drivers/pci/host/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/srcu.h>
  23 #include <linux/rculist.h>
  24 #include <linux/rcupdate.h>
  25
  26 #include <asm/irqdomain.h>
  27 #include <asm/device.h>
  28 #include <asm/msi.h>
  29 #include <asm/msidef.h>
  30
  31 #define VMD_CFGBAR      0
  32 #define VMD_MEMBAR1     2
  33 #define VMD_MEMBAR2     4
  34
  35 /*
  36  * Lock for manipulating VMD IRQ lists.
  37  */
  38 static DEFINE_RAW_SPINLOCK(list_lock);
  39
  40 /**
  41  * struct vmd_irq - private data to map driver IRQ to the VMD shared vector
  42  * @node:       list item for parent traversal.
  43  * @irq:        back pointer to parent.
  44  * @enabled:    true if driver enabled IRQ
  45  * @virq:       the virtual IRQ value provided to the requesting driver.
  46  *
  47  * Every MSI/MSI-X IRQ requested for a device in a VMD domain will be mapped to
  48  * a VMD IRQ using this structure.
  49  */
  50 struct vmd_irq {
  51         struct list_head        node;
  52         struct vmd_irq_list     *irq;
  53         bool                    enabled;
  54         unsigned int            virq;
  55 };
  56
  57 /**
  58  * struct vmd_irq_list - list of driver requested IRQs mapping to a VMD vector
  59  * @irq_list:   the list of irq's the VMD one demuxes to.
  60  * @srcu:       SRCU struct for local synchronization.
  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 srcu_struct      srcu;
  67         unsigned int            count;
  68 };
  69
  70 struct vmd_dev {
  71         struct pci_dev          *dev;
  72
  73         spinlock_t              cfg_lock;
  74         char __iomem            *cfgbar;
  75
  76         int msix_count;
  77         struct vmd_irq_list     *irqs;
  78
  79         struct pci_sysdata      sysdata;
  80         struct resource         resources[3];
  81         struct irq_domain       *irq_domain;
  82         struct pci_bus          *bus;
  83
  84 #ifdef CONFIG_X86_DEV_DMA_OPS
  85         struct dma_map_ops      dma_ops;
  86         struct dma_domain       dma_domain;
  87 #endif
  88 };
  89
  90 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
  91 {
  92         return container_of(bus->sysdata, struct vmd_dev, sysdata);
  93 }
  94
  95 static inline unsigned int index_from_irqs(struct vmd_dev *vmd,
  96                                            struct vmd_irq_list *irqs)
  97 {
  98         return irqs - vmd->irqs;
  99 }
 100
 101 /*
 102  * Drivers managing a device in a VMD domain allocate their own IRQs as before,
 103  * but the MSI entry for the hardware it's driving will be programmed with a
 104  * destination ID for the VMD MSI-X table.  The VMD muxes interrupts in its
 105  * domain into one of its own, and the VMD driver de-muxes these for the
 106  * handlers sharing that VMD IRQ.  The vmd irq_domain provides the operations
 107  * and irq_chip to set this up.
 108  */
 109 static void vmd_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
 110 {
 111         struct vmd_irq *vmdirq = data->chip_data;
 112         struct vmd_irq_list *irq = vmdirq->irq;
 113         struct vmd_dev *vmd = irq_data_get_irq_handler_data(data);
 114
 115         msg->address_hi = MSI_ADDR_BASE_HI;
 116         msg->address_lo = MSI_ADDR_BASE_LO |
 117                           MSI_ADDR_DEST_ID(index_from_irqs(vmd, irq));
 118         msg->data = 0;
 119 }
 120
 121 /*
 122  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 123  */
 124 static void vmd_irq_enable(struct irq_data *data)
 125 {
 126         struct vmd_irq *vmdirq = data->chip_data;
 127         unsigned long flags;
 128
 129         raw_spin_lock_irqsave(&list_lock, flags);
 130         WARN_ON(vmdirq->enabled);
 131         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 132         vmdirq->enabled = true;
 133         raw_spin_unlock_irqrestore(&list_lock, flags);
 134
 135         data->chip->irq_unmask(data);
 136 }
 137
 138 static void vmd_irq_disable(struct irq_data *data)
 139 {
 140         struct vmd_irq *vmdirq = data->chip_data;
 141         unsigned long flags;
 142
 143         data->chip->irq_mask(data);
 144
 145         raw_spin_lock_irqsave(&list_lock, flags);
 146         if (vmdirq->enabled) {
 147                 list_del_rcu(&vmdirq->node);
 148                 vmdirq->enabled = false;
 149         }
 150         raw_spin_unlock_irqrestore(&list_lock, flags);
 151 }
 152
 153 /*
 154  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 155  * other devices sharing the same vector.
