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 enum vmd_features {
  35         /*
  36          * Device may contain registers which hint the physical location of the
  37          * membars, in order to allow proper address translation during
  38          * resource assignment to enable guest virtualization
  39          */
  40         VMD_FEAT_HAS_MEMBAR_SHADOW      = (1 << 0),
  41
  42         /*
  43          * Device may provide root port configuration information which limits
  44          * bus numbering
  45          */
  46         VMD_FEAT_HAS_BUS_RESTRICTIONS   = (1 << 1),
  47 };
  48
  49 /*
  50  * Lock for manipulating VMD IRQ lists.
  51  */
  52 static DEFINE_RAW_SPINLOCK(list_lock);
  53
  54 /**
  55  * struct vmd_irq - private data to map driver IRQ to the VMD shared vector
  56  * @node:       list item for parent traversal.
  57  * @irq:        back pointer to parent.
  58  * @enabled:    true if driver enabled IRQ
  59  * @virq:       the virtual IRQ value provided to the requesting driver.
  60  *
  61  * Every MSI/MSI-X IRQ requested for a device in a VMD domain will be mapped to
  62  * a VMD IRQ using this structure.
  63  */
  64 struct vmd_irq {
  65         struct list_head        node;
  66         struct vmd_irq_list     *irq;
  67         bool                    enabled;
  68         unsigned int            virq;
  69 };
  70
  71 /**
  72  * struct vmd_irq_list - list of driver requested IRQs mapping to a VMD vector
  73  * @irq_list:   the list of irq's the VMD one demuxes to.
  74  * @srcu:       SRCU struct for local synchronization.
  75  * @count:      number of child IRQs assigned to this vector; used to track
  76  *              sharing.
  77  */
  78 struct vmd_irq_list {
  79         struct list_head        irq_list;
  80         struct srcu_struct      srcu;
  81         unsigned int            count;
  82 };
  83
  84 struct vmd_dev {
  85         struct pci_dev          *dev;
  86
  87         spinlock_t              cfg_lock;
  88         char __iomem            *cfgbar;
  89
  90         int msix_count;
  91         struct vmd_irq_list     *irqs;
  92
  93         struct pci_sysdata      sysdata;
  94         struct resource         resources[3];
  95         struct irq_domain       *irq_domain;
  96         struct pci_bus          *bus;
  97
  98 #ifdef CONFIG_X86_DEV_DMA_OPS
  99         struct dma_map_ops      dma_ops;
 100         struct dma_domain       dma_domain;
 101 #endif
 102 };
 103
 104 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
 105 {
 106         return container_of(bus->sysdata, struct vmd_dev, sysdata);
 107 }
 108
 109 static inline unsigned int index_from_irqs(struct vmd_dev *vmd,
 110                                            struct vmd_irq_list *irqs)
 111 {
 112         return irqs - vmd->irqs;
 113 }
 114
 115 /*
 116  * Drivers managing a device in a VMD domain allocate their own IRQs as before,
 117  * but the MSI entry for the hardware it's driving will be programmed with a
 118  * destination ID for the VMD MSI-X table.  The VMD muxes interrupts in its
 119  * domain into one of its own, and the VMD driver de-muxes these for the
 120  * handlers sharing that VMD IRQ.  The vmd irq_domain provides the operations
 121  * and irq_chip to set this up.
 122  */
 123 static void vmd_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
 124 {
 125         struct vmd_irq *vmdirq = data->chip_data;
 126         struct vmd_irq_list *irq = vmdirq->irq;
 127         struct vmd_dev *vmd = irq_data_get_irq_handler_data(data);
 128
 129         msg->address_hi = MSI_ADDR_BASE_HI;
 130         msg->address_lo = MSI_ADDR_BASE_LO |
 131                           MSI_ADDR_DEST_ID(index_from_irqs(vmd, irq));
 132         msg->data = 0;
 133 }
 134
 135 /*
 136  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 137  */
 138 static void vmd_irq_enable(struct irq_data *data)
 139 {
 140         struct vmd_irq *vmdirq = data->chip_data;
 141         unsigned long flags;
 142
 143         raw_spin_lock_irqsave(&list_lock, flags);
 144         WARN_ON(vmdirq->enabled);
 145         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 146         vmdirq->enabled = true;
 147         raw_spin_unlock_irqrestore(&list_lock, flags);
 148
 149         data->chip->irq_unmask(data);
 150 }
 151
 152 static void vmd_irq_disable(struct irq_data *data)
 153 {
 154         struct vmd_irq *vmdirq = data->chip_data;
 155         unsigned long flags;
 156
 157         data->chip->irq_mask(data);
 158
 159         raw_spin_lock_irqsave(&list_lock, flags);
 160         if (vmdirq->enabled) {
 161                 list_del_rcu(&vmdirq->node);
 162                 vmdirq->enabled = false;
 163         }
 164         raw_spin_unlock_irqrestore(&list_lock, flags);
 165 }
 166
 167 /*
 168  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 169  * other devices sharing the same vector.
