drivers/pci/controller/vmd.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Volume Management Device driver
   4  * Copyright (c) 2015, Intel Corporation.
   5  */
   6
   7 #include <linux/device.h>
   8 #include <linux/interrupt.h>
   9 #include <linux/iommu.h>
  10 #include <linux/irq.h>
  11 #include <linux/kernel.h>
  12 #include <linux/module.h>
  13 #include <linux/msi.h>
  14 #include <linux/pci.h>
  15 #include <linux/pci-acpi.h>
  16 #include <linux/pci-ecam.h>
  17 #include <linux/srcu.h>
  18 #include <linux/rculist.h>
  19 #include <linux/rcupdate.h>
  20
  21 #include <asm/irqdomain.h>
  22
  23 #define VMD_CFGBAR      0
  24 #define VMD_MEMBAR1     2
  25 #define VMD_MEMBAR2     4
  26
  27 #define PCI_REG_VMCAP           0x40
  28 #define BUS_RESTRICT_CAP(vmcap) (vmcap & 0x1)
  29 #define PCI_REG_VMCONFIG        0x44
  30 #define BUS_RESTRICT_CFG(vmcfg) ((vmcfg >> 8) & 0x3)
  31 #define VMCONFIG_MSI_REMAP      0x2
  32 #define PCI_REG_VMLOCK          0x70
  33 #define MB2_SHADOW_EN(vmlock)   (vmlock & 0x2)
  34
  35 #define MB2_SHADOW_OFFSET       0x2000
  36 #define MB2_SHADOW_SIZE         16
  37
  38 enum vmd_features {
  39         /*
  40          * Device may contain registers which hint the physical location of the
  41          * membars, in order to allow proper address translation during
  42          * resource assignment to enable guest virtualization
  43          */
  44         VMD_FEAT_HAS_MEMBAR_SHADOW              = (1 << 0),
  45
  46         /*
  47          * Device may provide root port configuration information which limits
  48          * bus numbering
  49          */
  50         VMD_FEAT_HAS_BUS_RESTRICTIONS           = (1 << 1),
  51
  52         /*
  53          * Device contains physical location shadow registers in
  54          * vendor-specific capability space
  55          */
  56         VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP        = (1 << 2),
  57
  58         /*
  59          * Device may use MSI-X vector 0 for software triggering and will not
  60          * be used for MSI remapping
  61          */
  62         VMD_FEAT_OFFSET_FIRST_VECTOR            = (1 << 3),
  63
  64         /*
  65          * Device can bypass remapping MSI-X transactions into its MSI-X table,
  66          * avoiding the requirement of a VMD MSI domain for child device
  67          * interrupt handling.
  68          */
  69         VMD_FEAT_CAN_BYPASS_MSI_REMAP           = (1 << 4),
  70 };
  71
  72 static DEFINE_IDA(vmd_instance_ida);
  73
  74 /*
  75  * Lock for manipulating VMD IRQ lists.
  76  */
  77 static DEFINE_RAW_SPINLOCK(list_lock);
  78
  79 /**
  80  * struct vmd_irq - private data to map driver IRQ to the VMD shared vector
  81  * @node:       list item for parent traversal.
  82  * @irq:        back pointer to parent.
  83  * @enabled:    true if driver enabled IRQ
  84  * @virq:       the virtual IRQ value provided to the requesting driver.
  85  *
  86  * Every MSI/MSI-X IRQ requested for a device in a VMD domain will be mapped to
  87  * a VMD IRQ using this structure.
  88  */
  89 struct vmd_irq {
  90         struct list_head        node;
  91         struct vmd_irq_list     *irq;
  92         bool                    enabled;
  93         unsigned int            virq;
  94 };
  95
  96 /**
  97  * struct vmd_irq_list - list of driver requested IRQs mapping to a VMD vector
  98  * @irq_list:   the list of irq's the VMD one demuxes to.
  99  * @srcu:       SRCU struct for local synchronization.
 100  * @count:      number of child IRQs assigned to this vector; used to track
 101  *              sharing.
