZ_TX_DESC_WPTR_DWORD, write_ptr);
 343         efx_writed_page(tx_queue->efx, &reg,
 344                         FR_AZ_TX_DESC_UPD_DWORD_P0, tx_queue->queue);
 345 }
 346
 347 /* Write pointer and first descriptor for TX descriptor ring */
 348 static inline void efx_push_tx_desc(struct efx_tx_queue *tx_queue,
 349                                     const efx_qword_t *txd)
 350 {
 351         unsigned write_ptr;
 352         efx_oword_t reg;
 353
 354         BUILD_BUG_ON(FRF_AZ_TX_DESC_LBN != 0);
 355         BUILD_BUG_ON(FR_AA_TX_DESC_UPD_KER != FR_BZ_TX_DESC_UPD_P0);
 356
 357         write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
 358         EFX_POPULATE_OWORD_2(reg, FRF_AZ_TX_DESC_PUSH_CMD, true,
 359                              FRF_AZ_TX_DESC_WPTR, write_ptr);
 360         reg.qword[0] = *txd;
 361         efx_writeo_page(tx_queue->efx, &reg,
 362                         FR_BZ_TX_DESC_UPD_P0, tx_queue->queue);
 363 }
 364
 365 static inline bool
 366 efx_may_push_tx_desc(struct efx_tx_queue *tx_queue, unsigned int write_count)
 367 {
 368         unsigned empty_read_count = ACCESS_ONCE(tx_queue->empty_read_count);
 369
 370         if (empty_read_count == 0)
 371                 return false;
 372
 373         tx_queue->empty_read_count = 0;
 374         return ((empty_read_count ^ write_count) & ~EFX_EMPTY_COUNT_VALID) == 0;
 375 }
 376
 377 /* For each entry inserted into the software descriptor ring, create a
 378  * descriptor in the hardware TX descriptor ring (in host memory), and
 379  * write a doorbell.
 380  */
 381 void efx_nic_push_buffers(struct efx_tx_queue *tx_queue)
 382 {
 383
 384         struct efx_tx_buffer *buffer;
 385         efx_qword_t *txd;
 386         unsigned write_ptr;
 387         unsigned old_write_count = tx_queue->write_count;
 388
 389         BUG_ON(tx_queue->write_count == tx_queue->insert_count);
 390
 391         do {
 392                 write_ptr = tx_queue->write_count & tx_queue->ptr_mask;
 393                 buffer = &tx_queue->buffer[write_ptr];
 394                 txd = efx_tx_desc(tx_queue, write_ptr);
 395                 ++tx_queue->write_count;
 396
 397                 /* Create TX descriptor ring entry */
 398                 EFX_POPULATE_QWORD_4(*txd,
 399                                      FSF_AZ_TX_KER_CONT, buffer->continuation,
 400                                      FSF_AZ_TX_KER_BYTE_COUNT, buffer->len,
 401                                      FSF_AZ_TX_KER_BUF_REGION, 0,
 402                                      FSF_AZ_TX_KER_BUF_ADDR, buffer->dma_addr);
 403         } while (tx_queue->write_count != tx_queue->insert_count);
 404
 405         wmb(); /* Ensure descriptors are written before they are fetched */
 406
 407         if (efx_may_push_tx_desc(tx_queue, old_write_count)) {
 408                 txd = efx_tx_desc(tx_queue,
 409                                   old_write_count & tx_queue->ptr_mask);
 410                 efx_push_tx_desc(tx_queue, txd);
 411                 ++tx_queue->pushes;
 412         } else {
 413                 efx_notify_tx_desc(tx_queue);
 414         }
 415 }
 416
 417 /* Allocate hardware resources for a TX queue */
 418 int efx_nic_probe_tx(struct efx_tx_queue *tx_queue)
 419 {
 420         struct efx_nic *efx = tx_queue->efx;
 421         unsigned entries;
 422
 423         entries = tx_queue->ptr_mask + 1;
 424         return efx_alloc_special_buffer(efx, &tx_queue->txd,
 425                                         entries * sizeof(efx_qword_t));
 426 }
 427
 428 void efx_nic_init_tx(struct efx_tx_queue *tx_queue)
 429 {
 430         struct efx_nic *efx = tx_queue->efx;
 431         efx_oword_t reg;
 432
 433         tx_queue->flushed = FLUSH_NONE;
 434
 435         /* Pin TX descriptor ring */
 436         efx_init_special_buffer(efx, &tx_queue->txd);
 437
 438         /* Push TX descriptor ring to card */
 439         EFX_POPULATE_OWORD_10(reg,
 440                               FRF_AZ_TX_DESCQ_EN, 1,
 441                               FRF_AZ_TX_ISCSI_DDIG_EN, 0,
 442                               FRF_AZ_TX_ISCSI_HDIG_EN, 0,
 443                               FRF_AZ_TX_DESCQ_BUF_BASE_ID, tx_queue->txd.index,
 444                               FRF_AZ_TX_DESCQ_EVQ_ID,
 445                               tx_queue->channel->channel,
 446                               FRF_AZ_TX_DESCQ_OWNER_ID, 0,
 447                               FRF_AZ_TX_DESCQ_LABEL, tx_queue->queue,
 448                               FRF_AZ_TX_DESCQ_SIZE,
 449                               __ffs(tx_queue->txd.entries),
 450                               FRF_AZ_TX_DESCQ_TYPE, 0,
 451                               FRF_BZ_TX_NON_IP_DROP_DIS, 1);
 452
 453         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
 454                 int csum = tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD;
 455                 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_IP_CHKSM_DIS, !csum);
 456                 EFX_SET_OWORD_FIELD(reg, FRF_BZ_TX_TCP_CHKSM_DIS,
 457                                     !csum);
 458         }
 459
 460         efx_writeo_table(efx, &reg, efx->type->txd_ptr_tbl_base,
 461                          tx_queue->queue);
 462
 463         if (efx_nic_rev(efx) < EFX_REV_FALCON_B0) {
 464                 /* Only 128 bits in this register */
 465                 BUILD_BUG_ON(EFX_MAX_TX_QUEUES > 128);
 466
 467                 efx_reado(efx, &reg, FR_AA_TX_CHKSM_CFG);
 468                 if (tx_queue->queue & EFX_TXQ_TYPE_OFFLOAD)
 469                         clear_bit_le(tx_queue->queue, (void *)&reg);
 470                 else
 471                         set_bit_le(tx_queue->queue, (void *)&reg);
 472                 efx_writeo(efx, &reg, FR_AA_TX_CHKSM_CFG);
 473         }
 474
 475         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
 476                 EFX_POPULATE_OWORD_1(reg,
 477                                      FRF_BZ_TX_PACE,
 478                                      (tx_queue->queue & EFX_TXQ_TYPE_HIGHPRI) ?
 479                                      FFE_BZ_TX_PACE_OFF :
 480                                      FFE_BZ_TX_PACE_RESERVED);
 481                 efx_writeo_table(efx, &reg, FR_BZ_TX_PACE_TBL,
 482                                  tx_queue->queue);
 483         }
 484 }
 485
 486 static void efx_flush_tx_queue(struct efx_tx_queue *tx_queue)
 487 {
 488         struct efx_nic *efx = tx_queue->efx;
 489         efx_oword_t tx_flush_descq;
 490
 491         tx_queue->flushed = FLUSH_PENDING;
 492
 493         /* Post a flush command */
 494         EFX_POPULATE_OWORD_2(tx_flush_descq,
 495                              FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
 496                              FRF_AZ_TX_FLUSH_DESCQ, tx_queue->queue);
 497         efx_writeo(efx, &tx_flush_descq, FR_AZ_TX_FLUSH_DESCQ);
 498 }
 499
 500 void efx_nic_fini_tx(struct efx_tx_queue *tx_queue)
 501 {
 502         struct efx_nic *efx = tx_queue->efx;
 503         efx_oword_t tx_desc_ptr;
 504
 505         /* The queue should have been flushed */
 506         WARN_ON(tx_queue->flushed != FLUSH_DONE);
 507
 508         /* Remove TX descriptor ring from card */
 509         EFX_ZERO_OWORD(tx_desc_ptr);
 510         efx_writeo_table(efx, &tx_desc_ptr, efx->type->txd_ptr_tbl_base,
 511                          tx_queue->queue);
 512
 513         /* Unpin TX descriptor ring */
 514         efx_fini_special_buffer(efx, &tx_queue->txd);
 515 }
 516
 517 /* Free buffers backing TX queue */
 518 void efx_nic_remove_tx(struct efx_tx_queue *tx_queue)
 519 {
 520         efx_free_special_buffer(tx_queue->efx, &tx_queue->txd);
 521 }
 522
 523 /**************************************************************************
 524  *
 525  * RX path
 526  *
 527  **************************************************************************/
 528
 529 /* Returns a pointer to the specified descriptor in the RX descriptor queue */
 530 static inline efx_qword_t *
 531 efx_rx_desc(struct efx_rx_queue *rx_queue, unsigned int index)
 532 {
 533         return ((efx_qword_t *) (rx_queue->rxd.addr)) + index;
 534 }
 535
 536 /* This creates an entry in the RX descriptor queue */
 537 static inline void
 538 efx_build_rx_desc(struct efx_rx_queue *rx_queue, unsigned index)
 539 {
 540         struct efx_rx_buffer *rx_buf;
 541         efx_qword_t *rxd;
 542
 543         rxd = efx_rx_desc(rx_queue, index);
 544         rx_buf = efx_rx_buffer(rx_queue, index);
 545         EFX_POPULATE_QWORD_3(*rxd,
 546                              FSF_AZ_RX_KER_BUF_SIZE,
 547                              rx_buf->len -
 548                              rx_queue->efx->type->rx_buffer_padding,
 549                              FSF_AZ_RX_KER_BUF_REGION, 0,
 550                              FSF_AZ_RX_KER_BUF_ADDR, rx_buf->dma_addr);
 551 }
 552
 553 /* This writes to the RX_DESC_WPTR register for the specified receive
 554  * descriptor ring.
