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  * @vmd_vector: the h/w IRQ assigned to the VMD.
  60  * @index:      index into the VMD MSI-X table; used for message routing.
  61  * @count:      number of child IRQs assigned to this vector; used to track
  62  *              sharing.
  63  */
  64 struct vmd_irq_list {
  65         struct list_head        irq_list;
  66         struct vmd_dev          *vmd;
  67         unsigned int            vmd_vector;
  68         unsigned int            index;
  69         unsigned int            count;
  70 };
  71
  72 struct vmd_dev {
  73         struct pci_dev          *dev;
  74
  75         spinlock_t              cfg_lock;
  76         char __iomem            *cfgbar;
  77
  78         int msix_count;
  79         struct msix_entry       *msix_entries;
  80         struct vmd_irq_list     *irqs;
  81
  82         struct pci_sysdata      sysdata;
  83         struct resource         resources[3];
  84         struct irq_domain       *irq_domain;
  85         struct pci_bus          *bus;
  86
  87 #ifdef CONFIG_X86_DEV_DMA_OPS
  88         struct dma_map_ops      dma_ops;
  89         struct dma_domain       dma_domain;
  90 #endif
  91 };
  92
  93 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
  94 {
  95         return container_of(bus->sysdata, struct vmd_dev, sysdata);
  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
 111         msg->address_hi = MSI_ADDR_BASE_HI;
 112         msg->address_lo = MSI_ADDR_BASE_LO | MSI_ADDR_DEST_ID(irq->index);
 113         msg->data = 0;
 114 }
 115
 116 /*
 117  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 118  */
 119 static void vmd_irq_enable(struct irq_data *data)
 120 {
 121         struct vmd_irq *vmdirq = data->chip_data;
 122         unsigned long flags;
 123
 124         raw_spin_lock_irqsave(&list_lock, flags);
 125         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 126         raw_spin_unlock_irqrestore(&list_lock, flags);
 127
 128         data->chip->irq_unmask(data);
 129 }
 130
 131 static void vmd_irq_disable(struct irq_data *data)
 132 {
 133         struct vmd_irq *vmdirq = data->chip_data;
 134         unsigned long flags;
 135
 136         data->chip->irq_mask(data);
 137
 138         raw_spin_lock_irqsave(&list_lock, flags);
 139         list_del_rcu(&vmdirq->node);
 140         INIT_LIST_HEAD_RCU(&vmdirq->node);
 141         raw_spin_unlock_irqrestore(&list_lock, flags);
 142 }
 143
 144 /*
 145  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 146  * other devices sharing the same vector.
 147  */
 148 static int vmd_irq_set_affinity(struct irq_data *data,
 149                                 const struct cpumask *dest, bool force)
 150 {
 151         return -EINVAL;
 152 }
 153
 154 static struct irq_chip vmd_msi_controller = {
 155         .name                   = "VMD-MSI",
 156         .irq_enable             = vmd_irq_enable,
 157         .irq_disable            = vmd_irq_disable,
 158         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 159         .irq_set_affinity       = vmd_irq_set_affinity,
 160 };
 161
 162 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 163                                      msi_alloc_info_t *arg)
 164 {
 165         return 0;
 166 }
 167
 168 /*
 169  * XXX: We can be even smarter selecting the best IRQ once we solve the
 170  * affinity problem.
