drivers/thunderbolt/nhi.c

   1 /*
   2  * Thunderbolt Cactus Ridge driver - NHI driver
   3  *
   4  * The NHI (native host interface) is the pci device that allows us to send and
   5  * receive frames from the thunderbolt bus.
   6  *
   7  * Copyright (c) 2014 Andreas Noever <andreas.noever@gmail.com>
   8  */
   9
  10 #include <linux/pm_runtime.h>
  11 #include <linux/slab.h>
  12 #include <linux/errno.h>
  13 #include <linux/pci.h>
  14 #include <linux/interrupt.h>
  15 #include <linux/module.h>
  16 #include <linux/dmi.h>
  17
  18 #include "nhi.h"
  19 #include "nhi_regs.h"
  20 #include "tb.h"
  21
  22 #define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
  23
  24 /*
  25  * Minimal number of vectors when we use MSI-X. Two for control channel
  26  * Rx/Tx and the rest four are for cross domain DMA paths.
  27  */
  28 #define MSIX_MIN_VECS           6
  29 #define MSIX_MAX_VECS           16
  30
  31 static int ring_interrupt_index(struct tb_ring *ring)
  32 {
  33         int bit = ring->hop;
  34         if (!ring->is_tx)
  35                 bit += ring->nhi->hop_count;
  36         return bit;
  37 }
  38
  39 /**
  40  * ring_interrupt_active() - activate/deactivate interrupts for a single ring
  41  *
  42  * ring->nhi->lock must be held.
  43  */
  44 static void ring_interrupt_active(struct tb_ring *ring, bool active)
  45 {
  46         int reg = REG_RING_INTERRUPT_BASE +
  47                   ring_interrupt_index(ring) / 32 * 4;
  48         int bit = ring_interrupt_index(ring) & 31;
  49         int mask = 1 << bit;
  50         u32 old, new;
  51
  52         if (ring->irq > 0) {
  53                 u32 step, shift, ivr, misc;
  54                 void __iomem *ivr_base;
  55                 int index;
  56
  57                 if (ring->is_tx)
  58                         index = ring->hop;
  59                 else
  60                         index = ring->hop + ring->nhi->hop_count;
  61
  62                 /*
  63                  * Ask the hardware to clear interrupt status bits automatically
  64                  * since we already know which interrupt was triggered.
  65                  */
  66                 misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
  67                 if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
  68                         misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
  69                         iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
  70                 }
  71
  72                 ivr_base = ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE;
  73                 step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
  74                 shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
  75                 ivr = ioread32(ivr_base + step);
  76                 ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
  77                 if (active)
  78                         ivr |= ring->vector << shift;
  79                 iowrite32(ivr, ivr_base + step);
  80         }
  81
  82         old = ioread32(ring->nhi->iobase + reg);
  83         if (active)
  84                 new = old | mask;
  85         else
  86                 new = old & ~mask;
  87
  88         dev_info(&ring->nhi->pdev->dev,
  89                  "%s interrupt at register %#x bit %d (%#x -> %#x)\n",
  90                  active ? "enabling" : "disabling", reg, bit, old, new);
  91
  92         if (new == old)
  93                 dev_WARN(&ring->nhi->pdev->dev,
  94                                          "interrupt for %s %d is already %s\n",
  95                                          RING_TYPE(ring), ring->hop,
  96                                          active ? "enabled" : "disabled");
  97         iowrite32(new, ring->nhi->iobase + reg);
  98 }
  99
 100 /**
 101  * nhi_disable_interrupts() - disable interrupts for all rings
 102  *
 103  * Use only during init and shutdown.
