arch/x86/pci/vmd.c

   1 /*
   2  * Volume Management Device driver
   3  * Copyright (c) 2015, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14
  15 #include <linux/device.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/irq.h>
  18 #include <linux/kernel.h>
  19 #include <linux/module.h>
  20 #include <linux/msi.h>
  21 #include <linux/pci.h>
  22 #include <linux/rculist.h>
  23 #include <linux/rcupdate.h>
  24
  25 #include <asm/irqdomain.h>
  26 #include <asm/device.h>
  27 #include <asm/msi.h>
  28 #include <asm/msidef.h>
  29
  30 #define VMD_CFGBAR      0
  31 #define VMD_MEMBAR1     2
  32 #define VMD_MEMBAR2     4
  33
  34 /*
  35  * Lock for manipulating VMD IRQ lists.
  36  */
  37 static DEFINE_RAW_SPINLOCK(list_lock);
  38
  39 /**
  40  * struct vmd_irq - private data to map driver IRQ to the VMD shared vector
  41  * @node:       list item for parent traversal.
  42  * @rcu:        RCU callback item for freeing.
  43  * @irq:        back pointer to parent.
  44  * @virq:       the virtual IRQ value provided to the requesting driver.
  45  *
  46  * Every MSI/MSI-X IRQ requested for a device in a VMD domain will be mapped to
  47  * a VMD IRQ using this structure.
  48  */
  49 struct vmd_irq {
  50         struct list_head        node;
  51         struct rcu_head         rcu;
  52         struct vmd_irq_list     *irq;
  53         unsigned int            virq;
  54 };
  55
  56 /**
  57  * struct vmd_irq_list - list of driver requested IRQs mapping to a VMD vector
  58  * @irq_list:   the list of irq's the VMD one demuxes to.
  59  * @count:      number of child IRQs assigned to this vector; used to track
  60  *              sharing.
  61  */
  62 struct vmd_irq_list {
  63         struct list_head        irq_list;
  64         unsigned int            count;
  65 };
  66
  67 struct vmd_dev {
  68         struct pci_dev          *dev;
  69
  70         spinlock_t              cfg_lock;
  71         char __iomem            *cfgbar;
  72
  73         int msix_count;
  74         struct vmd_irq_list     *irqs;
  75
  76         struct pci_sysdata      sysdata;
  77         struct resource         resources[3];
  78         struct irq_domain       *irq_domain;
  79         struct pci_bus          *bus;
  80
  81 #ifdef CONFIG_X86_DEV_DMA_OPS
  82         struct dma_map_ops      dma_ops;
  83         struct dma_domain       dma_domain;
  84 #endif
  85 };
  86
  87 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
  88 {
  89         return container_of(bus->sysdata, struct vmd_dev, sysdata);
  90 }
  91
  92 static inline unsigned int index_from_irqs(struct vmd_dev *vmd,
  93                                            struct vmd_irq_list *irqs)
  94 {
  95         return irqs - vmd->irqs;
  96 }
  97
  98 /*
  99  * Drivers managing a device in a VMD domain allocate their own IRQs as before,
 100  * but the MSI entry for the hardware it's driving will be programmed with a
 101  * destination ID for the VMD MSI-X table.  The VMD muxes interrupts in its
 102  * domain into one of its own, and the VMD driver de-muxes these for the
 103  * handlers sharing that VMD IRQ.  The vmd irq_domain provides the operations
 104  * and irq_chip to set this up.
 105  */
 106 static void vmd_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
 107 {
 108         struct vmd_irq *vmdirq = data->chip_data;
 109         struct vmd_irq_list *irq = vmdirq->irq;
 110         struct vmd_dev *vmd = irq_data_get_irq_handler_data(data);
 111
 112         msg->address_hi = MSI_ADDR_BASE_HI;
 113         msg->address_lo = MSI_ADDR_BASE_LO |
 114                           MSI_ADDR_DEST_ID(index_from_irqs(vmd, irq));
 115         msg->data = 0;
 116 }
 117
 118 /*
 119  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 120  */
 121 static void vmd_irq_enable(struct irq_data *data)
 122 {
 123         struct vmd_irq *vmdirq = data->chip_data;
 124         unsigned long flags;
 125
 126         raw_spin_lock_irqsave(&list_lock, flags);
 127         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 128         raw_spin_unlock_irqrestore(&list_lock, flags);
 129
 130         data->chip->irq_unmask(data);
 131 }
 132
 133 static void vmd_irq_disable(struct irq_data *data)
 134 {
 135         struct vmd_irq *vmdirq = data->chip_data;
 136         unsigned long flags;
 137
 138         data->chip->irq_mask(data);
 139
 140         raw_spin_lock_irqsave(&list_lock, flags);
 141         list_del_rcu(&vmdirq->node);
 142         INIT_LIST_HEAD_RCU(&vmdirq->node);
 143         raw_spin_unlock_irqrestore(&list_lock, flags);
 144 }
 145
 146 /*
 147  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 148  * other devices sharing the same vector.
