hypervisor/pci.c

   1 /*
   2  * Jailhouse, a Linux-based partitioning hypervisor
   3  *
   4  * Copyright (c) Siemens AG, 2014
   5  *
   6  * Authors:
   7  *  Ivan Kolchin <ivan.kolchin@siemens.com>
   8  *  Jan Kiszka <jan.kiszka@siemens.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
  14 #include <jailhouse/control.h>
  15 #include <jailhouse/mmio.h>
  16 #include <jailhouse/pci.h>
  17 #include <jailhouse/printk.h>
  18 #include <jailhouse/utils.h>
  19
  20 #define MSIX_VECTOR_CTRL_DWORD          3
  21
  22 #define for_each_configured_pci_device(dev, cell)                       \
  23         for ((dev) = (cell)->pci_devices;                               \
  24              (dev) - (cell)->pci_devices < (cell)->config->num_pci_devices; \
  25              (dev)++)
  26
  27 #define for_each_pci_cap(cap, dev, counter)                             \
  28         for ((cap) = jailhouse_cell_pci_caps((dev)->cell->config) +     \
  29                 (dev)->info->caps_start, (counter) = 0;                 \
  30              (counter) < (dev)->info->num_caps;                         \
  31              (cap)++, (counter)++)
  32
  33 /* entry for PCI config space whitelist (granting access) */
  34 struct pci_cfg_access {
  35         u32 reg_num; /* Register number (4-byte aligned) */
  36         u32 mask; /* Bit set: access allowed */
  37 };
  38
  39 /* --- Whilelists for writing to PCI config space registers --- */
  40 /* Type 1: Endpoints */
  41 static const struct pci_cfg_access endpoint_write_access[] = {
  42         { 0x04, 0xffffffff }, /* Command, Status */
  43         { 0x0c, 0xff00ffff }, /* BIST, Latency Timer, Cacheline */
  44         { 0x3c, 0x000000ff }, /* Int Line */
  45 };
  46 /* Type 2: Bridges */
  47 static const struct pci_cfg_access bridge_write_access[] = {
  48         { 0x04, 0xffffffff }, /* Command, Status */
  49         { 0x0c, 0xff00ffff }, /* BIST, Latency Timer, Cacheline */
  50         { 0x3c, 0xffff00ff }, /* Int Line, Bridge Control */
  51 };
  52
  53 static void *pci_space;
  54 static u64 mmcfg_start, mmcfg_end;
  55 static u8 end_bus;
  56
  57 static void *pci_get_device_mmcfg_base(u16 bdf)
  58 {
  59         return pci_space + ((unsigned long)bdf << 12);
  60 }
  61
  62 /**
  63  * Read from PCI config space.
  64  * @param bdf           16-bit bus/device/function ID of target.
  65  * @param address       Config space access address.
  66  * @param size          Access size (1, 2 or 4 bytes).
  67  *
  68  * @return Read value.
  69  *
  70  * @see pci_write_config
  71  */
  72 u32 pci_read_config(u16 bdf, u16 address, unsigned int size)
  73 {
  74         void *mmcfg_addr = pci_get_device_mmcfg_base(bdf) + address;
  75
  76         if (!pci_space || PCI_BUS(bdf) > end_bus)
  77                 return arch_pci_read_config(bdf, address, size);
  78
  79         if (size == 1)
  80                 return mmio_read8(mmcfg_addr);
  81         else if (size == 2)
  82                 return mmio_read16(mmcfg_addr);
  83         else
  84                 return mmio_read32(mmcfg_addr);
  85 }
  86
  87 /**
  88  * Write to PCI config space.
  89  * @param bdf           16-bit bus/device/function ID of target.
  90  * @param address       Config space access address.
  91  * @param value         Value to be written.
  92  * @param size          Access size (1, 2 or 4 bytes).
  93  *
  94  * @see pci_read_config
  95  */
  96 void pci_write_config(u16 bdf, u16 address, u32 value, unsigned int size)
  97 {
  98         void *mmcfg_addr = pci_get_device_mmcfg_base(bdf) + address;
  99
 100         if (!pci_space || PCI_BUS(bdf) > end_bus)
 101                 return arch_pci_write_config(bdf, address, value, size);
 102
 103         if (size == 1)
 104                 mmio_write8(mmcfg_addr, value);
 105         else if (size == 2)
 106                 mmio_write16(mmcfg_addr, value);
 107         else
 108                 mmio_write32(mmcfg_addr, value);
 109 }
 110
 111 /**
 112  * Look up device owned by a cell.
