hw/msix.c

   1 /*
   2  * MSI-X device support
   3  *
   4  * This module includes support for MSI-X in pci devices.
   5  *
   6  * Author: Michael S. Tsirkin <mst@redhat.com>
   7  *
   8  *  Copyright (c) 2009, Red Hat Inc, Michael S. Tsirkin (mst@redhat.com)
   9  *
  10  * This work is licensed under the terms of the GNU GPL, version 2.  See
  11  * the COPYING file in the top-level directory.
  12  *
  13  * Contributions after 2012-01-13 are licensed under the terms of the
  14  * GNU GPL, version 2 or (at your option) any later version.
  15  */
  16
  17 #include "hw.h"
  18 #include "msi.h"
  19 #include "msix.h"
  20 #include "pci.h"
  21 #include "range.h"
  22
  23 #define MSIX_CAP_LENGTH 12
  24
  25 /* MSI enable bit and maskall bit are in byte 1 in FLAGS register */
  26 #define MSIX_CONTROL_OFFSET (PCI_MSIX_FLAGS + 1)
  27 #define MSIX_ENABLE_MASK (PCI_MSIX_FLAGS_ENABLE >> 8)
  28 #define MSIX_MASKALL_MASK (PCI_MSIX_FLAGS_MASKALL >> 8)
  29
  30 static MSIMessage msix_get_message(PCIDevice *dev, unsigned vector)
  31 {
  32     uint8_t *table_entry = dev->msix_table + vector * PCI_MSIX_ENTRY_SIZE;
  33     MSIMessage msg;
  34
  35     msg.address = pci_get_quad(table_entry + PCI_MSIX_ENTRY_LOWER_ADDR);
  36     msg.data = pci_get_long(table_entry + PCI_MSIX_ENTRY_DATA);
  37     return msg;
  38 }
  39
  40 /*
  41  * Special API for POWER to configure the vectors through
  42  * a side channel. Should never be used by devices.
  43  */
  44 void msix_set_message(PCIDevice *dev, int vector, struct MSIMessage msg)
  45 {
  46     uint8_t *table_entry = dev->msix_table + vector * PCI_MSIX_ENTRY_SIZE;
  47
  48     pci_set_quad(table_entry + PCI_MSIX_ENTRY_LOWER_ADDR, msg.address);
  49     pci_set_long(table_entry + PCI_MSIX_ENTRY_DATA, msg.data);
  50     table_entry[PCI_MSIX_ENTRY_VECTOR_CTRL] &= ~PCI_MSIX_ENTRY_CTRL_MASKBIT;
  51 }
  52
  53 static uint8_t msix_pending_mask(int vector)
  54 {
  55     return 1 << (vector % 8);
  56 }
  57
  58 static uint8_t *msix_pending_byte(PCIDevice *dev, int vector)
  59 {
  60     return dev->msix_pba + vector / 8;
  61 }
  62
  63 static int msix_is_pending(PCIDevice *dev, int vector)
  64 {
  65     return *msix_pending_byte(dev, vector) & msix_pending_mask(vector);
  66 }
  67
  68 static void msix_set_pending(PCIDevice *dev, int vector)
  69 {
  70     *msix_pending_byte(dev, vector) |= msix_pending_mask(vector);
  71 }
  72
  73 static void msix_clr_pending(PCIDevice *dev, int vector)
  74 {
  75     *msix_pending_byte(dev, vector) &= ~msix_pending_mask(vector);
  76 }
  77
  78 static bool msix_vector_masked(PCIDevice *dev, int vector, bool fmask)
  79 {
  80     unsigned offset = vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
  81     return fmask || dev->msix_table[offset] & PCI_MSIX_ENTRY_CTRL_MASKBIT;
  82 }
  83
  84 static bool msix_is_masked(PCIDevice *dev, int vector)
  85 {
  86     return msix_vector_masked(dev, vector, dev->msix_function_masked);
  87 }
  88
  89 static void msix_fire_vector_notifier(PCIDevice *dev,
  90                                       unsigned int vector, bool is_masked)
  91 {
  92     MSIMessage msg;
  93     int ret;
  94
  95     if (!dev->msix_vector_use_notifier) {
  96         return;
  97     }
  98     if (is_masked) {
  99         dev->msix_vector_release_notifier(dev, vector);
 100     } else {
 101         msg = msix_get_message(dev, vector);
