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/irq.h>
  10 #include <linux/kernel.h>
  11 #include <linux/module.h>
  12 #include <linux/msi.h>
  13 #include <linux/pci.h>
  14 #include <linux/srcu.h>
  15 #include <linux/rculist.h>
  16 #include <linux/rcupdate.h>
  17
  18 #include <asm/irqdomain.h>
  19 #include <asm/device.h>
  20 #include <asm/msi.h>
  21 #include <asm/msidef.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 PCI_REG_VMLOCK          0x70
  32 #define MB2_SHADOW_EN(vmlock)   (vmlock & 0x2)
  33
  34 #define MB2_SHADOW_OFFSET       0x2000
  35 #define MB2_SHADOW_SIZE         16
  36
  37 enum vmd_features {
  38         /*
  39          * Device may contain registers which hint the physical location of the
  40          * membars, in order to allow proper address translation during
  41          * resource assignment to enable guest virtualization
  42          */
  43         VMD_FEAT_HAS_MEMBAR_SHADOW      = (1 << 0),
  44
  45         /*
  46          * Device may provide root port configuration information which limits
  47          * bus numbering
  48          */
  49         VMD_FEAT_HAS_BUS_RESTRICTIONS   = (1 << 1),
  50 };
  51
  52 /*
  53  * Lock for manipulating VMD IRQ lists.
  54  */
  55 static DEFINE_RAW_SPINLOCK(list_lock);
  56
  57 /**
  58  * struct vmd_irq - private data to map driver IRQ to the VMD shared vector
  59  * @node:       list item for parent traversal.
  60  * @irq:        back pointer to parent.
  61  * @enabled:    true if driver enabled IRQ
  62  * @virq:       the virtual IRQ value provided to the requesting driver.
  63  *
  64  * Every MSI/MSI-X IRQ requested for a device in a VMD domain will be mapped to
  65  * a VMD IRQ using this structure.
  66  */
  67 struct vmd_irq {
  68         struct list_head        node;
  69         struct vmd_irq_list     *irq;
  70         bool                    enabled;
  71         unsigned int            virq;
  72 };
  73
  74 /**
  75  * struct vmd_irq_list - list of driver requested IRQs mapping to a VMD vector
  76  * @irq_list:   the list of irq's the VMD one demuxes to.
  77  * @srcu:       SRCU struct for local synchronization.
  78  * @count:      number of child IRQs assigned to this vector; used to track
  79  *              sharing.
  80  */
  81 struct vmd_irq_list {
  82         struct list_head        irq_list;
  83         struct srcu_struct      srcu;
  84         unsigned int            count;
  85 };
  86
  87 struct vmd_dev {
  88         struct pci_dev          *dev;
  89
  90         spinlock_t              cfg_lock;
  91         char __iomem            *cfgbar;
  92
  93         int msix_count;
  94         struct vmd_irq_list     *irqs;
  95
  96         struct pci_sysdata      sysdata;
  97         struct resource         resources[3];
  98         struct irq_domain       *irq_domain;
  99         struct pci_bus          *bus;
 100         u8                      busn_start;
 101
 102 #ifdef CONFIG_X86_DEV_DMA_OPS
 103         struct dma_map_ops      dma_ops;
 104         struct dma_domain       dma_domain;
 105 #endif
 106 };
 107
 108 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
 109 {
 110         return container_of(bus->sysdata, struct vmd_dev, sysdata);
 111 }
 112
 113 static inline unsigned int index_from_irqs(struct vmd_dev *vmd,
 114                                            struct vmd_irq_list *irqs)
 115 {
 116         return irqs - vmd->irqs;
 117 }
 118
 119 /*
 120  * Drivers managing a device in a VMD domain allocate their own IRQs as before,
 121  * but the MSI entry for the hardware it's driving will be programmed with a
 122  * destination ID for the VMD MSI-X table.  The VMD muxes interrupts in its
 123  * domain into one of its own, and the VMD driver de-muxes these for the
 124  * handlers sharing that VMD IRQ.  The vmd irq_domain provides the operations
 125  * and irq_chip to set this up.