 102  * @virq:       The underlying VMD Linux interrupt number
 103  */
 104 struct vmd_irq_list {
 105         struct list_head        irq_list;
 106         struct srcu_struct      srcu;
 107         unsigned int            count;
 108         unsigned int            virq;
 109 };
 110
 111 struct vmd_dev {
 112         struct pci_dev          *dev;
 113
 114         spinlock_t              cfg_lock;
 115         void __iomem            *cfgbar;
 116
 117         int msix_count;
 118         struct vmd_irq_list     *irqs;
 119
 120         struct pci_sysdata      sysdata;
 121         struct resource         resources[3];
 122         struct irq_domain       *irq_domain;
 123         struct pci_bus          *bus;
 124         u8                      busn_start;
 125         u8                      first_vec;
 126         char                    *name;
 127         int                     instance;
 128 };
 129
 130 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
 131 {
 132         return container_of(bus->sysdata, struct vmd_dev, sysdata);
 133 }
 134
 135 static inline unsigned int index_from_irqs(struct vmd_dev *vmd,
 136                                            struct vmd_irq_list *irqs)
 137 {
 138         return irqs - vmd->irqs;
 139 }
 140
 141 /*
 142  * Drivers managing a device in a VMD domain allocate their own IRQs as before,
 143  * but the MSI entry for the hardware it's driving will be programmed with a
 144  * destination ID for the VMD MSI-X table.  The VMD muxes interrupts in its
 145  * domain into one of its own, and the VMD driver de-muxes these for the
 146  * handlers sharing that VMD IRQ.  The vmd irq_domain provides the operations
 147  * and irq_chip to set this up.
 148  */
 149 static void vmd_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
 150 {
 151         struct vmd_irq *vmdirq = data->chip_data;
 152         struct vmd_irq_list *irq = vmdirq->irq;
 153         struct vmd_dev *vmd = irq_data_get_irq_handler_data(data);
 154
 155         memset(msg, 0, sizeof(*msg));
 156         msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH;
 157         msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW;
 158         msg->arch_addr_lo.destid_0_7 = index_from_irqs(vmd, irq);
 159 }
 160
 161 /*
 162  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 163  */
 164 static void vmd_irq_enable(struct irq_data *data)
 165 {
 166         struct vmd_irq *vmdirq = data->chip_data;
 167         unsigned long flags;
 168
 169         raw_spin_lock_irqsave(&list_lock, flags);
 170         WARN_ON(vmdirq->enabled);
 171         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 172         vmdirq->enabled = true;
 173         raw_spin_unlock_irqrestore(&list_lock, flags);
 174
 175         data->chip->irq_unmask(data);
 176 }
 177
 178 static void vmd_irq_disable(struct irq_data *data)
 179 {
 180         struct vmd_irq *vmdirq = data->chip_data;
 181         unsigned long flags;
 182
 183         data->chip->irq_mask(data);
 184
 185         raw_spin_lock_irqsave(&list_lock, flags);
 186         if (vmdirq->enabled) {
 187                 list_del_rcu(&vmdirq->node);
 188                 vmdirq->enabled = false;
 189         }
 190         raw_spin_unlock_irqrestore(&list_lock, flags);
 191 }
 192
 193 /*
 194  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 195  * other devices sharing the same vector.
 196  */
 197 static int vmd_irq_set_affinity(struct irq_data *data,
 198                                 const struct cpumask *dest, bool force)
 199 {
 200         return -EINVAL;
 201 }
 202
 203 static struct irq_chip vmd_msi_controller = {
 204         .name                   = "VMD-MSI",
 205         .irq_enable             = vmd_irq_enable,
 206         .irq_disable            = vmd_irq_disable,
 207         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 208         .irq_set_affinity       = vmd_irq_set_affinity,
 209 };
 210
 211 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 212                                      msi_alloc_info_t *arg)
 213 {
 214         return 0;
 215 }
 216
 217 /*
 218  * XXX: We can be even smarter selecting the best IRQ once we solve the
 219  * affinity problem.
 220  */
 221 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 222 {
 223         unsigned long flags;
 224         int i, best;
 225
 226         if (vmd->msix_count == 1 + vmd->first_vec)
 227                 return &vmd->irqs[vmd->first_vec];
 228
 229         /*
 230          * White list for fast-interrupt handlers. All others will share the
 231          * "slow" interrupt vector.