 104  */
 105 static void nhi_disable_interrupts(struct tb_nhi *nhi)
 106 {
 107         int i = 0;
 108         /* disable interrupts */
 109         for (i = 0; i < RING_INTERRUPT_REG_COUNT(nhi); i++)
 110                 iowrite32(0, nhi->iobase + REG_RING_INTERRUPT_BASE + 4 * i);
 111
 112         /* clear interrupt status bits */
 113         for (i = 0; i < RING_NOTIFY_REG_COUNT(nhi); i++)
 114                 ioread32(nhi->iobase + REG_RING_NOTIFY_BASE + 4 * i);
 115 }
 116
 117 /* ring helper methods */
 118
 119 static void __iomem *ring_desc_base(struct tb_ring *ring)
 120 {
 121         void __iomem *io = ring->nhi->iobase;
 122         io += ring->is_tx ? REG_TX_RING_BASE : REG_RX_RING_BASE;
 123         io += ring->hop * 16;
 124         return io;
 125 }
 126
 127 static void __iomem *ring_options_base(struct tb_ring *ring)
 128 {
 129         void __iomem *io = ring->nhi->iobase;
 130         io += ring->is_tx ? REG_TX_OPTIONS_BASE : REG_RX_OPTIONS_BASE;
 131         io += ring->hop * 32;
 132         return io;
 133 }
 134
 135 static void ring_iowrite16desc(struct tb_ring *ring, u32 value, u32 offset)
 136 {
 137         iowrite16(value, ring_desc_base(ring) + offset);
 138 }
 139
 140 static void ring_iowrite32desc(struct tb_ring *ring, u32 value, u32 offset)
 141 {
 142         iowrite32(value, ring_desc_base(ring) + offset);
 143 }
 144
 145 static void ring_iowrite64desc(struct tb_ring *ring, u64 value, u32 offset)
 146 {
 147         iowrite32(value, ring_desc_base(ring) + offset);
 148         iowrite32(value >> 32, ring_desc_base(ring) + offset + 4);
 149 }
 150
 151 static void ring_iowrite32options(struct tb_ring *ring, u32 value, u32 offset)
 152 {
 153         iowrite32(value, ring_options_base(ring) + offset);
 154 }
 155
 156 static bool ring_full(struct tb_ring *ring)
 157 {
 158         return ((ring->head + 1) % ring->size) == ring->tail;
 159 }
 160
 161 static bool ring_empty(struct tb_ring *ring)
 162 {
 163         return ring->head == ring->tail;
 164 }
 165
 166 /**
 167  * ring_write_descriptors() - post frames from ring->queue to the controller
 168  *
 169  * ring->lock is held.
 170  */
 171 static void ring_write_descriptors(struct tb_ring *ring)
 172 {
 173         struct ring_frame *frame, *n;
 174         struct ring_desc *descriptor;
 175         list_for_each_entry_safe(frame, n, &ring->queue, list) {
 176                 if (ring_full(ring))
 177                         break;
 178                 list_move_tail(&frame->list, &ring->in_flight);
 179                 descriptor = &ring->descriptors[ring->head];
 180                 descriptor->phys = frame->buffer_phy;
 181                 descriptor->time = 0;
 182                 descriptor->flags = RING_DESC_POSTED | RING_DESC_INTERRUPT;
 183                 if (ring->is_tx) {
 184                         descriptor->length = frame->size;
 185                         descriptor->eof = frame->eof;
 186                         descriptor->sof = frame->sof;
 187                 }
 188                 ring->head = (ring->head + 1) % ring->size;
 189                 ring_iowrite16desc(ring, ring->head, ring->is_tx ? 10 : 8);
 190         }
 191 }
 192
 193 /**
 194  * ring_work() - progress completed frames
 195  *
 196  * If the ring is shutting down then all frames are marked as canceled and
 197  * their callbacks are invoked.
 198  *
 199  * Otherwise we collect all completed frame from the ring buffer, write new
 200  * frame to the ring buffer and invoke the callbacks for the completed frames.