 171  */
 172 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 173 {
 174         int i, best = 1;
 175         unsigned long flags;
 176
 177         if (!desc->msi_attrib.is_msix || vmd->msix_count == 1)
 178                 return &vmd->irqs[0];
 179
 180         raw_spin_lock_irqsave(&list_lock, flags);
 181         for (i = 1; i < vmd->msix_count; i++)
 182                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 183                         best = i;
 184         vmd->irqs[best].count++;
 185         raw_spin_unlock_irqrestore(&list_lock, flags);
 186
 187         return &vmd->irqs[best];
 188 }
 189
 190 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 191                         unsigned int virq, irq_hw_number_t hwirq,
 192                         msi_alloc_info_t *arg)
 193 {
 194         struct msi_desc *desc = arg->desc;
 195         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 196         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 197
 198         if (!vmdirq)
 199                 return -ENOMEM;
 200
 201         INIT_LIST_HEAD(&vmdirq->node);
 202         vmdirq->irq = vmd_next_irq(vmd, desc);
 203         vmdirq->virq = virq;
 204
 205         irq_domain_set_info(domain, virq, vmdirq->irq->vmd_vector, info->chip,
 206                             vmdirq, handle_untracked_irq, vmd, NULL);
 207         return 0;
 208 }
 209
 210 static void vmd_msi_free(struct irq_domain *domain,
 211                         struct msi_domain_info *info, unsigned int virq)
 212 {
 213         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 214         unsigned long flags;
 215
 216         /* XXX: Potential optimization to rebalance */
 217         raw_spin_lock_irqsave(&list_lock, flags);
 218         vmdirq->irq->count--;
 219         raw_spin_unlock_irqrestore(&list_lock, flags);
 220
 221         kfree_rcu(vmdirq, rcu);
 222 }
 223
 224 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 225                            int nvec, msi_alloc_info_t *arg)
 226 {
 227         struct pci_dev *pdev = to_pci_dev(dev);
 228         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 229
 230         if (nvec > vmd->msix_count)
 231                 return vmd->msix_count;
 232
 233         memset(arg, 0, sizeof(*arg));
 234         return 0;
 235 }
 236
 237 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 238 {
 239         arg->desc = desc;
 240 }
 241
 242 static struct msi_domain_ops vmd_msi_domain_ops = {
 243         .get_hwirq      = vmd_get_hwirq,
 244         .msi_init       = vmd_msi_init,
 245         .msi_free       = vmd_msi_free,
 246         .msi_prepare    = vmd_msi_prepare,
 247         .set_desc       = vmd_set_desc,
 248 };
 249
 250 static struct msi_domain_info vmd_msi_domain_info = {
 251         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 252                           MSI_FLAG_PCI_MSIX,
 253         .ops            = &vmd_msi_domain_ops,
 254         .chip           = &vmd_msi_controller,
 255 };
 256
 257 #ifdef CONFIG_X86_DEV_DMA_OPS
 258 /*
 259  * VMD replaces the requester ID with its own.  DMA mappings for devices in a
 260  * VMD domain need to be mapped for the VMD, not the device requiring
 261  * the mapping.
 262  */
 263 static struct device *to_vmd_dev(struct device *dev)
 264 {
 265         struct pci_dev *pdev = to_pci_dev(dev);
 266         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 267
 268         return &vmd->dev->dev;
 269 }
 270
 271 static struct dma_map_ops *vmd_dma_ops(struct device *dev)
 272 {
 273         return get_dma_ops(to_vmd_dev(dev));
 274 }
 275
 276 static void *vmd_alloc(struct device *dev, size_t size, dma_addr_t *addr,
 277                        gfp_t flag, unsigned long attrs)
 278 {
 279         return vmd_dma_ops(dev)->alloc(to_vmd_dev(dev), size, addr, flag,
 280                                        attrs);
 281 }
 282
 283 static void vmd_free(struct device *dev, size_t size, void *vaddr,
 284                      dma_addr_t addr, unsigned long attrs)
 285 {
 286         return vmd_dma_ops(dev)->free(to_vmd_dev(dev), size, vaddr, addr,
 287                                       attrs);
 288 }
 289
 290 static int vmd_mmap(struct device *dev, struct vm_area_struct *vma,
 291                     void *cpu_addr, dma_addr_t addr, size_t size,
 292                     unsigned long attrs)
 293 {
 294         return vmd_dma_ops(dev)->mmap(to_vmd_dev(dev), vma, cpu_addr, addr,