 149  */
 150 static int vmd_irq_set_affinity(struct irq_data *data,
 151                                 const struct cpumask *dest, bool force)
 152 {
 153         return -EINVAL;
 154 }
 155
 156 static struct irq_chip vmd_msi_controller = {
 157         .name                   = "VMD-MSI",
 158         .irq_enable             = vmd_irq_enable,
 159         .irq_disable            = vmd_irq_disable,
 160         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 161         .irq_set_affinity       = vmd_irq_set_affinity,
 162 };
 163
 164 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 165                                      msi_alloc_info_t *arg)
 166 {
 167         return 0;
 168 }
 169
 170 /*
 171  * XXX: We can be even smarter selecting the best IRQ once we solve the
 172  * affinity problem.
 173  */
 174 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 175 {
 176         int i, best = 1;
 177         unsigned long flags;
 178
 179         if (!desc->msi_attrib.is_msix || vmd->msix_count == 1)
 180                 return &vmd->irqs[0];
 181
 182         raw_spin_lock_irqsave(&list_lock, flags);
 183         for (i = 1; i < vmd->msix_count; i++)
 184                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 185                         best = i;
 186         vmd->irqs[best].count++;
 187         raw_spin_unlock_irqrestore(&list_lock, flags);
 188
 189         return &vmd->irqs[best];
 190 }
 191
 192 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 193                         unsigned int virq, irq_hw_number_t hwirq,
 194                         msi_alloc_info_t *arg)
 195 {
 196         struct msi_desc *desc = arg->desc;
 197         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 198         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 199         unsigned int index, vector;
 200
 201         if (!vmdirq)
 202                 return -ENOMEM;
 203
 204         INIT_LIST_HEAD(&vmdirq->node);
 205         vmdirq->irq = vmd_next_irq(vmd, desc);
 206         vmdirq->virq = virq;
 207         index = index_from_irqs(vmd, vmdirq->irq);
 208         vector = pci_irq_vector(vmd->dev, index);
 209
 210         irq_domain_set_info(domain, virq, vector, info->chip, vmdirq,
 211                             handle_untracked_irq, vmd, NULL);
 212         return 0;
 213 }
 214
 215 static void vmd_msi_free(struct irq_domain *domain,
 216                         struct msi_domain_info *info, unsigned int virq)
 217 {
 218         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 219         unsigned long flags;
 220
 221         /* XXX: Potential optimization to rebalance */
 222         raw_spin_lock_irqsave(&list_lock, flags);
 223         vmdirq->irq->count--;
 224         raw_spin_unlock_irqrestore(&list_lock, flags);
 225
 226         kfree_rcu(vmdirq, rcu);
 227 }
 228
 229 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 230                            int nvec, msi_alloc_info_t *arg)
 231 {
 232         struct pci_dev *pdev = to_pci_dev(dev);
 233         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 234
 235         if (nvec > vmd->msix_count)
 236                 return vmd->msix_count;
 237
 238         memset(arg, 0, sizeof(*arg));
 239         return 0;
 240 }
 241
 242 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 243 {
 244         arg->desc = desc;
 245 }
 246
 247 static struct msi_domain_ops vmd_msi_domain_ops = {
 248         .get_hwirq      = vmd_get_hwirq,
 249         .msi_init       = vmd_msi_init,
 250         .msi_free       = vmd_msi_free,
 251         .msi_prepare    = vmd_msi_prepare,
 252         .set_desc       = vmd_set_desc,
 253 };
 254
 255 static struct msi_domain_info vmd_msi_domain_info = {
 256         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 257                           MSI_FLAG_PCI_MSIX,
 258         .ops            = &vmd_msi_domain_ops,
 259         .chip           = &vmd_msi_controller,
 260 };
 261
 262 #ifdef CONFIG_X86_DEV_DMA_OPS
 263 /*
 264  * VMD replaces the requester ID with its own.  DMA mappings for devices in a
 265  * VMD domain need to be mapped for the VMD, not the device requiring
 266  * the mapping.