 113  * @param[in] cell      Owning cell.
 114  * @param bdf           16-bit bus/device/function ID.
 115  *
 116  * @return Pointer to owned PCI device or NULL.
 117  */
 118 struct pci_device *pci_get_assigned_device(const struct cell *cell, u16 bdf)
 119 {
 120         const struct jailhouse_pci_device *dev_info =
 121                 jailhouse_cell_pci_devices(cell->config);
 122         u32 n;
 123
 124         /* We iterate over the static device information to increase cache
 125          * locality. */
 126         for (n = 0; n < cell->config->num_pci_devices; n++)
 127                 if (dev_info[n].bdf == bdf)
 128                         return cell->pci_devices[n].cell ?
 129                                 &cell->pci_devices[n] : NULL;
 130
 131         return NULL;
 132 }
 133
 134 /**
 135  * Look up capability at given config space address.
 136  * @param device        The device to be accessed.
 137  * @param address       Config space access address.
 138  *
 139  * @return Corresponding capability structure or NULL if none found.
 140  *
 141  * @private
 142  */
 143 static const struct jailhouse_pci_capability *
 144 pci_find_capability(struct pci_device *device, u16 address)
 145 {
 146         const struct jailhouse_pci_capability *cap =
 147                 jailhouse_cell_pci_caps(device->cell->config) +
 148                 device->info->caps_start;
 149         u32 n;
 150
 151         for (n = 0; n < device->info->num_caps; n++, cap++)
 152                 if (cap->start <= address && cap->start + cap->len > address)
 153                         return cap;
 154
 155         return NULL;
 156 }
 157
 158 /**
 159  * Moderate config space read access.
 160  * @param device        The device to be accessed. If NULL, access will be
 161  *                      emulated, returning a value of -1.
 162  * @param address       Config space address.
 163  * @param size          Access size (1, 2 or 4 bytes).
 164  * @param value         Pointer to buffer to receive the emulated value if
 165  *                      PCI_ACCESS_DONE is returned.
 166  *
 167  * @return PCI_ACCESS_PERFORM or PCI_ACCESS_DONE.
 168  *
 169  * @see pci_cfg_write_moderate
 170  */
 171 enum pci_access pci_cfg_read_moderate(struct pci_device *device, u16 address,
 172                                       unsigned int size, u32 *value)
 173 {
 174         const struct jailhouse_pci_capability *cap;
 175         unsigned int cap_offs;
 176
 177         if (!device) {
 178                 *value = -1;
 179                 return PCI_ACCESS_DONE;
 180         }
 181
 182         if (address < PCI_CONFIG_HEADER_SIZE)
 183                 return PCI_ACCESS_PERFORM;
 184
 185         cap = pci_find_capability(device, address);
 186         if (!cap)
 187                 return PCI_ACCESS_PERFORM;
 188
 189         cap_offs = address - cap->start;
 190         if (cap->id == PCI_CAP_MSI && cap_offs >= 4 &&
 191             (cap_offs < 10 || (device->info->msi_64bits && cap_offs < 14))) {
 192                 *value = device->msi_registers.raw[cap_offs / 4] >>
 193                         ((cap_offs % 4) * 8);
 194                 return PCI_ACCESS_DONE;
 195         }
 196
 197         return PCI_ACCESS_PERFORM;
 198 }
 199
 200 static int pci_update_msix(struct pci_device *device,
 201                            const struct jailhouse_pci_capability *cap)
 202 {
 203         unsigned int n;
 204         int result;
 205
 206         for (n = 0; n < device->info->num_msix_vectors; n++) {
 207                 result = arch_pci_update_msix_vector(device, n);
 208                 if (result < 0)
 209                         return result;
 210         }
 211         return 0;
 212 }
 213
 214 /**
 215  * Moderate config space write access.
 216  * @param device        The device to be accessed. If NULL, access will be
 217  *                      rejected.
 218  * @param address       Config space address.
 219  * @param size          Access size (1, 2 or 4 bytes).
 220  * @param value         Value to be written.
 221  *
 222  * @return PCI_ACCESS_REJECT, PCI_ACCESS_PERFORM or PCI_ACCESS_DONE.