 126  */
 127 static void vmd_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
 128 {
 129         struct vmd_irq *vmdirq = data->chip_data;
 130         struct vmd_irq_list *irq = vmdirq->irq;
 131         struct vmd_dev *vmd = irq_data_get_irq_handler_data(data);
 132
 133         msg->address_hi = MSI_ADDR_BASE_HI;
 134         msg->address_lo = MSI_ADDR_BASE_LO |
 135                           MSI_ADDR_DEST_ID(index_from_irqs(vmd, irq));
 136         msg->data = 0;
 137 }
 138
 139 /*
 140  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 141  */
 142 static void vmd_irq_enable(struct irq_data *data)
 143 {
 144         struct vmd_irq *vmdirq = data->chip_data;
 145         unsigned long flags;
 146
 147         raw_spin_lock_irqsave(&list_lock, flags);
 148         WARN_ON(vmdirq->enabled);
 149         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 150         vmdirq->enabled = true;
 151         raw_spin_unlock_irqrestore(&list_lock, flags);
 152
 153         data->chip->irq_unmask(data);
 154 }
 155
 156 static void vmd_irq_disable(struct irq_data *data)
 157 {
 158         struct vmd_irq *vmdirq = data->chip_data;
 159         unsigned long flags;
 160
 161         data->chip->irq_mask(data);
 162
 163         raw_spin_lock_irqsave(&list_lock, flags);
 164         if (vmdirq->enabled) {
 165                 list_del_rcu(&vmdirq->node);
 166                 vmdirq->enabled = false;
 167         }
 168         raw_spin_unlock_irqrestore(&list_lock, flags);
 169 }
 170
 171 /*
 172  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 173  * other devices sharing the same vector.
 174  */
 175 static int vmd_irq_set_affinity(struct irq_data *data,
 176                                 const struct cpumask *dest, bool force)
 177 {
 178         return -EINVAL;
 179 }
 180
 181 static struct irq_chip vmd_msi_controller = {
 182         .name                   = "VMD-MSI",
 183         .irq_enable             = vmd_irq_enable,
 184         .irq_disable            = vmd_irq_disable,
 185         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 186         .irq_set_affinity       = vmd_irq_set_affinity,
 187 };
 188
 189 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 190                                      msi_alloc_info_t *arg)
 191 {
 192         return 0;
 193 }
 194
 195 /*
 196  * XXX: We can be even smarter selecting the best IRQ once we solve the
 197  * affinity problem.
 198  */
 199 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 200 {
 201         int i, best = 1;
 202         unsigned long flags;
 203
 204         if (vmd->msix_count == 1)
 205                 return &vmd->irqs[0];
 206
 207         /*
 208          * White list for fast-interrupt handlers. All others will share the
 209          * "slow" interrupt vector.