 156  */
 157 static int vmd_irq_set_affinity(struct irq_data *data,
 158                                 const struct cpumask *dest, bool force)
 159 {
 160         return -EINVAL;
 161 }
 162
 163 static struct irq_chip vmd_msi_controller = {
 164         .name                   = "VMD-MSI",
 165         .irq_enable             = vmd_irq_enable,
 166         .irq_disable            = vmd_irq_disable,
 167         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 168         .irq_set_affinity       = vmd_irq_set_affinity,
 169 };
 170
 171 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 172                                      msi_alloc_info_t *arg)
 173 {
 174         return 0;
 175 }
 176
 177 /*
 178  * XXX: We can be even smarter selecting the best IRQ once we solve the
 179  * affinity problem.
 180  */
 181 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 182 {
 183         int i, best = 1;
 184         unsigned long flags;
 185
 186         if (vmd->msix_count == 1)
 187                 return &vmd->irqs[0];
 188
 189         /*
 190          * White list for fast-interrupt handlers. All others will share the
 191          * "slow" interrupt vector.
 192          */
 193         switch (msi_desc_to_pci_dev(desc)->class) {
 194         case PCI_CLASS_STORAGE_EXPRESS:
 195                 break;
 196         default:
 197                 return &vmd->irqs[0];
 198         }
 199
 200         raw_spin_lock_irqsave(&list_lock, flags);
 201         for (i = 1; i < vmd->msix_count; i++)
 202                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 203                         best = i;
 204         vmd->irqs[best].count++;
 205         raw_spin_unlock_irqrestore(&list_lock, flags);
 206
 207         return &vmd->irqs[best];
 208 }
 209
 210 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 211                         unsigned int virq, irq_hw_number_t hwirq,
 212                         msi_alloc_info_t *arg)
 213 {
 214         struct msi_desc *desc = arg->desc;
 215         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 216         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 217         unsigned int index, vector;
 218
 219         if (!vmdirq)
 220                 return -ENOMEM;
 221
 222         INIT_LIST_HEAD(&vmdirq->node);
 223         vmdirq->irq = vmd_next_irq(vmd, desc);
 224         vmdirq->virq = virq;
 225         index = index_from_irqs(vmd, vmdirq->irq);
 226         vector = pci_irq_vector(vmd->dev, index);
 227
 228         irq_domain_set_info(domain, virq, vector, info->chip, vmdirq,
 229                             handle_untracked_irq, vmd, NULL);
 230         return 0;
 231 }
 232
 233 static void vmd_msi_free(struct irq_domain *domain,
 234                         struct msi_domain_info *info, unsigned int virq)
 235 {
 236         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 237         unsigned long flags;
 238
 239         synchronize_srcu(&vmdirq->irq->srcu);
 240
 241         /* XXX: Potential optimization to rebalance */
 242         raw_spin_lock_irqsave(&list_lock, flags);
 243         vmdirq->irq->count--;
 244         raw_spin_unlock_irqrestore(&list_lock, flags);
 245
 246         kfree(vmdirq);
 247 }
 248
 249 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 250                            int nvec, msi_alloc_info_t *arg)
 251 {
 252         struct pci_dev *pdev = to_pci_dev(dev);
 253         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 254
 255         if (nvec > vmd->msix_count)
 256                 return vmd->msix_count;
 257
 258         memset(arg, 0, sizeof(*arg));
 259         return 0;
 260 }
 261
 262 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 263 {
 264         arg->desc = desc;
 265 }
 266
 267 static struct msi_domain_ops vmd_msi_domain_ops = {
 268         .get_hwirq      = vmd_get_hwirq,
 269         .msi_init       = vmd_msi_init,
 270         .msi_free       = vmd_msi_free,
 271         .msi_prepare    = vmd_msi_prepare,
 272         .set_desc       = vmd_set_desc,
 273 };
 274
 275 static struct msi_domain_info vmd_msi_domain_info = {
 276         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 277                           MSI_FLAG_PCI_MSIX,
 278         .ops            = &vmd_msi_domain_ops,
 279         .chip           = &vmd_msi_controller,
 280 };
 281
 282 #ifdef CONFIG_X86_DEV_DMA_OPS
 283 /*
 284  * VMD replaces the requester ID with its own.  DMA mappings for devices in a
 285  * VMD domain need to be mapped for the VMD, not the device requiring
 286  * the mapping.