 267  */
 268 static struct device *to_vmd_dev(struct device *dev)
 269 {
 270         struct pci_dev *pdev = to_pci_dev(dev);
 271         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 272
 273         return &vmd->dev->dev;
 274 }
 275
 276 static struct dma_map_ops *vmd_dma_ops(struct device *dev)
 277 {
 278         return get_dma_ops(to_vmd_dev(dev));
 279 }
 280
 281 static void *vmd_alloc(struct device *dev, size_t size, dma_addr_t *addr,
 282                        gfp_t flag, unsigned long attrs)
 283 {
 284         return vmd_dma_ops(dev)->alloc(to_vmd_dev(dev), size, addr, flag,
 285                                        attrs);
 286 }
 287
 288 static void vmd_free(struct device *dev, size_t size, void *vaddr,
 289                      dma_addr_t addr, unsigned long attrs)
 290 {
 291         return vmd_dma_ops(dev)->free(to_vmd_dev(dev), size, vaddr, addr,
 292                                       attrs);
 293 }
 294
 295 static int vmd_mmap(struct device *dev, struct vm_area_struct *vma,
 296                     void *cpu_addr, dma_addr_t addr, size_t size,
 297                     unsigned long attrs)
 298 {
 299         return vmd_dma_ops(dev)->mmap(to_vmd_dev(dev), vma, cpu_addr, addr,
 300                                       size, attrs);
 301 }
 302
 303 static int vmd_get_sgtable(struct device *dev, struct sg_table *sgt,
 304                            void *cpu_addr, dma_addr_t addr, size_t size,
 305                            unsigned long attrs)
 306 {
 307         return vmd_dma_ops(dev)->get_sgtable(to_vmd_dev(dev), sgt, cpu_addr,
 308                                              addr, size, attrs);
 309 }
 310
 311 static dma_addr_t vmd_map_page(struct device *dev, struct page *page,
 312                                unsigned long offset, size_t size,
 313                                enum dma_data_direction dir,
 314                                unsigned long attrs)
 315 {
 316         return vmd_dma_ops(dev)->map_page(to_vmd_dev(dev), page, offset, size,
 317                                           dir, attrs);
 318 }
 319
 320 static void vmd_unmap_page(struct device *dev, dma_addr_t addr, size_t size,
 321                            enum dma_data_direction dir, unsigned long attrs)
 322 {
 323         vmd_dma_ops(dev)->unmap_page(to_vmd_dev(dev), addr, size, dir, attrs);
 324 }
 325
 326 static int vmd_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 327                       enum dma_data_direction dir, unsigned long attrs)
 328 {
 329         return vmd_dma_ops(dev)->map_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 330 }
 331
 332 static void vmd_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 333                          enum dma_data_direction dir, unsigned long attrs)
 334 {
 335         vmd_dma_ops(dev)->unmap_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 336 }
 337
 338 static void vmd_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 339                                     size_t size, enum dma_data_direction dir)
 340 {
 341         vmd_dma_ops(dev)->sync_single_for_cpu(to_vmd_dev(dev), addr, size, dir);
 342 }
 343
 344 static void vmd_sync_single_for_device(struct device *dev, dma_addr_t addr,
 345                                        size_t size, enum dma_data_direction dir)
 346 {
 347         vmd_dma_ops(dev)->sync_single_for_device(to_vmd_dev(dev), addr, size,
 348                                                  dir);
 349 }
 350
 351 static void vmd_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 352                                 int nents, enum dma_data_direction dir)
 353 {
 354         vmd_dma_ops(dev)->sync_sg_for_cpu(to_vmd_dev(dev), sg, nents, dir);
 355 }
 356
 357 static void vmd_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 358                                    int nents, enum dma_data_direction dir)
 359 {
 360         vmd_dma_ops(dev)->sync_sg_for_device(to_vmd_dev(dev), sg, nents, dir);
 361 }
 362
 363 static int vmd_mapping_error(struct device *dev, dma_addr_t addr)
 364 {
 365         return vmd_dma_ops(dev)->mapping_error(to_vmd_dev(dev), addr);
 366 }
 367
 368 static int vmd_dma_supported(struct device *dev, u64 mask)