 210          */
 211         switch (msi_desc_to_pci_dev(desc)->class) {
 212         case PCI_CLASS_STORAGE_EXPRESS:
 213                 break;
 214         default:
 215                 return &vmd->irqs[0];
 216         }
 217
 218         raw_spin_lock_irqsave(&list_lock, flags);
 219         for (i = 1; i < vmd->msix_count; i++)
 220                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 221                         best = i;
 222         vmd->irqs[best].count++;
 223         raw_spin_unlock_irqrestore(&list_lock, flags);
 224
 225         return &vmd->irqs[best];
 226 }
 227
 228 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 229                         unsigned int virq, irq_hw_number_t hwirq,
 230                         msi_alloc_info_t *arg)
 231 {
 232         struct msi_desc *desc = arg->desc;
 233         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 234         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 235         unsigned int index, vector;
 236
 237         if (!vmdirq)
 238                 return -ENOMEM;
 239
 240         INIT_LIST_HEAD(&vmdirq->node);
 241         vmdirq->irq = vmd_next_irq(vmd, desc);
 242         vmdirq->virq = virq;
 243         index = index_from_irqs(vmd, vmdirq->irq);
 244         vector = pci_irq_vector(vmd->dev, index);
 245
 246         irq_domain_set_info(domain, virq, vector, info->chip, vmdirq,
 247                             handle_untracked_irq, vmd, NULL);
 248         return 0;
 249 }
 250
 251 static void vmd_msi_free(struct irq_domain *domain,
 252                         struct msi_domain_info *info, unsigned int virq)
 253 {
 254         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 255         unsigned long flags;
 256
 257         synchronize_srcu(&vmdirq->irq->srcu);
 258
 259         /* XXX: Potential optimization to rebalance */
 260         raw_spin_lock_irqsave(&list_lock, flags);
 261         vmdirq->irq->count--;
 262         raw_spin_unlock_irqrestore(&list_lock, flags);
 263
 264         kfree(vmdirq);
 265 }
 266
 267 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 268                            int nvec, msi_alloc_info_t *arg)
 269 {
 270         struct pci_dev *pdev = to_pci_dev(dev);
 271         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 272
 273         if (nvec > vmd->msix_count)
 274                 return vmd->msix_count;
 275
 276         memset(arg, 0, sizeof(*arg));
 277         return 0;
 278 }
 279
 280 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 281 {
 282         arg->desc = desc;
 283 }
 284
 285 static struct msi_domain_ops vmd_msi_domain_ops = {
 286         .get_hwirq      = vmd_get_hwirq,
 287         .msi_init       = vmd_msi_init,
 288         .msi_free       = vmd_msi_free,
 289         .msi_prepare    = vmd_msi_prepare,
 290         .set_desc       = vmd_set_desc,
 291 };
 292
 293 static struct msi_domain_info vmd_msi_domain_info = {
 294         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 295                           MSI_FLAG_PCI_MSIX,
 296         .ops            = &vmd_msi_domain_ops,
 297         .chip           = &vmd_msi_controller,
 298 };
 299
 300 #ifdef CONFIG_X86_DEV_DMA_OPS
 301 /*
 302  * VMD replaces the requester ID with its own.  DMA mappings for devices in a
 303  * VMD domain need to be mapped for the VMD, not the device requiring
 304  * the mapping.
 305  */
 306 static struct device *to_vmd_dev(struct device *dev)
 307 {
 308         struct pci_dev *pdev = to_pci_dev(dev);
 309         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 310
 311         return &vmd->dev->dev;
 312 }
 313
 314 static const struct dma_map_ops *vmd_dma_ops(struct device *dev)
 315 {
 316         return get_dma_ops(to_vmd_dev(dev));
 317 }
 318
 319 static void *vmd_alloc(struct device *dev, size_t size, dma_addr_t *addr,
 320                        gfp_t flag, unsigned long attrs)
 321 {
 322         return vmd_dma_ops(dev)->alloc(to_vmd_dev(dev), size, addr, flag,
 323                                        attrs);
 324 }
 325
 326 static void vmd_free(struct device *dev, size_t size, void *vaddr,
 327                      dma_addr_t addr, unsigned long attrs)
 328 {
 329         return vmd_dma_ops(dev)->free(to_vmd_dev(dev), size, vaddr, addr,
 330                                       attrs);
 331 }
 332
 333 static int vmd_mmap(struct device *dev, struct vm_area_struct *vma,
 334                     void *cpu_addr, dma_addr_t addr, size_t size,
 335                     unsigned long attrs)
 336 {