 102         ret = dev->msix_vector_use_notifier(dev, vector, msg);
 103         assert(ret >= 0);
 104     }
 105 }
 106
 107 static void msix_handle_mask_update(PCIDevice *dev, int vector, bool was_masked)
 108 {
 109     bool is_masked = msix_is_masked(dev, vector);
 110
 111     if (is_masked == was_masked) {
 112         return;
 113     }
 114
 115     msix_fire_vector_notifier(dev, vector, is_masked);
 116
 117     if (!is_masked && msix_is_pending(dev, vector)) {
 118         msix_clr_pending(dev, vector);
 119         msix_notify(dev, vector);
 120     }
 121 }
 122
 123 static void msix_update_function_masked(PCIDevice *dev)
 124 {
 125     dev->msix_function_masked = !msix_enabled(dev) ||
 126         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] & MSIX_MASKALL_MASK);
 127 }
 128
 129 /* Handle MSI-X capability config write. */
 130 void msix_write_config(PCIDevice *dev, uint32_t addr,
 131                        uint32_t val, int len)
 132 {
 133     unsigned enable_pos = dev->msix_cap + MSIX_CONTROL_OFFSET;
 134     int vector;
 135     bool was_masked;
 136
 137     if (!msix_present(dev) || !range_covers_byte(addr, len, enable_pos)) {
 138         return;
 139     }
 140
 141     was_masked = dev->msix_function_masked;
 142     msix_update_function_masked(dev);
 143
 144     if (!msix_enabled(dev)) {
 145         return;
 146     }
 147
 148     pci_device_deassert_intx(dev);
 149
 150     if (dev->msix_function_masked == was_masked) {
 151         return;
 152     }
 153
 154     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
 155         msix_handle_mask_update(dev, vector,
 156                                 msix_vector_masked(dev, vector, was_masked));
 157     }
 158 }
 159
 160 static uint64_t msix_table_mmio_read(void *opaque, hwaddr addr,
 161                                      unsigned size)
 162 {
 163     PCIDevice *dev = opaque;
 164
 165     return pci_get_long(dev->msix_table + addr);
 166 }
 167
 168 static void msix_table_mmio_write(void *opaque, hwaddr addr,
 169                                   uint64_t val, unsigned size)
 170 {
 171     PCIDevice *dev = opaque;
 172     int vector = addr / PCI_MSIX_ENTRY_SIZE;
 173     bool was_masked;
 174
 175     was_masked = msix_is_masked(dev, vector);
 176     pci_set_long(dev->msix_table + addr, val);
 177     msix_handle_mask_update(dev, vector, was_masked);
 178 }
 179
 180 static const MemoryRegionOps msix_table_mmio_ops = {
 181     .read = msix_table_mmio_read,
 182     .write = msix_table_mmio_write,
 183     /* TODO: MSIX should be LITTLE_ENDIAN. */
 184     .endianness = DEVICE_NATIVE_ENDIAN,
 185     .valid = {
 186         .min_access_size = 4,
 187         .max_access_size = 4,
 188     },
 189 };
 190
 191 static uint64_t msix_pba_mmio_read(void *opaque, hwaddr addr,
 192                                    unsigned size)
 193 {
 194     PCIDevice *dev = opaque;
 195
 196     return pci_get_long(dev->msix_pba + addr);
 197 }
 198
 199 static const MemoryRegionOps msix_pba_mmio_ops = {
 200     .read = msix_pba_mmio_read,
 201     /* TODO: MSIX should be LITTLE_ENDIAN. */
 202     .endianness = DEVICE_NATIVE_ENDIAN,
 203     .valid = {
 204         .min_access_size = 4,
 205         .max_access_size = 4,
 206     },
 207 };
 208
 209 static void msix_mask_all(struct PCIDevice *dev, unsigned nentries)
 210 {
 211     int vector;
 212
 213     for (vector = 0; vector < nentries; ++vector) {