 201  */
 202 static void ring_work(struct work_struct *work)
 203 {
 204         struct tb_ring *ring = container_of(work, typeof(*ring), work);
 205         struct ring_frame *frame;
 206         bool canceled = false;
 207         LIST_HEAD(done);
 208         mutex_lock(&ring->lock);
 209
 210         if (!ring->running) {
 211                 /*  Move all frames to done and mark them as canceled. */
 212                 list_splice_tail_init(&ring->in_flight, &done);
 213                 list_splice_tail_init(&ring->queue, &done);
 214                 canceled = true;
 215                 goto invoke_callback;
 216         }
 217
 218         while (!ring_empty(ring)) {
 219                 if (!(ring->descriptors[ring->tail].flags
 220                                 & RING_DESC_COMPLETED))
 221                         break;
 222                 frame = list_first_entry(&ring->in_flight, typeof(*frame),
 223                                          list);
 224                 list_move_tail(&frame->list, &done);
 225                 if (!ring->is_tx) {
 226                         frame->size = ring->descriptors[ring->tail].length;
 227                         frame->eof = ring->descriptors[ring->tail].eof;
 228                         frame->sof = ring->descriptors[ring->tail].sof;
 229                         frame->flags = ring->descriptors[ring->tail].flags;
 230                         if (frame->sof != 0)
 231                                 dev_WARN(&ring->nhi->pdev->dev,
 232                                          "%s %d got unexpected SOF: %#x\n",
 233                                          RING_TYPE(ring), ring->hop,
 234                                          frame->sof);
 235                         /*
 236                          * known flags:
 237                          * raw not enabled, interupt not set: 0x2=0010
 238                          * raw enabled: 0xa=1010
 239                          * raw not enabled: 0xb=1011
 240                          * partial frame (>MAX_FRAME_SIZE): 0xe=1110
 241                          */
 242                         if (frame->flags != 0xa)
 243                                 dev_WARN(&ring->nhi->pdev->dev,
 244                                          "%s %d got unexpected flags: %#x\n",
 245                                          RING_TYPE(ring), ring->hop,
 246                                          frame->flags);
 247                 }
 248                 ring->tail = (ring->tail + 1) % ring->size;
 249         }
 250         ring_write_descriptors(ring);
 251
 252 invoke_callback:
 253         mutex_unlock(&ring->lock); /* allow callbacks to schedule new work */
 254         while (!list_empty(&done)) {
 255                 frame = list_first_entry(&done, typeof(*frame), list);
 256                 /*
 257                  * The callback may reenqueue or delete frame.
 258                  * Do not hold on to it.
 259                  */
 260                 list_del_init(&frame->list);
 261                 frame->callback(ring, frame, canceled);
 262         }
 263 }
 264
 265 int __ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
 266 {
 267         int ret = 0;
 268         mutex_lock(&ring->lock);
 269         if (ring->running) {
 270                 list_add_tail(&frame->list, &ring->queue);
 271                 ring_write_descriptors(ring);
 272         } else {
 273                 ret = -ESHUTDOWN;
 274         }
 275         mutex_unlock(&ring->lock);
 276         return ret;
 277 }
 278
 279 static irqreturn_t ring_msix(int irq, void *data)
 280 {
 281         struct tb_ring *ring = data;
 282
 283         schedule_work(&ring->work);
 284         return IRQ_HANDLED;
 285 }
 286
 287 static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
 288 {
 289         struct tb_nhi *nhi = ring->nhi;
 290         unsigned long irqflags;
 291         int ret;
 292
 293         if (!nhi->pdev->msix_enabled)
 294                 return 0;
 295
 296         ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
 297         if (ret < 0)
 298                 return ret;
 299
 300         ring->vector = ret;
 301
 302         ring->irq = pci_irq_vector(ring->nhi->pdev, ring->vector);
 303         if (ring->irq < 0)
 304                 return ring->irq;
 305
 306         irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
 307         return request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
 308 }
 309
 310 static void ring_release_msix(struct tb_ring *ring)
