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 (!desc->msi_attrib.is_msix || vmd->msix_count == 1)
 187                 return &vmd->irqs[0];
 188
 189         raw_spin_lock_irqsave(&list_lock, flags);
 190         for (i = 1; i < vmd->msix_count; i++)
 191                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 192                         best = i;
 193         vmd->irqs[best].count++;
 194         raw_spin_unlock_irqrestore(&list_lock, flags);
 195
 196         return &vmd->irqs[best];
 197 }
 198
 199 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 200                         unsigned int virq, irq_hw_number_t hwirq,
 201                         msi_alloc_info_t *arg)
 202 {
 203         struct msi_desc *desc = arg->desc;
 204         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 205         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 206         unsigned int index, vector;
 207
 208         if (!vmdirq)
 209                 return -ENOMEM;
 210
 211         INIT_LIST_HEAD(&vmdirq->node);
 212         vmdirq->irq = vmd_next_irq(vmd, desc);
 213         vmdirq->virq = virq;
 214         index = index_from_irqs(vmd, vmdirq->irq);
 215         vector = pci_irq_vector(vmd->dev, index);
 216
 217         irq_domain_set_info(domain, virq, vector, info->chip, vmdirq,
 218                             handle_untracked_irq, vmd, NULL);
 219         return 0;
 220 }
 221
 222 static void vmd_msi_free(struct irq_domain *domain,
 223                         struct msi_domain_info *info, unsigned int virq)
 224 {
 225         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 226         unsigned long flags;
 227
 228         synchronize_srcu(&vmdirq->irq->srcu);
 229
 230         /* XXX: Potential optimization to rebalance */
 231         raw_spin_lock_irqsave(&list_lock, flags);
 232         vmdirq->irq->count--;
 233         raw_spin_unlock_irqrestore(&list_lock, flags);
 234
 235         kfree(vmdirq);
 236 }
 237
 238 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 239                            int nvec, msi_alloc_info_t *arg)
 240 {
 241         struct pci_dev *pdev = to_pci_dev(dev);
 242         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 243
 244         if (nvec > vmd->msix_count)
 245                 return vmd->msix_count;
 246
 247         memset(arg, 0, sizeof(*arg));
 248         return 0;
 249 }
 250
 251 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 252 {
 253         arg->desc = desc;
 254 }
 255
 256 static struct msi_domain_ops vmd_msi_domain_ops = {
 257         .get_hwirq      = vmd_get_hwirq,
 258         .msi_init       = vmd_msi_init,
 259         .msi_free       = vmd_msi_free,
 260         .msi_prepare    = vmd_msi_prepare,
 261         .set_desc       = vmd_set_desc,
 262 };
 263
 264 static struct msi_domain_info vmd_msi_domain_info = {
 265         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 266                           MSI_FLAG_PCI_MSIX,
 267         .ops            = &vmd_msi_domain_ops,
 268         .chip           = &vmd_msi_controller,
 269 };
 270
 271 #ifdef CONFIG_X86_DEV_DMA_OPS
 272 /*
 273  * VMD replaces the requester ID with its own.  DMA mappings for devices in a
 274  * VMD domain need to be mapped for the VMD, not the device requiring
 275  * the mapping.
