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 /* How much space does an MSIX table need. */
  31 /* The spec requires giving the table structure
  32  * a 4K aligned region all by itself. */
  33 #define MSIX_PAGE_SIZE 0x1000
  34 /* Reserve second half of the page for pending bits */
  35 #define MSIX_PAGE_PENDING (MSIX_PAGE_SIZE / 2)
  36 #define MSIX_MAX_ENTRIES 32
  37
  38 static MSIMessage msix_get_message(PCIDevice *dev, unsigned vector)
  39 {
  40     uint8_t *table_entry = dev->msix_table_page + vector * PCI_MSIX_ENTRY_SIZE;
  41     MSIMessage msg;
  42
  43     msg.address = pci_get_quad(table_entry + PCI_MSIX_ENTRY_LOWER_ADDR);
  44     msg.data = pci_get_long(table_entry + PCI_MSIX_ENTRY_DATA);
  45     return msg;
  46 }
  47
  48 /* Add MSI-X capability to the config space for the device. */
  49 /* Given a bar and its size, add MSI-X table on top of it
  50  * and fill MSI-X capability in the config space.
  51  * Original bar size must be a power of 2 or 0.
  52  * New bar size is returned. */
  53 static int msix_add_config(struct PCIDevice *pdev, unsigned short nentries,
  54                            unsigned bar_nr, unsigned bar_size)
  55 {
  56     int config_offset;
  57     uint8_t *config;
  58     uint32_t new_size;
  59
  60     if (nentries < 1 || nentries > PCI_MSIX_FLAGS_QSIZE + 1)
  61         return -EINVAL;
  62     if (bar_size > 0x80000000)
  63         return -ENOSPC;
  64
  65     /* Add space for MSI-X structures */
  66     if (!bar_size) {
  67         new_size = MSIX_PAGE_SIZE;
  68     } else if (bar_size < MSIX_PAGE_SIZE) {
  69         bar_size = MSIX_PAGE_SIZE;
  70         new_size = MSIX_PAGE_SIZE * 2;
  71     } else {
  72         new_size = bar_size * 2;
  73     }
  74
  75     pdev->msix_bar_size = new_size;
  76     config_offset = pci_add_capability(pdev, PCI_CAP_ID_MSIX,
  77                                        0, MSIX_CAP_LENGTH);
  78     if (config_offset < 0)
  79         return config_offset;
  80     config = pdev->config + config_offset;
  81
  82     pci_set_word(config + PCI_MSIX_FLAGS, nentries - 1);
  83     /* Table on top of BAR */
  84     pci_set_long(config + PCI_MSIX_TABLE, bar_size | bar_nr);
  85     /* Pending bits on top of that */
  86     pci_set_long(config + PCI_MSIX_PBA, (bar_size + MSIX_PAGE_PENDING) |
  87                  bar_nr);
  88     pdev->msix_cap = config_offset;
  89     /* Make flags bit writable. */
  90     pdev->wmask[config_offset + MSIX_CONTROL_OFFSET] |= MSIX_ENABLE_MASK |
  91             MSIX_MASKALL_MASK;
  92     pdev->msix_function_masked = true;
  93     return 0;
  94 }
  95
  96 static uint64_t msix_mmio_read(void *opaque, target_phys_addr_t addr,
  97                                unsigned size)
  98 {
  99     PCIDevice *dev = opaque;
 100     unsigned int offset = addr & (MSIX_PAGE_SIZE - 1) & ~0x3;
 101     void *page = dev->msix_table_page;
 102
 103     return pci_get_long(page + offset);
 104 }
 105
 106 static uint8_t msix_pending_mask(int vector)
 107 {
 108     return 1 << (vector % 8);
 109 }
 110
 111 static uint8_t *msix_pending_byte(PCIDevice *dev, int vector)
 112 {
 113     return dev->msix_table_page + MSIX_PAGE_PENDING + vector / 8;
 114 }
 115
 116 static int msix_is_pending(PCIDevice *dev, int vector)
 117 {
 118     return *msix_pending_byte(dev, vector) & msix_pending_mask(vector);
 119 }
 120
 121 static void msix_set_pending(PCIDevice *dev, int vector)
 122 {
 123     *msix_pending_byte(dev, vector) |= msix_pending_mask(vector);
 124 }
 125
 126 static void msix_clr_pending(PCIDevice *dev, int vector)
 127 {
 128     *msix_pending_byte(dev, vector) &= ~msix_pending_mask(vector);
 129 }
 130
 131 static bool msix_vector_masked(PCIDevice *dev, int vector, bool fmask)
 132 {
 133     unsigned offset = vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
 134     return fmask || dev->msix_table_page[offset] & PCI_MSIX_ENTRY_CTRL_MASKBIT;
 135 }
 136
 137 static bool msix_is_masked(PCIDevice *dev, int vector)
 138 {
 139     return msix_vector_masked(dev, vector, dev->msix_function_masked);