 170  */
 171 static int vmd_irq_set_affinity(struct irq_data *data,
 172                                 const struct cpumask *dest, bool force)
 173 {
 174         return -EINVAL;
 175 }
 176
 177 static struct irq_chip vmd_msi_controller = {
 178         .name                   = "VMD-MSI",
 179         .irq_enable             = vmd_irq_enable,
 180         .irq_disable            = vmd_irq_disable,
 181         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 182         .irq_set_affinity       = vmd_irq_set_affinity,
 183 };
 184
 185 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 186                                      msi_alloc_info_t *arg)
 187 {
 188         return 0;
 189 }
 190
 191 /*
 192  * XXX: We can be even smarter selecting the best IRQ once we solve the
 193  * affinity problem.
 194  */
 195 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 196 {
 197         int i, best = 1;
 198         unsigned long flags;
 199
 200         if (pci_is_bridge(msi_desc_to_pci_dev(desc)) || vmd->msix_count == 1)
 201                 return &vmd->irqs[0];
 202
 203         raw_spin_lock_irqsave(&list_lock, flags);
 204         for (i = 1; i < vmd->msix_count; i++)
 205                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 206                         best = i;
 207         vmd->irqs[best].count++;
 208         raw_spin_unlock_irqrestore(&list_lock, flags);
 209
 210         return &vmd->irqs[best];
 211 }
 212
 213 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 214                         unsigned int virq, irq_hw_number_t hwirq,
 215                         msi_alloc_info_t *arg)
 216 {
 217         struct msi_desc *desc = arg->desc;
 218         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 219         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 220         unsigned int index, vector;
 221
 222         if (!vmdirq)
 223                 return -ENOMEM;
 224
 225         INIT_LIST_HEAD(&vmdirq->node);
 226         vmdirq->irq = vmd_next_irq(vmd, desc);
 227         vmdirq->virq = virq;
 228         index = index_from_irqs(vmd, vmdirq->irq);
 229         vector = pci_irq_vector(vmd->dev, index);
 230
 231         irq_domain_set_info(domain, virq, vector, info->chip, vmdirq,
 232                             handle_untracked_irq, vmd, NULL);
 233         return 0;
 234 }
 235
 236 static void vmd_msi_free(struct irq_domain *domain,
 237                         struct msi_domain_info *info, unsigned int virq)
 238 {
 239         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 240         unsigned long flags;
 241
 242         synchronize_srcu(&vmdirq->irq->srcu);
 243
 244         /* XXX: Potential optimization to rebalance */
 245         raw_spin_lock_irqsave(&list_lock, flags);
 246         vmdirq->irq->count--;
 247         raw_spin_unlock_irqrestore(&list_lock, flags);
 248
 249         kfree(vmdirq);
 250 }
 251
 252 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 253                            int nvec, msi_alloc_info_t *arg)
 254 {
 255         struct pci_dev *pdev = to_pci_dev(dev);
 256         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 257
 258         if (nvec > vmd->msix_count)
 259                 return vmd->msix_count;
 260
 261         memset(arg, 0, sizeof(*arg));
 262         return 0;
 263 }
 264
 265 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 266 {
 267         arg->desc = desc;
 268 }
 269
 270 static struct msi_domain_ops vmd_msi_domain_ops = {
 271         .get_hwirq      = vmd_get_hwirq,
 272         .msi_init       = vmd_msi_init,
 273         .msi_free       = vmd_msi_free,
 274         .msi_prepare    = vmd_msi_prepare,
 275         .set_desc       = vmd_set_desc,
 276 };
 277
 278 static struct msi_domain_info vmd_msi_domain_info = {
 279         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 280                           MSI_FLAG_PCI_MSIX,
 281         .ops            = &vmd_msi_domain_ops,
 282         .chip           = &vmd_msi_controller,
 283 };
 284
 285 #ifdef CONFIG_X86_DEV_DMA_OPS
 286 /*
 287  * VMD replaces the requester ID with its own.  DMA mappings for devices in a
 288  * VMD domain need to be mapped for the VMD, not the device requiring
 289  * the mapping.