 223  *
 224  * @see pci_cfg_read_moderate
 225  */
 226 enum pci_access pci_cfg_write_moderate(struct pci_device *device, u16 address,
 227                                        unsigned int size, u32 value)
 228 {
 229         const struct jailhouse_pci_capability *cap;
 230         /* initialize list to work around wrong compiler warning */
 231         const struct pci_cfg_access *list = NULL;
 232         unsigned int bias_shift = (address % 4) * 8;
 233         u32 mask = BYTE_MASK(size) << bias_shift;
 234         unsigned int n, cap_offs, len = 0;
 235
 236         if (!device)
 237                 return PCI_ACCESS_REJECT;
 238
 239         if (address < PCI_CONFIG_HEADER_SIZE) {
 240                 if (device->info->type == JAILHOUSE_PCI_TYPE_DEVICE) {
 241                         list = endpoint_write_access;
 242                         len = ARRAY_SIZE(endpoint_write_access);
 243                 } else if (device->info->type == JAILHOUSE_PCI_TYPE_BRIDGE) {
 244                         list = bridge_write_access;
 245                         len = ARRAY_SIZE(bridge_write_access);
 246                 }
 247
 248                 for (n = 0; n < len; n++) {
 249                         if (list[n].reg_num == (address & 0xffc) &&
 250                             (list[n].mask & mask) == mask)
 251                                 return PCI_ACCESS_PERFORM;
 252                 }
 253
 254                 // HACK to allow PCI bus rescanning in root-cell
 255                 if (device->info->type == JAILHOUSE_PCI_TYPE_BRIDGE &&
 256                     device->cell == &root_cell)
 257                         return PCI_ACCESS_DONE;
 258
 259                 return PCI_ACCESS_REJECT;
 260         }
 261
 262         cap = pci_find_capability(device, address);
 263         if (!cap || !(cap->flags & JAILHOUSE_PCICAPS_WRITE))
 264                 return PCI_ACCESS_REJECT;
 265
 266         value <<= bias_shift;
 267
 268         cap_offs = address - cap->start;
 269         if (cap->id == PCI_CAP_MSI &&
 270             (cap_offs < 10 || (device->info->msi_64bits && cap_offs < 14))) {
 271                 device->msi_registers.raw[cap_offs / 4] &= ~mask;
 272                 device->msi_registers.raw[cap_offs / 4] |= value;
 273
 274                 if (arch_pci_update_msi(device, cap) < 0)
 275                         return PCI_ACCESS_REJECT;
 276
 277                 /*
 278                  * Address and data words are emulated, the control word is
 279                  * written as-is.
 280                  */
 281                 if (cap_offs >= 4)
 282                         return PCI_ACCESS_DONE;
 283         } else if (cap->id == PCI_CAP_MSIX && cap_offs < 4) {
 284                 device->msix_registers.raw &= ~mask;
 285                 device->msix_registers.raw |= value;
 286
 287                 if (pci_update_msix(device, cap) < 0)
 288                         return PCI_ACCESS_REJECT;
 289         }
 290
 291         return PCI_ACCESS_PERFORM;
 292 }
 293
 294 /**
 295  * Initialization of PCI subsystem.
 296  *
 297  * @return 0 on success, negative error code otherwise.