 311 {
 312         if (ring->irq <= 0)
 313                 return;
 314
 315         free_irq(ring->irq, ring);
 316         ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
 317         ring->vector = 0;
 318         ring->irq = 0;
 319 }
 320
 321 static struct tb_ring *ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
 322                                   bool transmit, unsigned int flags)
 323 {
 324         struct tb_ring *ring = NULL;
 325         dev_info(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
 326                  transmit ? "TX" : "RX", hop, size);
 327
 328         mutex_lock(&nhi->lock);
 329         if (hop >= nhi->hop_count) {
 330                 dev_WARN(&nhi->pdev->dev, "invalid hop: %d\n", hop);
 331                 goto err;
 332         }
 333         if (transmit && nhi->tx_rings[hop]) {
 334                 dev_WARN(&nhi->pdev->dev, "TX hop %d already allocated\n", hop);
 335                 goto err;
 336         } else if (!transmit && nhi->rx_rings[hop]) {
 337                 dev_WARN(&nhi->pdev->dev, "RX hop %d already allocated\n", hop);
 338                 goto err;
 339         }
 340         ring = kzalloc(sizeof(*ring), GFP_KERNEL);
 341         if (!ring)
 342                 goto err;
 343
 344         mutex_init(&ring->lock);
 345         INIT_LIST_HEAD(&ring->queue);
 346         INIT_LIST_HEAD(&ring->in_flight);
 347         INIT_WORK(&ring->work, ring_work);
 348
 349         ring->nhi = nhi;
 350         ring->hop = hop;
 351         ring->is_tx = transmit;
 352         ring->size = size;
 353         ring->flags = flags;
 354         ring->head = 0;
 355         ring->tail = 0;
 356         ring->running = false;
 357
 358         if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
 359                 goto err;
 360
 361         ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
 362                         size * sizeof(*ring->descriptors),
 363                         &ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
 364         if (!ring->descriptors)
 365                 goto err;
 366
 367         if (transmit)
 368                 nhi->tx_rings[hop] = ring;
 369         else
 370                 nhi->rx_rings[hop] = ring;
 371         mutex_unlock(&nhi->lock);
 372         return ring;
 373
 374 err:
 375         if (ring)
 376                 mutex_destroy(&ring->lock);
 377         kfree(ring);
 378         mutex_unlock(&nhi->lock);
 379         return NULL;
 380 }
 381
 382 struct tb_ring *ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
 383                               unsigned int flags)
 384 {
 385         return ring_alloc(nhi, hop, size, true, flags);
 386 }
 387
 388 struct tb_ring *ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
 389                               unsigned int flags)
 390 {
 391         return ring_alloc(nhi, hop, size, false, flags);
 392 }
 393
 394 /**
 395  * ring_start() - enable a ring
 396  *
 397  * Must not be invoked in parallel with ring_stop().
 398  */
 399 void ring_start(struct tb_ring *ring)
 400 {
 401         mutex_lock(&ring->nhi->lock);
 402         mutex_lock(&ring->lock);
 403         if (ring->running) {
 404                 dev_WARN(&ring->nhi->pdev->dev, "ring already started\n");
 405                 goto err;
 406         }
 407         dev_info(&ring->nhi->pdev->dev, "starting %s %d\n",
 408                  RING_TYPE(ring), ring->hop);
 409
 410         ring_iowrite64desc(ring, ring->descriptors_dma, 0);
 411         if (ring->is_tx) {
 412                 ring_iowrite32desc(ring, ring->size, 12);
 413                 ring_iowrite32options(ring, 0, 4); /* time releated ? */
 414                 ring_iowrite32options(ring,
 415                                       RING_FLAG_ENABLE | RING_FLAG_RAW, 0);
 416         } else {
 417                 ring_iowrite32desc(ring,
 418                                    (TB_FRAME_SIZE << 16) | ring->size, 12);
 419                 ring_iowrite32options(ring, 0xffffffff, 4); /* SOF EOF mask */
 420                 ring_iowrite32options(ring,
 421                                       RING_FLAG_ENABLE | RING_FLAG_RAW, 0);
 422         }
 423         ring_interrupt_active(ring, true);
 424         ring->running = true;
 425 err:
 426         mutex_unlock(&ring->lock);
 427         mutex_unlock(&ring->nhi->lock);
 428 }
 429
 430
 431 /**
 432  * ring_stop() - shutdown a ring
 433  *
 434  * Must not be invoked from a callback.