 369 {
 370         return vmd_dma_ops(dev)->dma_supported(to_vmd_dev(dev), mask);
 371 }
 372
 373 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 374 static u64 vmd_get_required_mask(struct device *dev)
 375 {
 376         return vmd_dma_ops(dev)->get_required_mask(to_vmd_dev(dev));
 377 }
 378 #endif
 379
 380 static void vmd_teardown_dma_ops(struct vmd_dev *vmd)
 381 {
 382         struct dma_domain *domain = &vmd->dma_domain;
 383
 384         if (get_dma_ops(&vmd->dev->dev))
 385                 del_dma_domain(domain);
 386 }
 387
 388 #define ASSIGN_VMD_DMA_OPS(source, dest, fn)    \
 389         do {                                    \
 390                 if (source->fn)                 \
 391                         dest->fn = vmd_##fn;    \
 392         } while (0)
 393
 394 static void vmd_setup_dma_ops(struct vmd_dev *vmd)
 395 {
 396         const struct dma_map_ops *source = get_dma_ops(&vmd->dev->dev);
 397         struct dma_map_ops *dest = &vmd->dma_ops;
 398         struct dma_domain *domain = &vmd->dma_domain;
 399
 400         domain->domain_nr = vmd->sysdata.domain;
 401         domain->dma_ops = dest;
 402
 403         if (!source)
 404                 return;
 405         ASSIGN_VMD_DMA_OPS(source, dest, alloc);
 406         ASSIGN_VMD_DMA_OPS(source, dest, free);
 407         ASSIGN_VMD_DMA_OPS(source, dest, mmap);
 408         ASSIGN_VMD_DMA_OPS(source, dest, get_sgtable);
 409         ASSIGN_VMD_DMA_OPS(source, dest, map_page);
 410         ASSIGN_VMD_DMA_OPS(source, dest, unmap_page);
 411         ASSIGN_VMD_DMA_OPS(source, dest, map_sg);
 412         ASSIGN_VMD_DMA_OPS(source, dest, unmap_sg);
 413         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_cpu);
 414         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_device);
 415         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_cpu);
 416         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_device);
 417         ASSIGN_VMD_DMA_OPS(source, dest, mapping_error);
 418         ASSIGN_VMD_DMA_OPS(source, dest, dma_supported);
 419 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 420         ASSIGN_VMD_DMA_OPS(source, dest, get_required_mask);
 421 #endif
 422         add_dma_domain(domain);
 423 }
 424 #undef ASSIGN_VMD_DMA_OPS
 425 #else
 426 static void vmd_teardown_dma_ops(struct vmd_dev *vmd) {}
 427 static void vmd_setup_dma_ops(struct vmd_dev *vmd) {}
 428 #endif
 429
 430 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 431                                   unsigned int devfn, int reg, int len)
 432 {
 433         char __iomem *addr = vmd->cfgbar +
 434                              (bus->number << 20) + (devfn << 12) + reg;
 435
 436         if ((addr - vmd->cfgbar) + len >=
 437             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 438                 return NULL;
 439
 440         return addr;
 441 }
 442
 443 /*
 444  * CPU may deadlock if config space is not serialized on some versions of this
 445  * hardware, so all config space access is done under a spinlock.
 446  */
 447 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 448                         int len, u32 *value)
 449 {
 450         struct vmd_dev *vmd = vmd_from_bus(bus);
 451         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 452         unsigned long flags;
 453         int ret = 0;
 454
 455         if (!addr)
 456                 return -EFAULT;
 457
 458         spin_lock_irqsave(&vmd->cfg_lock, flags);
 459         switch (len) {
 460         case 1:
 461                 *value = readb(addr);
 462                 break;
 463         case 2:
 464                 *value = readw(addr);
 465                 break;
 466         case 4:
 467                 *value = readl(addr);
 468                 break;
 469         default:
 470                 ret = -EINVAL;
 471                 break;
 472         }
 473         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 474         return ret;
 475 }
 476
 477 /*
 478  * VMD h/w converts non-posted config writes to posted memory writes. The
 479  * read-back in this function forces the completion so it returns only after
 480  * the config space was written, as expected.