 295                                       size, attrs);
 296 }
 297
 298 static int vmd_get_sgtable(struct device *dev, struct sg_table *sgt,
 299                            void *cpu_addr, dma_addr_t addr, size_t size,
 300                            unsigned long attrs)
 301 {
 302         return vmd_dma_ops(dev)->get_sgtable(to_vmd_dev(dev), sgt, cpu_addr,
 303                                              addr, size, attrs);
 304 }
 305
 306 static dma_addr_t vmd_map_page(struct device *dev, struct page *page,
 307                                unsigned long offset, size_t size,
 308                                enum dma_data_direction dir,
 309                                unsigned long attrs)
 310 {
 311         return vmd_dma_ops(dev)->map_page(to_vmd_dev(dev), page, offset, size,
 312                                           dir, attrs);
 313 }
 314
 315 static void vmd_unmap_page(struct device *dev, dma_addr_t addr, size_t size,
 316                            enum dma_data_direction dir, unsigned long attrs)
 317 {
 318         vmd_dma_ops(dev)->unmap_page(to_vmd_dev(dev), addr, size, dir, attrs);
 319 }
 320
 321 static int vmd_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 322                       enum dma_data_direction dir, unsigned long attrs)
 323 {
 324         return vmd_dma_ops(dev)->map_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 325 }
 326
 327 static void vmd_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 328                          enum dma_data_direction dir, unsigned long attrs)
 329 {
 330         vmd_dma_ops(dev)->unmap_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 331 }
 332
 333 static void vmd_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 334                                     size_t size, enum dma_data_direction dir)
 335 {
 336         vmd_dma_ops(dev)->sync_single_for_cpu(to_vmd_dev(dev), addr, size, dir);
 337 }
 338
 339 static void vmd_sync_single_for_device(struct device *dev, dma_addr_t addr,
 340                                        size_t size, enum dma_data_direction dir)
 341 {
 342         vmd_dma_ops(dev)->sync_single_for_device(to_vmd_dev(dev), addr, size,
 343                                                  dir);
 344 }
 345
 346 static void vmd_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 347                                 int nents, enum dma_data_direction dir)
 348 {
 349         vmd_dma_ops(dev)->sync_sg_for_cpu(to_vmd_dev(dev), sg, nents, dir);
 350 }
 351
 352 static void vmd_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 353                                    int nents, enum dma_data_direction dir)
 354 {
 355         vmd_dma_ops(dev)->sync_sg_for_device(to_vmd_dev(dev), sg, nents, dir);
 356 }
 357
 358 static int vmd_mapping_error(struct device *dev, dma_addr_t addr)
 359 {
 360         return vmd_dma_ops(dev)->mapping_error(to_vmd_dev(dev), addr);
 361 }
 362
 363 static int vmd_dma_supported(struct device *dev, u64 mask)
 364 {
 365         return vmd_dma_ops(dev)->dma_supported(to_vmd_dev(dev), mask);
 366 }
 367
 368 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 369 static u64 vmd_get_required_mask(struct device *dev)
 370 {
 371         return vmd_dma_ops(dev)->get_required_mask(to_vmd_dev(dev));
 372 }
 373 #endif
 374
 375 static void vmd_teardown_dma_ops(struct vmd_dev *vmd)
 376 {
 377         struct dma_domain *domain = &vmd->dma_domain;
 378
 379         if (get_dma_ops(&vmd->dev->dev))
 380                 del_dma_domain(domain);
 381 }
 382
 383 #define ASSIGN_VMD_DMA_OPS(source, dest, fn)    \
 384         do {                                    \
 385                 if (source->fn)                 \
 386                         dest->fn = vmd_##fn;    \
 387         } while (0)
 388
 389 static void vmd_setup_dma_ops(struct vmd_dev *vmd)
 390 {
 391         const struct dma_map_ops *source = get_dma_ops(&vmd->dev->dev);
 392         struct dma_map_ops *dest = &vmd->dma_ops;
 393         struct dma_domain *domain = &vmd->dma_domain;
 394
 395         domain->domain_nr = vmd->sysdata.domain;
 396         domain->dma_ops = dest;
 397
 398         if (!source)
 399                 return;
 400         ASSIGN_VMD_DMA_OPS(source, dest, alloc);
 401         ASSIGN_VMD_DMA_OPS(source, dest, free);
 402         ASSIGN_VMD_DMA_OPS(source, dest, mmap);
 403         ASSIGN_VMD_DMA_OPS(source, dest, get_sgtable);