 298  */
 299 int pci_init(void)
 300 {
 301         unsigned int mmcfg_size;
 302         int err;
 303
 304         err = pci_cell_init(&root_cell);
 305         if (err)
 306                 return err;
 307
 308         mmcfg_start = system_config->platform_info.x86.mmconfig_base;
 309         if (mmcfg_start == 0)
 310                 return 0;
 311
 312         end_bus = system_config->platform_info.x86.mmconfig_end_bus;
 313         mmcfg_size = (end_bus + 1) * 256 * 4096;
 314         mmcfg_end = mmcfg_start + mmcfg_size - 4;
 315
 316         pci_space = page_alloc(&remap_pool, mmcfg_size / PAGE_SIZE);
 317         if (!pci_space)
 318                 return -ENOMEM;
 319
 320         return paging_create(&hv_paging_structs, mmcfg_start, mmcfg_size,
 321                              (unsigned long)pci_space,
 322                              PAGE_DEFAULT_FLAGS | PAGE_FLAG_UNCACHED,
 323                              PAGING_NON_COHERENT);
 324 }
 325
 326 static int pci_msix_access_handler(const struct cell *cell, bool is_write,
 327                                    u64 addr, u32 *value)
 328 {
 329         unsigned int dword = (addr % sizeof(union pci_msix_vector)) >> 2;
 330         struct pci_device *device = cell->msix_device_list;
 331         unsigned int index;
 332         u64 offs;
 333
 334         while (device) {
 335                 if (addr >= device->info->msix_address &&
 336                     addr < device->info->msix_address +
 337                            device->info->msix_region_size)
 338                         goto found;
 339                 device = device->next_msix_device;
 340         }
 341         return 0;
 342
 343 found:
 344         /* access must be DWORD-aligned */
 345         if (addr & 0x3)
 346                 goto invalid_access;
 347
 348         offs = addr - device->info->msix_address;
 349         index = offs / sizeof(union pci_msix_vector);
 350
 351         if (is_write) {
 352                 /*
 353                  * The PBA may share a page with the MSI-X table. Writing to
 354                  * PBA entries is undefined. We declare it as invalid.
 355                  */
 356                 if (index >= device->info->num_msix_vectors)
 357                         goto invalid_access;
 358                 if (dword == MSIX_VECTOR_CTRL_DWORD) {
 359                         mmio_write32(&device->msix_table[index].field.ctrl,
 360                                      *value);
 361                 } else {
 362                         device->msix_vectors[index].raw[dword] = *value;
 363                         if (arch_pci_update_msix_vector(device, index) < 0)
 364                                 goto invalid_access;
 365                 }
 366         } else {
 367                 if (index >= device->info->num_msix_vectors ||
 368                     dword == MSIX_VECTOR_CTRL_DWORD)
 369                         *value =
 370                             mmio_read32(((void *)device->msix_table) + offs);
 371                 else
 372                         *value = device->msix_vectors[index].raw[dword];
 373         }
 374         return 1;
 375
 376 invalid_access:
 377         panic_printk("FATAL: Invalid PCI MSIX BAR write, device "
 378                      "%02x:%02x.%x\n", PCI_BDF_PARAMS(device->info->bdf));
 379         return -1;
 380 }
 381
 382 /**
 383  * Handler for MMIO-accesses to PCI config space.
 384  * @param cell          Request issuing cell.
 385  * @param is_write      True if write access.
 386  * @param addr          Address accessed.
 387  * @param value         Pointer to value for reading/writing.
 388  *
 389  * @return 1 if handled successfully, 0 if unhandled, -1 on access error.
 390  */
 391 int pci_mmio_access_handler(const struct cell *cell, bool is_write,
 392                             u64 addr, u32 *value)
 393 {
 394         u32 mmcfg_offset, reg_addr;
 395         struct pci_device *device;
 396         enum pci_access access;
 397
 398         if (!pci_space || addr < mmcfg_start || addr > mmcfg_end)
 399                 return pci_msix_access_handler(cell, is_write, addr, value);
 400
 401         mmcfg_offset = addr - mmcfg_start;
 402         reg_addr = mmcfg_offset & 0xfff;
 403         /* access must be DWORD-aligned */
 404         if (reg_addr & 0x3)
 405                 goto invalid_access;
 406
 407         device = pci_get_assigned_device(cell, mmcfg_offset >> 12);
 408
 409         if (is_write) {
 410                 access = pci_cfg_write_moderate(device, reg_addr, 4, *value);
 411                 if (access == PCI_ACCESS_REJECT)
 412                         goto invalid_access;
 413                 if (access == PCI_ACCESS_PERFORM)
 414                         mmio_write32(pci_space + mmcfg_offset, *value);
 415         } else {
 416                 access = pci_cfg_read_moderate(device, reg_addr, 4, value);
 417                 if (access == PCI_ACCESS_PERFORM)
 418                         *value = mmio_read32(pci_space + mmcfg_offset);
 419         }
 420
 421         return 1;
 422
 423 invalid_access:
 424         panic_printk("FATAL: Invalid PCI MMCONFIG write, device %02x:%02x.%x, "
 425                      "reg: %\n", PCI_BDF_PARAMS(mmcfg_offset >> 12), reg_addr);
 426         return -1;
 427
 428 }
 429
 430 /**
 431  * Retrieve number of enabled MSI vector of a device.