 337         return vmd_dma_ops(dev)->mmap(to_vmd_dev(dev), vma, cpu_addr, addr,
 338                                       size, attrs);
 339 }
 340
 341 static int vmd_get_sgtable(struct device *dev, struct sg_table *sgt,
 342                            void *cpu_addr, dma_addr_t addr, size_t size,
 343                            unsigned long attrs)
 344 {
 345         return vmd_dma_ops(dev)->get_sgtable(to_vmd_dev(dev), sgt, cpu_addr,
 346                                              addr, size, attrs);
 347 }
 348
 349 static dma_addr_t vmd_map_page(struct device *dev, struct page *page,
 350                                unsigned long offset, size_t size,
 351                                enum dma_data_direction dir,
 352                                unsigned long attrs)
 353 {
 354         return vmd_dma_ops(dev)->map_page(to_vmd_dev(dev), page, offset, size,
 355                                           dir, attrs);
 356 }
 357
 358 static void vmd_unmap_page(struct device *dev, dma_addr_t addr, size_t size,
 359                            enum dma_data_direction dir, unsigned long attrs)
 360 {
 361         vmd_dma_ops(dev)->unmap_page(to_vmd_dev(dev), addr, size, dir, attrs);
 362 }
 363
 364 static int vmd_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 365                       enum dma_data_direction dir, unsigned long attrs)
 366 {
 367         return vmd_dma_ops(dev)->map_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 368 }
 369
 370 static void vmd_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 371                          enum dma_data_direction dir, unsigned long attrs)
 372 {
 373         vmd_dma_ops(dev)->unmap_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 374 }
 375
 376 static void vmd_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 377                                     size_t size, enum dma_data_direction dir)
 378 {
 379         vmd_dma_ops(dev)->sync_single_for_cpu(to_vmd_dev(dev), addr, size, dir);
 380 }
 381
 382 static void vmd_sync_single_for_device(struct device *dev, dma_addr_t addr,
 383                                        size_t size, enum dma_data_direction dir)
 384 {
 385         vmd_dma_ops(dev)->sync_single_for_device(to_vmd_dev(dev), addr, size,
 386                                                  dir);
 387 }
 388
 389 static void vmd_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 390                                 int nents, enum dma_data_direction dir)
 391 {
 392         vmd_dma_ops(dev)->sync_sg_for_cpu(to_vmd_dev(dev), sg, nents, dir);
 393 }
 394
 395 static void vmd_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 396                                    int nents, enum dma_data_direction dir)
 397 {
 398         vmd_dma_ops(dev)->sync_sg_for_device(to_vmd_dev(dev), sg, nents, dir);
 399 }
 400
 401 static int vmd_mapping_error(struct device *dev, dma_addr_t addr)
 402 {
 403         return vmd_dma_ops(dev)->mapping_error(to_vmd_dev(dev), addr);
 404 }
 405
 406 static int vmd_dma_supported(struct device *dev, u64 mask)
 407 {
 408         return vmd_dma_ops(dev)->dma_supported(to_vmd_dev(dev), mask);
 409 }
 410
 411 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 412 static u64 vmd_get_required_mask(struct device *dev)
 413 {
 414         return vmd_dma_ops(dev)->get_required_mask(to_vmd_dev(dev));
 415 }
 416 #endif
 417
 418 static void vmd_teardown_dma_ops(struct vmd_dev *vmd)
 419 {
 420         struct dma_domain *domain = &vmd->dma_domain;
 421
 422         if (get_dma_ops(&vmd->dev->dev))
 423                 del_dma_domain(domain);
 424 }
 425
 426 #define ASSIGN_VMD_DMA_OPS(source, dest, fn)    \
 427         do {                                    \
 428                 if (source->fn)                 \
 429                         dest->fn = vmd_##fn;    \
 430         } while (0)
 431
 432 static void vmd_setup_dma_ops(struct vmd_dev *vmd)
 433 {
 434         const struct dma_map_ops *source = get_dma_ops(&vmd->dev->dev);
 435         struct dma_map_ops *dest = &vmd->dma_ops;
 436         struct dma_domain *domain = &vmd->dma_domain;
 437
 438         domain->domain_nr = vmd->sysdata.domain;
 439         domain->dma_ops = dest;
 440
 441         if (!source)
 442                 return;
 443         ASSIGN_VMD_DMA_OPS(source, dest, alloc);
 444         ASSIGN_VMD_DMA_OPS(source, dest, free);
 445         ASSIGN_VMD_DMA_OPS(source, dest, mmap);
 446         ASSIGN_VMD_DMA_OPS(source, dest, get_sgtable);
 447         ASSIGN_VMD_DMA_OPS(source, dest, map_page);
 448         ASSIGN_VMD_DMA_OPS(source, dest, unmap_page);