 555  */
 556 void efx_nic_notify_rx_desc(struct efx_rx_queue *rx_queue)
 557 {
 558         struct efx_nic *efx = rx_queue->efx;
 559         efx_dword_t reg;
 560         unsigned write_ptr;
 561
 562         while (rx_queue->notified_count != rx_queue->added_count) {
 563                 efx_build_rx_desc(
 564                         rx_queue,
 565                         rx_queue->notified_count & rx_queue->ptr_mask);
 566                 ++rx_queue->notified_count;
 567         }
 568
 569         wmb();
 570         write_ptr = rx_queue->added_count & rx_queue->ptr_mask;
 571         EFX_POPULATE_DWORD_1(reg, FRF_AZ_RX_DESC_WPTR_DWORD, write_ptr);
 572         efx_writed_page(efx, &reg, FR_AZ_RX_DESC_UPD_DWORD_P0,
 573                         efx_rx_queue_index(rx_queue));
 574 }
 575
 576 int efx_nic_probe_rx(struct efx_rx_queue *rx_queue)
 577 {
 578         struct efx_nic *efx = rx_queue->efx;
 579         unsigned entries;
 580
 581         entries = rx_queue->ptr_mask + 1;
 582         return efx_alloc_special_buffer(efx, &rx_queue->rxd,
 583                                         entries * sizeof(efx_qword_t));
 584 }
 585
 586 void efx_nic_init_rx(struct efx_rx_queue *rx_queue)
 587 {
 588         efx_oword_t rx_desc_ptr;
 589         struct efx_nic *efx = rx_queue->efx;
 590         bool is_b0 = efx_nic_rev(efx) >= EFX_REV_FALCON_B0;
 591         bool iscsi_digest_en = is_b0;
 592
 593         netif_dbg(efx, hw, efx->net_dev,
 594                   "RX queue %d ring in special buffers %d-%d\n",
 595                   efx_rx_queue_index(rx_queue), rx_queue->rxd.index,
 596                   rx_queue->rxd.index + rx_queue->rxd.entries - 1);
 597
 598         rx_queue->flushed = FLUSH_NONE;
 599
 600         /* Pin RX descriptor ring */
 601         efx_init_special_buffer(efx, &rx_queue->rxd);
 602
 603         /* Push RX descriptor ring to card */
 604         EFX_POPULATE_OWORD_10(rx_desc_ptr,
 605                               FRF_AZ_RX_ISCSI_DDIG_EN, iscsi_digest_en,
 606                               FRF_AZ_RX_ISCSI_HDIG_EN, iscsi_digest_en,
 607                               FRF_AZ_RX_DESCQ_BUF_BASE_ID, rx_queue->rxd.index,
 608                               FRF_AZ_RX_DESCQ_EVQ_ID,
 609                               efx_rx_queue_channel(rx_queue)->channel,
 610                               FRF_AZ_RX_DESCQ_OWNER_ID, 0,
 611                               FRF_AZ_RX_DESCQ_LABEL,
 612                               efx_rx_queue_index(rx_queue),
 613                               FRF_AZ_RX_DESCQ_SIZE,
 614                               __ffs(rx_queue->rxd.entries),
 615                               FRF_AZ_RX_DESCQ_TYPE, 0 /* kernel queue */ ,
 616                               /* For >=B0 this is scatter so disable */
 617                               FRF_AZ_RX_DESCQ_JUMBO, !is_b0,
 618                               FRF_AZ_RX_DESCQ_EN, 1);
 619         efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
 620                          efx_rx_queue_index(rx_queue));
 621 }
 622
 623 static void efx_flush_rx_queue(struct efx_rx_queue *rx_queue)
 624 {
 625         struct efx_nic *efx = rx_queue->efx;
 626         efx_oword_t rx_flush_descq;
 627
 628         rx_queue->flushed = FLUSH_PENDING;
 629
 630         /* Post a flush command */
 631         EFX_POPULATE_OWORD_2(rx_flush_descq,
 632                              FRF_AZ_RX_FLUSH_DESCQ_CMD, 1,
 633                              FRF_AZ_RX_FLUSH_DESCQ,
 634                              efx_rx_queue_index(rx_queue));
 635         efx_writeo(efx, &rx_flush_descq, FR_AZ_RX_FLUSH_DESCQ);
 636 }
 637
 638 void efx_nic_fini_rx(struct efx_rx_queue *rx_queue)
 639 {
 640         efx_oword_t rx_desc_ptr;
 641         struct efx_nic *efx = rx_queue->efx;
 642
 643         /* The queue should already have been flushed */
 644         WARN_ON(rx_queue->flushed != FLUSH_DONE);
 645
 646         /* Remove RX descriptor ring from card */
 647         EFX_ZERO_OWORD(rx_desc_ptr);
 648         efx_writeo_table(efx, &rx_desc_ptr, efx->type->rxd_ptr_tbl_base,
 649                          efx_rx_queue_index(rx_queue));
 650
 651         /* Unpin RX descriptor ring */
 652         efx_fini_special_buffer(efx, &rx_queue->rxd);
 653 }
 654
 655 /* Free buffers backing RX queue */
 656 void efx_nic_remove_rx(struct efx_rx_queue *rx_queue)
 657 {
 658         efx_free_special_buffer(rx_queue->efx, &rx_queue->rxd);
 659 }
 660
 661 /**************************************************************************
 662  *
 663  * Event queue processing
 664  * Event queues are processed by per-channel tasklets.
 665  *
 666  **************************************************************************/
 667
 668 /* Update a channel's event queue's read pointer (RPTR) register
 669  *
 670  * This writes the EVQ_RPTR_REG register for the specified channel's
 671  * event queue.
 672  */
 673 void efx_nic_eventq_read_ack(struct efx_channel *channel)
 674 {
 675         efx_dword_t reg;
 676         struct efx_nic *efx = channel->efx;
 677
 678         EFX_POPULATE_DWORD_1(reg, FRF_AZ_EVQ_RPTR,
 679                              channel->eventq_read_ptr & channel->eventq_mask);
 680         efx_writed_table(efx, &reg, efx->type->evq_rptr_tbl_base,
 681                          channel->channel);
 682 }
 683
 684 /* Use HW to insert a SW defined event */
 685 static void efx_generate_event(struct efx_channel *channel, efx_qword_t *event)
 686 {
 687         efx_oword_t drv_ev_reg;
 688
 689         BUILD_BUG_ON(FRF_AZ_DRV_EV_DATA_LBN != 0 ||
 690                      FRF_AZ_DRV_EV_DATA_WIDTH != 64);
 691         drv_ev_reg.u32[0] = event->u32[0];
 692         drv_ev_reg.u32[1] = event->u32[1];
 693         drv_ev_reg.u32[2] = 0;
 694         drv_ev_reg.u32[3] = 0;
 695         EFX_SET_OWORD_FIELD(drv_ev_reg, FRF_AZ_DRV_EV_QID, channel->channel);
 696         efx_writeo(channel->efx, &drv_ev_reg, FR_AZ_DRV_EV);
 697 }
 698
 699 /* Handle a transmit completion event
 700  *
 701  * The NIC batches TX completion events; the message we receive is of
 702  * the form "complete all TX events up to this index".
 703  */
 704 static int
 705 efx_handle_tx_event(struct efx_channel *channel, efx_qword_t *event)
 706 {
 707         unsigned int tx_ev_desc_ptr;
 708         unsigned int tx_ev_q_label;
 709         struct efx_tx_queue *tx_queue;
 710         struct efx_nic *efx = channel->efx;
 711         int tx_packets = 0;
 712
 713         if (likely(EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_COMP))) {
 714                 /* Transmit completion */
 715                 tx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_DESC_PTR);
 716                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
 717                 tx_queue = efx_channel_get_tx_queue(
 718                         channel, tx_ev_q_label % EFX_TXQ_TYPES);
 719                 tx_packets = ((tx_ev_desc_ptr - tx_queue->read_count) &
 720                               tx_queue->ptr_mask);
 721                 channel->irq_mod_score += tx_packets;
 722                 efx_xmit_done(tx_queue, tx_ev_desc_ptr);
 723         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_WQ_FF_FULL)) {
 724                 /* Rewrite the FIFO write pointer */
 725                 tx_ev_q_label = EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_Q_LABEL);
 726                 tx_queue = efx_channel_get_tx_queue(
 727                         channel, tx_ev_q_label % EFX_TXQ_TYPES);
 728
 729                 if (efx_dev_registered(efx))
 730                         netif_tx_lock(efx->net_dev);
 731                 efx_notify_tx_desc(tx_queue);
 732                 if (efx_dev_registered(efx))
 733                         netif_tx_unlock(efx->net_dev);
 734         } else if (EFX_QWORD_FIELD(*event, FSF_AZ_TX_EV_PKT_ERR) &&
 735                    EFX_WORKAROUND_10727(efx)) {
 736                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
 737         } else {
 738                 netif_err(efx, tx_err, efx->net_dev,
 739                           "channel %d unexpected TX event "
 740                           EFX_QWORD_FMT"\n", channel->channel,
 741                           EFX_QWORD_VAL(*event));
 742         }
 743
 744         return tx_packets;
 745 }
 746
 747 /* Detect errors included in the rx_evt_pkt_ok bit. */
 748 static void efx_handle_rx_not_ok(struct efx_rx_queue *rx_queue,
 749                                  const efx_qword_t *event,
 750                                  bool *rx_ev_pkt_ok,
 751                                  bool *discard)
 752 {
 753         struct efx_channel *channel = efx_rx_queue_channel(rx_queue);
 754         struct efx_nic *efx = rx_queue->efx;
 755         bool rx_ev_buf_owner_id_err, rx_ev_ip_hdr_chksum_err;
 756         bool rx_ev_tcp_udp_chksum_err, rx_ev_eth_crc_err;
 757         bool rx_ev_frm_trunc, rx_ev_drib_nib, rx_ev_tobe_disc;
 758         bool rx_ev_other_err, rx_ev_pause_frm;
 759         bool rx_ev_hdr_type, rx_ev_mcast_pkt;
 760         unsigned rx_ev_pkt_type;
 761
 762         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
 763         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
 764         rx_ev_tobe_disc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_TOBE_DISC);
 765         rx_ev_pkt_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_TYPE);
 766         rx_ev_buf_owner_id_err = EFX_QWORD_FIELD(*event,
 767                                                  FSF_AZ_RX_EV_BUF_OWNER_ID_ERR);
 768         rx_ev_ip_hdr_chksum_err = EFX_QWORD_FIELD(*event,
 769                                                   FSF_AZ_RX_EV_IP_HDR_CHKSUM_ERR);
 770         rx_ev_tcp_udp_chksum_err = EFX_QWORD_FIELD(*event,
 771                                                    FSF_AZ_RX_EV_TCP_UDP_CHKSUM_ERR);
 772         rx_ev_eth_crc_err = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_ETH_CRC_ERR);
 773         rx_ev_frm_trunc = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_FRM_TRUNC);
 774         rx_ev_drib_nib = ((efx_nic_rev(efx) >= EFX_REV_FALCON_B0) ?