 404         ASSIGN_VMD_DMA_OPS(source, dest, map_page);
 405         ASSIGN_VMD_DMA_OPS(source, dest, unmap_page);
 406         ASSIGN_VMD_DMA_OPS(source, dest, map_sg);
 407         ASSIGN_VMD_DMA_OPS(source, dest, unmap_sg);
 408         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_cpu);
 409         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_device);
 410         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_cpu);
 411         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_device);
 412         ASSIGN_VMD_DMA_OPS(source, dest, mapping_error);
 413         ASSIGN_VMD_DMA_OPS(source, dest, dma_supported);
 414 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 415         ASSIGN_VMD_DMA_OPS(source, dest, get_required_mask);
 416 #endif
 417         add_dma_domain(domain);
 418 }
 419 #undef ASSIGN_VMD_DMA_OPS
 420 #else
 421 static void vmd_teardown_dma_ops(struct vmd_dev *vmd) {}
 422 static void vmd_setup_dma_ops(struct vmd_dev *vmd) {}
 423 #endif
 424
 425 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 426                                   unsigned int devfn, int reg, int len)
 427 {
 428         char __iomem *addr = vmd->cfgbar +
 429                              (bus->number << 20) + (devfn << 12) + reg;
 430
 431         if ((addr - vmd->cfgbar) + len >=
 432             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 433                 return NULL;
 434
 435         return addr;
 436 }
 437
 438 /*
 439  * CPU may deadlock if config space is not serialized on some versions of this
 440  * hardware, so all config space access is done under a spinlock.
 441  */
 442 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 443                         int len, u32 *value)
 444 {
 445         struct vmd_dev *vmd = vmd_from_bus(bus);
 446         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 447         unsigned long flags;
 448         int ret = 0;
 449
 450         if (!addr)
 451                 return -EFAULT;
 452
 453         spin_lock_irqsave(&vmd->cfg_lock, flags);
 454         switch (len) {
 455         case 1:
 456                 *value = readb(addr);
 457                 break;
 458         case 2:
 459                 *value = readw(addr);
 460                 break;
 461         case 4:
 462                 *value = readl(addr);
 463                 break;
 464         default:
 465                 ret = -EINVAL;
 466                 break;
 467         }
 468         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 469         return ret;
 470 }
 471
 472 /*
 473  * VMD h/w converts non-posted config writes to posted memory writes. The
 474  * read-back in this function forces the completion so it returns only after
 475  * the config space was written, as expected.
 476  */
 477 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 478                          int len, u32 value)
 479 {
 480         struct vmd_dev *vmd = vmd_from_bus(bus);
 481         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 482         unsigned long flags;
 483         int ret = 0;
 484
 485         if (!addr)
 486                 return -EFAULT;
 487
 488         spin_lock_irqsave(&vmd->cfg_lock, flags);
 489         switch (len) {
 490         case 1:
 491                 writeb(value, addr);
 492                 readb(addr);
 493                 break;
 494         case 2:
 495                 writew(value, addr);
 496                 readw(addr);
 497                 break;
 498         case 4:
 499                 writel(value, addr);
 500                 readl(addr);
 501                 break;
 502         default:
 503                 ret = -EINVAL;
 504                 break;
 505         }
 506         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 507         return ret;
 508 }
 509
 510 static struct pci_ops vmd_ops = {
 511         .read           = vmd_pci_read,
 512         .write          = vmd_pci_write,
 513 };
 514
 515 static void vmd_attach_resources(struct vmd_dev *vmd)
 516 {
 517         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 518         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 519 }
 520
 521 static void vmd_detach_resources(struct vmd_dev *vmd)
 522 {
 523         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 524         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 525 }
 526
 527 /*
 528  * VMD domains start at 0x1000 to not clash with ACPI _SEG domains.