 432  * @param device        The device to be examined.
 433  *
 434  * @return number of vectors.
 435  */
 436 unsigned int pci_enabled_msi_vectors(struct pci_device *device)
 437 {
 438         return device->msi_registers.msg32.enable ?
 439                 1 << device->msi_registers.msg32.mme : 0;
 440 }
 441
 442 static void pci_save_msi(struct pci_device *device,
 443                          const struct jailhouse_pci_capability *cap)
 444 {
 445         u16 bdf = device->info->bdf;
 446         unsigned int n;
 447
 448         for (n = 0; n < (device->info->msi_64bits ? 4 : 3); n++)
 449                 device->msi_registers.raw[n] =
 450                         pci_read_config(bdf, cap->start + n * 4, 4);
 451 }
 452
 453 static void pci_restore_msi(struct pci_device *device,
 454                             const struct jailhouse_pci_capability *cap)
 455 {
 456         unsigned int n;
 457
 458         for (n = 1; n < (device->info->msi_64bits ? 4 : 3); n++)
 459                 pci_write_config(device->info->bdf, cap->start + n * 4,
 460                                  device->msi_registers.raw[n], 4);
 461 }
 462
 463 static void pci_suppress_msix(struct pci_device *device,
 464                               const struct jailhouse_pci_capability *cap,
 465                               bool suppressed)
 466 {
 467         union pci_msix_registers regs = device->msix_registers;
 468
 469         if (suppressed)
 470                 regs.field.fmask = 1;
 471         pci_write_config(device->info->bdf, cap->start, regs.raw, 4);
 472 }
 473
 474 static void pci_save_msix(struct pci_device *device,
 475                           const struct jailhouse_pci_capability *cap)
 476 {
 477         unsigned int n, r;
 478
 479         device->msix_registers.raw =
 480                 pci_read_config(device->info->bdf, cap->start, 4);
 481
 482         for (n = 0; n < device->info->num_msix_vectors; n++)
 483                 for (r = 0; r < 3; r++)
 484                         device->msix_vectors[n].raw[r] =
 485                                 mmio_read32(&device->msix_table[n].raw[r]);
 486 }
 487
 488 static void pci_restore_msix(struct pci_device *device,
 489                              const struct jailhouse_pci_capability *cap)
 490 {
 491         unsigned int n, r;
 492
 493         for (n = 0; n < device->info->num_msix_vectors; n++)
 494                 for (r = 0; r < 3; r++)
 495                         mmio_write32(&device->msix_table[n].raw[r],
 496                                      device->msix_vectors[n].raw[r]);
 497         pci_suppress_msix(device, cap, false);
 498 }
 499
 500 /**
 501  * Prepare the handover of PCI devices to Jailhouse or back to Linux.
 502  */
 503 void pci_prepare_handover(void)
 504 {
 505         const struct jailhouse_pci_capability *cap;
 506         struct pci_device *device;
 507         unsigned int n;
 508
 509         if (!root_cell.pci_devices)
 510                 return;
 511
 512         for_each_configured_pci_device(device, &root_cell) {
 513                 if (device->cell)
 514                         for_each_pci_cap(cap, device, n)
 515                                 if (cap->id == PCI_CAP_MSI)
 516                                         arch_pci_suppress_msi(device, cap);
 517                                 else if (cap->id == PCI_CAP_MSIX)
 518                                         pci_suppress_msix(device, cap, true);
 519         }
 520 }
 521
 522 static void pci_add_virtual_device(struct cell *cell, struct pci_device *device)
 523 {
 524         device->cell = cell;
 525         device->next_virtual_device = cell->virtual_device_list;
 526         cell->virtual_device_list = device;
 527 }
 528
 529 static int pci_add_device(struct cell *cell, struct pci_device *device)
 530 {
 531         unsigned int size = device->info->msix_region_size;
 532         int err;
 533
 534         printk("Adding PCI device %02x:%02x.%x to cell \"%s\"\n",
 535                PCI_BDF_PARAMS(device->info->bdf), cell->config->name);
 536
 537         err = arch_pci_add_device(cell, device);
 538
 539         if (!err && device->info->msix_address) {