 775                           0 : EFX_QWORD_FIELD(*event, FSF_AA_RX_EV_DRIB_NIB));
 776         rx_ev_pause_frm = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PAUSE_FRM_ERR);
 777
 778         /* Every error apart from tobe_disc and pause_frm */
 779         rx_ev_other_err = (rx_ev_drib_nib | rx_ev_tcp_udp_chksum_err |
 780                            rx_ev_buf_owner_id_err | rx_ev_eth_crc_err |
 781                            rx_ev_frm_trunc | rx_ev_ip_hdr_chksum_err);
 782
 783         /* Count errors that are not in MAC stats.  Ignore expected
 784          * checksum errors during self-test. */
 785         if (rx_ev_frm_trunc)
 786                 ++channel->n_rx_frm_trunc;
 787         else if (rx_ev_tobe_disc)
 788                 ++channel->n_rx_tobe_disc;
 789         else if (!efx->loopback_selftest) {
 790                 if (rx_ev_ip_hdr_chksum_err)
 791                         ++channel->n_rx_ip_hdr_chksum_err;
 792                 else if (rx_ev_tcp_udp_chksum_err)
 793                         ++channel->n_rx_tcp_udp_chksum_err;
 794         }
 795
 796         /* The frame must be discarded if any of these are true. */
 797         *discard = (rx_ev_eth_crc_err | rx_ev_frm_trunc | rx_ev_drib_nib |
 798                     rx_ev_tobe_disc | rx_ev_pause_frm);
 799
 800         /* TOBE_DISC is expected on unicast mismatches; don't print out an
 801          * error message.  FRM_TRUNC indicates RXDP dropped the packet due
 802          * to a FIFO overflow.
 803          */
 804 #ifdef EFX_ENABLE_DEBUG
 805         if (rx_ev_other_err && net_ratelimit()) {
 806                 netif_dbg(efx, rx_err, efx->net_dev,
 807                           " RX queue %d unexpected RX event "
 808                           EFX_QWORD_FMT "%s%s%s%s%s%s%s%s\n",
 809                           efx_rx_queue_index(rx_queue), EFX_QWORD_VAL(*event),
 810                           rx_ev_buf_owner_id_err ? " [OWNER_ID_ERR]" : "",
 811                           rx_ev_ip_hdr_chksum_err ?
 812                           " [IP_HDR_CHKSUM_ERR]" : "",
 813                           rx_ev_tcp_udp_chksum_err ?
 814                           " [TCP_UDP_CHKSUM_ERR]" : "",
 815                           rx_ev_eth_crc_err ? " [ETH_CRC_ERR]" : "",
 816                           rx_ev_frm_trunc ? " [FRM_TRUNC]" : "",
 817                           rx_ev_drib_nib ? " [DRIB_NIB]" : "",
 818                           rx_ev_tobe_disc ? " [TOBE_DISC]" : "",
 819                           rx_ev_pause_frm ? " [PAUSE]" : "");
 820         }
 821 #endif
 822 }
 823
 824 /* Handle receive events that are not in-order. */
 825 static void
 826 efx_handle_rx_bad_index(struct efx_rx_queue *rx_queue, unsigned index)
 827 {
 828         struct efx_nic *efx = rx_queue->efx;
 829         unsigned expected, dropped;
 830
 831         expected = rx_queue->removed_count & rx_queue->ptr_mask;
 832         dropped = (index - expected) & rx_queue->ptr_mask;
 833         netif_info(efx, rx_err, efx->net_dev,
 834                    "dropped %d events (index=%d expected=%d)\n",
 835                    dropped, index, expected);
 836
 837         efx_schedule_reset(efx, EFX_WORKAROUND_5676(efx) ?
 838                            RESET_TYPE_RX_RECOVERY : RESET_TYPE_DISABLE);
 839 }
 840
 841 /* Handle a packet received event
 842  *
 843  * The NIC gives a "discard" flag if it's a unicast packet with the
 844  * wrong destination address
 845  * Also "is multicast" and "matches multicast filter" flags can be used to
 846  * discard non-matching multicast packets.
 847  */
 848 static void
 849 efx_handle_rx_event(struct efx_channel *channel, const efx_qword_t *event)
 850 {
 851         unsigned int rx_ev_desc_ptr, rx_ev_byte_cnt;
 852         unsigned int rx_ev_hdr_type, rx_ev_mcast_pkt;
 853         unsigned expected_ptr;
 854         bool rx_ev_pkt_ok, discard = false, checksummed;
 855         struct efx_rx_queue *rx_queue;
 856
 857         /* Basic packet information */
 858         rx_ev_byte_cnt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_BYTE_CNT);
 859         rx_ev_pkt_ok = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_PKT_OK);
 860         rx_ev_hdr_type = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_HDR_TYPE);
 861         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_JUMBO_CONT));
 862         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_SOP) != 1);
 863         WARN_ON(EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_Q_LABEL) !=
 864                 channel->channel);
 865
 866         rx_queue = efx_channel_get_rx_queue(channel);
 867
 868         rx_ev_desc_ptr = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_DESC_PTR);
 869         expected_ptr = rx_queue->removed_count & rx_queue->ptr_mask;
 870         if (unlikely(rx_ev_desc_ptr != expected_ptr))
 871                 efx_handle_rx_bad_index(rx_queue, rx_ev_desc_ptr);
 872
 873         if (likely(rx_ev_pkt_ok)) {
 874                 /* If packet is marked as OK and packet type is TCP/IP or
 875                  * UDP/IP, then we can rely on the hardware checksum.
 876                  */
 877                 checksummed =
 878                         rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_TCP ||
 879                         rx_ev_hdr_type == FSE_CZ_RX_EV_HDR_TYPE_IPV4V6_UDP;
 880         } else {
 881                 efx_handle_rx_not_ok(rx_queue, event, &rx_ev_pkt_ok, &discard);
 882                 checksummed = false;
 883         }
 884
 885         /* Detect multicast packets that didn't match the filter */
 886         rx_ev_mcast_pkt = EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_PKT);
 887         if (rx_ev_mcast_pkt) {
 888                 unsigned int rx_ev_mcast_hash_match =
 889                         EFX_QWORD_FIELD(*event, FSF_AZ_RX_EV_MCAST_HASH_MATCH);
 890
 891                 if (unlikely(!rx_ev_mcast_hash_match)) {
 892                         ++channel->n_rx_mcast_mismatch;
 893                         discard = true;
 894                 }
 895         }
 896
 897         channel->irq_mod_score += 2;
 898
 899         /* Handle received packet */
 900         efx_rx_packet(rx_queue, rx_ev_desc_ptr, rx_ev_byte_cnt,
 901                       checksummed, discard);
 902 }
 903
 904 static void
 905 efx_handle_generated_event(struct efx_channel *channel, efx_qword_t *event)
 906 {
 907         struct efx_nic *efx = channel->efx;
 908         unsigned code;
 909
 910         code = EFX_QWORD_FIELD(*event, FSF_AZ_DRV_GEN_EV_MAGIC);
 911         if (code == EFX_CHANNEL_MAGIC_TEST(channel))
 912                 ; /* ignore */
 913         else if (code == EFX_CHANNEL_MAGIC_FILL(channel))
 914                 /* The queue must be empty, so we won't receive any rx
 915                  * events, so efx_process_channel() won't refill the
 916                  * queue. Refill it here */
 917                 efx_fast_push_rx_descriptors(efx_channel_get_rx_queue(channel));
 918         else
 919                 netif_dbg(efx, hw, efx->net_dev, "channel %d received "
 920                           "generated event "EFX_QWORD_FMT"\n",
 921                           channel->channel, EFX_QWORD_VAL(*event));
 922 }
 923
 924 static void
 925 efx_handle_driver_event(struct efx_channel *channel, efx_qword_t *event)
 926 {
 927         struct efx_nic *efx = channel->efx;
 928         unsigned int ev_sub_code;
 929         unsigned int ev_sub_data;
 930
 931         ev_sub_code = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBCODE);
 932         ev_sub_data = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
 933
 934         switch (ev_sub_code) {
 935         case FSE_AZ_TX_DESCQ_FLS_DONE_EV:
 936                 netif_vdbg(efx, hw, efx->net_dev, "channel %d TXQ %d flushed\n",
 937                            channel->channel, ev_sub_data);
 938                 break;
 939         case FSE_AZ_RX_DESCQ_FLS_DONE_EV:
 940                 netif_vdbg(efx, hw, efx->net_dev, "channel %d RXQ %d flushed\n",
 941                            channel->channel, ev_sub_data);
 942                 break;
 943         case FSE_AZ_EVQ_INIT_DONE_EV:
 944                 netif_dbg(efx, hw, efx->net_dev,
 945                           "channel %d EVQ %d initialised\n",
 946                           channel->channel, ev_sub_data);
 947                 break;
 948         case FSE_AZ_SRM_UPD_DONE_EV:
 949                 netif_vdbg(efx, hw, efx->net_dev,
 950                            "channel %d SRAM update done\n", channel->channel);
 951                 break;
 952         case FSE_AZ_WAKE_UP_EV:
 953                 netif_vdbg(efx, hw, efx->net_dev,
 954                            "channel %d RXQ %d wakeup event\n",
 955                            channel->channel, ev_sub_data);
 956                 break;
 957         case FSE_AZ_TIMER_EV:
 958                 netif_vdbg(efx, hw, efx->net_dev,
 959                            "channel %d RX queue %d timer expired\n",
 960                            channel->channel, ev_sub_data);
 961                 break;
 962         case FSE_AA_RX_RECOVER_EV:
 963                 netif_err(efx, rx_err, efx->net_dev,
 964                           "channel %d seen DRIVER RX_RESET event. "
 965                         "Resetting.\n", channel->channel);
 966                 atomic_inc(&efx->rx_reset);