 449         ASSIGN_VMD_DMA_OPS(source, dest, map_sg);
 450         ASSIGN_VMD_DMA_OPS(source, dest, unmap_sg);
 451         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_cpu);
 452         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_device);
 453         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_cpu);
 454         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_device);
 455         ASSIGN_VMD_DMA_OPS(source, dest, mapping_error);
 456         ASSIGN_VMD_DMA_OPS(source, dest, dma_supported);
 457 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 458         ASSIGN_VMD_DMA_OPS(source, dest, get_required_mask);
 459 #endif
 460         add_dma_domain(domain);
 461 }
 462 #undef ASSIGN_VMD_DMA_OPS
 463 #else
 464 static void vmd_teardown_dma_ops(struct vmd_dev *vmd) {}
 465 static void vmd_setup_dma_ops(struct vmd_dev *vmd) {}
 466 #endif
 467
 468 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 469                                   unsigned int devfn, int reg, int len)
 470 {
 471         char __iomem *addr = vmd->cfgbar +
 472                              ((bus->number - vmd->busn_start) << 20) +
 473                              (devfn << 12) + reg;
 474
 475         if ((addr - vmd->cfgbar) + len >=
 476             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 477                 return NULL;
 478
 479         return addr;
 480 }
 481
 482 /*
 483  * CPU may deadlock if config space is not serialized on some versions of this
 484  * hardware, so all config space access is done under a spinlock.
 485  */
 486 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 487                         int len, u32 *value)
 488 {
 489         struct vmd_dev *vmd = vmd_from_bus(bus);
 490         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 491         unsigned long flags;
 492         int ret = 0;
 493
 494         if (!addr)
 495                 return -EFAULT;
 496
 497         spin_lock_irqsave(&vmd->cfg_lock, flags);
 498         switch (len) {
 499         case 1:
 500                 *value = readb(addr);
 501                 break;
 502         case 2:
 503                 *value = readw(addr);
 504                 break;
 505         case 4:
 506                 *value = readl(addr);
 507                 break;
 508         default:
 509                 ret = -EINVAL;
 510                 break;
 511         }
 512         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 513         return ret;
 514 }
 515
 516 /*
 517  * VMD h/w converts non-posted config writes to posted memory writes. The
 518  * read-back in this function forces the completion so it returns only after
 519  * the config space was written, as expected.
 520  */
 521 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 522                          int len, u32 value)
 523 {
 524         struct vmd_dev *vmd = vmd_from_bus(bus);
 525         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 526         unsigned long flags;
 527         int ret = 0;
 528
 529         if (!addr)
 530                 return -EFAULT;
 531
 532         spin_lock_irqsave(&vmd->cfg_lock, flags);
 533         switch (len) {
 534         case 1:
 535                 writeb(value, addr);
 536                 readb(addr);
 537                 break;
 538         case 2:
 539                 writew(value, addr);
 540                 readw(addr);
 541                 break;
 542         case 4:
 543                 writel(value, addr);
 544                 readl(addr);
 545                 break;
 546         default:
 547                 ret = -EINVAL;
 548                 break;
 549         }
 550         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 551         return ret;
 552 }
 553
 554 static struct pci_ops vmd_ops = {
 555         .read           = vmd_pci_read,
 556         .write          = vmd_pci_write,
 557 };
 558
 559 static void vmd_attach_resources(struct vmd_dev *vmd)
 560 {
 561         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 562         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 563 }
 564
 565 static void vmd_detach_resources(struct vmd_dev *vmd)
 566 {
 567         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 568         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 569 }
 570
 571 /*
 572  * VMD domains start at 0x10000 to not clash with ACPI _SEG domains.
 573  * Per ACPI r6.0, sec 6.5.6,  _SEG returns an integer, of which the lower
 574  * 16 bits are the PCI Segment Group (domain) number.  Other bits are
 575  * currently reserved.