 435  *
 436  * This method will disable the ring. Further calls to ring_tx/ring_rx will
 437  * return -ESHUTDOWN until ring_stop has been called.
 438  *
 439  * All enqueued frames will be canceled and their callbacks will be executed
 440  * with frame->canceled set to true (on the callback thread). This method
 441  * returns only after all callback invocations have finished.
 442  */
 443 void ring_stop(struct tb_ring *ring)
 444 {
 445         mutex_lock(&ring->nhi->lock);
 446         mutex_lock(&ring->lock);
 447         dev_info(&ring->nhi->pdev->dev, "stopping %s %d\n",
 448                  RING_TYPE(ring), ring->hop);
 449         if (!ring->running) {
 450                 dev_WARN(&ring->nhi->pdev->dev, "%s %d already stopped\n",
 451                          RING_TYPE(ring), ring->hop);
 452                 goto err;
 453         }
 454         ring_interrupt_active(ring, false);
 455
 456         ring_iowrite32options(ring, 0, 0);
 457         ring_iowrite64desc(ring, 0, 0);
 458         ring_iowrite16desc(ring, 0, ring->is_tx ? 10 : 8);
 459         ring_iowrite32desc(ring, 0, 12);
 460         ring->head = 0;
 461         ring->tail = 0;
 462         ring->running = false;
 463
 464 err:
 465         mutex_unlock(&ring->lock);
 466         mutex_unlock(&ring->nhi->lock);
 467
 468         /*
 469          * schedule ring->work to invoke callbacks on all remaining frames.
 470          */
 471         schedule_work(&ring->work);
 472         flush_work(&ring->work);
 473 }
 474
 475 /*
 476  * ring_free() - free ring
 477  *
 478  * When this method returns all invocations of ring->callback will have
 479  * finished.
 480  *
 481  * Ring must be stopped.
 482  *
 483  * Must NOT be called from ring_frame->callback!
 484  */
 485 void ring_free(struct tb_ring *ring)
 486 {
 487         mutex_lock(&ring->nhi->lock);
 488         /*
 489          * Dissociate the ring from the NHI. This also ensures that
 490          * nhi_interrupt_work cannot reschedule ring->work.
 491          */
 492         if (ring->is_tx)
 493                 ring->nhi->tx_rings[ring->hop] = NULL;
 494         else
 495                 ring->nhi->rx_rings[ring->hop] = NULL;
 496
 497         if (ring->running) {
 498                 dev_WARN(&ring->nhi->pdev->dev, "%s %d still running\n",
 499                          RING_TYPE(ring), ring->hop);
 500         }
 501
 502         ring_release_msix(ring);
 503
 504         dma_free_coherent(&ring->nhi->pdev->dev,
 505                           ring->size * sizeof(*ring->descriptors),
 506                           ring->descriptors, ring->descriptors_dma);
 507
 508         ring->descriptors = NULL;
 509         ring->descriptors_dma = 0;
 510
 511
 512         dev_info(&ring->nhi->pdev->dev,
 513                  "freeing %s %d\n",
 514                  RING_TYPE(ring),
 515                  ring->hop);
 516
 517         mutex_unlock(&ring->nhi->lock);
 518         /**
 519          * ring->work can no longer be scheduled (it is scheduled only
 520          * by nhi_interrupt_work, ring_stop and ring_msix). Wait for it
 521          * to finish before freeing the ring.