 967                 efx_schedule_reset(efx,
 968                                    EFX_WORKAROUND_6555(efx) ?
 969                                    RESET_TYPE_RX_RECOVERY :
 970                                    RESET_TYPE_DISABLE);
 971                 break;
 972         case FSE_BZ_RX_DSC_ERROR_EV:
 973                 netif_err(efx, rx_err, efx->net_dev,
 974                           "RX DMA Q %d reports descriptor fetch error."
 975                           " RX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
 976                 efx_schedule_reset(efx, RESET_TYPE_RX_DESC_FETCH);
 977                 break;
 978         case FSE_BZ_TX_DSC_ERROR_EV:
 979                 netif_err(efx, tx_err, efx->net_dev,
 980                           "TX DMA Q %d reports descriptor fetch error."
 981                           " TX Q %d is disabled.\n", ev_sub_data, ev_sub_data);
 982                 efx_schedule_reset(efx, RESET_TYPE_TX_DESC_FETCH);
 983                 break;
 984         default:
 985                 netif_vdbg(efx, hw, efx->net_dev,
 986                            "channel %d unknown driver event code %d "
 987                            "data %04x\n", channel->channel, ev_sub_code,
 988                            ev_sub_data);
 989                 break;
 990         }
 991 }
 992
 993 int efx_nic_process_eventq(struct efx_channel *channel, int budget)
 994 {
 995         struct efx_nic *efx = channel->efx;
 996         unsigned int read_ptr;
 997         efx_qword_t event, *p_event;
 998         int ev_code;
 999         int tx_packets = 0;
1000         int spent = 0;
1001
1002         read_ptr = channel->eventq_read_ptr;
1003
1004         for (;;) {
1005                 p_event = efx_event(channel, read_ptr);
1006                 event = *p_event;
1007
1008                 if (!efx_event_present(&event))
1009                         /* End of events */
1010                         break;
1011
1012                 netif_vdbg(channel->efx, intr, channel->efx->net_dev,
1013                            "channel %d event is "EFX_QWORD_FMT"\n",
1014                            channel->channel, EFX_QWORD_VAL(event));
1015
1016                 /* Clear this event by marking it all ones */
1017                 EFX_SET_QWORD(*p_event);
1018
1019                 ++read_ptr;
1020
1021                 ev_code = EFX_QWORD_FIELD(event, FSF_AZ_EV_CODE);
1022
1023                 switch (ev_code) {
1024                 case FSE_AZ_EV_CODE_RX_EV:
1025                         efx_handle_rx_event(channel, &event);
1026                         if (++spent == budget)
1027                                 goto out;
1028                         break;
1029                 case FSE_AZ_EV_CODE_TX_EV:
1030                         tx_packets += efx_handle_tx_event(channel, &event);
1031                         if (tx_packets > efx->txq_entries) {
1032                                 spent = budget;
1033                                 goto out;
1034                         }
1035                         break;
1036                 case FSE_AZ_EV_CODE_DRV_GEN_EV:
1037                         efx_handle_generated_event(channel, &event);
1038                         break;
1039                 case FSE_AZ_EV_CODE_DRIVER_EV:
1040                         efx_handle_driver_event(channel, &event);
1041                         break;
1042                 case FSE_CZ_EV_CODE_MCDI_EV:
1043                         efx_mcdi_process_event(channel, &event);
1044                         break;
1045                 case FSE_AZ_EV_CODE_GLOBAL_EV:
1046                         if (efx->type->handle_global_event &&
1047                             efx->type->handle_global_event(channel, &event))
1048                                 break;
1049                         /* else fall through */
1050                 default:
1051                         netif_err(channel->efx, hw, channel->efx->net_dev,
1052                                   "channel %d unknown event type %d (data "
1053                                   EFX_QWORD_FMT ")\n", channel->channel,
1054                                   ev_code, EFX_QWORD_VAL(event));
1055                 }
1056         }
1057
1058 out:
1059         channel->eventq_read_ptr = read_ptr;
1060         return spent;
1061 }
1062
1063 /* Check whether an event is present in the eventq at the current
1064  * read pointer.  Only useful for self-test.
1065  */
1066 bool efx_nic_event_present(struct efx_channel *channel)
1067 {
1068         return efx_event_present(efx_event(channel, channel->eventq_read_ptr));
1069 }
1070
1071 /* Allocate buffer table entries for event queue */
1072 int efx_nic_probe_eventq(struct efx_channel *channel)
1073 {
1074         struct efx_nic *efx = channel->efx;
1075         unsigned entries;
1076
1077         entries = channel->eventq_mask + 1;
1078         return efx_alloc_special_buffer(efx, &channel->eventq,
1079                                         entries * sizeof(efx_qword_t));
1080 }
1081
1082 void efx_nic_init_eventq(struct efx_channel *channel)
1083 {
1084         efx_oword_t reg;
1085         struct efx_nic *efx = channel->efx;
1086
1087         netif_dbg(efx, hw, efx->net_dev,
1088                   "channel %d event queue in special buffers %d-%d\n",
1089                   channel->channel, channel->eventq.index,
1090                   channel->eventq.index + channel->eventq.entries - 1);
1091
1092         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0) {
1093                 EFX_POPULATE_OWORD_3(reg,
1094                                      FRF_CZ_TIMER_Q_EN, 1,
1095                                      FRF_CZ_HOST_NOTIFY_MODE, 0,
1096                                      FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
1097                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1098         }
1099
1100         /* Pin event queue buffer */
1101         efx_init_special_buffer(efx, &channel->eventq);
1102
1103         /* Fill event queue with all ones (i.e. empty events) */
1104         memset(channel->eventq.addr, 0xff, channel->eventq.len);
1105
1106         /* Push event queue to card */
1107         EFX_POPULATE_OWORD_3(reg,
1108                              FRF_AZ_EVQ_EN, 1,
1109                              FRF_AZ_EVQ_SIZE, __ffs(channel->eventq.entries),
1110                              FRF_AZ_EVQ_BUF_BASE_ID, channel->eventq.index);
1111         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1112                          channel->channel);
1113
1114         efx->type->push_irq_moderation(channel);
1115 }
1116
1117 void efx_nic_fini_eventq(struct efx_channel *channel)
1118 {
1119         efx_oword_t reg;
1120         struct efx_nic *efx = channel->efx;
1121
1122         /* Remove event queue from card */
1123         EFX_ZERO_OWORD(reg);
1124         efx_writeo_table(efx, &reg, efx->type->evq_ptr_tbl_base,
1125                          channel->channel);
1126         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1127                 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, channel->channel);
1128
1129         /* Unpin event queue */
1130         efx_fini_special_buffer(efx, &channel->eventq);
1131 }
1132
1133 /* Free buffers backing event queue */
1134 void efx_nic_remove_eventq(struct efx_channel *channel)
1135 {
1136         efx_free_special_buffer(channel->efx, &channel->eventq);
1137 }
1138
1139
1140 void efx_nic_generate_test_event(struct efx_channel *channel)
1141 {
1142         unsigned int magic = EFX_CHANNEL_MAGIC_TEST(channel);
1143         efx_qword_t test_event;
1144
1145         EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1146                              FSE_AZ_EV_CODE_DRV_GEN_EV,
1147                              FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1148         efx_generate_event(channel, &test_event);
1149 }
1150
1151 void efx_nic_generate_fill_event(struct efx_channel *channel)
1152 {
1153         unsigned int magic = EFX_CHANNEL_MAGIC_FILL(channel);
1154         efx_qword_t test_event;
1155
1156         EFX_POPULATE_QWORD_2(test_event, FSF_AZ_EV_CODE,
1157                              FSE_AZ_EV_CODE_DRV_GEN_EV,
1158                              FSF_AZ_DRV_GEN_EV_MAGIC, magic);
1159         efx_generate_event(channel, &test_event);
1160 }
1161
1162 /**************************************************************************
1163  *
1164  * Flush handling
1165  *
1166  **************************************************************************/
1167
1168
1169 static void efx_poll_flush_events(struct efx_nic *efx)
1170 {
1171         struct efx_channel *channel = efx_get_channel(efx, 0);
1172         struct efx_tx_queue *tx_queue;
1173         struct efx_rx_queue *rx_queue;
1174         unsigned int read_ptr = channel->eventq_read_ptr;
1175         unsigned int end_ptr = read_ptr + channel->eventq_mask - 1;
1176
1177         do {
1178                 efx_qword_t *event = efx_event(channel, read_ptr);
1179                 int ev_code, ev_sub_code, ev_queue;
1180                 bool ev_failed;
1181
1182                 if (!efx_event_present(event))
1183                         break;
1184
1185                 ev_code = EFX_QWORD_FIELD(*event, FSF_AZ_EV_CODE);
1186                 ev_sub_code = EFX_QWORD_FIELD(*event,
1187                                               FSF_AZ_DRIVER_EV_SUBCODE);
1188                 if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1189                     ev_sub_code == FSE_AZ_TX_DESCQ_FLS_DONE_EV) {
1190                         ev_queue = EFX_QWORD_FIELD(*event,
1191                                                    FSF_AZ_DRIVER_EV_SUBDATA);
1192                         if (ev_queue < EFX_TXQ_TYPES * efx->n_tx_channels) {
1193                                 tx_queue = efx_get_tx_queue(
1194                                         efx, ev_queue / EFX_TXQ_TYPES,
1195                                         ev_queue % EFX_TXQ_TYPES);
1196                                 tx_queue->flushed = FLUSH_DONE;
1197                         }
1198                 } else if (ev_code == FSE_AZ_EV_CODE_DRIVER_EV &&