 290  */
 291 static struct device *to_vmd_dev(struct device *dev)
 292 {
 293         struct pci_dev *pdev = to_pci_dev(dev);
 294         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 295
 296         return &vmd->dev->dev;
 297 }
 298
 299 static const struct dma_map_ops *vmd_dma_ops(struct device *dev)
 300 {
 301         return get_dma_ops(to_vmd_dev(dev));
 302 }
 303
 304 static void *vmd_alloc(struct device *dev, size_t size, dma_addr_t *addr,
 305                        gfp_t flag, unsigned long attrs)
 306 {
 307         return vmd_dma_ops(dev)->alloc(to_vmd_dev(dev), size, addr, flag,
 308                                        attrs);
 309 }
 310
 311 static void vmd_free(struct device *dev, size_t size, void *vaddr,
 312                      dma_addr_t addr, unsigned long attrs)
 313 {
 314         return vmd_dma_ops(dev)->free(to_vmd_dev(dev), size, vaddr, addr,
 315                                       attrs);
 316 }
 317
 318 static int vmd_mmap(struct device *dev, struct vm_area_struct *vma,
 319                     void *cpu_addr, dma_addr_t addr, size_t size,
 320                     unsigned long attrs)
 321 {
 322         return vmd_dma_ops(dev)->mmap(to_vmd_dev(dev), vma, cpu_addr, addr,
 323                                       size, attrs);
 324 }
 325
 326 static int vmd_get_sgtable(struct device *dev, struct sg_table *sgt,
 327                            void *cpu_addr, dma_addr_t addr, size_t size,
 328                            unsigned long attrs)
 329 {
 330         return vmd_dma_ops(dev)->get_sgtable(to_vmd_dev(dev), sgt, cpu_addr,
 331                                              addr, size, attrs);
 332 }
 333
 334 static dma_addr_t vmd_map_page(struct device *dev, struct page *page,
 335                                unsigned long offset, size_t size,
 336                                enum dma_data_direction dir,
 337                                unsigned long attrs)
 338 {
 339         return vmd_dma_ops(dev)->map_page(to_vmd_dev(dev), page, offset, size,
 340                                           dir, attrs);
 341 }
 342
 343 static void vmd_unmap_page(struct device *dev, dma_addr_t addr, size_t size,
 344                            enum dma_data_direction dir, unsigned long attrs)
 345 {
 346         vmd_dma_ops(dev)->unmap_page(to_vmd_dev(dev), addr, size, dir, attrs);
 347 }
 348
 349 static int vmd_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 350                       enum dma_data_direction dir, unsigned long attrs)
 351 {
 352         return vmd_dma_ops(dev)->map_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 353 }
 354
 355 static void vmd_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 356                          enum dma_data_direction dir, unsigned long attrs)
 357 {
 358         vmd_dma_ops(dev)->unmap_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 359 }
 360
 361 static void vmd_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 362                                     size_t size, enum dma_data_direction dir)
 363 {
 364         vmd_dma_ops(dev)->sync_single_for_cpu(to_vmd_dev(dev), addr, size, dir);
 365 }
 366
 367 static void vmd_sync_single_for_device(struct device *dev, dma_addr_t addr,
 368                                        size_t size, enum dma_data_direction dir)
 369 {
 370         vmd_dma_ops(dev)->sync_single_for_device(to_vmd_dev(dev), addr, size,
 371                                                  dir);
 372 }
 373
 374 static void vmd_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 375                                 int nents, enum dma_data_direction dir)
 376 {
 377         vmd_dma_ops(dev)->sync_sg_for_cpu(to_vmd_dev(dev), sg, nents, dir);
 378 }
 379
 380 static void vmd_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 381                                    int nents, enum dma_data_direction dir)
 382 {
 383         vmd_dma_ops(dev)->sync_sg_for_device(to_vmd_dev(dev), sg, nents, dir);
 384 }
 385
 386 static int vmd_mapping_error(struct device *dev, dma_addr_t addr)
 387 {
 388         return vmd_dma_ops(dev)->mapping_error(to_vmd_dev(dev), addr);
 389 }
 390
 391 static int vmd_dma_supported(struct device *dev, u64 mask)