 287  */
 288 static struct device *to_vmd_dev(struct device *dev)
 289 {
 290         struct pci_dev *pdev = to_pci_dev(dev);
 291         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 292
 293         return &vmd->dev->dev;
 294 }
 295
 296 static const struct dma_map_ops *vmd_dma_ops(struct device *dev)
 297 {
 298         return get_dma_ops(to_vmd_dev(dev));
 299 }
 300
 301 static void *vmd_alloc(struct device *dev, size_t size, dma_addr_t *addr,
 302                        gfp_t flag, unsigned long attrs)
 303 {
 304         return vmd_dma_ops(dev)->alloc(to_vmd_dev(dev), size, addr, flag,
 305                                        attrs);
 306 }
 307
 308 static void vmd_free(struct device *dev, size_t size, void *vaddr,
 309                      dma_addr_t addr, unsigned long attrs)
 310 {
 311         return vmd_dma_ops(dev)->free(to_vmd_dev(dev), size, vaddr, addr,
 312                                       attrs);
 313 }
 314
 315 static int vmd_mmap(struct device *dev, struct vm_area_struct *vma,
 316                     void *cpu_addr, dma_addr_t addr, size_t size,
 317                     unsigned long attrs)
 318 {
 319         return vmd_dma_ops(dev)->mmap(to_vmd_dev(dev), vma, cpu_addr, addr,
 320                                       size, attrs);
 321 }
 322
 323 static int vmd_get_sgtable(struct device *dev, struct sg_table *sgt,
 324                            void *cpu_addr, dma_addr_t addr, size_t size,
 325                            unsigned long attrs)
 326 {
 327         return vmd_dma_ops(dev)->get_sgtable(to_vmd_dev(dev), sgt, cpu_addr,
 328                                              addr, size, attrs);
 329 }
 330
 331 static dma_addr_t vmd_map_page(struct device *dev, struct page *page,
 332                                unsigned long offset, size_t size,
 333                                enum dma_data_direction dir,
 334                                unsigned long attrs)
 335 {
 336         return vmd_dma_ops(dev)->map_page(to_vmd_dev(dev), page, offset, size,
 337                                           dir, attrs);
 338 }
 339
 340 static void vmd_unmap_page(struct device *dev, dma_addr_t addr, size_t size,
 341                            enum dma_data_direction dir, unsigned long attrs)
 342 {
 343         vmd_dma_ops(dev)->unmap_page(to_vmd_dev(dev), addr, size, dir, attrs);
 344 }
 345
 346 static int vmd_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 347                       enum dma_data_direction dir, unsigned long attrs)
 348 {
 349         return vmd_dma_ops(dev)->map_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 350 }
 351
 352 static void vmd_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 353                          enum dma_data_direction dir, unsigned long attrs)
 354 {
 355         vmd_dma_ops(dev)->unmap_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 356 }
 357
 358 static void vmd_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 359                                     size_t size, enum dma_data_direction dir)
 360 {
 361         vmd_dma_ops(dev)->sync_single_for_cpu(to_vmd_dev(dev), addr, size, dir);
 362 }
 363
 364 static void vmd_sync_single_for_device(struct device *dev, dma_addr_t addr,
 365                                        size_t size, enum dma_data_direction dir)
 366 {
 367         vmd_dma_ops(dev)->sync_single_for_device(to_vmd_dev(dev), addr, size,
 368                                                  dir);
 369 }
 370
 371 static void vmd_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 372                                 int nents, enum dma_data_direction dir)