 576  */
 577 static int vmd_find_free_domain(void)
 578 {
 579         int domain = 0xffff;
 580         struct pci_bus *bus = NULL;
 581
 582         while ((bus = pci_find_next_bus(bus)) != NULL)
 583                 domain = max_t(int, domain, pci_domain_nr(bus));
 584         return domain + 1;
 585 }
 586
 587 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 588 {
 589         struct pci_sysdata *sd = &vmd->sysdata;
 590         struct fwnode_handle *fn;
 591         struct resource *res;
 592         u32 upper_bits;
 593         unsigned long flags;
 594         LIST_HEAD(resources);
 595         resource_size_t offset[2] = {0};
 596         resource_size_t membar2_offset = 0x2000;
 597
 598         /*
 599          * Shadow registers may exist in certain VMD device ids which allow
 600          * guests to correctly assign host physical addresses to the root ports
 601          * and child devices. These registers will either return the host value
 602          * or 0, depending on an enable bit in the VMD device.
 603          */
 604         if (features & VMD_FEAT_HAS_MEMBAR_SHADOW) {
 605                 u32 vmlock;
 606                 int ret;
 607
 608                 membar2_offset = MB2_SHADOW_OFFSET + MB2_SHADOW_SIZE;
 609                 ret = pci_read_config_dword(vmd->dev, PCI_REG_VMLOCK, &vmlock);
 610                 if (ret || vmlock == ~0)
 611                         return -ENODEV;
 612
 613                 if (MB2_SHADOW_EN(vmlock)) {
 614                         void __iomem *membar2;
 615
 616                         membar2 = pci_iomap(vmd->dev, VMD_MEMBAR2, 0);
 617                         if (!membar2)
 618                                 return -ENOMEM;
 619                         offset[0] = vmd->dev->resource[VMD_MEMBAR1].start -
 620                                         (readq(membar2 + MB2_SHADOW_OFFSET) &
 621                                          PCI_BASE_ADDRESS_MEM_MASK);
 622                         offset[1] = vmd->dev->resource[VMD_MEMBAR2].start -
 623                                         (readq(membar2 + MB2_SHADOW_OFFSET + 8) &
 624                                          PCI_BASE_ADDRESS_MEM_MASK);
 625                         pci_iounmap(vmd->dev, membar2);
 626                 }
 627         }
 628
 629         /*
 630          * Certain VMD devices may have a root port configuration option which
 631          * limits the bus range to between 0-127 or 128-255
 632          */
 633         if (features & VMD_FEAT_HAS_BUS_RESTRICTIONS) {
 634                 u32 vmcap, vmconfig;
 635
 636                 pci_read_config_dword(vmd->dev, PCI_REG_VMCAP, &vmcap);
 637                 pci_read_config_dword(vmd->dev, PCI_REG_VMCONFIG, &vmconfig);
 638                 if (BUS_RESTRICT_CAP(vmcap) &&
 639                     (BUS_RESTRICT_CFG(vmconfig) == 0x1))
 640                         vmd->busn_start = 128;
 641         }
 642
 643         res = &vmd->dev->resource[VMD_CFGBAR];
 644         vmd->resources[0] = (struct resource) {
 645                 .name  = "VMD CFGBAR",
 646                 .start = vmd->busn_start,
 647                 .end   = vmd->busn_start + (resource_size(res) >> 20) - 1,
 648                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 649         };
 650
 651         /*
 652          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 653          * put 32-bit resources in the window.
 654          *
 655          * There's no hardware reason why a 64-bit window *couldn't*
 656          * contain a 32-bit resource, but pbus_size_mem() computes the
 657          * bridge window size assuming a 64-bit window will contain no
 658          * 32-bit resources.  __pci_assign_resource() enforces that
 659          * artificial restriction to make sure everything will fit.
 660          *
 661          * The only way we could use a 64-bit non-prefechable MEMBAR is
 662          * if its address is <4GB so that we can convert it to a 32-bit
 663          * resource.  To be visible to the host OS, all VMD endpoints must
 664          * be initially configured by platform BIOS, which includes setting
 665          * up these resources.  We can assume the device is configured
 666          * according to the platform needs.