 140 }
 141
 142 static void msix_fire_vector_notifier(PCIDevice *dev,
 143                                       unsigned int vector, bool is_masked)
 144 {
 145     MSIMessage msg;
 146     int ret;
 147
 148     if (!dev->msix_vector_use_notifier) {
 149         return;
 150     }
 151     if (is_masked) {
 152         dev->msix_vector_release_notifier(dev, vector);
 153     } else {
 154         msg = msix_get_message(dev, vector);
 155         ret = dev->msix_vector_use_notifier(dev, vector, msg);
 156         assert(ret >= 0);
 157     }
 158 }
 159
 160 static void msix_handle_mask_update(PCIDevice *dev, int vector, bool was_masked)
 161 {
 162     bool is_masked = msix_is_masked(dev, vector);
 163
 164     if (is_masked == was_masked) {
 165         return;
 166     }
 167
 168     msix_fire_vector_notifier(dev, vector, is_masked);
 169
 170     if (!is_masked && msix_is_pending(dev, vector)) {
 171         msix_clr_pending(dev, vector);
 172         msix_notify(dev, vector);
 173     }
 174 }
 175
 176 static void msix_update_function_masked(PCIDevice *dev)
 177 {
 178     dev->msix_function_masked = !msix_enabled(dev) ||
 179         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] & MSIX_MASKALL_MASK);
 180 }
 181
 182 /* Handle MSI-X capability config write. */
 183 void msix_write_config(PCIDevice *dev, uint32_t addr,
 184                        uint32_t val, int len)
 185 {
 186     unsigned enable_pos = dev->msix_cap + MSIX_CONTROL_OFFSET;
 187     int vector;
 188     bool was_masked;
 189
 190     if (!range_covers_byte(addr, len, enable_pos)) {
 191         return;
 192     }
 193
 194     was_masked = dev->msix_function_masked;
 195     msix_update_function_masked(dev);
 196
 197     if (!msix_enabled(dev)) {
 198         return;
 199     }
 200
 201     pci_device_deassert_intx(dev);
 202
 203     if (dev->msix_function_masked == was_masked) {
 204         return;
 205     }
 206
 207     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
 208         msix_handle_mask_update(dev, vector,
 209                                 msix_vector_masked(dev, vector, was_masked));
 210     }
 211 }
 212
 213 static void msix_mmio_write(void *opaque, target_phys_addr_t addr,
 214                             uint64_t val, unsigned size)
 215 {
 216     PCIDevice *dev = opaque;
 217     unsigned int offset = addr & (MSIX_PAGE_SIZE - 1) & ~0x3;
 218     int vector = offset / PCI_MSIX_ENTRY_SIZE;
 219     bool was_masked;
 220
 221     /* MSI-X page includes a read-only PBA and a writeable Vector Control. */
 222     if (vector >= dev->msix_entries_nr) {
 223         return;
 224     }
 225
 226     was_masked = msix_is_masked(dev, vector);
 227     pci_set_long(dev->msix_table_page + offset, val);
 228     msix_handle_mask_update(dev, vector, was_masked);
 229 }
 230
 231 static const MemoryRegionOps msix_mmio_ops = {
 232     .read = msix_mmio_read,
 233     .write = msix_mmio_write,
 234     .endianness = DEVICE_NATIVE_ENDIAN,
 235     .valid = {
 236         .min_access_size = 4,
 237         .max_access_size = 4,
 238     },
 239 };
 240
 241 static void msix_mmio_setup(PCIDevice *d, MemoryRegion *bar)
 242 {
 243     uint8_t *config = d->config + d->msix_cap;
 244     uint32_t table = pci_get_long(config + PCI_MSIX_TABLE);
 245     uint32_t offset = table & ~(MSIX_PAGE_SIZE - 1);
 246     /* TODO: for assigned devices, we'll want to make it possible to map
 247      * pending bits separately in case they are in a separate bar. */
 248
 249     memory_region_add_subregion(bar, offset, &d->msix_mmio);
 250 }
 251
 252 static void msix_mask_all(struct PCIDevice *dev, unsigned nentries)
 253 {
 254     int vector;
 255
 256     for (vector = 0; vector < nentries; ++vector) {
 257         unsigned offset =
 258             vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
 259         bool was_masked = msix_is_masked(dev, vector);
 260
 261         dev->msix_table_page[offset] |= PCI_MSIX_ENTRY_CTRL_MASKBIT;
 262         msix_handle_mask_update(dev, vector, was_masked);