 522          */
 523         flush_work(&ring->work);
 524         mutex_destroy(&ring->lock);
 525         kfree(ring);
 526 }
 527
 528 static void nhi_interrupt_work(struct work_struct *work)
 529 {
 530         struct tb_nhi *nhi = container_of(work, typeof(*nhi), interrupt_work);
 531         int value = 0; /* Suppress uninitialized usage warning. */
 532         int bit;
 533         int hop = -1;
 534         int type = 0; /* current interrupt type 0: TX, 1: RX, 2: RX overflow */
 535         struct tb_ring *ring;
 536
 537         mutex_lock(&nhi->lock);
 538
 539         /*
 540          * Starting at REG_RING_NOTIFY_BASE there are three status bitfields
 541          * (TX, RX, RX overflow). We iterate over the bits and read a new
 542          * dwords as required. The registers are cleared on read.
 543          */
 544         for (bit = 0; bit < 3 * nhi->hop_count; bit++) {
 545                 if (bit % 32 == 0)
 546                         value = ioread32(nhi->iobase
 547                                          + REG_RING_NOTIFY_BASE
 548                                          + 4 * (bit / 32));
 549                 if (++hop == nhi->hop_count) {
 550                         hop = 0;
 551                         type++;
 552                 }
 553                 if ((value & (1 << (bit % 32))) == 0)
 554                         continue;
 555                 if (type == 2) {
 556                         dev_warn(&nhi->pdev->dev,
 557                                  "RX overflow for ring %d\n",
 558                                  hop);
 559                         continue;
 560                 }
 561                 if (type == 0)
 562                         ring = nhi->tx_rings[hop];
 563                 else
 564                         ring = nhi->rx_rings[hop];
 565                 if (ring == NULL) {
 566                         dev_warn(&nhi->pdev->dev,
 567                                  "got interrupt for inactive %s ring %d\n",
 568                                  type ? "RX" : "TX",
 569                                  hop);
 570                         continue;
 571                 }
 572                 /* we do not check ring->running, this is done in ring->work */
 573                 schedule_work(&ring->work);
 574         }
 575         mutex_unlock(&nhi->lock);
 576 }
 577
 578 static irqreturn_t nhi_msi(int irq, void *data)
 579 {
 580         struct tb_nhi *nhi = data;
 581         schedule_work(&nhi->interrupt_work);
 582         return IRQ_HANDLED;
 583 }
 584
 585 static int nhi_suspend_noirq(struct device *dev)
 586 {
 587         struct pci_dev *pdev = to_pci_dev(dev);
 588         struct tb *tb = pci_get_drvdata(pdev);
 589
 590         return tb_domain_suspend_noirq(tb);
 591 }
 592
 593 static int nhi_resume_noirq(struct device *dev)
 594 {
 595         struct pci_dev *pdev = to_pci_dev(dev);
 596         struct tb *tb = pci_get_drvdata(pdev);
 597
 598         return tb_domain_resume_noirq(tb);
 599 }
 600
 601 static void nhi_shutdown(struct tb_nhi *nhi)
 602 {
 603         int i;
 604         dev_info(&nhi->pdev->dev, "shutdown\n");
 605
 606         for (i = 0; i < nhi->hop_count; i++) {
 607                 if (nhi->tx_rings[i])
 608                         dev_WARN(&nhi->pdev->dev,
 609                                  "TX ring %d is still active\n", i);
 610                 if (nhi->rx_rings[i])
 611                         dev_WARN(&nhi->pdev->dev,
 612                                  "RX ring %d is still active\n", i);
 613         }
 614         nhi_disable_interrupts(nhi);
 615         /*
 616          * We have to release the irq before calling flush_work. Otherwise an
 617          * already executing IRQ handler could call schedule_work again.