 540                 device->msix_table = page_alloc(&remap_pool, size / PAGE_SIZE);
 541                 if (!device->msix_table) {
 542                         err = -ENOMEM;
 543                         goto error_remove_dev;
 544                 }
 545
 546                 err = paging_create(&hv_paging_structs,
 547                                     device->info->msix_address, size,
 548                                     (unsigned long)device->msix_table,
 549                                     PAGE_DEFAULT_FLAGS | PAGE_FLAG_UNCACHED,
 550                                     PAGING_NON_COHERENT);
 551                 if (err)
 552                         goto error_page_free;
 553
 554                 device->next_msix_device = cell->msix_device_list;
 555                 cell->msix_device_list = device;
 556         }
 557         return 0;
 558
 559 error_page_free:
 560         page_free(&remap_pool, device->msix_table, size / PAGE_SIZE);
 561 error_remove_dev:
 562         arch_pci_remove_device(device);
 563         return err;
 564 }
 565
 566 static void pci_remove_virtual_device(struct pci_device *device)
 567 {
 568         struct pci_device *prev = device->cell->virtual_device_list;
 569
 570         if (prev == device) {
 571                 device->cell->virtual_device_list = device->next_virtual_device;
 572         } else {
 573                 while (prev->next_virtual_device != device)
 574                         prev = prev->next_virtual_device;
 575                 prev->next_virtual_device = device->next_virtual_device;
 576         }
 577 }
 578
 579 static void pci_remove_device(struct pci_device *device)
 580 {
 581         unsigned int size = device->info->msix_region_size;
 582         struct pci_device *prev_msix_device;
 583
 584         printk("Removing PCI device %02x:%02x.%x from cell \"%s\"\n",
 585                PCI_BDF_PARAMS(device->info->bdf), device->cell->config->name);
 586         arch_pci_remove_device(device);
 587         pci_write_config(device->info->bdf, PCI_CFG_COMMAND,
 588                          PCI_CMD_INTX_OFF, 2);
 589
 590         if (!device->msix_table)
 591                 return;
 592
 593         /* cannot fail, destruction of same size as construction */
 594         paging_destroy(&hv_paging_structs, (unsigned long)device->msix_table,
 595                        size, PAGING_NON_COHERENT);
 596         page_free(&remap_pool, device->msix_table, size / PAGE_SIZE);
 597
 598         prev_msix_device = device->cell->msix_device_list;
 599         if (prev_msix_device == device) {
 600                 device->cell->msix_device_list = device->next_msix_device;
 601         } else {
 602                 while (prev_msix_device->next_msix_device != device)
 603                         prev_msix_device = prev_msix_device->next_msix_device;
 604                 prev_msix_device->next_msix_device = device->next_msix_device;
 605         }
 606 }
 607
 608 /**
 609  * Perform PCI-specific initialization for a new cell.
 610  * @param cell  Cell to be initialized.
 611  *
 612  * @return 0 on success, negative error code otherwise.
 613  *
 614  * @see pci_cell_exit
 615  */
 616 int pci_cell_init(struct cell *cell)
 617 {
 618         unsigned int devlist_pages = PAGES(cell->config->num_pci_devices *
 619                                            sizeof(struct pci_device));
 620         const struct jailhouse_pci_device *dev_infos =
 621                 jailhouse_cell_pci_devices(cell->config);
 622         const struct jailhouse_pci_capability *cap;
 623         struct pci_device *device, *root_device;
 624         unsigned int ndev, ncap;
 625         int err;
 626
 627         cell->pci_devices = page_alloc(&mem_pool, devlist_pages);
 628         if (!cell->pci_devices)
 629                 return -ENOMEM;
 630
 631         /*
 632          * We order device states in the same way as the static information
 633          * so that we can use the index of the latter to find the former. For
 634          * the other way around and for obtaining the owner cell, we use more
 635          * handy pointers. The cell pointer also encodes active ownership.