 373 {
 374         vmd_dma_ops(dev)->sync_sg_for_cpu(to_vmd_dev(dev), sg, nents, dir);
 375 }
 376
 377 static void vmd_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 378                                    int nents, enum dma_data_direction dir)
 379 {
 380         vmd_dma_ops(dev)->sync_sg_for_device(to_vmd_dev(dev), sg, nents, dir);
 381 }
 382
 383 static int vmd_mapping_error(struct device *dev, dma_addr_t addr)
 384 {
 385         return vmd_dma_ops(dev)->mapping_error(to_vmd_dev(dev), addr);
 386 }
 387
 388 static int vmd_dma_supported(struct device *dev, u64 mask)
 389 {
 390         return vmd_dma_ops(dev)->dma_supported(to_vmd_dev(dev), mask);
 391 }
 392
 393 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 394 static u64 vmd_get_required_mask(struct device *dev)
 395 {
 396         return vmd_dma_ops(dev)->get_required_mask(to_vmd_dev(dev));
 397 }
 398 #endif
 399
 400 static void vmd_teardown_dma_ops(struct vmd_dev *vmd)
 401 {
 402         struct dma_domain *domain = &vmd->dma_domain;
 403
 404         if (get_dma_ops(&vmd->dev->dev))
 405                 del_dma_domain(domain);
 406 }
 407
 408 #define ASSIGN_VMD_DMA_OPS(source, dest, fn)    \
 409         do {                                    \
 410                 if (source->fn)                 \
 411                         dest->fn = vmd_##fn;    \
 412         } while (0)
 413
 414 static void vmd_setup_dma_ops(struct vmd_dev *vmd)
 415 {
 416         const struct dma_map_ops *source = get_dma_ops(&vmd->dev->dev);
 417         struct dma_map_ops *dest = &vmd->dma_ops;
 418         struct dma_domain *domain = &vmd->dma_domain;
 419
 420         domain->domain_nr = vmd->sysdata.domain;
 421         domain->dma_ops = dest;
 422
 423         if (!source)
 424                 return;
 425         ASSIGN_VMD_DMA_OPS(source, dest, alloc);
 426         ASSIGN_VMD_DMA_OPS(source, dest, free);
 427         ASSIGN_VMD_DMA_OPS(source, dest, mmap);
 428         ASSIGN_VMD_DMA_OPS(source, dest, get_sgtable);
 429         ASSIGN_VMD_DMA_OPS(source, dest, map_page);
 430         ASSIGN_VMD_DMA_OPS(source, dest, unmap_page);
 431         ASSIGN_VMD_DMA_OPS(source, dest, map_sg);
 432         ASSIGN_VMD_DMA_OPS(source, dest, unmap_sg);
 433         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_cpu);
 434         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_device);
 435         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_cpu);
 436         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_device);
 437         ASSIGN_VMD_DMA_OPS(source, dest, mapping_error);
 438         ASSIGN_VMD_DMA_OPS(source, dest, dma_supported);
 439 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 440         ASSIGN_VMD_DMA_OPS(source, dest, get_required_mask);
 441 #endif
 442         add_dma_domain(domain);
 443 }
 444 #undef ASSIGN_VMD_DMA_OPS
 445 #else
 446 static void vmd_teardown_dma_ops(struct vmd_dev *vmd) {}
 447 static void vmd_setup_dma_ops(struct vmd_dev *vmd) {}
 448 #endif
 449
 450 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 451                                   unsigned int devfn, int reg, int len)
 452 {
 453         char __iomem *addr = vmd->cfgbar +
 454                              (bus->number << 20) + (devfn << 12) + reg;
 455
 456         if ((addr - vmd->cfgbar) + len >=
 457             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 458                 return NULL;
 459
 460         return addr;
 461 }
 462
 463 /*
 464  * CPU may deadlock if config space is not serialized on some versions of this
 465  * hardware, so all config space access is done under a spinlock.