 263     }
 264 }
 265
 266 /* Initialize the MSI-X structures. Note: if MSI-X is supported, BAR size is
 267  * modified, it should be retrieved with msix_bar_size. */
 268 int msix_init(struct PCIDevice *dev, unsigned short nentries,
 269               MemoryRegion *bar,
 270               unsigned bar_nr, unsigned bar_size)
 271 {
 272     int ret;
 273
 274     /* Nothing to do if MSI is not supported by interrupt controller */
 275     if (!msi_supported) {
 276         return -ENOTSUP;
 277     }
 278     if (nentries > MSIX_MAX_ENTRIES)
 279         return -EINVAL;
 280
 281     dev->msix_entry_used = g_malloc0(MSIX_MAX_ENTRIES *
 282                                         sizeof *dev->msix_entry_used);
 283
 284     dev->msix_table_page = g_malloc0(MSIX_PAGE_SIZE);
 285     msix_mask_all(dev, nentries);
 286
 287     memory_region_init_io(&dev->msix_mmio, &msix_mmio_ops, dev,
 288                           "msix", MSIX_PAGE_SIZE);
 289
 290     dev->msix_entries_nr = nentries;
 291     ret = msix_add_config(dev, nentries, bar_nr, bar_size);
 292     if (ret)
 293         goto err_config;
 294
 295     dev->cap_present |= QEMU_PCI_CAP_MSIX;
 296     msix_mmio_setup(dev, bar);
 297     return 0;
 298
 299 err_config:
 300     dev->msix_entries_nr = 0;
 301     memory_region_destroy(&dev->msix_mmio);
 302     g_free(dev->msix_table_page);
 303     dev->msix_table_page = NULL;
 304     g_free(dev->msix_entry_used);
 305     dev->msix_entry_used = NULL;
 306     return ret;
 307 }
 308
 309 static void msix_free_irq_entries(PCIDevice *dev)
 310 {
 311     int vector;
 312
 313     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
 314         dev->msix_entry_used[vector] = 0;
 315         msix_clr_pending(dev, vector);
 316     }
 317 }
 318
 319 /* Clean up resources for the device. */
 320 int msix_uninit(PCIDevice *dev, MemoryRegion *bar)
 321 {
 322     if (!(dev->cap_present & QEMU_PCI_CAP_MSIX))
 323         return 0;
 324     pci_del_capability(dev, PCI_CAP_ID_MSIX, MSIX_CAP_LENGTH);
 325     dev->msix_cap = 0;
 326     msix_free_irq_entries(dev);
 327     dev->msix_entries_nr = 0;
 328     memory_region_del_subregion(bar, &dev->msix_mmio);
 329     memory_region_destroy(&dev->msix_mmio);
 330     g_free(dev->msix_table_page);
 331     dev->msix_table_page = NULL;
 332     g_free(dev->msix_entry_used);
 333     dev->msix_entry_used = NULL;
 334     dev->cap_present &= ~QEMU_PCI_CAP_MSIX;
 335     return 0;
 336 }
 337
 338 void msix_save(PCIDevice *dev, QEMUFile *f)
 339 {
 340     unsigned n = dev->msix_entries_nr;
 341
 342     if (!(dev->cap_present & QEMU_PCI_CAP_MSIX)) {
 343         return;
 344     }
 345
 346     qemu_put_buffer(f, dev->msix_table_page, n * PCI_MSIX_ENTRY_SIZE);
 347     qemu_put_buffer(f, dev->msix_table_page + MSIX_PAGE_PENDING, (n + 7) / 8);
 348 }
 349
 350 /* Should be called after restoring the config space. */
 351 void msix_load(PCIDevice *dev, QEMUFile *f)
 352 {
 353     unsigned n = dev->msix_entries_nr;
 354     unsigned int vector;
 355
 356     if (!(dev->cap_present & QEMU_PCI_CAP_MSIX)) {
 357         return;
 358     }
 359
 360     msix_free_irq_entries(dev);
 361     qemu_get_buffer(f, dev->msix_table_page, n * PCI_MSIX_ENTRY_SIZE);
 362     qemu_get_buffer(f, dev->msix_table_page + MSIX_PAGE_PENDING, (n + 7) / 8);
 363     msix_update_function_masked(dev);
 364
 365     for (vector = 0; vector < n; vector++) {
 366         msix_handle_mask_update(dev, vector, true);
 367     }
 368 }
 369
 370 /* Does device support MSI-X? */
 371 int msix_present(PCIDevice *dev)
 372 {
 373     return dev->cap_present & QEMU_PCI_CAP_MSIX;
 374 }
 375
 376 /* Is MSI-X enabled? */
 377 int msix_enabled(PCIDevice *dev)
 378 {
 379     return (dev->cap_present & QEMU_PCI_CAP_MSIX) &&
 380         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
 381          MSIX_ENABLE_MASK);
 382 }
 383
 384 /* Size of bar where MSI-X table resides, or 0 if MSI-X not supported. */
 385 uint32_t msix_bar_size(PCIDevice *dev)
 386 {
 387     return (dev->cap_present & QEMU_PCI_CAP_MSIX) ?