 232          */
 233         switch (msi_desc_to_pci_dev(desc)->class) {
 234         case PCI_CLASS_STORAGE_EXPRESS:
 235                 break;
 236         default:
 237                 return &vmd->irqs[vmd->first_vec];
 238         }
 239
 240         raw_spin_lock_irqsave(&list_lock, flags);
 241         best = vmd->first_vec + 1;
 242         for (i = best; i < vmd->msix_count; i++)
 243                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 244                         best = i;
 245         vmd->irqs[best].count++;
 246         raw_spin_unlock_irqrestore(&list_lock, flags);
 247
 248         return &vmd->irqs[best];
 249 }
 250
 251 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 252                         unsigned int virq, irq_hw_number_t hwirq,
 253                         msi_alloc_info_t *arg)
 254 {
 255         struct msi_desc *desc = arg->desc;
 256         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 257         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 258
 259         if (!vmdirq)
 260                 return -ENOMEM;
 261
 262         INIT_LIST_HEAD(&vmdirq->node);
 263         vmdirq->irq = vmd_next_irq(vmd, desc);
 264         vmdirq->virq = virq;
 265
 266         irq_domain_set_info(domain, virq, vmdirq->irq->virq, info->chip, vmdirq,
 267                             handle_untracked_irq, vmd, NULL);
 268         return 0;
 269 }
 270
 271 static void vmd_msi_free(struct irq_domain *domain,
 272                         struct msi_domain_info *info, unsigned int virq)
 273 {
 274         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 275         unsigned long flags;
 276
 277         synchronize_srcu(&vmdirq->irq->srcu);
 278
 279         /* XXX: Potential optimization to rebalance */
 280         raw_spin_lock_irqsave(&list_lock, flags);
 281         vmdirq->irq->count--;
 282         raw_spin_unlock_irqrestore(&list_lock, flags);
 283
 284         kfree(vmdirq);
 285 }
 286
 287 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 288                            int nvec, msi_alloc_info_t *arg)
 289 {
 290         struct pci_dev *pdev = to_pci_dev(dev);
 291         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 292
 293         if (nvec > vmd->msix_count)
 294                 return vmd->msix_count;
 295
 296         memset(arg, 0, sizeof(*arg));
 297         return 0;
 298 }
 299
 300 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 301 {
 302         arg->desc = desc;
 303 }
 304
 305 static struct msi_domain_ops vmd_msi_domain_ops = {
 306         .get_hwirq      = vmd_get_hwirq,
 307         .msi_init       = vmd_msi_init,
 308         .msi_free       = vmd_msi_free,
 309         .msi_prepare    = vmd_msi_prepare,
 310         .set_desc       = vmd_set_desc,
 311 };
 312
 313 static struct msi_domain_info vmd_msi_domain_info = {
 314         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 315                           MSI_FLAG_PCI_MSIX,
 316         .ops            = &vmd_msi_domain_ops,
 317         .chip           = &vmd_msi_controller,
 318 };
 319
 320 static void vmd_set_msi_remapping(struct vmd_dev *vmd, bool enable)
 321 {
 322         u16 reg;
 323
 324         pci_read_config_word(vmd->dev, PCI_REG_VMCONFIG, &reg);
 325         reg = enable ? (reg & ~VMCONFIG_MSI_REMAP) :
 326                        (reg | VMCONFIG_MSI_REMAP);
 327         pci_write_config_word(vmd->dev, PCI_REG_VMCONFIG, reg);
 328 }
 329
 330 static int vmd_create_irq_domain(struct vmd_dev *vmd)
 331 {
 332         struct fwnode_handle *fn;
 333
 334         fn = irq_domain_alloc_named_id_fwnode("VMD-MSI", vmd->sysdata.domain);
 335         if (!fn)
 336                 return -ENODEV;
 337
 338         vmd->irq_domain = pci_msi_create_irq_domain(fn, &vmd_msi_domain_info, NULL);
 339         if (!vmd->irq_domain) {
 340                 irq_domain_free_fwnode(fn);
 341                 return -ENODEV;
 342         }
 343
 344         return 0;
 345 }
 346
 347 static void vmd_remove_irq_domain(struct vmd_dev *vmd)
 348 {
 349         /*
 350          * Some production BIOS won't enable remapping between soft reboots.
 351          * Ensure remapping is restored before unloading the driver.
 352          */
 353         if (!vmd->msix_count)
 354                 vmd_set_msi_remapping(vmd, true);
 355
 356         if (vmd->irq_domain) {
 357                 struct fwnode_handle *fn = vmd->irq_domain->fwnode;
 358
 359                 irq_domain_remove(vmd->irq_domain);
 360                 irq_domain_free_fwnode(fn);
 361         }
 362 }
 363
 364 static void __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 365                                   unsigned int devfn, int reg, int len)
 366 {
 367         unsigned int busnr_ecam = bus->number - vmd->busn_start;
 368         u32 offset = PCIE_ECAM_OFFSET(busnr_ecam, devfn, reg);
 369
 370         if (offset + len >= resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 371                 return NULL;
 372
 373         return vmd->cfgbar + offset;
 374 }
 375
 376 /*
 377  * CPU may deadlock if config space is not serialized on some versions of this
 378  * hardware, so all config space access is done under a spinlock.
 379  */
 380 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 381                         int len, u32 *value)
 382 {
 383         struct vmd_dev *vmd = vmd_from_bus(bus);
 384         void __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 385         unsigned long flags;
 386         int ret = 0;
 387
 388         if (!addr)
 389                 return -EFAULT;
 390
 391         spin_lock_irqsave(&vmd->cfg_lock, flags);
 392         switch (len) {
 393         case 1:
 394                 *value = readb(addr);
 395                 break;
 396         case 2:
 397                 *value = readw(addr);
 398                 break;
 399         case 4:
 400                 *value = readl(addr);
 401                 break;
 402         default:
 403                 ret = -EINVAL;
 404                 break;
 405         }
 406         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 407         return ret;
 408 }
 409
 410 /*
 411  * VMD h/w converts non-posted config writes to posted memory writes. The
 412  * read-back in this function forces the completion so it returns only after
 413  * the config space was written, as expected.