 466  */
 467 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 468                         int len, u32 *value)
 469 {
 470         struct vmd_dev *vmd = vmd_from_bus(bus);
 471         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 472         unsigned long flags;
 473         int ret = 0;
 474
 475         if (!addr)
 476                 return -EFAULT;
 477
 478         spin_lock_irqsave(&vmd->cfg_lock, flags);
 479         switch (len) {
 480         case 1:
 481                 *value = readb(addr);
 482                 break;
 483         case 2:
 484                 *value = readw(addr);
 485                 break;
 486         case 4:
 487                 *value = readl(addr);
 488                 break;
 489         default:
 490                 ret = -EINVAL;
 491                 break;
 492         }
 493         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 494         return ret;
 495 }
 496
 497 /*
 498  * VMD h/w converts non-posted config writes to posted memory writes. The
 499  * read-back in this function forces the completion so it returns only after
 500  * the config space was written, as expected.
 501  */
 502 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 503                          int len, u32 value)
 504 {
 505         struct vmd_dev *vmd = vmd_from_bus(bus);
 506         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 507         unsigned long flags;
 508         int ret = 0;
 509
 510         if (!addr)
 511                 return -EFAULT;
 512
 513         spin_lock_irqsave(&vmd->cfg_lock, flags);
 514         switch (len) {
 515         case 1:
 516                 writeb(value, addr);
 517                 readb(addr);
 518                 break;
 519         case 2:
 520                 writew(value, addr);
 521                 readw(addr);
 522                 break;
 523         case 4:
 524                 writel(value, addr);
 525                 readl(addr);
 526                 break;
 527         default:
 528                 ret = -EINVAL;
 529                 break;
 530         }
 531         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 532         return ret;
 533 }
 534
 535 static struct pci_ops vmd_ops = {
 536         .read           = vmd_pci_read,
 537         .write          = vmd_pci_write,
 538 };
 539
 540 static void vmd_attach_resources(struct vmd_dev *vmd)
 541 {
 542         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 543         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 544 }
 545
 546 static void vmd_detach_resources(struct vmd_dev *vmd)
 547 {
 548         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 549         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 550 }
 551
 552 /*
 553  * VMD domains start at 0x10000 to not clash with ACPI _SEG domains.
 554  * Per ACPI r6.0, sec 6.5.6,  _SEG returns an integer, of which the lower
 555  * 16 bits are the PCI Segment Group (domain) number.  Other bits are
 556  * currently reserved.
 557  */
 558 static int vmd_find_free_domain(void)
 559 {
 560         int domain = 0xffff;
 561         struct pci_bus *bus = NULL;
 562
 563         while ((bus = pci_find_next_bus(bus)) != NULL)
 564                 domain = max_t(int, domain, pci_domain_nr(bus));
 565         return domain + 1;
 566 }
 567
 568 static int vmd_enable_domain(struct vmd_dev *vmd)
 569 {
 570         struct pci_sysdata *sd = &vmd->sysdata;
 571         struct fwnode_handle *fn;
 572         struct resource *res;
 573         u32 upper_bits;
 574         unsigned long flags;
 575         LIST_HEAD(resources);
 576
 577         res = &vmd->dev->resource[VMD_CFGBAR];
 578         vmd->resources[0] = (struct resource) {
 579                 .name  = "VMD CFGBAR",
 580                 .start = 0,
 581                 .end   = (resource_size(res) >> 20) - 1,
 582                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 583         };
 584
 585         /*
 586          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 587          * put 32-bit resources in the window.
 588          *
 589          * There's no hardware reason why a 64-bit window *couldn't*
 590          * contain a 32-bit resource, but pbus_size_mem() computes the
 591          * bridge window size assuming a 64-bit window will contain no
 592          * 32-bit resources.  __pci_assign_resource() enforces that
 593          * artificial restriction to make sure everything will fit.