 529  */
 530 static int vmd_find_free_domain(void)
 531 {
 532         int domain = 0xffff;
 533         struct pci_bus *bus = NULL;
 534
 535         while ((bus = pci_find_next_bus(bus)) != NULL)
 536                 domain = max_t(int, domain, pci_domain_nr(bus));
 537         return domain + 1;
 538 }
 539
 540 static int vmd_enable_domain(struct vmd_dev *vmd)
 541 {
 542         struct pci_sysdata *sd = &vmd->sysdata;
 543         struct resource *res;
 544         u32 upper_bits;
 545         unsigned long flags;
 546         LIST_HEAD(resources);
 547
 548         res = &vmd->dev->resource[VMD_CFGBAR];
 549         vmd->resources[0] = (struct resource) {
 550                 .name  = "VMD CFGBAR",
 551                 .start = 0,
 552                 .end   = (resource_size(res) >> 20) - 1,
 553                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 554         };
 555
 556         /*
 557          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 558          * put 32-bit resources in the window.
 559          *
 560          * There's no hardware reason why a 64-bit window *couldn't*
 561          * contain a 32-bit resource, but pbus_size_mem() computes the
 562          * bridge window size assuming a 64-bit window will contain no
 563          * 32-bit resources.  __pci_assign_resource() enforces that
 564          * artificial restriction to make sure everything will fit.
 565          *
 566          * The only way we could use a 64-bit non-prefechable MEMBAR is
 567          * if its address is <4GB so that we can convert it to a 32-bit
 568          * resource.  To be visible to the host OS, all VMD endpoints must
 569          * be initially configured by platform BIOS, which includes setting
 570          * up these resources.  We can assume the device is configured
 571          * according to the platform needs.
 572          */
 573         res = &vmd->dev->resource[VMD_MEMBAR1];
 574         upper_bits = upper_32_bits(res->end);
 575         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 576         if (!upper_bits)
 577                 flags &= ~IORESOURCE_MEM_64;
 578         vmd->resources[1] = (struct resource) {
 579                 .name  = "VMD MEMBAR1",
 580                 .start = res->start,
 581                 .end   = res->end,
 582                 .flags = flags,
 583                 .parent = res,
 584         };
 585
 586         res = &vmd->dev->resource[VMD_MEMBAR2];
 587         upper_bits = upper_32_bits(res->end);
 588         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 589         if (!upper_bits)
 590                 flags &= ~IORESOURCE_MEM_64;
 591         vmd->resources[2] = (struct resource) {
 592                 .name  = "VMD MEMBAR2",
 593                 .start = res->start + 0x2000,
 594                 .end   = res->end,
 595                 .flags = flags,
 596                 .parent = res,
 597         };
 598
 599         sd->domain = vmd_find_free_domain();
 600         if (sd->domain < 0)
 601                 return sd->domain;
 602
 603         sd->node = pcibus_to_node(vmd->dev->bus);
 604
 605         vmd->irq_domain = pci_msi_create_irq_domain(NULL, &vmd_msi_domain_info,
 606                                                     x86_vector_domain);
 607         if (!vmd->irq_domain)
 608                 return -ENODEV;
 609
 610         pci_add_resource(&resources, &vmd->resources[0]);
 611         pci_add_resource(&resources, &vmd->resources[1]);
 612         pci_add_resource(&resources, &vmd->resources[2]);
 613         vmd->bus = pci_create_root_bus(&vmd->dev->dev, 0, &vmd_ops, sd,
 614                                        &resources);
 615         if (!vmd->bus) {
 616                 pci_free_resource_list(&resources);
 617                 irq_domain_remove(vmd->irq_domain);
 618                 return -ENODEV;
 619         }
 620
 621         vmd_attach_resources(vmd);
 622         vmd_setup_dma_ops(vmd);
 623         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 624         pci_rescan_bus(vmd->bus);
 625
 626         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 627                                "domain"), "Can't create symlink to domain\n");
 628         return 0;
 629 }
 630
 631 static irqreturn_t vmd_irq(int irq, void *data)
 632 {
 633         struct vmd_irq_list *irqs = data;
 634         struct vmd_irq *vmdirq;
 635