 414  */
 415 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 416                          int len, u32 value)
 417 {
 418         struct vmd_dev *vmd = vmd_from_bus(bus);
 419         void __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 420         unsigned long flags;
 421         int ret = 0;
 422
 423         if (!addr)
 424                 return -EFAULT;
 425
 426         spin_lock_irqsave(&vmd->cfg_lock, flags);
 427         switch (len) {
 428         case 1:
 429                 writeb(value, addr);
 430                 readb(addr);
 431                 break;
 432         case 2:
 433                 writew(value, addr);
 434                 readw(addr);
 435                 break;
 436         case 4:
 437                 writel(value, addr);
 438                 readl(addr);
 439                 break;
 440         default:
 441                 ret = -EINVAL;
 442                 break;
 443         }
 444         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 445         return ret;
 446 }
 447
 448 static struct pci_ops vmd_ops = {
 449         .read           = vmd_pci_read,
 450         .write          = vmd_pci_write,
 451 };
 452
 453 #ifdef CONFIG_ACPI
 454 static struct acpi_device *vmd_acpi_find_companion(struct pci_dev *pci_dev)
 455 {
 456         struct pci_host_bridge *bridge;
 457         u32 busnr, addr;
 458
 459         if (pci_dev->bus->ops != &vmd_ops)
 460                 return NULL;
 461
 462         bridge = pci_find_host_bridge(pci_dev->bus);
 463         busnr = pci_dev->bus->number - bridge->bus->number;
 464         /*
 465          * The address computation below is only applicable to relative bus
 466          * numbers below 32.
 467          */
 468         if (busnr > 31)
 469                 return NULL;
 470
 471         addr = (busnr << 24) | ((u32)pci_dev->devfn << 16) | 0x8000FFFFU;
 472
 473         dev_dbg(&pci_dev->dev, "Looking for ACPI companion (address 0x%x)\n",
 474                 addr);
 475
 476         return acpi_find_child_device(ACPI_COMPANION(bridge->dev.parent), addr,
 477                                       false);
 478 }
 479
 480 static bool hook_installed;
 481
 482 static void vmd_acpi_begin(void)
 483 {
 484         if (pci_acpi_set_companion_lookup_hook(vmd_acpi_find_companion))
 485                 return;
 486
 487         hook_installed = true;
 488 }
 489
 490 static void vmd_acpi_end(void)
 491 {
 492         if (!hook_installed)
 493                 return;
 494
 495         pci_acpi_clear_companion_lookup_hook();
 496         hook_installed = false;
 497 }
 498 #else
 499 static inline void vmd_acpi_begin(void) { }
 500 static inline void vmd_acpi_end(void) { }
 501 #endif /* CONFIG_ACPI */
 502
 503 static void vmd_domain_reset(struct vmd_dev *vmd)
 504 {
 505         u16 bus, max_buses = resource_size(&vmd->resources[0]);
 506         u8 dev, functions, fn, hdr_type;
 507         char __iomem *base;
 508
 509         for (bus = 0; bus < max_buses; bus++) {
 510                 for (dev = 0; dev < 32; dev++) {
 511                         base = vmd->cfgbar + PCIE_ECAM_OFFSET(bus,
 512                                                 PCI_DEVFN(dev, 0), 0);
 513
 514                         hdr_type = readb(base + PCI_HEADER_TYPE) &
 515                                          PCI_HEADER_TYPE_MASK;
 516
 517                         functions = (hdr_type & 0x80) ? 8 : 1;
 518                         for (fn = 0; fn < functions; fn++) {
 519                                 base = vmd->cfgbar + PCIE_ECAM_OFFSET(bus,
 520                                                 PCI_DEVFN(dev, fn), 0);
 521
 522                                 hdr_type = readb(base + PCI_HEADER_TYPE) &
 523                                                 PCI_HEADER_TYPE_MASK;
 524
 525                                 if (hdr_type != PCI_HEADER_TYPE_BRIDGE ||
 526                                     (readw(base + PCI_CLASS_DEVICE) !=
 527                                      PCI_CLASS_BRIDGE_PCI))
 528                                         continue;
 529
 530                                 memset_io(base + PCI_IO_BASE, 0,
 531                                           PCI_ROM_ADDRESS1 - PCI_IO_BASE);
 532                         }
 533                 }
 534         }
 535 }
 536
 537 static void vmd_attach_resources(struct vmd_dev *vmd)
 538 {
 539         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 540         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 541 }
 542
 543 static void vmd_detach_resources(struct vmd_dev *vmd)
 544 {
 545         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 546         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 547 }
 548
 549 /*
 550  * VMD domains start at 0x10000 to not clash with ACPI _SEG domains.