 481  */
 482 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 483                          int len, u32 value)
 484 {
 485         struct vmd_dev *vmd = vmd_from_bus(bus);
 486         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 487         unsigned long flags;
 488         int ret = 0;
 489
 490         if (!addr)
 491                 return -EFAULT;
 492
 493         spin_lock_irqsave(&vmd->cfg_lock, flags);
 494         switch (len) {
 495         case 1:
 496                 writeb(value, addr);
 497                 readb(addr);
 498                 break;
 499         case 2:
 500                 writew(value, addr);
 501                 readw(addr);
 502                 break;
 503         case 4:
 504                 writel(value, addr);
 505                 readl(addr);
 506                 break;
 507         default:
 508                 ret = -EINVAL;
 509                 break;
 510         }
 511         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 512         return ret;
 513 }
 514
 515 static struct pci_ops vmd_ops = {
 516         .read           = vmd_pci_read,
 517         .write          = vmd_pci_write,
 518 };
 519
 520 static void vmd_attach_resources(struct vmd_dev *vmd)
 521 {
 522         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 523         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 524 }
 525
 526 static void vmd_detach_resources(struct vmd_dev *vmd)
 527 {
 528         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 529         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 530 }
 531
 532 /*
 533  * VMD domains start at 0x1000 to not clash with ACPI _SEG domains.
 534  */
 535 static int vmd_find_free_domain(void)
 536 {
 537         int domain = 0xffff;
 538         struct pci_bus *bus = NULL;
 539
 540         while ((bus = pci_find_next_bus(bus)) != NULL)
 541                 domain = max_t(int, domain, pci_domain_nr(bus));
 542         return domain + 1;
 543 }
 544
 545 static int vmd_enable_domain(struct vmd_dev *vmd)
 546 {
 547         struct pci_sysdata *sd = &vmd->sysdata;
 548         struct resource *res;
 549         u32 upper_bits;
 550         unsigned long flags;
 551         LIST_HEAD(resources);
 552
 553         res = &vmd->dev->resource[VMD_CFGBAR];
 554         vmd->resources[0] = (struct resource) {
 555                 .name  = "VMD CFGBAR",
 556                 .start = 0,
 557                 .end   = (resource_size(res) >> 20) - 1,
 558                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 559         };
 560
 561         /*
 562          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 563          * put 32-bit resources in the window.
 564          *
 565          * There's no hardware reason why a 64-bit window *couldn't*
 566          * contain a 32-bit resource, but pbus_size_mem() computes the
 567          * bridge window size assuming a 64-bit window will contain no
 568          * 32-bit resources.  __pci_assign_resource() enforces that
 569          * artificial restriction to make sure everything will fit.
 570          *
 571          * The only way we could use a 64-bit non-prefechable MEMBAR is
 572          * if its address is <4GB so that we can convert it to a 32-bit
 573          * resource.  To be visible to the host OS, all VMD endpoints must
 574          * be initially configured by platform BIOS, which includes setting
 575          * up these resources.  We can assume the device is configured
 576          * according to the platform needs.
 577          */
 578         res = &vmd->dev->resource[VMD_MEMBAR1];
 579         upper_bits = upper_32_bits(res->end);
 580         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 581         if (!upper_bits)
 582                 flags &= ~IORESOURCE_MEM_64;
 583         vmd->resources[1] = (struct resource) {
 584                 .name  = "VMD MEMBAR1",
 585                 .start = res->start,
 586                 .end   = res->end,
 587                 .flags = flags,
 588                 .parent = res,
 589         };
 590
 591         res = &vmd->dev->resource[VMD_MEMBAR2];
 592         upper_bits = upper_32_bits(res->end);
 593         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 594         if (!upper_bits)
 595                 flags &= ~IORESOURCE_MEM_64;
 596         vmd->resources[2] = (struct resource) {
 597                 .name  = "VMD MEMBAR2",
 598                 .start = res->start + 0x2000,
 599                 .end   = res->end,
 600                 .flags = flags,
 601                 .parent = res,
 602         };
 603
 604         sd->domain = vmd_find_free_domain();
 605         if (sd->domain < 0)
 606                 return sd->domain;
 607
 608         sd->node = pcibus_to_node(vmd->dev->bus);
 609
 610         vmd->irq_domain = pci_msi_create_irq_domain(NULL, &vmd_msi_domain_info,
 611                                                     x86_vector_domain);
 612         if (!vmd->irq_domain)