 636         rcu_read_lock();
 637         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 638                 generic_handle_irq(vmdirq->virq);
 639         rcu_read_unlock();
 640
 641         return IRQ_HANDLED;
 642 }
 643
 644 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 645 {
 646         struct vmd_dev *vmd;
 647         int i, err;
 648
 649         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 650                 return -ENOMEM;
 651
 652         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 653         if (!vmd)
 654                 return -ENOMEM;
 655
 656         vmd->dev = dev;
 657         err = pcim_enable_device(dev);
 658         if (err < 0)
 659                 return err;
 660
 661         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 662         if (!vmd->cfgbar)
 663                 return -ENOMEM;
 664
 665         pci_set_master(dev);
 666         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 667             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 668                 return -ENODEV;
 669
 670         vmd->msix_count = pci_msix_vec_count(dev);
 671         if (vmd->msix_count < 0)
 672                 return -ENODEV;
 673
 674         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 675                                  GFP_KERNEL);
 676         if (!vmd->irqs)
 677                 return -ENOMEM;
 678
 679         vmd->msix_entries = devm_kcalloc(&dev->dev, vmd->msix_count,
 680                                          sizeof(*vmd->msix_entries),
 681                                          GFP_KERNEL);
 682         if (!vmd->msix_entries)
 683                 return -ENOMEM;
 684         for (i = 0; i < vmd->msix_count; i++)
 685                 vmd->msix_entries[i].entry = i;
 686
 687         vmd->msix_count = pci_enable_msix_range(vmd->dev, vmd->msix_entries, 1,
 688                                                 vmd->msix_count);
 689         if (vmd->msix_count < 0)
 690                 return vmd->msix_count;
 691
 692         for (i = 0; i < vmd->msix_count; i++) {
 693                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 694                 vmd->irqs[i].vmd_vector = vmd->msix_entries[i].vector;
 695                 vmd->irqs[i].index = i;
 696
 697                 err = devm_request_irq(&dev->dev, vmd->irqs[i].vmd_vector,
 698                                        vmd_irq, 0, "vmd", &vmd->irqs[i]);
 699                 if (err)
 700                         return err;
 701         }
 702
 703         spin_lock_init(&vmd->cfg_lock);
 704         pci_set_drvdata(dev, vmd);
 705         err = vmd_enable_domain(vmd);
 706         if (err)
 707                 return err;
 708
 709         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 710                  vmd->sysdata.domain);
 711         return 0;
 712 }
 713
 714 static void vmd_remove(struct pci_dev *dev)
 715 {
 716         struct vmd_dev *vmd = pci_get_drvdata(dev);
 717
 718         vmd_detach_resources(vmd);
 719         pci_set_drvdata(dev, NULL);
 720         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 721         pci_stop_root_bus(vmd->bus);
 722         pci_remove_root_bus(vmd->bus);
 723         vmd_teardown_dma_ops(vmd);
 724         irq_domain_remove(vmd->irq_domain);
 725 }
 726
 727 #ifdef CONFIG_PM
 728 static int vmd_suspend(struct device *dev)
 729 {
 730         struct pci_dev *pdev = to_pci_dev(dev);
 731
 732         pci_save_state(pdev);
 733         return 0;
 734 }
 735
 736 static int vmd_resume(struct device *dev)
 737 {
 738         struct pci_dev *pdev = to_pci_dev(dev);
 739
 740         pci_restore_state(pdev);
 741         return 0;
 742 }
 743 #endif
 744 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 745
 746 static const struct pci_device_id vmd_ids[] = {
 747         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x201d),},
 748         {0,}
 749 };
 750 MODULE_DEVICE_TABLE(pci, vmd_ids);
 751
 752 static struct pci_driver vmd_drv = {
 753         .name           = "vmd",
 754         .id_table       = vmd_ids,
 755         .probe          = vmd_probe,
 756         .remove         = vmd_remove,
 757         .driver         = {
 758                 .pm     = &vmd_dev_pm_ops,
 759         },
 760 };
 761 module_pci_driver(vmd_drv);
 762
 763 MODULE_AUTHOR("Intel Corporation");
 764 MODULE_LICENSE("GPL v2");
 765 MODULE_VERSION("0.6");