 214         unsigned offset =
 215             vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
 216         bool was_masked = msix_is_masked(dev, vector);
 217
 218         dev->msix_table[offset] |= PCI_MSIX_ENTRY_CTRL_MASKBIT;
 219         msix_handle_mask_update(dev, vector, was_masked);
 220     }
 221 }
 222
 223 /* Initialize the MSI-X structures */
 224 int msix_init(struct PCIDevice *dev, unsigned short nentries,
 225               MemoryRegion *table_bar, uint8_t table_bar_nr,
 226               unsigned table_offset, MemoryRegion *pba_bar,
 227               uint8_t pba_bar_nr, unsigned pba_offset, uint8_t cap_pos)
 228 {
 229     int cap;
 230     unsigned table_size, pba_size;
 231     uint8_t *config;
 232
 233     /* Nothing to do if MSI is not supported by interrupt controller */
 234     if (!msi_supported) {
 235         return -ENOTSUP;
 236     }
 237
 238     if (nentries < 1 || nentries > PCI_MSIX_FLAGS_QSIZE + 1) {
 239         return -EINVAL;
 240     }
 241
 242     table_size = nentries * PCI_MSIX_ENTRY_SIZE;
 243     pba_size = QEMU_ALIGN_UP(nentries, 64) / 8;
 244
 245     /* Sanity test: table & pba don't overlap, fit within BARs, min aligned */
 246     if ((table_bar_nr == pba_bar_nr &&
 247          ranges_overlap(table_offset, table_size, pba_offset, pba_size)) ||
 248         table_offset + table_size > memory_region_size(table_bar) ||
 249         pba_offset + pba_size > memory_region_size(pba_bar) ||
 250         (table_offset | pba_offset) & PCI_MSIX_FLAGS_BIRMASK) {
 251         return -EINVAL;
 252     }
 253
 254     cap = pci_add_capability(dev, PCI_CAP_ID_MSIX, cap_pos, MSIX_CAP_LENGTH);
 255     if (cap < 0) {
 256         return cap;
 257     }
 258
 259     dev->msix_cap = cap;
 260     dev->cap_present |= QEMU_PCI_CAP_MSIX;
 261     config = dev->config + cap;
 262
 263     pci_set_word(config + PCI_MSIX_FLAGS, nentries - 1);
 264     dev->msix_entries_nr = nentries;
 265     dev->msix_function_masked = true;
 266
 267     pci_set_long(config + PCI_MSIX_TABLE, table_offset | table_bar_nr);
 268     pci_set_long(config + PCI_MSIX_PBA, pba_offset | pba_bar_nr);
 269
 270     /* Make flags bit writable. */
 271     dev->wmask[cap + MSIX_CONTROL_OFFSET] |= MSIX_ENABLE_MASK |
 272                                              MSIX_MASKALL_MASK;
 273
 274     dev->msix_table = g_malloc0(table_size);
 275     dev->msix_pba = g_malloc0(pba_size);
 276     dev->msix_entry_used = g_malloc0(nentries * sizeof *dev->msix_entry_used);
 277
 278     msix_mask_all(dev, nentries);
 279
 280     memory_region_init_io(&dev->msix_table_mmio, &msix_table_mmio_ops, dev,
 281                           "msix-table", table_size);
 282     memory_region_add_subregion(table_bar, table_offset, &dev->msix_table_mmio);
 283     memory_region_init_io(&dev->msix_pba_mmio, &msix_pba_mmio_ops, dev,
 284                           "msix-pba", pba_size);
 285     memory_region_add_subregion(pba_bar, pba_offset, &dev->msix_pba_mmio);
 286
 287     return 0;
 288 }
 289
 290 int msix_init_exclusive_bar(PCIDevice *dev, unsigned short nentries,
 291                             uint8_t bar_nr)
 292 {
 293     int ret;
 294     char *name;
 295
 296     /*
 297      * Migration compatibility dictates that this remains a 4k
 298      * BAR with the vector table in the lower half and PBA in
 299      * the upper half.  Do not use these elsewhere!