1199                            ev_sub_code == FSE_AZ_RX_DESCQ_FLS_DONE_EV) {
1200                         ev_queue = EFX_QWORD_FIELD(
1201                                 *event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1202                         ev_failed = EFX_QWORD_FIELD(
1203                                 *event, FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1204                         if (ev_queue < efx->n_rx_channels) {
1205                                 rx_queue = efx_get_rx_queue(efx, ev_queue);
1206                                 rx_queue->flushed =
1207                                         ev_failed ? FLUSH_FAILED : FLUSH_DONE;
1208                         }
1209                 }
1210
1211                 /* We're about to destroy the queue anyway, so
1212                  * it's ok to throw away every non-flush event */
1213                 EFX_SET_QWORD(*event);
1214
1215                 ++read_ptr;
1216         } while (read_ptr != end_ptr);
1217
1218         channel->eventq_read_ptr = read_ptr;
1219 }
1220
1221 /* Handle tx and rx flushes at the same time, since they run in
1222  * parallel in the hardware and there's no reason for us to
1223  * serialise them */
1224 int efx_nic_flush_queues(struct efx_nic *efx)
1225 {
1226         struct efx_channel *channel;
1227         struct efx_rx_queue *rx_queue;
1228         struct efx_tx_queue *tx_queue;
1229         int i, tx_pending, rx_pending;
1230
1231         /* If necessary prepare the hardware for flushing */
1232         efx->type->prepare_flush(efx);
1233
1234         /* Flush all tx queues in parallel */
1235         efx_for_each_channel(channel, efx) {
1236                 efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1237                         if (tx_queue->initialised)
1238                                 efx_flush_tx_queue(tx_queue);
1239                 }
1240         }
1241
1242         /* The hardware supports four concurrent rx flushes, each of which may
1243          * need to be retried if there is an outstanding descriptor fetch */
1244         for (i = 0; i < EFX_FLUSH_POLL_COUNT; ++i) {
1245                 rx_pending = tx_pending = 0;
1246                 efx_for_each_channel(channel, efx) {
1247                         efx_for_each_channel_rx_queue(rx_queue, channel) {
1248                                 if (rx_queue->flushed == FLUSH_PENDING)
1249                                         ++rx_pending;
1250                         }
1251                 }
1252                 efx_for_each_channel(channel, efx) {
1253                         efx_for_each_channel_rx_queue(rx_queue, channel) {
1254                                 if (rx_pending == EFX_RX_FLUSH_COUNT)
1255                                         break;
1256                                 if (rx_queue->flushed == FLUSH_FAILED ||
1257                                     rx_queue->flushed == FLUSH_NONE) {
1258                                         efx_flush_rx_queue(rx_queue);
1259                                         ++rx_pending;
1260                                 }
1261                         }
1262                         efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1263                                 if (tx_queue->initialised &&
1264                                     tx_queue->flushed != FLUSH_DONE) {
1265                                         efx_oword_t txd_ptr_tbl;
1266
1267                                         efx_reado_table(efx, &txd_ptr_tbl,
1268                                                         FR_BZ_TX_DESC_PTR_TBL,
1269                                                         tx_queue->queue);
1270                                         if (EFX_OWORD_FIELD(txd_ptr_tbl,
1271                                                             FRF_AZ_TX_DESCQ_FLUSH) ||
1272                                             EFX_OWORD_FIELD(txd_ptr_tbl,
1273                                                             FRF_AZ_TX_DESCQ_EN))
1274                                                 ++tx_pending;
1275                                         else
1276                                                 tx_queue->flushed = FLUSH_DONE;
1277                                 }
1278                         }
1279                 }
1280
1281                 if (rx_pending == 0 && tx_pending == 0) {
1282                         efx->type->finish_flush(efx);
1283                         return 0;
1284                 }
1285
1286                 msleep(EFX_FLUSH_INTERVAL);
1287                 efx_poll_flush_events(efx);
1288         }
1289
1290         /* Mark the queues as all flushed. We're going to return failure
1291          * leading to a reset, or fake up success anyway */
1292         efx_for_each_channel(channel, efx) {
1293                 efx_for_each_possible_channel_tx_queue(tx_queue, channel) {
1294                         if (tx_queue->initialised &&
1295                             tx_queue->flushed != FLUSH_DONE)
1296                                 netif_err(efx, hw, efx->net_dev,
1297                                           "tx queue %d flush command timed out\n",
1298                                           tx_queue->queue);
1299                         tx_queue->flushed = FLUSH_DONE;
1300                 }
1301                 efx_for_each_channel_rx_queue(rx_queue, channel) {
1302                         if (rx_queue->flushed != FLUSH_DONE)
1303                                 netif_err(efx, hw, efx->net_dev,
1304                                           "rx queue %d flush command timed out\n",
1305                                           efx_rx_queue_index(rx_queue));
1306                         rx_queue->flushed = FLUSH_DONE;
1307                 }
1308         }
1309
1310         efx->type->finish_flush(efx);
1311         return -ETIMEDOUT;
1312 }
1313
1314 /**************************************************************************
1315  *
1316  * Hardware interrupts
1317  * The hardware interrupt handler does very little work; all the event
1318  * queue processing is carried out by per-channel tasklets.
1319  *
1320  **************************************************************************/
1321
1322 /* Enable/disable/generate interrupts */
1323 static inline void efx_nic_interrupts(struct efx_nic *efx,
1324                                       bool enabled, bool force)
1325 {
1326         efx_oword_t int_en_reg_ker;
1327
1328         EFX_POPULATE_OWORD_3(int_en_reg_ker,
1329                              FRF_AZ_KER_INT_LEVE_SEL, efx->fatal_irq_level,
1330                              FRF_AZ_KER_INT_KER, force,
1331                              FRF_AZ_DRV_INT_EN_KER, enabled);
1332         efx_writeo(efx, &int_en_reg_ker, FR_AZ_INT_EN_KER);
1333 }
1334
1335 void efx_nic_enable_interrupts(struct efx_nic *efx)
1336 {
1337         struct efx_channel *channel;
1338
1339         EFX_ZERO_OWORD(*((efx_oword_t *) efx->irq_status.addr));
1340         wmb(); /* Ensure interrupt vector is clear before interrupts enabled */
1341
1342         /* Enable interrupts */
1343         efx_nic_interrupts(efx, true, false);
1344
1345         /* Force processing of all the channels to get the EVQ RPTRs up to
1346            date */
1347         efx_for_each_channel(channel, efx)
1348                 efx_schedule_channel(channel);
1349 }
1350
1351 void efx_nic_disable_interrupts(struct efx_nic *efx)
1352 {
1353         /* Disable interrupts */
1354         efx_nic_interrupts(efx, false, false);
1355 }
1356
1357 /* Generate a test interrupt
1358  * Interrupt must already have been enabled, otherwise nasty things
1359  * may happen.
1360  */
1361 void efx_nic_generate_interrupt(struct efx_nic *efx)
1362 {
1363         efx_nic_interrupts(efx, true, true);
1364 }
1365
1366 /* Process a fatal interrupt
1367  * Disable bus mastering ASAP and schedule a reset
1368  */
1369 irqreturn_t efx_nic_fatal_interrupt(struct efx_nic *efx)
1370 {
1371         struct falcon_nic_data *nic_data = efx->nic_data;
1372         efx_oword_t *int_ker = efx->irq_status.addr;
1373         efx_oword_t fatal_intr;
1374         int error, mem_perr;
1375
1376         efx_reado(efx, &fatal_intr, FR_AZ_FATAL_INTR_KER);
1377         error = EFX_OWORD_FIELD(fatal_intr, FRF_AZ_FATAL_INTR);
1378
1379         netif_err(efx, hw, efx->net_dev, "SYSTEM ERROR "EFX_OWORD_FMT" status "
1380                   EFX_OWORD_FMT ": %s\n", EFX_OWORD_VAL(*int_ker),
1381                   EFX_OWORD_VAL(fatal_intr),
1382                   error ? "disabling bus mastering" : "no recognised error");
1383
1384         /* If this is a memory parity error dump which blocks are offending */
1385         mem_perr = (EFX_OWORD_FIELD(fatal_intr, FRF_AZ_MEM_PERR_INT_KER) ||
1386                     EFX_OWORD_FIELD(fatal_intr, FRF_AZ_SRM_PERR_INT_KER));
1387         if (mem_perr) {
1388                 efx_oword_t reg;
1389                 efx_reado(efx, &reg, FR_AZ_MEM_STAT);
1390                 netif_err(efx, hw, efx->net_dev,
1391                           "SYSTEM ERROR: memory parity error "EFX_OWORD_FMT"\n",
1392                           EFX_OWORD_VAL(reg));
1393         }
1394
1395         /* Disable both devices */
1396         pci_clear_master(efx->pci_dev);
1397         if (efx_nic_is_dual_func(efx))
1398                 pci_clear_master(nic_data->pci_dev2);
1399         efx_nic_disable_interrupts(efx);
1400
1401         /* Count errors and reset or disable the NIC accordingly */
1402         if (efx->int_error_count == 0 ||
1403             time_after(jiffies, efx->int_error_expire)) {
1404                 efx->int_error_count = 0;
1405                 efx->int_error_expire =
1406                         jiffies + EFX_INT_ERROR_EXPIRE * HZ;
1407         }
1408         if (++efx->int_error_count < EFX_MAX_INT_ERRORS) {
1409                 netif_err(efx, hw, efx->net_dev,
1410                           "SYSTEM ERROR - reset scheduled\n");
1411                 efx_schedule_reset(efx, RESET_TYPE_INT_ERROR);
1412         } else {
1413                 netif_err(efx, hw, efx->net_dev,
1414                           "SYSTEM ERROR - max number of errors seen."