 551  * Per ACPI r6.0, sec 6.5.6,  _SEG returns an integer, of which the lower
 552  * 16 bits are the PCI Segment Group (domain) number.  Other bits are
 553  * currently reserved.
 554  */
 555 static int vmd_find_free_domain(void)
 556 {
 557         int domain = 0xffff;
 558         struct pci_bus *bus = NULL;
 559
 560         while ((bus = pci_find_next_bus(bus)) != NULL)
 561                 domain = max_t(int, domain, pci_domain_nr(bus));
 562         return domain + 1;
 563 }
 564
 565 static int vmd_get_phys_offsets(struct vmd_dev *vmd, bool native_hint,
 566                                 resource_size_t *offset1,
 567                                 resource_size_t *offset2)
 568 {
 569         struct pci_dev *dev = vmd->dev;
 570         u64 phys1, phys2;
 571
 572         if (native_hint) {
 573                 u32 vmlock;
 574                 int ret;
 575
 576                 ret = pci_read_config_dword(dev, PCI_REG_VMLOCK, &vmlock);
 577                 if (ret || PCI_POSSIBLE_ERROR(vmlock))
 578                         return -ENODEV;
 579
 580                 if (MB2_SHADOW_EN(vmlock)) {
 581                         void __iomem *membar2;
 582
 583                         membar2 = pci_iomap(dev, VMD_MEMBAR2, 0);
 584                         if (!membar2)
 585                                 return -ENOMEM;
 586                         phys1 = readq(membar2 + MB2_SHADOW_OFFSET);
 587                         phys2 = readq(membar2 + MB2_SHADOW_OFFSET + 8);
 588                         pci_iounmap(dev, membar2);
 589                 } else
 590                         return 0;
 591         } else {
 592                 /* Hypervisor-Emulated Vendor-Specific Capability */
 593                 int pos = pci_find_capability(dev, PCI_CAP_ID_VNDR);
 594                 u32 reg, regu;
 595
 596                 pci_read_config_dword(dev, pos + 4, &reg);
 597
 598                 /* "SHDW" */
 599                 if (pos && reg == 0x53484457) {
 600                         pci_read_config_dword(dev, pos + 8, &reg);
 601                         pci_read_config_dword(dev, pos + 12, &regu);
 602                         phys1 = (u64) regu << 32 | reg;
 603
 604                         pci_read_config_dword(dev, pos + 16, &reg);
 605                         pci_read_config_dword(dev, pos + 20, &regu);
 606                         phys2 = (u64) regu << 32 | reg;
 607                 } else
 608                         return 0;
 609         }
 610
 611         *offset1 = dev->resource[VMD_MEMBAR1].start -
 612                         (phys1 & PCI_BASE_ADDRESS_MEM_MASK);
 613         *offset2 = dev->resource[VMD_MEMBAR2].start -
 614                         (phys2 & PCI_BASE_ADDRESS_MEM_MASK);
 615
 616         return 0;
 617 }
 618
 619 static int vmd_get_bus_number_start(struct vmd_dev *vmd)
 620 {
 621         struct pci_dev *dev = vmd->dev;
 622         u16 reg;
 623
 624         pci_read_config_word(dev, PCI_REG_VMCAP, &reg);
 625         if (BUS_RESTRICT_CAP(reg)) {
 626                 pci_read_config_word(dev, PCI_REG_VMCONFIG, &reg);
 627
 628                 switch (BUS_RESTRICT_CFG(reg)) {
 629                 case 0:
 630                         vmd->busn_start = 0;
 631                         break;
 632                 case 1:
 633                         vmd->busn_start = 128;
 634                         break;
 635                 case 2:
 636                         vmd->busn_start = 224;
 637                         break;
 638                 default:
 639                         pci_err(dev, "Unknown Bus Offset Setting (%d)\n",
 640                                 BUS_RESTRICT_CFG(reg));
 641                         return -ENODEV;
 642                 }
 643         }
 644
 645         return 0;
 646 }
 647
 648 static irqreturn_t vmd_irq(int irq, void *data)
 649 {
 650         struct vmd_irq_list *irqs = data;
 651         struct vmd_irq *vmdirq;
 652         int idx;
 653
 654         idx = srcu_read_lock(&irqs->srcu);
 655         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 656                 generic_handle_irq(vmdirq->virq);
 657         srcu_read_unlock(&irqs->srcu, idx);
 658
 659         return IRQ_HANDLED;
 660 }
 661
 662 static int vmd_alloc_irqs(struct vmd_dev *vmd)
 663 {
 664         struct pci_dev *dev = vmd->dev;
 665         int i, err;
 666
 667         vmd->msix_count = pci_msix_vec_count(dev);
 668         if (vmd->msix_count < 0)
 669                 return -ENODEV;
 670
 671         vmd->msix_count = pci_alloc_irq_vectors(dev, vmd->first_vec + 1,
 672                                                 vmd->msix_count, PCI_IRQ_MSIX);
 673         if (vmd->msix_count < 0)
 674                 return vmd->msix_count;
 675
 676         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 677                                  GFP_KERNEL);
 678         if (!vmd->irqs)
 679                 return -ENOMEM;
 680
 681         for (i = 0; i < vmd->msix_count; i++) {
 682                 err = init_srcu_struct(&vmd->irqs[i].srcu);
 683                 if (err)
 684                         return err;
 685
 686                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 687                 vmd->irqs[i].virq = pci_irq_vector(dev, i);
 688                 err = devm_request_irq(&dev->dev, vmd->irqs[i].virq,
 689                                        vmd_irq, IRQF_NO_THREAD,
 690                                        vmd->name, &vmd->irqs[i]);
 691                 if (err)
 692                         return err;
 693         }
 694
 695         return 0;
 696 }
 697
 698 /*
 699  * Since VMD is an aperture to regular PCIe root ports, only allow it to
 700  * control features that the OS is allowed to control on the physical PCI bus.