 392 {
 393         return vmd_dma_ops(dev)->dma_supported(to_vmd_dev(dev), mask);
 394 }
 395
 396 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 397 static u64 vmd_get_required_mask(struct device *dev)
 398 {
 399         return vmd_dma_ops(dev)->get_required_mask(to_vmd_dev(dev));
 400 }
 401 #endif
 402
 403 static void vmd_teardown_dma_ops(struct vmd_dev *vmd)
 404 {
 405         struct dma_domain *domain = &vmd->dma_domain;
 406
 407         if (get_dma_ops(&vmd->dev->dev))
 408                 del_dma_domain(domain);
 409 }
 410
 411 #define ASSIGN_VMD_DMA_OPS(source, dest, fn)    \
 412         do {                                    \
 413                 if (source->fn)                 \
 414                         dest->fn = vmd_##fn;    \
 415         } while (0)
 416
 417 static void vmd_setup_dma_ops(struct vmd_dev *vmd)
 418 {
 419         const struct dma_map_ops *source = get_dma_ops(&vmd->dev->dev);
 420         struct dma_map_ops *dest = &vmd->dma_ops;
 421         struct dma_domain *domain = &vmd->dma_domain;
 422
 423         domain->domain_nr = vmd->sysdata.domain;
 424         domain->dma_ops = dest;
 425
 426         if (!source)
 427                 return;
 428         ASSIGN_VMD_DMA_OPS(source, dest, alloc);
 429         ASSIGN_VMD_DMA_OPS(source, dest, free);
 430         ASSIGN_VMD_DMA_OPS(source, dest, mmap);
 431         ASSIGN_VMD_DMA_OPS(source, dest, get_sgtable);
 432         ASSIGN_VMD_DMA_OPS(source, dest, map_page);
 433         ASSIGN_VMD_DMA_OPS(source, dest, unmap_page);
 434         ASSIGN_VMD_DMA_OPS(source, dest, map_sg);
 435         ASSIGN_VMD_DMA_OPS(source, dest, unmap_sg);
 436         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_cpu);
 437         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_device);
 438         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_cpu);
 439         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_device);
 440         ASSIGN_VMD_DMA_OPS(source, dest, mapping_error);
 441         ASSIGN_VMD_DMA_OPS(source, dest, dma_supported);
 442 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 443         ASSIGN_VMD_DMA_OPS(source, dest, get_required_mask);
 444 #endif
 445         add_dma_domain(domain);
 446 }
 447 #undef ASSIGN_VMD_DMA_OPS
 448 #else
 449 static void vmd_teardown_dma_ops(struct vmd_dev *vmd) {}
 450 static void vmd_setup_dma_ops(struct vmd_dev *vmd) {}
 451 #endif
 452
 453 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 454                                   unsigned int devfn, int reg, int len)
 455 {
 456         char __iomem *addr = vmd->cfgbar +
 457                              (bus->number << 20) + (devfn << 12) + reg;
 458
 459         if ((addr - vmd->cfgbar) + len >=
 460             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 461                 return NULL;
 462
 463         return addr;
 464 }
 465
 466 /*
 467  * CPU may deadlock if config space is not serialized on some versions of this
 468  * hardware, so all config space access is done under a spinlock.
 469  */
 470 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 471                         int len, u32 *value)
 472 {
 473         struct vmd_dev *vmd = vmd_from_bus(bus);
 474         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 475         unsigned long flags;
 476         int ret = 0;
 477
 478         if (!addr)
 479                 return -EFAULT;
 480
 481         spin_lock_irqsave(&vmd->cfg_lock, flags);
 482         switch (len) {
 483         case 1:
 484                 *value = readb(addr);
 485                 break;
 486         case 2:
 487                 *value = readw(addr);
 488                 break;
 489         case 4:
 490                 *value = readl(addr);
 491                 break;
 492         default:
 493                 ret = -EINVAL;
 494                 break;
 495         }
 496         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 497         return ret;
 498 }
 499
 500 /*
 501  * VMD h/w converts non-posted config writes to posted memory writes. The
 502  * read-back in this function forces the completion so it returns only after
 503  * the config space was written, as expected.