 300      */
 301 #define MSIX_EXCLUSIVE_BAR_SIZE 4096
 302 #define MSIX_EXCLUSIVE_BAR_TABLE_OFFSET 0
 303 #define MSIX_EXCLUSIVE_BAR_PBA_OFFSET (MSIX_EXCLUSIVE_BAR_SIZE / 2)
 304 #define MSIX_EXCLUSIVE_CAP_OFFSET 0
 305
 306     if (nentries * PCI_MSIX_ENTRY_SIZE > MSIX_EXCLUSIVE_BAR_PBA_OFFSET) {
 307         return -EINVAL;
 308     }
 309
 310     name = g_strdup_printf("%s-msix", dev->name);
 311     memory_region_init(&dev->msix_exclusive_bar, name, MSIX_EXCLUSIVE_BAR_SIZE);
 312     g_free(name);
 313
 314     ret = msix_init(dev, nentries, &dev->msix_exclusive_bar, bar_nr,
 315                     MSIX_EXCLUSIVE_BAR_TABLE_OFFSET, &dev->msix_exclusive_bar,
 316                     bar_nr, MSIX_EXCLUSIVE_BAR_PBA_OFFSET,
 317                     MSIX_EXCLUSIVE_CAP_OFFSET);
 318     if (ret) {
 319         memory_region_destroy(&dev->msix_exclusive_bar);
 320         return ret;
 321     }
 322
 323     pci_register_bar(dev, bar_nr, PCI_BASE_ADDRESS_SPACE_MEMORY,
 324                      &dev->msix_exclusive_bar);
 325
 326     return 0;
 327 }
 328
 329 static void msix_free_irq_entries(PCIDevice *dev)
 330 {
 331     int vector;
 332
 333     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
 334         dev->msix_entry_used[vector] = 0;
 335         msix_clr_pending(dev, vector);
 336     }
 337 }
 338
 339 static void msix_clear_all_vectors(PCIDevice *dev)
 340 {
 341     int vector;
 342
 343     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
 344         msix_clr_pending(dev, vector);
 345     }
 346 }
 347
 348 /* Clean up resources for the device. */
 349 void msix_uninit(PCIDevice *dev, MemoryRegion *table_bar, MemoryRegion *pba_bar)
 350 {
 351     if (!msix_present(dev)) {
 352         return;
 353     }
 354     pci_del_capability(dev, PCI_CAP_ID_MSIX, MSIX_CAP_LENGTH);
 355     dev->msix_cap = 0;
 356     msix_free_irq_entries(dev);
 357     dev->msix_entries_nr = 0;
 358     memory_region_del_subregion(pba_bar, &dev->msix_pba_mmio);
 359     memory_region_destroy(&dev->msix_pba_mmio);
 360     g_free(dev->msix_pba);
 361     dev->msix_pba = NULL;
 362     memory_region_del_subregion(table_bar, &dev->msix_table_mmio);
 363     memory_region_destroy(&dev->msix_table_mmio);
 364     g_free(dev->msix_table);
 365     dev->msix_table = NULL;
 366     g_free(dev->msix_entry_used);
 367     dev->msix_entry_used = NULL;
 368     dev->cap_present &= ~QEMU_PCI_CAP_MSIX;
 369 }
 370
 371 void msix_uninit_exclusive_bar(PCIDevice *dev)
 372 {
 373     if (msix_present(dev)) {
 374         msix_uninit(dev, &dev->msix_exclusive_bar, &dev->msix_exclusive_bar);
 375         memory_region_destroy(&dev->msix_exclusive_bar);
 376     }
 377 }
 378
 379 void msix_save(PCIDevice *dev, QEMUFile *f)
 380 {
 381     unsigned n = dev->msix_entries_nr;
 382
 383     if (!msix_present(dev)) {
 384         return;
 385     }
 386
 387     qemu_put_buffer(f, dev->msix_table, n * PCI_MSIX_ENTRY_SIZE);
 388     qemu_put_buffer(f, dev->msix_pba, (n + 7) / 8);
 389 }
 390
 391 /* Should be called after restoring the config space. */
 392 void msix_load(PCIDevice *dev, QEMUFile *f)
 393 {
 394     unsigned n = dev->msix_entries_nr;
 395     unsigned int vector;
 396
 397     if (!msix_present(dev)) {
 398         return;
 399     }
 400
 401     msix_clear_all_vectors(dev);
 402     qemu_get_buffer(f, dev->msix_table, n * PCI_MSIX_ENTRY_SIZE);
 403     qemu_get_buffer(f, dev->msix_pba, (n + 7) / 8);
 404     msix_update_function_masked(dev);
 405
 406     for (vector = 0; vector < n; vector++) {
 407         msix_handle_mask_update(dev, vector, true);
 408     }
 409 }
 410
 411 /* Does device support MSI-X? */
 412 int msix_present(PCIDevice *dev)
 413 {
 414     return dev->cap_present & QEMU_PCI_CAP_MSIX;
 415 }
 416
 417 /* Is MSI-X enabled? */
 418 int msix_enabled(PCIDevice *dev)
 419 {
 420     return (dev->cap_present & QEMU_PCI_CAP_MSIX) &&
 421         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
 422          MSIX_ENABLE_MASK);
 423 }
 424
 425 /* Send an MSI-X message */
 426 void msix_notify(PCIDevice *dev, unsigned vector)
 427 {
 428     MSIMessage msg;
 429
 430     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector])
 431         return;
 432     if (msix_is_masked(dev, vector)) {
 433         msix_set_pending(dev, vector);
 434         return;
 435     }
 436
 437     msg = msix_get_message(dev, vector);
 438
 439     stl_le_phys(msg.address, msg.data);
 440 }
 441
 442 void msix_reset(PCIDevice *dev)
 443 {
 444     if (!msix_present(dev)) {
 445         return;
 446     }
 447     msix_clear_all_vectors(dev);
 448     dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &=
 449             ~dev->wmask[dev->msix_cap + MSIX_CONTROL_OFFSET];
 450     memset(dev->msix_table, 0, dev->msix_entries_nr * PCI_MSIX_ENTRY_SIZE);
 451     memset(dev->msix_pba, 0, QEMU_ALIGN_UP(dev->msix_entries_nr, 64) / 8);
 452     msix_mask_all(dev, dev->msix_entries_nr);
 453 }
 454
 455 /* PCI spec suggests that devices make it possible for software to configure
 456  * less vectors than supported by the device, but does not specify a standard
 457  * mechanism for devices to do so.