 701  */
 702 static void vmd_copy_host_bridge_flags(struct pci_host_bridge *root_bridge,
 703                                        struct pci_host_bridge *vmd_bridge)
 704 {
 705         vmd_bridge->native_pcie_hotplug = root_bridge->native_pcie_hotplug;
 706         vmd_bridge->native_shpc_hotplug = root_bridge->native_shpc_hotplug;
 707         vmd_bridge->native_aer = root_bridge->native_aer;
 708         vmd_bridge->native_pme = root_bridge->native_pme;
 709         vmd_bridge->native_ltr = root_bridge->native_ltr;
 710         vmd_bridge->native_dpc = root_bridge->native_dpc;
 711 }
 712
 713 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 714 {
 715         struct pci_sysdata *sd = &vmd->sysdata;
 716         struct resource *res;
 717         u32 upper_bits;
 718         unsigned long flags;
 719         LIST_HEAD(resources);
 720         resource_size_t offset[2] = {0};
 721         resource_size_t membar2_offset = 0x2000;
 722         struct pci_bus *child;
 723         int ret;
 724
 725         /*
 726          * Shadow registers may exist in certain VMD device ids which allow
 727          * guests to correctly assign host physical addresses to the root ports
 728          * and child devices. These registers will either return the host value
 729          * or 0, depending on an enable bit in the VMD device.
 730          */
 731         if (features & VMD_FEAT_HAS_MEMBAR_SHADOW) {
 732                 membar2_offset = MB2_SHADOW_OFFSET + MB2_SHADOW_SIZE;
 733                 ret = vmd_get_phys_offsets(vmd, true, &offset[0], &offset[1]);
 734                 if (ret)
 735                         return ret;
 736         } else if (features & VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP) {
 737                 ret = vmd_get_phys_offsets(vmd, false, &offset[0], &offset[1]);
 738                 if (ret)
 739                         return ret;
 740         }
 741
 742         /*
 743          * Certain VMD devices may have a root port configuration option which
 744          * limits the bus range to between 0-127, 128-255, or 224-255
 745          */
 746         if (features & VMD_FEAT_HAS_BUS_RESTRICTIONS) {
 747                 ret = vmd_get_bus_number_start(vmd);
 748                 if (ret)
 749                         return ret;
 750         }
 751
 752         res = &vmd->dev->resource[VMD_CFGBAR];
 753         vmd->resources[0] = (struct resource) {
 754                 .name  = "VMD CFGBAR",
 755                 .start = vmd->busn_start,
 756                 .end   = vmd->busn_start + (resource_size(res) >> 20) - 1,
 757                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 758         };
 759
 760         /*
 761          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 762          * put 32-bit resources in the window.
 763          *
 764          * There's no hardware reason why a 64-bit window *couldn't*
 765          * contain a 32-bit resource, but pbus_size_mem() computes the
 766          * bridge window size assuming a 64-bit window will contain no
 767          * 32-bit resources.  __pci_assign_resource() enforces that
 768          * artificial restriction to make sure everything will fit.
 769          *
 770          * The only way we could use a 64-bit non-prefetchable MEMBAR is
 771          * if its address is <4GB so that we can convert it to a 32-bit
 772          * resource.  To be visible to the host OS, all VMD endpoints must
 773          * be initially configured by platform BIOS, which includes setting
 774          * up these resources.  We can assume the device is configured
 775          * according to the platform needs.