1415                           "NIC will be disabled\n");
1416                 efx_schedule_reset(efx, RESET_TYPE_DISABLE);
1417         }
1418
1419         return IRQ_HANDLED;
1420 }
1421
1422 /* Handle a legacy interrupt
1423  * Acknowledges the interrupt and schedule event queue processing.
1424  */
1425 static irqreturn_t efx_legacy_interrupt(int irq, void *dev_id)
1426 {
1427         struct efx_nic *efx = dev_id;
1428         efx_oword_t *int_ker = efx->irq_status.addr;
1429         irqreturn_t result = IRQ_NONE;
1430         struct efx_channel *channel;
1431         efx_dword_t reg;
1432         u32 queues;
1433         int syserr;
1434
1435         /* Could this be ours?  If interrupts are disabled then the
1436          * channel state may not be valid.
1437          */
1438         if (!efx->legacy_irq_enabled)
1439                 return result;
1440
1441         /* Read the ISR which also ACKs the interrupts */
1442         efx_readd(efx, &reg, FR_BZ_INT_ISR0);
1443         queues = EFX_EXTRACT_DWORD(reg, 0, 31);
1444
1445         /* Check to see if we have a serious error condition */
1446         if (queues & (1U << efx->fatal_irq_level)) {
1447                 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1448                 if (unlikely(syserr))
1449                         return efx_nic_fatal_interrupt(efx);
1450         }
1451
1452         if (queues != 0) {
1453                 if (EFX_WORKAROUND_15783(efx))
1454                         efx->irq_zero_count = 0;
1455
1456                 /* Schedule processing of any interrupting queues */
1457                 efx_for_each_channel(channel, efx) {
1458                         if (queues & 1)
1459                                 efx_schedule_channel(channel);
1460                         queues >>= 1;
1461                 }
1462                 result = IRQ_HANDLED;
1463
1464         } else if (EFX_WORKAROUND_15783(efx)) {
1465                 efx_qword_t *event;
1466
1467                 /* We can't return IRQ_HANDLED more than once on seeing ISR=0
1468                  * because this might be a shared interrupt. */
1469                 if (efx->irq_zero_count++ == 0)
1470                         result = IRQ_HANDLED;
1471
1472                 /* Ensure we schedule or rearm all event queues */
1473                 efx_for_each_channel(channel, efx) {
1474                         event = efx_event(channel, channel->eventq_read_ptr);
1475                         if (efx_event_present(event))
1476                                 efx_schedule_channel(channel);
1477                         else
1478                                 efx_nic_eventq_read_ack(channel);
1479                 }
1480         }
1481
1482         if (result == IRQ_HANDLED) {
1483                 efx->last_irq_cpu = raw_smp_processor_id();
1484                 netif_vdbg(efx, intr, efx->net_dev,
1485                            "IRQ %d on CPU %d status " EFX_DWORD_FMT "\n",
1486                            irq, raw_smp_processor_id(), EFX_DWORD_VAL(reg));
1487         }
1488
1489         return result;
1490 }
1491
1492 /* Handle an MSI interrupt
1493  *
1494  * Handle an MSI hardware interrupt.  This routine schedules event
1495  * queue processing.  No interrupt acknowledgement cycle is necessary.
1496  * Also, we never need to check that the interrupt is for us, since
1497  * MSI interrupts cannot be shared.
1498  */
1499 static irqreturn_t efx_msi_interrupt(int irq, void *dev_id)
1500 {
1501         struct efx_channel *channel = *(struct efx_channel **)dev_id;
1502         struct efx_nic *efx = channel->efx;
1503         efx_oword_t *int_ker = efx->irq_status.addr;
1504         int syserr;
1505
1506         efx->last_irq_cpu = raw_smp_processor_id();
1507         netif_vdbg(efx, intr, efx->net_dev,
1508                    "IRQ %d on CPU %d status " EFX_OWORD_FMT "\n",
1509                    irq, raw_smp_processor_id(), EFX_OWORD_VAL(*int_ker));
1510
1511         /* Check to see if we have a serious error condition */
1512         if (channel->channel == efx->fatal_irq_level) {
1513                 syserr = EFX_OWORD_FIELD(*int_ker, FSF_AZ_NET_IVEC_FATAL_INT);
1514                 if (unlikely(syserr))
1515                         return efx_nic_fatal_interrupt(efx);
1516         }
1517
1518         /* Schedule processing of the channel */
1519         efx_schedule_channel(channel);
1520
1521         return IRQ_HANDLED;
1522 }
1523
1524
1525 /* Setup RSS indirection table.
1526  * This maps from the hash value of the packet to RXQ
1527  */
1528 void efx_nic_push_rx_indir_table(struct efx_nic *efx)
1529 {
1530         size_t i = 0;
1531         efx_dword_t dword;
1532
1533         if (efx_nic_rev(efx) < EFX_REV_FALCON_B0)
1534                 return;
1535
1536         BUILD_BUG_ON(ARRAY_SIZE(efx->rx_indir_table) !=
1537                      FR_BZ_RX_INDIRECTION_TBL_ROWS);
1538
1539         for (i = 0; i < FR_BZ_RX_INDIRECTION_TBL_ROWS; i++) {
1540                 EFX_POPULATE_DWORD_1(dword, FRF_BZ_IT_QUEUE,
1541                                      efx->rx_indir_table[i]);
1542                 efx_writed_table(efx, &dword, FR_BZ_RX_INDIRECTION_TBL, i);
1543         }
1544 }
1545
1546 /* Hook interrupt handler(s)
1547  * Try MSI and then legacy interrupts.
1548  */
1549 int efx_nic_init_interrupt(struct efx_nic *efx)
1550 {
1551         struct efx_channel *channel;
1552         int rc;
1553
1554         if (!EFX_INT_MODE_USE_MSI(efx)) {
1555                 irq_handler_t handler;
1556                 if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1557                         handler = efx_legacy_interrupt;
1558                 else
1559                         handler = falcon_legacy_interrupt_a1;
1560
1561                 rc = request_irq(efx->legacy_irq, handler, IRQF_SHARED,
1562                                  efx->name, efx);
1563                 if (rc) {
1564                         netif_err(efx, drv, efx->net_dev,
1565                                   "failed to hook legacy IRQ %d\n",
1566                                   efx->pci_dev->irq);
1567                         goto fail1;
1568                 }
1569                 return 0;
1570         }
1571
1572         /* Hook MSI or MSI-X interrupt */
1573         efx_for_each_channel(channel, efx) {
1574                 rc = request_irq(channel->irq, efx_msi_interrupt,
1575                                  IRQF_PROBE_SHARED, /* Not shared */
1576                                  efx->channel_name[channel->channel],
1577                                  &efx->channel[channel->channel]);
1578                 if (rc) {
1579                         netif_err(efx, drv, efx->net_dev,
1580                                   "failed to hook IRQ %d\n", channel->irq);
1581                         goto fail2;
1582                 }
1583         }
1584
1585         return 0;
1586
1587  fail2:
1588         efx_for_each_channel(channel, efx)
1589                 free_irq(channel->irq, &efx->channel[channel->channel]);
1590  fail1:
1591         return rc;
1592 }
1593
1594 void efx_nic_fini_interrupt(struct efx_nic *efx)
1595 {
1596         struct efx_channel *channel;
1597         efx_oword_t reg;
1598
1599         /* Disable MSI/MSI-X interrupts */
1600         efx_for_each_channel(channel, efx) {
1601                 if (channel->irq)
1602                         free_irq(channel->irq, &efx->channel[channel->channel]);
1603         }
1604
1605         /* ACK legacy interrupt */
1606         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1607                 efx_reado(efx, &reg, FR_BZ_INT_ISR0);
1608         else
1609                 falcon_irq_ack_a1(efx);
1610
1611         /* Disable legacy interrupt */
1612         if (efx->legacy_irq)
1613                 free_irq(efx->legacy_irq, efx);
1614 }
1615
1616 u32 efx_nic_fpga_ver(struct efx_nic *efx)
1617 {
1618         efx_oword_t altera_build;
1619         efx_reado(efx, &altera_build, FR_AZ_ALTERA_BUILD);
1620         return EFX_OWORD_FIELD(altera_build, FRF_AZ_ALTERA_BUILD_VER);
1621 }
1622
1623 void efx_nic_init_common(struct efx_nic *efx)
1624 {
1625         efx_oword_t temp;
1626
1627         /* Set positions of descriptor caches in SRAM. */
1628         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_TX_DC_BASE_ADR,
1629                              efx->type->tx_dc_base / 8);
1630         efx_writeo(efx, &temp, FR_AZ_SRM_TX_DC_CFG);
1631         EFX_POPULATE_OWORD_1(temp, FRF_AZ_SRM_RX_DC_BASE_ADR,
1632                              efx->type->rx_dc_base / 8);
1633         efx_writeo(efx, &temp, FR_AZ_SRM_RX_DC_CFG);
1634
1635         /* Set TX descriptor cache size. */
1636         BUILD_BUG_ON(TX_DC_ENTRIES != (8 << TX_DC_ENTRIES_ORDER));
1637         EFX_POPULATE_OWORD_1(temp, FRF_AZ_TX_DC_SIZE, TX_DC_ENTRIES_ORDER);
1638         efx_writeo(efx, &temp, FR_AZ_TX_DC_CFG);
1639
1640         /* Set RX descriptor cache size.  Set low watermark to size-8, as
1641          * this allows most efficient prefetching.