 667          */
 668         res = &vmd->dev->resource[VMD_MEMBAR1];
 669         upper_bits = upper_32_bits(res->end);
 670         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 671         if (!upper_bits)
 672                 flags &= ~IORESOURCE_MEM_64;
 673         vmd->resources[1] = (struct resource) {
 674                 .name  = "VMD MEMBAR1",
 675                 .start = res->start,
 676                 .end   = res->end,
 677                 .flags = flags,
 678                 .parent = res,
 679         };
 680
 681         res = &vmd->dev->resource[VMD_MEMBAR2];
 682         upper_bits = upper_32_bits(res->end);
 683         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 684         if (!upper_bits)
 685                 flags &= ~IORESOURCE_MEM_64;
 686         vmd->resources[2] = (struct resource) {
 687                 .name  = "VMD MEMBAR2",
 688                 .start = res->start + membar2_offset,
 689                 .end   = res->end,
 690                 .flags = flags,
 691                 .parent = res,
 692         };
 693
 694         sd->vmd_domain = true;
 695         sd->domain = vmd_find_free_domain();
 696         if (sd->domain < 0)
 697                 return sd->domain;
 698
 699         sd->node = pcibus_to_node(vmd->dev->bus);
 700
 701         fn = irq_domain_alloc_named_id_fwnode("VMD-MSI", vmd->sysdata.domain);
 702         if (!fn)
 703                 return -ENODEV;
 704
 705         vmd->irq_domain = pci_msi_create_irq_domain(fn, &vmd_msi_domain_info,
 706                                                     x86_vector_domain);
 707         if (!vmd->irq_domain) {
 708                 irq_domain_free_fwnode(fn);
 709                 return -ENODEV;
 710         }
 711
 712         pci_add_resource(&resources, &vmd->resources[0]);
 713         pci_add_resource_offset(&resources, &vmd->resources[1], offset[0]);
 714         pci_add_resource_offset(&resources, &vmd->resources[2], offset[1]);
 715
 716         vmd->bus = pci_create_root_bus(&vmd->dev->dev, vmd->busn_start,
 717                                         &vmd_ops, sd, &resources);
 718         if (!vmd->bus) {
 719                 pci_free_resource_list(&resources);
 720                 irq_domain_remove(vmd->irq_domain);
 721                 irq_domain_free_fwnode(fn);
 722                 return -ENODEV;
 723         }
 724
 725         vmd_attach_resources(vmd);
 726         vmd_setup_dma_ops(vmd);
 727         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 728         pci_rescan_bus(vmd->bus);
 729
 730         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 731                                "domain"), "Can't create symlink to domain\n");
 732         return 0;
 733 }
 734
 735 static irqreturn_t vmd_irq(int irq, void *data)
 736 {
 737         struct vmd_irq_list *irqs = data;
 738         struct vmd_irq *vmdirq;
 739         int idx;
 740
 741         idx = srcu_read_lock(&irqs->srcu);
 742         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 743                 generic_handle_irq(vmdirq->virq);
 744         srcu_read_unlock(&irqs->srcu, idx);
 745
 746         return IRQ_HANDLED;
 747 }
 748
 749 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 750 {
 751         struct vmd_dev *vmd;
 752         int i, err;
 753
 754         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 755                 return -ENOMEM;
 756
 757         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 758         if (!vmd)
 759                 return -ENOMEM;
 760
 761         vmd->dev = dev;
 762         err = pcim_enable_device(dev);
 763         if (err < 0)
 764                 return err;
 765
 766         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 767         if (!vmd->cfgbar)
 768                 return -ENOMEM;
 769
 770         pci_set_master(dev);
 771         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 772             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 773                 return -ENODEV;
 774
 775         vmd->msix_count = pci_msix_vec_count(dev);
 776         if (vmd->msix_count < 0)
 777                 return -ENODEV;
 778
 779         vmd->msix_count = pci_alloc_irq_vectors(dev, 1, vmd->msix_count,