 504  */
 505 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 506                          int len, u32 value)
 507 {
 508         struct vmd_dev *vmd = vmd_from_bus(bus);
 509         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 510         unsigned long flags;
 511         int ret = 0;
 512
 513         if (!addr)
 514                 return -EFAULT;
 515
 516         spin_lock_irqsave(&vmd->cfg_lock, flags);
 517         switch (len) {
 518         case 1:
 519                 writeb(value, addr);
 520                 readb(addr);
 521                 break;
 522         case 2:
 523                 writew(value, addr);
 524                 readw(addr);
 525                 break;
 526         case 4:
 527                 writel(value, addr);
 528                 readl(addr);
 529                 break;
 530         default:
 531                 ret = -EINVAL;
 532                 break;
 533         }
 534         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 535         return ret;
 536 }
 537
 538 static struct pci_ops vmd_ops = {
 539         .read           = vmd_pci_read,
 540         .write          = vmd_pci_write,
 541 };
 542
 543 static void vmd_attach_resources(struct vmd_dev *vmd)
 544 {
 545         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 546         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 547 }
 548
 549 static void vmd_detach_resources(struct vmd_dev *vmd)
 550 {
 551         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 552         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 553 }
 554
 555 /*
 556  * VMD domains start at 0x10000 to not clash with ACPI _SEG domains.
 557  * Per ACPI r6.0, sec 6.5.6,  _SEG returns an integer, of which the lower
 558  * 16 bits are the PCI Segment Group (domain) number.  Other bits are
 559  * currently reserved.
 560  */
 561 static int vmd_find_free_domain(void)
 562 {
 563         int domain = 0xffff;
 564         struct pci_bus *bus = NULL;
 565
 566         while ((bus = pci_find_next_bus(bus)) != NULL)
 567                 domain = max_t(int, domain, pci_domain_nr(bus));
 568         return domain + 1;
 569 }
 570
 571 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 572 {
 573         struct pci_sysdata *sd = &vmd->sysdata;
 574         struct fwnode_handle *fn;
 575         struct resource *res;
 576         u32 upper_bits;
 577         unsigned long flags;
 578         LIST_HEAD(resources);
 579         resource_size_t offset[2] = {0};
 580         resource_size_t membar2_offset = 0x2000, busn_start = 0;
 581
 582         /*
 583          * Shadow registers may exist in certain VMD device ids which allow
 584          * guests to correctly assign host physical addresses to the root ports
 585          * and child devices. These registers will either return the host value
 586          * or 0, depending on an enable bit in the VMD device.
 587          */
 588         if (features & VMD_FEAT_HAS_MEMBAR_SHADOW) {
 589                 u32 vmlock;
 590                 int ret;
 591
 592                 membar2_offset = 0x2018;
 593                 ret = pci_read_config_dword(vmd->dev, PCI_REG_VMLOCK, &vmlock);
 594                 if (ret || vmlock == ~0)
 595                         return -ENODEV;
 596
 597                 if (MB2_SHADOW_EN(vmlock)) {
 598                         void __iomem *membar2;
 599
 600                         membar2 = pci_iomap(vmd->dev, VMD_MEMBAR2, 0);
 601                         if (!membar2)
 602                                 return -ENOMEM;
 603                         offset[0] = vmd->dev->resource[VMD_MEMBAR1].start -
 604                                                 readq(membar2 + 0x2008);
 605                         offset[1] = vmd->dev->resource[VMD_MEMBAR2].start -
 606                                                 readq(membar2 + 0x2010);
 607                         pci_iounmap(vmd->dev, membar2);
 608                 }
 609         }
 610
 611         /*
 612          * Certain VMD devices may have a root port configuration option which
 613          * limits the bus range to between 0-127 or 128-255
 614          */
 615         if (features & VMD_FEAT_HAS_BUS_RESTRICTIONS) {
 616                 u32 vmcap, vmconfig;
 617
 618                 pci_read_config_dword(vmd->dev, PCI_REG_VMCAP, &vmcap);
 619                 pci_read_config_dword(vmd->dev, PCI_REG_VMCONFIG, &vmconfig);
 620                 if (BUS_RESTRICT_CAP(vmcap) &&
 621                     (BUS_RESTRICT_CFG(vmconfig) == 0x1))
 622                         busn_start = 128;
 623         }
 624
 625         res = &vmd->dev->resource[VMD_CFGBAR];
 626         vmd->resources[0] = (struct resource) {
 627                 .name  = "VMD CFGBAR",
 628                 .start = busn_start,
 629                 .end   = busn_start + (resource_size(res) >> 20) - 1,
 630                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 631         };
 632
 633         /*
 634          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 635          * put 32-bit resources in the window.
