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