 613                 return -ENODEV;
 614
 615         pci_add_resource(&resources, &vmd->resources[0]);
 616         pci_add_resource(&resources, &vmd->resources[1]);
 617         pci_add_resource(&resources, &vmd->resources[2]);
 618         vmd->bus = pci_create_root_bus(&vmd->dev->dev, 0, &vmd_ops, sd,
 619                                        &resources);
 620         if (!vmd->bus) {
 621                 pci_free_resource_list(&resources);
 622                 irq_domain_remove(vmd->irq_domain);
 623                 return -ENODEV;
 624         }
 625
 626         vmd_attach_resources(vmd);
 627         vmd_setup_dma_ops(vmd);
 628         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 629         pci_rescan_bus(vmd->bus);
 630
 631         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 632                                "domain"), "Can't create symlink to domain\n");
 633         return 0;
 634 }
 635
 636 static irqreturn_t vmd_irq(int irq, void *data)
 637 {
 638         struct vmd_irq_list *irqs = data;
 639         struct vmd_irq *vmdirq;
 640
 641         rcu_read_lock();
 642         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 643                 generic_handle_irq(vmdirq->virq);
 644         rcu_read_unlock();
 645
 646         return IRQ_HANDLED;
 647 }
 648
 649 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 650 {
 651         struct vmd_dev *vmd;
 652         int i, err;
 653
 654         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 655                 return -ENOMEM;
 656
 657         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 658         if (!vmd)
 659                 return -ENOMEM;
 660
 661         vmd->dev = dev;
 662         err = pcim_enable_device(dev);
 663         if (err < 0)
 664                 return err;
 665
 666         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 667         if (!vmd->cfgbar)
 668                 return -ENOMEM;
 669
 670         pci_set_master(dev);
 671         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 672             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 673                 return -ENODEV;
 674
 675         vmd->msix_count = pci_msix_vec_count(dev);
 676         if (vmd->msix_count < 0)
 677                 return -ENODEV;
 678
 679         vmd->msix_count = pci_alloc_irq_vectors(dev, 1, vmd->msix_count,
 680                                         PCI_IRQ_MSIX | PCI_IRQ_AFFINITY);
 681         if (vmd->msix_count < 0)
 682                 return vmd->msix_count;
 683
 684         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 685                                  GFP_KERNEL);
 686         if (!vmd->irqs)
 687                 return -ENOMEM;
 688
 689         for (i = 0; i < vmd->msix_count; i++) {
 690                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 691                 err = devm_request_irq(&dev->dev, pci_irq_vector(dev, i),
 692                                        vmd_irq, 0, "vmd", &vmd->irqs[i]);
 693                 if (err)
 694                         return err;
 695         }
 696
 697         spin_lock_init(&vmd->cfg_lock);
 698         pci_set_drvdata(dev, vmd);
 699         err = vmd_enable_domain(vmd);
 700         if (err)
 701                 return err;
 702
 703         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 704                  vmd->sysdata.domain);
 705         return 0;
 706 }
 707
 708 static void vmd_remove(struct pci_dev *dev)
 709 {
 710         struct vmd_dev *vmd = pci_get_drvdata(dev);
 711
 712         vmd_detach_resources(vmd);
 713         pci_set_drvdata(dev, NULL);
 714         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 715         pci_stop_root_bus(vmd->bus);
 716         pci_remove_root_bus(vmd->bus);
 717         vmd_teardown_dma_ops(vmd);
 718         irq_domain_remove(vmd->irq_domain);
 719 }
 720
 721 #ifdef CONFIG_PM
 722 static int vmd_suspend(struct device *dev)
 723 {
 724         struct pci_dev *pdev = to_pci_dev(dev);
 725
 726         pci_save_state(pdev);
 727         return 0;
 728 }
 729
 730 static int vmd_resume(struct device *dev)
 731 {
 732         struct pci_dev *pdev = to_pci_dev(dev);
 733
 734         pci_restore_state(pdev);
 735         return 0;
 736 }
 737 #endif
 738 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 739
 740 static const struct pci_device_id vmd_ids[] = {
 741         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x201d),},
 742         {0,}
 743 };
 744 MODULE_DEVICE_TABLE(pci, vmd_ids);
 745
 746 static struct pci_driver vmd_drv = {
 747         .name           = "vmd",
 748         .id_table       = vmd_ids,
 749         .probe          = vmd_probe,
 750         .remove         = vmd_remove,
 751         .driver         = {
 752                 .pm     = &vmd_dev_pm_ops,
 753         },
 754 };
 755 module_pci_driver(vmd_drv);
 756
 757 MODULE_AUTHOR("Intel Corporation");
 758 MODULE_LICENSE("GPL v2");
 759 MODULE_VERSION("0.6");