 636          *
 637          * There's no hardware reason why a 64-bit window *couldn't*
 638          * contain a 32-bit resource, but pbus_size_mem() computes the
 639          * bridge window size assuming a 64-bit window will contain no
 640          * 32-bit resources.  __pci_assign_resource() enforces that
 641          * artificial restriction to make sure everything will fit.
 642          *
 643          * The only way we could use a 64-bit non-prefechable MEMBAR is
 644          * if its address is <4GB so that we can convert it to a 32-bit
 645          * resource.  To be visible to the host OS, all VMD endpoints must
 646          * be initially configured by platform BIOS, which includes setting
 647          * up these resources.  We can assume the device is configured
 648          * according to the platform needs.
 649          */
 650         res = &vmd->dev->resource[VMD_MEMBAR1];
 651         upper_bits = upper_32_bits(res->end);
 652         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 653         if (!upper_bits)
 654                 flags &= ~IORESOURCE_MEM_64;
 655         vmd->resources[1] = (struct resource) {
 656                 .name  = "VMD MEMBAR1",
 657                 .start = res->start,
 658                 .end   = res->end,
 659                 .flags = flags,
 660                 .parent = res,
 661         };
 662
 663         res = &vmd->dev->resource[VMD_MEMBAR2];
 664         upper_bits = upper_32_bits(res->end);
 665         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 666         if (!upper_bits)
 667                 flags &= ~IORESOURCE_MEM_64;
 668         vmd->resources[2] = (struct resource) {
 669                 .name  = "VMD MEMBAR2",
 670                 .start = res->start + membar2_offset,
 671                 .end   = res->end,
 672                 .flags = flags,
 673                 .parent = res,
 674         };
 675
 676         sd->vmd_domain = true;
 677         sd->domain = vmd_find_free_domain();
 678         if (sd->domain < 0)
 679                 return sd->domain;
 680
 681         sd->node = pcibus_to_node(vmd->dev->bus);
 682
 683         fn = irq_domain_alloc_named_id_fwnode("VMD-MSI", vmd->sysdata.domain);
 684         if (!fn)
 685                 return -ENODEV;
 686
 687         vmd->irq_domain = pci_msi_create_irq_domain(fn, &vmd_msi_domain_info,
 688                                                     x86_vector_domain);
 689         irq_domain_free_fwnode(fn);
 690         if (!vmd->irq_domain)
 691                 return -ENODEV;
 692
 693         pci_add_resource(&resources, &vmd->resources[0]);
 694         pci_add_resource_offset(&resources, &vmd->resources[1], offset[0]);
 695         pci_add_resource_offset(&resources, &vmd->resources[2], offset[1]);
 696
 697         vmd->bus = pci_create_root_bus(&vmd->dev->dev, busn_start, &vmd_ops,
 698                                        sd, &resources);
 699         if (!vmd->bus) {
 700                 pci_free_resource_list(&resources);
 701                 irq_domain_remove(vmd->irq_domain);
 702                 return -ENODEV;
 703         }
 704
 705         vmd_attach_resources(vmd);
 706         vmd_setup_dma_ops(vmd);
 707         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 708         pci_rescan_bus(vmd->bus);
 709
 710         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 711                                "domain"), "Can't create symlink to domain\n");
 712         return 0;
 713 }
 714
 715 static irqreturn_t vmd_irq(int irq, void *data)
 716 {
 717         struct vmd_irq_list *irqs = data;
 718         struct vmd_irq *vmdirq;
 719         int idx;
 720
 721         idx = srcu_read_lock(&irqs->srcu);
 722         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 723                 generic_handle_irq(vmdirq->virq);
 724         srcu_read_unlock(&irqs->srcu, idx);
 725
 726         return IRQ_HANDLED;
 727 }
 728
 729 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 730 {
 731         struct vmd_dev *vmd;
 732         int i, err;
 733
 734         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 735                 return -ENOMEM;
 736
 737         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 738         if (!vmd)
 739                 return -ENOMEM;
 740
 741         vmd->dev = dev;
 742         err = pcim_enable_device(dev);
 743         if (err < 0)
 744                 return err;
 745