 458  *
 459  * We support this by asking devices to declare vectors software is going to
 460  * actually use, and checking this on the notification path. Devices that
 461  * don't want to follow the spec suggestion can declare all vectors as used. */
 462
 463 /* Mark vector as used. */
 464 int msix_vector_use(PCIDevice *dev, unsigned vector)
 465 {
 466     if (vector >= dev->msix_entries_nr)
 467         return -EINVAL;
 468     dev->msix_entry_used[vector]++;
 469     return 0;
 470 }
 471
 472 /* Mark vector as unused. */
 473 void msix_vector_unuse(PCIDevice *dev, unsigned vector)
 474 {
 475     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector]) {
 476         return;
 477     }
 478     if (--dev->msix_entry_used[vector]) {
 479         return;
 480     }
 481     msix_clr_pending(dev, vector);
 482 }
 483
 484 void msix_unuse_all_vectors(PCIDevice *dev)
 485 {
 486     if (!msix_present(dev)) {
 487         return;
 488     }
 489     msix_free_irq_entries(dev);
 490 }
 491
 492 unsigned int msix_nr_vectors_allocated(const PCIDevice *dev)
 493 {
 494     return dev->msix_entries_nr;
 495 }
 496
 497 static int msix_set_notifier_for_vector(PCIDevice *dev, unsigned int vector)
 498 {
 499     MSIMessage msg;
 500
 501     if (msix_is_masked(dev, vector)) {
 502         return 0;
 503     }
 504     msg = msix_get_message(dev, vector);
 505     return dev->msix_vector_use_notifier(dev, vector, msg);
 506 }
 507
 508 static void msix_unset_notifier_for_vector(PCIDevice *dev, unsigned int vector)
 509 {
 510     if (msix_is_masked(dev, vector)) {
 511         return;
 512     }
 513     dev->msix_vector_release_notifier(dev, vector);
 514 }
 515
 516 int msix_set_vector_notifiers(PCIDevice *dev,
 517                               MSIVectorUseNotifier use_notifier,
 518                               MSIVectorReleaseNotifier release_notifier)
 519 {
 520     int vector, ret;
 521
 522     assert(use_notifier && release_notifier);
 523
 524     dev->msix_vector_use_notifier = use_notifier;
 525     dev->msix_vector_release_notifier = release_notifier;
 526
 527     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
 528         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
 529         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
 530             ret = msix_set_notifier_for_vector(dev, vector);
 531             if (ret < 0) {
 532                 goto undo;
 533             }
 534         }
 535     }
 536     return 0;
 537
 538 undo:
 539     while (--vector >= 0) {
 540         msix_unset_notifier_for_vector(dev, vector);
 541     }
 542     dev->msix_vector_use_notifier = NULL;
 543     dev->msix_vector_release_notifier = NULL;
 544     return ret;
 545 }
 546
 547 void msix_unset_vector_notifiers(PCIDevice *dev)
 548 {
 549     int vector;
 550
 551     assert(dev->msix_vector_use_notifier &&
 552            dev->msix_vector_release_notifier);
 553
 554     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
 555         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
 556         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
 557             msix_unset_notifier_for_vector(dev, vector);
 558         }
 559     }
 560     dev->msix_vector_use_notifier = NULL;
 561     dev->msix_vector_release_notifier = NULL;
 562 }