 618          */
 619         if (!nhi->pdev->msix_enabled) {
 620                 devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
 621                 flush_work(&nhi->interrupt_work);
 622         }
 623         mutex_destroy(&nhi->lock);
 624         ida_destroy(&nhi->msix_ida);
 625 }
 626
 627 static int nhi_init_msi(struct tb_nhi *nhi)
 628 {
 629         struct pci_dev *pdev = nhi->pdev;
 630         int res, irq, nvec;
 631
 632         /* In case someone left them on. */
 633         nhi_disable_interrupts(nhi);
 634
 635         ida_init(&nhi->msix_ida);
 636
 637         /*
 638          * The NHI has 16 MSI-X vectors or a single MSI. We first try to
 639          * get all MSI-X vectors and if we succeed, each ring will have
 640          * one MSI-X. If for some reason that does not work out, we
 641          * fallback to a single MSI.
 642          */
 643         nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
 644                                      PCI_IRQ_MSIX);
 645         if (nvec < 0) {
 646                 nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
 647                 if (nvec < 0)
 648                         return nvec;
 649
 650                 INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
 651
 652                 irq = pci_irq_vector(nhi->pdev, 0);
 653                 if (irq < 0)
 654                         return irq;
 655
 656                 res = devm_request_irq(&pdev->dev, irq, nhi_msi,
 657                                        IRQF_NO_SUSPEND, "thunderbolt", nhi);
 658                 if (res) {
 659                         dev_err(&pdev->dev, "request_irq failed, aborting\n");
 660                         return res;
 661                 }
 662         }
 663
 664         return 0;
 665 }
 666
 667 static int nhi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 668 {
 669         struct tb_nhi *nhi;
 670         struct tb *tb;
 671         int res;
 672
 673         res = pcim_enable_device(pdev);
 674         if (res) {
 675                 dev_err(&pdev->dev, "cannot enable PCI device, aborting\n");
 676                 return res;
 677         }
 678
 679         res = pcim_iomap_regions(pdev, 1 << 0, "thunderbolt");
 680         if (res) {
 681                 dev_err(&pdev->dev, "cannot obtain PCI resources, aborting\n");
 682                 return res;
 683         }
 684
 685         nhi = devm_kzalloc(&pdev->dev, sizeof(*nhi), GFP_KERNEL);
 686         if (!nhi)
 687                 return -ENOMEM;
 688
 689         nhi->pdev = pdev;
 690         /* cannot fail - table is allocated bin pcim_iomap_regions */
 691         nhi->iobase = pcim_iomap_table(pdev)[0];
 692         nhi->hop_count = ioread32(nhi->iobase + REG_HOP_COUNT) & 0x3ff;
 693         if (nhi->hop_count != 12 && nhi->hop_count != 32)
 694                 dev_warn(&pdev->dev, "unexpected hop count: %d\n",
 695                          nhi->hop_count);
 696
 697         nhi->tx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
 698                                      sizeof(*nhi->tx_rings), GFP_KERNEL);
 699         nhi->rx_rings = devm_kcalloc(&pdev->dev, nhi->hop_count,
 700                                      sizeof(*nhi->rx_rings), GFP_KERNEL);
 701         if (!nhi->tx_rings || !nhi->rx_rings)
 702                 return -ENOMEM;
 703
 704         res = nhi_init_msi(nhi);
 705         if (res) {
 706                 dev_err(&pdev->dev, "cannot enable MSI, aborting\n");
 707                 return res;
 708         }
 709
 710         mutex_init(&nhi->lock);
 711
 712         pci_set_master(pdev);
 713
 714         /* magic value - clock related? */
 715         iowrite32(3906250 / 10000, nhi->iobase + 0x38c00);
 716
 717         dev_info(&nhi->pdev->dev, "NHI initialized, starting thunderbolt\n");
 718         tb = tb_probe(nhi);
 719         if (!tb)
 720                 return -ENODEV;
 721
 722         res = tb_domain_add(tb);
 723         if (res) {
 724                 /*
 725                  * At this point the RX/TX rings might already have been
 726                  * activated. Do a proper shutdown.