 594          *
 595          * The only way we could use a 64-bit non-prefechable MEMBAR is
 596          * if its address is <4GB so that we can convert it to a 32-bit
 597          * resource.  To be visible to the host OS, all VMD endpoints must
 598          * be initially configured by platform BIOS, which includes setting
 599          * up these resources.  We can assume the device is configured
 600          * according to the platform needs.
 601          */
 602         res = &vmd->dev->resource[VMD_MEMBAR1];
 603         upper_bits = upper_32_bits(res->end);
 604         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 605         if (!upper_bits)
 606                 flags &= ~IORESOURCE_MEM_64;
 607         vmd->resources[1] = (struct resource) {
 608                 .name  = "VMD MEMBAR1",
 609                 .start = res->start,
 610                 .end   = res->end,
 611                 .flags = flags,
 612                 .parent = res,
 613         };
 614
 615         res = &vmd->dev->resource[VMD_MEMBAR2];
 616         upper_bits = upper_32_bits(res->end);
 617         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 618         if (!upper_bits)
 619                 flags &= ~IORESOURCE_MEM_64;
 620         vmd->resources[2] = (struct resource) {
 621                 .name  = "VMD MEMBAR2",
 622                 .start = res->start + 0x2000,
 623                 .end   = res->end,
 624                 .flags = flags,
 625                 .parent = res,
 626         };
 627
 628         sd->vmd_domain = true;
 629         sd->domain = vmd_find_free_domain();
 630         if (sd->domain < 0)
 631                 return sd->domain;
 632
 633         sd->node = pcibus_to_node(vmd->dev->bus);
 634
 635         fn = irq_domain_alloc_named_id_fwnode("VMD-MSI", vmd->sysdata.domain);
 636         if (!fn)
 637                 return -ENODEV;
 638
 639         vmd->irq_domain = pci_msi_create_irq_domain(fn, &vmd_msi_domain_info,
 640                                                     x86_vector_domain);
 641         if (!vmd->irq_domain) {
 642                 irq_domain_free_fwnode(fn);
 643                 return -ENODEV;
 644         }
 645
 646         pci_add_resource(&resources, &vmd->resources[0]);
 647         pci_add_resource(&resources, &vmd->resources[1]);
 648         pci_add_resource(&resources, &vmd->resources[2]);
 649         vmd->bus = pci_create_root_bus(&vmd->dev->dev, 0, &vmd_ops, sd,
 650                                        &resources);
 651         if (!vmd->bus) {
 652                 pci_free_resource_list(&resources);
 653                 irq_domain_remove(vmd->irq_domain);
 654                 irq_domain_free_fwnode(fn);
 655                 return -ENODEV;
 656         }
 657
 658         vmd_attach_resources(vmd);
 659         vmd_setup_dma_ops(vmd);
 660         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 661         pci_rescan_bus(vmd->bus);
 662
 663         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 664                                "domain"), "Can't create symlink to domain\n");
 665         return 0;
 666 }
 667
 668 static irqreturn_t vmd_irq(int irq, void *data)
 669 {
 670         struct vmd_irq_list *irqs = data;
 671         struct vmd_irq *vmdirq;
 672         int idx;
 673
 674         idx = srcu_read_lock(&irqs->srcu);
 675         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 676                 generic_handle_irq(vmdirq->virq);
 677         srcu_read_unlock(&irqs->srcu, idx);
 678
 679         return IRQ_HANDLED;
 680 }
 681
 682 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 683 {
 684         struct vmd_dev *vmd;
 685         int i, err;
 686
 687         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 688                 return -ENOMEM;
 689
 690         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 691         if (!vmd)
 692                 return -ENOMEM;
 693
 694         vmd->dev = dev;
 695         err = pcim_enable_device(dev);
 696         if (err < 0)
 697                 return err;
 698
 699         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 700         if (!vmd->cfgbar)
 701                 return -ENOMEM;