 780                                         PCI_IRQ_MSIX);
 781         if (vmd->msix_count < 0)
 782                 return vmd->msix_count;
 783
 784         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 785                                  GFP_KERNEL);
 786         if (!vmd->irqs)
 787                 return -ENOMEM;
 788
 789         for (i = 0; i < vmd->msix_count; i++) {
 790                 err = init_srcu_struct(&vmd->irqs[i].srcu);
 791                 if (err)
 792                         return err;
 793
 794                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 795                 err = devm_request_irq(&dev->dev, pci_irq_vector(dev, i),
 796                                        vmd_irq, IRQF_NO_THREAD,
 797                                        "vmd", &vmd->irqs[i]);
 798                 if (err)
 799                         return err;
 800         }
 801
 802         spin_lock_init(&vmd->cfg_lock);
 803         pci_set_drvdata(dev, vmd);
 804         err = vmd_enable_domain(vmd, (unsigned long) id->driver_data);
 805         if (err)
 806                 return err;
 807
 808         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 809                  vmd->sysdata.domain);
 810         return 0;
 811 }
 812
 813 static void vmd_cleanup_srcu(struct vmd_dev *vmd)
 814 {
 815         int i;
 816
 817         for (i = 0; i < vmd->msix_count; i++)
 818                 cleanup_srcu_struct(&vmd->irqs[i].srcu);
 819 }
 820
 821 static void vmd_remove(struct pci_dev *dev)
 822 {
 823         struct vmd_dev *vmd = pci_get_drvdata(dev);
 824         struct fwnode_handle *fn = vmd->irq_domain->fwnode;
 825
 826         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 827         pci_stop_root_bus(vmd->bus);
 828         pci_remove_root_bus(vmd->bus);
 829         vmd_cleanup_srcu(vmd);
 830         vmd_teardown_dma_ops(vmd);
 831         vmd_detach_resources(vmd);
 832         irq_domain_remove(vmd->irq_domain);
 833         irq_domain_free_fwnode(fn);
 834 }
 835
 836 #ifdef CONFIG_PM_SLEEP
 837 static int vmd_suspend(struct device *dev)
 838 {
 839         struct pci_dev *pdev = to_pci_dev(dev);
 840         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 841         int i;
 842
 843         for (i = 0; i < vmd->msix_count; i++)
 844                 devm_free_irq(dev, pci_irq_vector(pdev, i), &vmd->irqs[i]);
 845
 846         pci_save_state(pdev);
 847         return 0;
 848 }
 849
 850 static int vmd_resume(struct device *dev)
 851 {
 852         struct pci_dev *pdev = to_pci_dev(dev);
 853         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 854         int err, i;
 855
 856         for (i = 0; i < vmd->msix_count; i++) {
 857                 err = devm_request_irq(dev, pci_irq_vector(pdev, i),
 858                                        vmd_irq, IRQF_NO_THREAD,
 859                                        "vmd", &vmd->irqs[i]);
 860                 if (err)
 861                         return err;
 862         }
 863
 864         pci_restore_state(pdev);
 865         return 0;
 866 }
 867 #endif
 868 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 869
 870 static const struct pci_device_id vmd_ids[] = {
 871         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_201D),},
 872         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_28C0),
 873                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW |
 874                                 VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 875         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B),
 876                 .driver_data = VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 877         {0,}
 878 };
 879 MODULE_DEVICE_TABLE(pci, vmd_ids);
 880
 881 static struct pci_driver vmd_drv = {
 882         .name           = "vmd",
 883         .id_table       = vmd_ids,
 884         .probe          = vmd_probe,
 885         .remove         = vmd_remove,
 886         .driver         = {
 887                 .pm     = &vmd_dev_pm_ops,
 888         },
 889 };
 890 module_pci_driver(vmd_drv);
 891
 892 MODULE_AUTHOR("Intel Corporation");
 893 MODULE_LICENSE("GPL v2");
 894 MODULE_VERSION("0.6");