 727                  */
 728                 tb_domain_put(tb);
 729                 nhi_shutdown(nhi);
 730                 return -EIO;
 731         }
 732         pci_set_drvdata(pdev, tb);
 733
 734         return 0;
 735 }
 736
 737 static void nhi_remove(struct pci_dev *pdev)
 738 {
 739         struct tb *tb = pci_get_drvdata(pdev);
 740         struct tb_nhi *nhi = tb->nhi;
 741
 742         tb_domain_remove(tb);
 743         nhi_shutdown(nhi);
 744 }
 745
 746 /*
 747  * The tunneled pci bridges are siblings of us. Use resume_noirq to reenable
 748  * the tunnels asap. A corresponding pci quirk blocks the downstream bridges
 749  * resume_noirq until we are done.
 750  */
 751 static const struct dev_pm_ops nhi_pm_ops = {
 752         .suspend_noirq = nhi_suspend_noirq,
 753         .resume_noirq = nhi_resume_noirq,
 754         .freeze_noirq = nhi_suspend_noirq, /*
 755                                             * we just disable hotplug, the
 756                                             * pci-tunnels stay alive.
 757                                             */
 758         .restore_noirq = nhi_resume_noirq,
 759 };
 760
 761 static struct pci_device_id nhi_ids[] = {
 762         /*
 763          * We have to specify class, the TB bridges use the same device and
 764          * vendor (sub)id on gen 1 and gen 2 controllers.
 765          */
 766         {
 767                 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
 768                 .vendor = PCI_VENDOR_ID_INTEL,
 769                 .device = PCI_DEVICE_ID_INTEL_LIGHT_RIDGE,
 770                 .subvendor = 0x2222, .subdevice = 0x1111,
 771         },
 772         {
 773                 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
 774                 .vendor = PCI_VENDOR_ID_INTEL,
 775                 .device = PCI_DEVICE_ID_INTEL_CACTUS_RIDGE_4C,
 776                 .subvendor = 0x2222, .subdevice = 0x1111,
 777         },
 778         {
 779                 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
 780                 .vendor = PCI_VENDOR_ID_INTEL,
 781                 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_2C_NHI,
 782                 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
 783         },
 784         {
 785                 .class = PCI_CLASS_SYSTEM_OTHER << 8, .class_mask = ~0,
 786                 .vendor = PCI_VENDOR_ID_INTEL,
 787                 .device = PCI_DEVICE_ID_INTEL_FALCON_RIDGE_4C_NHI,
 788                 .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID,
 789         },
 790         { 0,}
 791 };
 792
 793 MODULE_DEVICE_TABLE(pci, nhi_ids);
 794 MODULE_LICENSE("GPL");
 795
 796 static struct pci_driver nhi_driver = {
 797         .name = "thunderbolt",
 798         .id_table = nhi_ids,
 799         .probe = nhi_probe,
 800         .remove = nhi_remove,
 801         .driver.pm = &nhi_pm_ops,
 802 };
 803
 804 static int __init nhi_init(void)
 805 {
 806         int ret;
 807
 808         if (!dmi_match(DMI_BOARD_VENDOR, "Apple Inc."))
 809                 return -ENOSYS;
 810         ret = tb_domain_init();
 811         if (ret)
 812                 return ret;
 813         ret = pci_register_driver(&nhi_driver);
 814         if (ret)
 815                 tb_domain_exit();
 816         return ret;
 817 }
 818
 819 static void __exit nhi_unload(void)
 820 {
 821         pci_unregister_driver(&nhi_driver);
 822         tb_domain_exit();
 823 }
 824
 825 module_init(nhi_init);
 826 module_exit(nhi_unload);