 702
 703         pci_set_master(dev);
 704         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 705             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 706                 return -ENODEV;
 707
 708         vmd->msix_count = pci_msix_vec_count(dev);
 709         if (vmd->msix_count < 0)
 710                 return -ENODEV;
 711
 712         vmd->msix_count = pci_alloc_irq_vectors(dev, 1, vmd->msix_count,
 713                                         PCI_IRQ_MSIX);
 714         if (vmd->msix_count < 0)
 715                 return vmd->msix_count;
 716
 717         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 718                                  GFP_KERNEL);
 719         if (!vmd->irqs)
 720                 return -ENOMEM;
 721
 722         for (i = 0; i < vmd->msix_count; i++) {
 723                 err = init_srcu_struct(&vmd->irqs[i].srcu);
 724                 if (err)
 725                         return err;
 726
 727                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 728                 err = devm_request_irq(&dev->dev, pci_irq_vector(dev, i),
 729                                        vmd_irq, IRQF_NO_THREAD,
 730                                        "vmd", &vmd->irqs[i]);
 731                 if (err)
 732                         return err;
 733         }
 734
 735         spin_lock_init(&vmd->cfg_lock);
 736         pci_set_drvdata(dev, vmd);
 737         err = vmd_enable_domain(vmd);
 738         if (err)
 739                 return err;
 740
 741         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 742                  vmd->sysdata.domain);
 743         return 0;
 744 }
 745
 746 static void vmd_cleanup_srcu(struct vmd_dev *vmd)
 747 {
 748         int i;
 749
 750         for (i = 0; i < vmd->msix_count; i++)
 751                 cleanup_srcu_struct(&vmd->irqs[i].srcu);
 752 }
 753
 754 static void vmd_remove(struct pci_dev *dev)
 755 {
 756         struct vmd_dev *vmd = pci_get_drvdata(dev);
 757         struct fwnode_handle *fn = vmd->irq_domain->fwnode;
 758
 759         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 760         pci_stop_root_bus(vmd->bus);
 761         pci_remove_root_bus(vmd->bus);
 762         vmd_cleanup_srcu(vmd);
 763         vmd_teardown_dma_ops(vmd);
 764         vmd_detach_resources(vmd);
 765         irq_domain_remove(vmd->irq_domain);
 766         irq_domain_free_fwnode(fn);
 767 }
 768
 769 #ifdef CONFIG_PM_SLEEP
 770 static int vmd_suspend(struct device *dev)
 771 {
 772         struct pci_dev *pdev = to_pci_dev(dev);
 773         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 774         int i;
 775
 776         for (i = 0; i < vmd->msix_count; i++)
 777                 devm_free_irq(dev, pci_irq_vector(pdev, i), &vmd->irqs[i]);
 778
 779         pci_save_state(pdev);
 780         return 0;
 781 }
 782
 783 static int vmd_resume(struct device *dev)
 784 {
 785         struct pci_dev *pdev = to_pci_dev(dev);
 786         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 787         int err, i;
 788
 789         for (i = 0; i < vmd->msix_count; i++) {
 790                 err = devm_request_irq(dev, pci_irq_vector(pdev, i),
 791                                        vmd_irq, IRQF_NO_THREAD,
 792                                        "vmd", &vmd->irqs[i]);
 793                 if (err)
 794                         return err;
 795         }
 796
 797         pci_restore_state(pdev);
 798         return 0;
 799 }
 800 #endif
 801 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 802
 803 static const struct pci_device_id vmd_ids[] = {
 804         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x201d),},
 805         {0,}
 806 };
 807 MODULE_DEVICE_TABLE(pci, vmd_ids);
 808
 809 static struct pci_driver vmd_drv = {
 810         .name           = "vmd",
 811         .id_table       = vmd_ids,
 812         .probe          = vmd_probe,
 813         .remove         = vmd_remove,
 814         .driver         = {
 815                 .pm     = &vmd_dev_pm_ops,
 816         },
 817 };
 818 module_pci_driver(vmd_drv);
 819
 820 MODULE_AUTHOR("Intel Corporation");
 821 MODULE_LICENSE("GPL v2");
 822 MODULE_VERSION("0.6");