1642          */
1643         BUILD_BUG_ON(RX_DC_ENTRIES != (8 << RX_DC_ENTRIES_ORDER));
1644         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_SIZE, RX_DC_ENTRIES_ORDER);
1645         efx_writeo(efx, &temp, FR_AZ_RX_DC_CFG);
1646         EFX_POPULATE_OWORD_1(temp, FRF_AZ_RX_DC_PF_LWM, RX_DC_ENTRIES - 8);
1647         efx_writeo(efx, &temp, FR_AZ_RX_DC_PF_WM);
1648
1649         /* Program INT_KER address */
1650         EFX_POPULATE_OWORD_2(temp,
1651                              FRF_AZ_NORM_INT_VEC_DIS_KER,
1652                              EFX_INT_MODE_USE_MSI(efx),
1653                              FRF_AZ_INT_ADR_KER, efx->irq_status.dma_addr);
1654         efx_writeo(efx, &temp, FR_AZ_INT_ADR_KER);
1655
1656         if (EFX_WORKAROUND_17213(efx) && !EFX_INT_MODE_USE_MSI(efx))
1657                 /* Use an interrupt level unused by event queues */
1658                 efx->fatal_irq_level = 0x1f;
1659         else
1660                 /* Use a valid MSI-X vector */
1661                 efx->fatal_irq_level = 0;
1662
1663         /* Enable all the genuinely fatal interrupts.  (They are still
1664          * masked by the overall interrupt mask, controlled by
1665          * falcon_interrupts()).
1666          *
1667          * Note: All other fatal interrupts are enabled
1668          */
1669         EFX_POPULATE_OWORD_3(temp,
1670                              FRF_AZ_ILL_ADR_INT_KER_EN, 1,
1671                              FRF_AZ_RBUF_OWN_INT_KER_EN, 1,
1672                              FRF_AZ_TBUF_OWN_INT_KER_EN, 1);
1673         if (efx_nic_rev(efx) >= EFX_REV_SIENA_A0)
1674                 EFX_SET_OWORD_FIELD(temp, FRF_CZ_SRAM_PERR_INT_P_KER_EN, 1);
1675         EFX_INVERT_OWORD(temp);
1676         efx_writeo(efx, &temp, FR_AZ_FATAL_INTR_KER);
1677
1678         efx_nic_push_rx_indir_table(efx);
1679
1680         /* Disable the ugly timer-based TX DMA backoff and allow TX DMA to be
1681          * controlled by the RX FIFO fill level. Set arbitration to one pkt/Q.
1682          */
1683         efx_reado(efx, &temp, FR_AZ_TX_RESERVED);
1684         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER, 0xfe);
1685         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_RX_SPACER_EN, 1);
1686         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_ONE_PKT_PER_Q, 1);
1687         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PUSH_EN, 1);
1688         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_DIS_NON_IP_EV, 1);
1689         /* Enable SW_EV to inherit in char driver - assume harmless here */
1690         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_SOFT_EVT_EN, 1);
1691         /* Prefetch threshold 2 => fetch when descriptor cache half empty */
1692         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_THRESHOLD, 2);
1693         /* Disable hardware watchdog which can misfire */
1694         EFX_SET_OWORD_FIELD(temp, FRF_AZ_TX_PREF_WD_TMR, 0x3fffff);
1695         /* Squash TX of packets of 16 bytes or less */
1696         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0)
1697                 EFX_SET_OWORD_FIELD(temp, FRF_BZ_TX_FLUSH_MIN_LEN_EN, 1);
1698         efx_writeo(efx, &temp, FR_AZ_TX_RESERVED);
1699
1700         if (efx_nic_rev(efx) >= EFX_REV_FALCON_B0) {
1701                 EFX_POPULATE_OWORD_4(temp,
1702                                      /* Default values */
1703                                      FRF_BZ_TX_PACE_SB_NOT_AF, 0x15,
1704                                      FRF_BZ_TX_PACE_SB_AF, 0xb,
1705                                      FRF_BZ_TX_PACE_FB_BASE, 0,
1706                                      /* Allow large pace values in the
1707                                       * fast bin. */
1708                                      FRF_BZ_TX_PACE_BIN_TH,
1709                                      FFE_BZ_TX_PACE_RESERVED);
1710                 efx_writeo(efx, &temp, FR_BZ_TX_PACE);
1711         }
1712 }
1713
1714 /* Register dump */
1715
1716 #define REGISTER_REVISION_A     1
1717 #define REGISTER_REVISION_B     2
1718 #define REGISTER_REVISION_C     3
1719 #define REGISTER_REVISION_Z     3       /* latest revision */
1720
1721 struct efx_nic_reg {
1722         u32 offset:24;
1723         u32 min_revision:2, max_revision:2;
1724 };
1725
1726 #define REGISTER(name, min_rev, max_rev) {                              \
1727         FR_ ## min_rev ## max_rev ## _ ## name,                         \
1728         REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev    \
1729 }
1730 #define REGISTER_AA(name) REGISTER(name, A, A)
1731 #define REGISTER_AB(name) REGISTER(name, A, B)
1732 #define REGISTER_AZ(name) REGISTER(name, A, Z)
1733 #define REGISTER_BB(name) REGISTER(name, B, B)
1734 #define REGISTER_BZ(name) REGISTER(name, B, Z)
1735 #define REGISTER_CZ(name) REGISTER(name, C, Z)
1736
1737 static const struct efx_nic_reg efx_nic_regs[] = {
1738         REGISTER_AZ(ADR_REGION),
1739         REGISTER_AZ(INT_EN_KER),
1740         REGISTER_BZ(INT_EN_CHAR),
1741         REGISTER_AZ(INT_ADR_KER),
1742         REGISTER_BZ(INT_ADR_CHAR),
1743         /* INT_ACK_KER is WO */
1744         /* INT_ISR0 is RC */
1745         REGISTER_AZ(HW_INIT),
1746         REGISTER_CZ(USR_EV_CFG),
1747         REGISTER_AB(EE_SPI_HCMD),
1748         REGISTER_AB(EE_SPI_HADR),
1749         REGISTER_AB(EE_SPI_HDATA),
1750         REGISTER_AB(EE_BASE_PAGE),
1751         REGISTER_AB(EE_VPD_CFG0),
1752         /* EE_VPD_SW_CNTL and EE_VPD_SW_DATA are not used */
1753         /* PMBX_DBG_IADDR and PBMX_DBG_IDATA are indirect */
1754         /* PCIE_CORE_INDIRECT is indirect */
1755         REGISTER_AB(NIC_STAT),
1756         REGISTER_AB(GPIO_CTL),
1757         REGISTER_AB(GLB_CTL),
1758         /* FATAL_INTR_KER and FATAL_INTR_CHAR are partly RC */
1759         REGISTER_BZ(DP_CTRL),
1760         REGISTER_AZ(MEM_STAT),
1761         REGISTER_AZ(CS_DEBUG),
1762         REGISTER_AZ(ALTERA_BUILD),
1763         REGISTER_AZ(CSR_SPARE),
1764         REGISTER_AB(PCIE_SD_CTL0123),
1765         REGISTER_AB(PCIE_SD_CTL45),
1766         REGISTER_AB(PCIE_PCS_CTL_STAT),
1767         /* DEBUG_DATA_OUT is not used */
1768         /* DRV_EV is WO */
1769         REGISTER_AZ(EVQ_CTL),
1770         REGISTER_AZ(EVQ_CNT1),
1771         REGISTER_AZ(EVQ_CNT2),
1772         REGISTER_AZ(BUF_TBL_CFG),
1773         REGISTER_AZ(SRM_RX_DC_CFG),
1774         REGISTER_AZ(SRM_TX_DC_CFG),
1775         REGISTER_AZ(SRM_CFG),
1776         /* BUF_TBL_UPD is WO */
1777         REGISTER_AZ(SRM_UPD_EVQ),
1778         REGISTER_AZ(SRAM_PARITY),
1779         REGISTER_AZ(RX_CFG),
1780         REGISTER_BZ(RX_FILTER_CTL),
1781         /* RX_FLUSH_DESCQ is WO */
1782         REGISTER_AZ(RX_DC_CFG),
1783         REGISTER_AZ(RX_DC_PF_WM),
1784         REGISTER_BZ(RX_RSS_TKEY),
1785         /* RX_NODESC_DROP is RC */
1786         REGISTER_AA(RX_SELF_RST),
1787         /* RX_DEBUG, RX_PUSH_DROP are not used */
1788         REGISTER_CZ(RX_RSS_IPV6_REG1),
1789         REGISTER_CZ(RX_RSS_IPV6_REG2),
1790         REGISTER_CZ(RX_RSS_IPV6_REG3),
1791         /* TX_FLUSH_DESCQ is WO */
1792         REGISTER_AZ(TX_DC_CFG),
1793         REGISTER_AA(TX_CHKSM_CFG),
1794         REGISTER_AZ(TX_CFG),
1795         /* TX_PUSH_DROP is not used */
1796         REGISTER_AZ(TX_RESERVED),
1797         REGISTER_BZ(TX_PACE),
1798         /* TX_PACE_DROP_QID is RC */
1799         REGISTER_BB(TX_VLAN),
1800         REGISTER_BZ(TX_IPFIL_PORTEN),
1801         REGISTER_AB(MD_TXD),
1802         REGISTER_AB(MD_RXD),
1803         REGISTER_AB(MD_CS),
1804         REGISTER_AB(MD_PHY_ADR),
1805         REGISTER_AB(MD_ID),
1806         /* MD_STAT is RC */
1807         REGISTER_AB(MAC_STAT_DMA),
1808         REGISTER_AB(MAC_CTRL),
1809         REGISTER_BB(GEN_MODE),
1810         REGISTER_AB(MAC_MC_HASH_REG0),
1811         REGISTER_AB(MAC_MC_HASH_REG1),
1812         REGISTER_AB(GM_CFG1),
1813         REGISTER_AB(GM_CFG2),