 388         dev->msix_bar_size : 0;
 389 }
 390
 391 /* Send an MSI-X message */
 392 void msix_notify(PCIDevice *dev, unsigned vector)
 393 {
 394     MSIMessage msg;
 395
 396     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector])
 397         return;
 398     if (msix_is_masked(dev, vector)) {
 399         msix_set_pending(dev, vector);
 400         return;
 401     }
 402
 403     msg = msix_get_message(dev, vector);
 404
 405     stl_le_phys(msg.address, msg.data);
 406 }
 407
 408 void msix_reset(PCIDevice *dev)
 409 {
 410     if (!(dev->cap_present & QEMU_PCI_CAP_MSIX))
 411         return;
 412     msix_free_irq_entries(dev);
 413     dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &=
 414             ~dev->wmask[dev->msix_cap + MSIX_CONTROL_OFFSET];
 415     memset(dev->msix_table_page, 0, MSIX_PAGE_SIZE);
 416     msix_mask_all(dev, dev->msix_entries_nr);
 417 }
 418
 419 /* PCI spec suggests that devices make it possible for software to configure
 420  * less vectors than supported by the device, but does not specify a standard
 421  * mechanism for devices to do so.
 422  *
 423  * We support this by asking devices to declare vectors software is going to
 424  * actually use, and checking this on the notification path. Devices that
 425  * don't want to follow the spec suggestion can declare all vectors as used. */
 426
 427 /* Mark vector as used. */
 428 int msix_vector_use(PCIDevice *dev, unsigned vector)
 429 {
 430     if (vector >= dev->msix_entries_nr)
 431         return -EINVAL;
 432     dev->msix_entry_used[vector]++;
 433     return 0;
 434 }
 435
 436 /* Mark vector as unused. */
 437 void msix_vector_unuse(PCIDevice *dev, unsigned vector)
 438 {
 439     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector]) {
 440         return;
 441     }
 442     if (--dev->msix_entry_used[vector]) {
 443         return;
 444     }
 445     msix_clr_pending(dev, vector);
 446 }
 447
 448 void msix_unuse_all_vectors(PCIDevice *dev)
 449 {
 450     if (!(dev->cap_present & QEMU_PCI_CAP_MSIX))
 451         return;
 452     msix_free_irq_entries(dev);
 453 }
 454
 455 unsigned int msix_nr_vectors_allocated(const PCIDevice *dev)
 456 {
 457     return dev->msix_entries_nr;
 458 }
 459
 460 static int msix_set_notifier_for_vector(PCIDevice *dev, unsigned int vector)
 461 {
 462     MSIMessage msg;
 463
 464     if (msix_is_masked(dev, vector)) {
 465         return 0;
 466     }
 467     msg = msix_get_message(dev, vector);
 468     return dev->msix_vector_use_notifier(dev, vector, msg);
 469 }
 470
 471 static void msix_unset_notifier_for_vector(PCIDevice *dev, unsigned int vector)
 472 {
 473     if (msix_is_masked(dev, vector)) {
 474         return;
 475     }
 476     dev->msix_vector_release_notifier(dev, vector);
 477 }
 478
 479 int msix_set_vector_notifiers(PCIDevice *dev,
 480                               MSIVectorUseNotifier use_notifier,
 481                               MSIVectorReleaseNotifier release_notifier)
 482 {
 483     int vector, ret;
 484
 485     assert(use_notifier && release_notifier);
 486
 487     dev->msix_vector_use_notifier = use_notifier;
 488     dev->msix_vector_release_notifier = release_notifier;
 489
 490     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
 491         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
 492         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
 493             ret = msix_set_notifier_for_vector(dev, vector);
 494             if (ret < 0) {
 495                 goto undo;
 496             }
 497         }
 498     }
 499     return 0;
 500
 501 undo:
 502     while (--vector >= 0) {
 503         msix_unset_notifier_for_vector(dev, vector);
 504     }
 505     dev->msix_vector_use_notifier = NULL;
 506     dev->msix_vector_release_notifier = NULL;
 507     return ret;
 508 }
 509
 510 void msix_unset_vector_notifiers(PCIDevice *dev)
 511 {
 512     int vector;
 513
 514     assert(dev->msix_vector_use_notifier &&
 515            dev->msix_vector_release_notifier);
 516
 517     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
 518         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
 519         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
 520             msix_unset_notifier_for_vector(dev, vector);
 521         }
 522     }
 523     dev->msix_vector_use_notifier = NULL;
 524     dev->msix_vector_release_notifier = NULL;
 525 }