 776          */
 777         res = &vmd->dev->resource[VMD_MEMBAR1];
 778         upper_bits = upper_32_bits(res->end);
 779         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 780         if (!upper_bits)
 781                 flags &= ~IORESOURCE_MEM_64;
 782         vmd->resources[1] = (struct resource) {
 783                 .name  = "VMD MEMBAR1",
 784                 .start = res->start,
 785                 .end   = res->end,
 786                 .flags = flags,
 787                 .parent = res,
 788         };
 789
 790         res = &vmd->dev->resource[VMD_MEMBAR2];
 791         upper_bits = upper_32_bits(res->end);
 792         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 793         if (!upper_bits)
 794                 flags &= ~IORESOURCE_MEM_64;
 795         vmd->resources[2] = (struct resource) {
 796                 .name  = "VMD MEMBAR2",
 797                 .start = res->start + membar2_offset,
 798                 .end   = res->end,
 799                 .flags = flags,
 800                 .parent = res,
 801         };
 802
 803         sd->vmd_dev = vmd->dev;
 804         sd->domain = vmd_find_free_domain();
 805         if (sd->domain < 0)
 806                 return sd->domain;
 807
 808         sd->node = pcibus_to_node(vmd->dev->bus);
 809
 810         /*
 811          * Currently MSI remapping must be enabled in guest passthrough mode
 812          * due to some missing interrupt remapping plumbing. This is probably
 813          * acceptable because the guest is usually CPU-limited and MSI
 814          * remapping doesn't become a performance bottleneck.
 815          */
 816         if (iommu_capable(vmd->dev->dev.bus, IOMMU_CAP_INTR_REMAP) ||
 817             !(features & VMD_FEAT_CAN_BYPASS_MSI_REMAP) ||
 818             offset[0] || offset[1]) {
 819                 ret = vmd_alloc_irqs(vmd);
 820                 if (ret)
 821                         return ret;
 822
 823                 vmd_set_msi_remapping(vmd, true);
 824
 825                 ret = vmd_create_irq_domain(vmd);
 826                 if (ret)
 827                         return ret;
 828
 829                 /*
 830                  * Override the IRQ domain bus token so the domain can be
 831                  * distinguished from a regular PCI/MSI domain.
 832                  */
 833                 irq_domain_update_bus_token(vmd->irq_domain, DOMAIN_BUS_VMD_MSI);
 834         } else {
 835                 vmd_set_msi_remapping(vmd, false);
 836         }
 837
 838         pci_add_resource(&resources, &vmd->resources[0]);
 839         pci_add_resource_offset(&resources, &vmd->resources[1], offset[0]);
 840         pci_add_resource_offset(&resources, &vmd->resources[2], offset[1]);
 841
 842         vmd->bus = pci_create_root_bus(&vmd->dev->dev, vmd->busn_start,
 843                                        &vmd_ops, sd, &resources);
 844         if (!vmd->bus) {
 845                 pci_free_resource_list(&resources);
 846                 vmd_remove_irq_domain(vmd);
 847                 return -ENODEV;
 848         }
 849
 850         vmd_copy_host_bridge_flags(pci_find_host_bridge(vmd->dev->bus),
 851                                    to_pci_host_bridge(vmd->bus->bridge));
 852
 853         vmd_attach_resources(vmd);
 854         if (vmd->irq_domain)
 855                 dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 856
 857         vmd_acpi_begin();
 858
 859         pci_scan_child_bus(vmd->bus);
 860         vmd_domain_reset(vmd);
 861         list_for_each_entry(child, &vmd->bus->children, node)
 862                 pci_reset_bus(child->self);
 863         pci_assign_unassigned_bus_resources(vmd->bus);
 864
 865         /*
 866          * VMD root buses are virtual and don't return true on pci_is_pcie()
 867          * and will fail pcie_bus_configure_settings() early. It can instead be
 868          * run on each of the real root ports.