 746         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 747         if (!vmd->cfgbar)
 748                 return -ENOMEM;
 749
 750         pci_set_master(dev);
 751         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 752             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 753                 return -ENODEV;
 754
 755         vmd->msix_count = pci_msix_vec_count(dev);
 756         if (vmd->msix_count < 0)
 757                 return -ENODEV;
 758
 759         vmd->msix_count = pci_alloc_irq_vectors(dev, 1, vmd->msix_count,
 760                                         PCI_IRQ_MSIX);
 761         if (vmd->msix_count < 0)
 762                 return vmd->msix_count;
 763
 764         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 765                                  GFP_KERNEL);
 766         if (!vmd->irqs)
 767                 return -ENOMEM;
 768
 769         for (i = 0; i < vmd->msix_count; i++) {
 770                 err = init_srcu_struct(&vmd->irqs[i].srcu);
 771                 if (err)
 772                         return err;
 773
 774                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 775                 err = devm_request_irq(&dev->dev, pci_irq_vector(dev, i),
 776                                        vmd_irq, IRQF_NO_THREAD,
 777                                        "vmd", &vmd->irqs[i]);
 778                 if (err)
 779                         return err;
 780         }
 781
 782         spin_lock_init(&vmd->cfg_lock);
 783         pci_set_drvdata(dev, vmd);
 784         err = vmd_enable_domain(vmd, (unsigned long) id->driver_data);
 785         if (err)
 786                 return err;
 787
 788         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 789                  vmd->sysdata.domain);
 790         return 0;
 791 }
 792
 793 static void vmd_cleanup_srcu(struct vmd_dev *vmd)
 794 {
 795         int i;
 796
 797         for (i = 0; i < vmd->msix_count; i++)
 798                 cleanup_srcu_struct(&vmd->irqs[i].srcu);
 799 }
 800
 801 static void vmd_remove(struct pci_dev *dev)
 802 {
 803         struct vmd_dev *vmd = pci_get_drvdata(dev);
 804
 805         vmd_detach_resources(vmd);
 806         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 807         pci_stop_root_bus(vmd->bus);
 808         pci_remove_root_bus(vmd->bus);
 809         vmd_cleanup_srcu(vmd);
 810         vmd_teardown_dma_ops(vmd);
 811         irq_domain_remove(vmd->irq_domain);
 812 }
 813
 814 #ifdef CONFIG_PM_SLEEP
 815 static int vmd_suspend(struct device *dev)
 816 {
 817         struct pci_dev *pdev = to_pci_dev(dev);
 818         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 819         int i;
 820
 821         for (i = 0; i < vmd->msix_count; i++)
 822                 devm_free_irq(dev, pci_irq_vector(pdev, i), &vmd->irqs[i]);
 823
 824         pci_save_state(pdev);
 825         return 0;
 826 }
 827
 828 static int vmd_resume(struct device *dev)
 829 {
 830         struct pci_dev *pdev = to_pci_dev(dev);
 831         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 832         int err, i;
 833
 834         for (i = 0; i < vmd->msix_count; i++) {
 835                 err = devm_request_irq(dev, pci_irq_vector(pdev, i),
 836                                        vmd_irq, IRQF_NO_THREAD,
 837                                        "vmd", &vmd->irqs[i]);
 838                 if (err)
 839                         return err;
 840         }
 841
 842         pci_restore_state(pdev);
 843         return 0;
 844 }
 845 #endif
 846 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 847
 848 static const struct pci_device_id vmd_ids[] = {
 849         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_201D),},
 850         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_28C0),
 851                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW |
 852                                 VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 853         {0,}
 854 };
 855 MODULE_DEVICE_TABLE(pci, vmd_ids);
 856
 857 static struct pci_driver vmd_drv = {
 858         .name           = "vmd",
 859         .id_table       = vmd_ids,
 860         .probe          = vmd_probe,
 861         .remove         = vmd_remove,
 862         .driver         = {
 863                 .pm     = &vmd_dev_pm_ops,
 864         },
 865 };
 866 module_pci_driver(vmd_drv);
 867
 868 MODULE_AUTHOR("Intel Corporation");
 869 MODULE_LICENSE("GPL v2");
 870 MODULE_VERSION("0.6");