 636          */
 637         for (ndev = 0; ndev < cell->config->num_pci_devices; ndev++) {
 638                 if (dev_infos[ndev].num_msix_vectors > PCI_MAX_MSIX_VECTORS) {
 639                         err = -ERANGE;
 640                         goto error;
 641                 }
 642
 643                 device = &cell->pci_devices[ndev];
 644                 device->info = &dev_infos[ndev];
 645
 646                 root_device = pci_get_assigned_device(&root_cell,
 647                                                       dev_infos[ndev].bdf);
 648                 if (root_device) {
 649                         pci_remove_device(root_device);
 650                         root_device->cell = NULL;
 651                 }
 652
 653                 err = pci_add_device(cell, device);
 654                 if (err)
 655                         goto error;
 656
 657                 device->cell = cell;
 658
 659                 for_each_pci_cap(cap, device, ncap)
 660                         if (cap->id == PCI_CAP_MSI)
 661                                 pci_save_msi(device, cap);
 662                         else if (cap->id == PCI_CAP_MSIX)
 663                                 pci_save_msix(device, cap);
 664         }
 665
 666         if (cell == &root_cell)
 667                 pci_prepare_handover();
 668
 669         return 0;
 670 error:
 671         pci_cell_exit(cell);
 672         return err;
 673 }
 674
 675 static void pci_return_device_to_root_cell(struct pci_device *device)
 676 {
 677         struct pci_device *root_device;
 678
 679         for_each_configured_pci_device(root_device, &root_cell)
 680                 if (root_device->info->domain == device->info->domain &&
 681                     root_device->info->bdf == device->info->bdf) {
 682                         if (pci_add_device(&root_cell, root_device) < 0)
 683                                 printk("WARNING: Failed to re-assign PCI "
 684                                        "device to root cell\n");
 685                         else
 686                                 root_device->cell = &root_cell;
 687                         break;
 688                 }
 689 }
 690
 691 /**
 692  * Perform PCI-specific cleanup for a cell under destruction.
 693  * @param cell  Cell to be destructed.
 694  *
 695  * @see pci_cell_init
 696  */
 697 void pci_cell_exit(struct cell *cell)
 698 {
 699         unsigned int devlist_pages = PAGES(cell->config->num_pci_devices *
 700                                            sizeof(struct pci_device));
 701         struct pci_device *device;
 702
 703         /*
 704          * Do not destroy the root cell. We will shut down the complete
 705          * hypervisor instead.
 706          */
 707         if (cell == &root_cell)
 708                 return;
 709
 710         for_each_configured_pci_device(device, cell)
 711                 if (device->cell) {
 712                         pci_remove_device(device);
 713                         pci_return_device_to_root_cell(device);
 714                 }
 715
 716         page_free(&mem_pool, cell->pci_devices, devlist_pages);
 717 }
 718
 719 /**
 720  * Apply PCI-specific configuration changes.
 721  * @param cell_added_removed    Cell that was added or removed to/from the
 722  *                              system or NULL.
 723  *
 724  * @see arch_config_commit
 725  */
 726 void pci_config_commit(struct cell *cell_added_removed)
 727 {
 728         const struct jailhouse_pci_capability *cap;
 729         struct pci_device *device;
 730         unsigned int n;
 731         int err = 0;
 732
 733         if (!cell_added_removed)
 734                 return;
 735
 736         for_each_configured_pci_device(device, &root_cell)
 737                 if (device->cell)
 738                         for_each_pci_cap(cap, device, n) {
 739                                 if (cap->id == PCI_CAP_MSI) {
 740                                         err = arch_pci_update_msi(device, cap);
 741                                 } else if (cap->id == PCI_CAP_MSIX) {
 742                                         err = pci_update_msix(device, cap);
 743                                         pci_suppress_msix(device, cap, false);
 744                                 }
 745                                 if (err)
 746                                         goto error;
 747                         }
 748         return;
 749
 750 error:
 751         panic_printk("FATAL: Unsupported MSI/MSI-X state, device %02x:%02x.%x,"
 752                      " cap %d\n", PCI_BDF_PARAMS(device->info->bdf), cap->id);
 753         panic_stop();
 754 }
 755
 756 /**
 757  * Shut down the PCI layer during hypervisor deactivation.
 758  */
 759 void pci_shutdown(void)
 760 {
 761         const struct jailhouse_pci_capability *cap;
 762         struct pci_device *device;
 763         unsigned int n;
 764
 765         if (!root_cell.pci_devices)
 766                 return;
 767
 768         for_each_configured_pci_device(device, &root_cell)
 769                 if (device->cell)
 770                         for_each_pci_cap(cap, device, n)
 771                                 if (cap->id == PCI_CAP_MSI)
 772                                         pci_restore_msi(device, cap);
 773                                 else if (cap->id == PCI_CAP_MSIX)
 774                                         pci_restore_msix(device, cap);
 775 }