 869          */
 870         list_for_each_entry(child, &vmd->bus->children, node)
 871                 pcie_bus_configure_settings(child);
 872
 873         pci_bus_add_devices(vmd->bus);
 874
 875         vmd_acpi_end();
 876
 877         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 878                                "domain"), "Can't create symlink to domain\n");
 879         return 0;
 880 }
 881
 882 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 883 {
 884         unsigned long features = (unsigned long) id->driver_data;
 885         struct vmd_dev *vmd;
 886         int err;
 887
 888         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 889                 return -ENOMEM;
 890
 891         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 892         if (!vmd)
 893                 return -ENOMEM;
 894
 895         vmd->dev = dev;
 896         vmd->instance = ida_simple_get(&vmd_instance_ida, 0, 0, GFP_KERNEL);
 897         if (vmd->instance < 0)
 898                 return vmd->instance;
 899
 900         vmd->name = kasprintf(GFP_KERNEL, "vmd%d", vmd->instance);
 901         if (!vmd->name) {
 902                 err = -ENOMEM;
 903                 goto out_release_instance;
 904         }
 905
 906         err = pcim_enable_device(dev);
 907         if (err < 0)
 908                 goto out_release_instance;
 909
 910         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 911         if (!vmd->cfgbar) {
 912                 err = -ENOMEM;
 913                 goto out_release_instance;
 914         }
 915
 916         pci_set_master(dev);
 917         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 918             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32))) {
 919                 err = -ENODEV;
 920                 goto out_release_instance;
 921         }
 922
 923         if (features & VMD_FEAT_OFFSET_FIRST_VECTOR)
 924                 vmd->first_vec = 1;
 925
 926         spin_lock_init(&vmd->cfg_lock);
 927         pci_set_drvdata(dev, vmd);
 928         err = vmd_enable_domain(vmd, features);
 929         if (err)
 930                 goto out_release_instance;
 931
 932         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 933                  vmd->sysdata.domain);
 934         return 0;
 935
 936  out_release_instance:
 937         ida_simple_remove(&vmd_instance_ida, vmd->instance);
 938         kfree(vmd->name);
 939         return err;
 940 }
 941
 942 static void vmd_cleanup_srcu(struct vmd_dev *vmd)
 943 {
 944         int i;
 945
 946         for (i = 0; i < vmd->msix_count; i++)
 947                 cleanup_srcu_struct(&vmd->irqs[i].srcu);
 948 }
 949
 950 static void vmd_remove(struct pci_dev *dev)
 951 {
 952         struct vmd_dev *vmd = pci_get_drvdata(dev);
 953
 954         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 955         pci_stop_root_bus(vmd->bus);
 956         pci_remove_root_bus(vmd->bus);
 957         vmd_cleanup_srcu(vmd);
 958         vmd_detach_resources(vmd);
 959         vmd_remove_irq_domain(vmd);
 960         ida_simple_remove(&vmd_instance_ida, vmd->instance);
 961         kfree(vmd->name);
 962 }
 963
 964 #ifdef CONFIG_PM_SLEEP
 965 static int vmd_suspend(struct device *dev)
 966 {
 967         struct pci_dev *pdev = to_pci_dev(dev);
 968         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 969         int i;
 970
 971         for (i = 0; i < vmd->msix_count; i++)
 972                 devm_free_irq(dev, vmd->irqs[i].virq, &vmd->irqs[i]);
 973
 974         return 0;
 975 }
 976
 977 static int vmd_resume(struct device *dev)
 978 {
 979         struct pci_dev *pdev = to_pci_dev(dev);
 980         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 981         int err, i;
 982
 983         for (i = 0; i < vmd->msix_count; i++) {
 984                 err = devm_request_irq(dev, vmd->irqs[i].virq,
 985                                        vmd_irq, IRQF_NO_THREAD,
 986                                        vmd->name, &vmd->irqs[i]);
 987                 if (err)
 988                         return err;
 989         }
 990
 991         return 0;
 992 }
 993 #endif
 994 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 995
 996 static const struct pci_device_id vmd_ids[] = {
 997         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_201D),
 998                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP,},
 999         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_28C0),
1000                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW |
1001                                 VMD_FEAT_HAS_BUS_RESTRICTIONS |
1002                                 VMD_FEAT_CAN_BYPASS_MSI_REMAP,},
1003         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x467f),
1004                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
1005                                 VMD_FEAT_HAS_BUS_RESTRICTIONS |
1006                                 VMD_FEAT_OFFSET_FIRST_VECTOR,},
1007         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x4c3d),
1008                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
1009                                 VMD_FEAT_HAS_BUS_RESTRICTIONS |
1010                                 VMD_FEAT_OFFSET_FIRST_VECTOR,},
1011         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0xa77f),
1012                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
1013                                 VMD_FEAT_HAS_BUS_RESTRICTIONS |
1014                                 VMD_FEAT_OFFSET_FIRST_VECTOR,},
1015         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B),
1016                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
1017                                 VMD_FEAT_HAS_BUS_RESTRICTIONS |
1018                                 VMD_FEAT_OFFSET_FIRST_VECTOR,},
1019         {0,}
1020 };
1021 MODULE_DEVICE_TABLE(pci, vmd_ids);
1022
1023 static struct pci_driver vmd_drv = {
1024         .name           = "vmd",
1025         .id_table       = vmd_ids,
1026         .probe          = vmd_probe,
1027         .remove         = vmd_remove,
1028         .driver         = {
1029                 .pm     = &vmd_dev_pm_ops,
1030         },
1031 };
1032 module_pci_driver(vmd_drv);
1033
1034 MODULE_AUTHOR("Intel Corporation");
1035 MODULE_LICENSE("GPL v2");
1036 MODULE_VERSION("0.6");