1814         /* GM_IPG and GM_HD are not used */
1815         REGISTER_AB(GM_MAX_FLEN),
1816         /* GM_TEST is not used */
1817         REGISTER_AB(GM_ADR1),
1818         REGISTER_AB(GM_ADR2),
1819         REGISTER_AB(GMF_CFG0),
1820         REGISTER_AB(GMF_CFG1),
1821         REGISTER_AB(GMF_CFG2),
1822         REGISTER_AB(GMF_CFG3),
1823         REGISTER_AB(GMF_CFG4),
1824         REGISTER_AB(GMF_CFG5),
1825         REGISTER_BB(TX_SRC_MAC_CTL),
1826         REGISTER_AB(XM_ADR_LO),
1827         REGISTER_AB(XM_ADR_HI),
1828         REGISTER_AB(XM_GLB_CFG),
1829         REGISTER_AB(XM_TX_CFG),
1830         REGISTER_AB(XM_RX_CFG),
1831         REGISTER_AB(XM_MGT_INT_MASK),
1832         REGISTER_AB(XM_FC),
1833         REGISTER_AB(XM_PAUSE_TIME),
1834         REGISTER_AB(XM_TX_PARAM),
1835         REGISTER_AB(XM_RX_PARAM),
1836         /* XM_MGT_INT_MSK (note no 'A') is RC */
1837         REGISTER_AB(XX_PWR_RST),
1838         REGISTER_AB(XX_SD_CTL),
1839         REGISTER_AB(XX_TXDRV_CTL),
1840         /* XX_PRBS_CTL, XX_PRBS_CHK and XX_PRBS_ERR are not used */
1841         /* XX_CORE_STAT is partly RC */
1842 };
1843
1844 struct efx_nic_reg_table {
1845         u32 offset:24;
1846         u32 min_revision:2, max_revision:2;
1847         u32 step:6, rows:21;
1848 };
1849
1850 #define REGISTER_TABLE_DIMENSIONS(_, offset, min_rev, max_rev, step, rows) { \
1851         offset,                                                         \
1852         REGISTER_REVISION_ ## min_rev, REGISTER_REVISION_ ## max_rev,   \
1853         step, rows                                                      \
1854 }
1855 #define REGISTER_TABLE(name, min_rev, max_rev)                          \
1856         REGISTER_TABLE_DIMENSIONS(                                      \
1857                 name, FR_ ## min_rev ## max_rev ## _ ## name,           \
1858                 min_rev, max_rev,                                       \
1859                 FR_ ## min_rev ## max_rev ## _ ## name ## _STEP,        \
1860                 FR_ ## min_rev ## max_rev ## _ ## name ## _ROWS)
1861 #define REGISTER_TABLE_AA(name) REGISTER_TABLE(name, A, A)
1862 #define REGISTER_TABLE_AZ(name) REGISTER_TABLE(name, A, Z)
1863 #define REGISTER_TABLE_BB(name) REGISTER_TABLE(name, B, B)
1864 #define REGISTER_TABLE_BZ(name) REGISTER_TABLE(name, B, Z)
1865 #define REGISTER_TABLE_BB_CZ(name)                                      \
1866         REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, B, B,           \
1867                                   FR_BZ_ ## name ## _STEP,              \
1868                                   FR_BB_ ## name ## _ROWS),             \
1869         REGISTER_TABLE_DIMENSIONS(name, FR_BZ_ ## name, C, Z,           \
1870                                   FR_BZ_ ## name ## _STEP,              \
1871                                   FR_CZ_ ## name ## _ROWS)
1872 #define REGISTER_TABLE_CZ(name) REGISTER_TABLE(name, C, Z)
1873
1874 static const struct efx_nic_reg_table efx_nic_reg_tables[] = {
1875         /* DRIVER is not used */
1876         /* EVQ_RPTR, TIMER_COMMAND, USR_EV and {RX,TX}_DESC_UPD are WO */
1877         REGISTER_TABLE_BB(TX_IPFIL_TBL),
1878         REGISTER_TABLE_BB(TX_SRC_MAC_TBL),
1879         REGISTER_TABLE_AA(RX_DESC_PTR_TBL_KER),
1880         REGISTER_TABLE_BB_CZ(RX_DESC_PTR_TBL),
1881         REGISTER_TABLE_AA(TX_DESC_PTR_TBL_KER),
1882         REGISTER_TABLE_BB_CZ(TX_DESC_PTR_TBL),
1883         REGISTER_TABLE_AA(EVQ_PTR_TBL_KER),
1884         REGISTER_TABLE_BB_CZ(EVQ_PTR_TBL),
1885         /* We can't reasonably read all of the buffer table (up to 8MB!).
1886          * However this driver will only use a few entries.  Reading
1887          * 1K entries allows for some expansion of queue count and
1888          * size before we need to change the version. */
1889         REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL_KER, FR_AA_BUF_FULL_TBL_KER,
1890                                   A, A, 8, 1024),
1891         REGISTER_TABLE_DIMENSIONS(BUF_FULL_TBL, FR_BZ_BUF_FULL_TBL,
1892                                   B, Z, 8, 1024),
1893         REGISTER_TABLE_CZ(RX_MAC_FILTER_TBL0),
1894         REGISTER_TABLE_BB_CZ(TIMER_TBL),
1895         REGISTER_TABLE_BB_CZ(TX_PACE_TBL),
1896         REGISTER_TABLE_BZ(RX_INDIRECTION_TBL),
1897         /* TX_FILTER_TBL0 is huge and not used by this driver */
1898         REGISTER_TABLE_CZ(TX_MAC_FILTER_TBL0),
1899         REGISTER_TABLE_CZ(MC_TREG_SMEM),
1900         /* MSIX_PBA_TABLE is not mapped */
1901         /* SRM_DBG is not mapped (and is redundant with BUF_FLL_TBL) */
1902         REGISTER_TABLE_BZ(RX_FILTER_TBL0),
1903 };
1904
1905 size_t efx_nic_get_regs_len(struct efx_nic *efx)
1906 {
1907         const struct efx_nic_reg *reg;
1908         const struct efx_nic_reg_table *table;
1909         size_t len = 0;
1910
1911         for (reg = efx_nic_regs;
1912              reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1913              reg++)
1914                 if (efx->type->revision >= reg->min_revision &&
1915                     efx->type->revision <= reg->max_revision)
1916                         len += sizeof(efx_oword_t);
1917
1918         for (table = efx_nic_reg_tables;
1919              table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1920              table++)
1921                 if (efx->type->revision >= table->min_revision &&
1922                     efx->type->revision <= table->max_revision)
1923                         len += table->rows * min_t(size_t, table->step, 16);
1924
1925         return len;
1926 }
1927
1928 void efx_nic_get_regs(struct efx_nic *efx, void *buf)
1929 {
1930         const struct efx_nic_reg *reg;
1931         const struct efx_nic_reg_table *table;
1932
1933         for (reg = efx_nic_regs;
1934              reg < efx_nic_regs + ARRAY_SIZE(efx_nic_regs);
1935              reg++) {
1936                 if (efx->type->revision >= reg->min_revision &&
1937                     efx->type->revision <= reg->max_revision) {
1938                         efx_reado(efx, (efx_oword_t *)buf, reg->offset);
1939                         buf += sizeof(efx_oword_t);
1940                 }
1941         }
1942
1943         for (table = efx_nic_reg_tables;
1944              table < efx_nic_reg_tables + ARRAY_SIZE(efx_nic_reg_tables);
1945              table++) {
1946                 size_t size, i;
1947
1948                 if (!(efx->type->revision >= table->min_revision &&
1949                       efx->type->revision <= table->max_revision))
1950                         continue;
1951
1952                 size = min_t(size_t, table->step, 16);
1953
1954                 for (i = 0; i < table->rows; i++) {
1955                         switch (table->step) {
1956                         case 4: /* 32-bit register or SRAM */
1957                                 efx_readd_table(efx, buf, table->offset, i);
1958                                 break;
1959                         case 8: /* 64-bit SRAM */
1960                                 efx_sram_readq(efx,
1961                                                efx->membase + table->offset,
1962                                                buf, i);
1963                                 break;
1964                         case 16: /* 128-bit register */
1965                                 efx_reado_table(efx, buf, table->offset, i);
1966                                 break;
1967                         case 32: /* 128-bit register, interleaved */
1968                                 efx_reado_table(efx, buf, table->offset, 2 * i);
1969                                 break;
1970                         default:
1971                                 WARN_ON(1);
1972                                 return;
1973                         }
1974                         buf += size;
1975                 }
1976         }
1977 }