 276  */
 277 static struct device *to_vmd_dev(struct device *dev)
 278 {
 279         struct pci_dev *pdev = to_pci_dev(dev);
 280         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 281
 282         return &vmd->dev->dev;
 283 }
 284
 285 static const struct dma_map_ops *vmd_dma_ops(struct device *dev)
 286 {
 287         return get_dma_ops(to_vmd_dev(dev));
 288 }
 289
 290 static void *vmd_alloc(struct device *dev, size_t size, dma_addr_t *addr,
 291                        gfp_t flag, unsigned long attrs)
 292 {
 293         return vmd_dma_ops(dev)->alloc(to_vmd_dev(dev), size, addr, flag,
 294                                        attrs);
 295 }
 296
 297 static void vmd_free(struct device *dev, size_t size, void *vaddr,
 298                      dma_addr_t addr, unsigned long attrs)
 299 {
 300         return vmd_dma_ops(dev)->free(to_vmd_dev(dev), size, vaddr, addr,
 301                                       attrs);
 302 }
 303
 304 static int vmd_mmap(struct device *dev, struct vm_area_struct *vma,
 305                     void *cpu_addr, dma_addr_t addr, size_t size,
 306                     unsigned long attrs)
 307 {
 308         return vmd_dma_ops(dev)->mmap(to_vmd_dev(dev), vma, cpu_addr, addr,
 309                                       size, attrs);
 310 }
 311
 312 static int vmd_get_sgtable(struct device *dev, struct sg_table *sgt,
 313                            void *cpu_addr, dma_addr_t addr, size_t size,
 314                            unsigned long attrs)
 315 {
 316         return vmd_dma_ops(dev)->get_sgtable(to_vmd_dev(dev), sgt, cpu_addr,
 317                                              addr, size, attrs);
 318 }
 319
 320 static dma_addr_t vmd_map_page(struct device *dev, struct page *page,
 321                                unsigned long offset, size_t size,
 322                                enum dma_data_direction dir,
 323                                unsigned long attrs)
 324 {
 325         return vmd_dma_ops(dev)->map_page(to_vmd_dev(dev), page, offset, size,
 326                                           dir, attrs);
 327 }
 328
 329 static void vmd_unmap_page(struct device *dev, dma_addr_t addr, size_t size,
 330                            enum dma_data_direction dir, unsigned long attrs)
 331 {
 332         vmd_dma_ops(dev)->unmap_page(to_vmd_dev(dev), addr, size, dir, attrs);
 333 }
 334
 335 static int vmd_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 336                       enum dma_data_direction dir, unsigned long attrs)
 337 {
 338         return vmd_dma_ops(dev)->map_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 339 }
 340
 341 static void vmd_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 342                          enum dma_data_direction dir, unsigned long attrs)
 343 {
 344         vmd_dma_ops(dev)->unmap_sg(to_vmd_dev(dev), sg, nents, dir, attrs);
 345 }
 346
 347 static void vmd_sync_single_for_cpu(struct device *dev, dma_addr_t addr,
 348                                     size_t size, enum dma_data_direction dir)
 349 {
 350         vmd_dma_ops(dev)->sync_single_for_cpu(to_vmd_dev(dev), addr, size, dir);
 351 }
 352
 353 static void vmd_sync_single_for_device(struct device *dev, dma_addr_t addr,
 354                                        size_t size, enum dma_data_direction dir)
 355 {
 356         vmd_dma_ops(dev)->sync_single_for_device(to_vmd_dev(dev), addr, size,
 357                                                  dir);
 358 }
 359
 360 static void vmd_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 361                                 int nents, enum dma_data_direction dir)
 362 {
 363         vmd_dma_ops(dev)->sync_sg_for_cpu(to_vmd_dev(dev), sg, nents, dir);
 364 }
 365
 366 static void vmd_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 367                                    int nents, enum dma_data_direction dir)
 368 {
 369         vmd_dma_ops(dev)->sync_sg_for_device(to_vmd_dev(dev), sg, nents, dir);
 370 }
 371
 372 static int vmd_mapping_error(struct device *dev, dma_addr_t addr)
 373 {
 374         return vmd_dma_ops(dev)->mapping_error(to_vmd_dev(dev), addr);
 375 }
 376
 377 static int vmd_dma_supported(struct device *dev, u64 mask)
 378 {
 379         return vmd_dma_ops(dev)->dma_supported(to_vmd_dev(dev), mask);
 380 }
 381
 382 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 383 static u64 vmd_get_required_mask(struct device *dev)
 384 {
 385         return vmd_dma_ops(dev)->get_required_mask(to_vmd_dev(dev));
 386 }
 387 #endif
 388
 389 static void vmd_teardown_dma_ops(struct vmd_dev *vmd)
 390 {
 391         struct dma_domain *domain = &vmd->dma_domain;
 392
 393         if (get_dma_ops(&vmd->dev->dev))
 394                 del_dma_domain(domain);
 395 }
 396
 397 #define ASSIGN_VMD_DMA_OPS(source, dest, fn)    \
 398         do {                                    \
 399                 if (source->fn)                 \
 400                         dest->fn = vmd_##fn;    \
 401         } while (0)
 402
 403 static void vmd_setup_dma_ops(struct vmd_dev *vmd)
 404 {
 405         const struct dma_map_ops *source = get_dma_ops(&vmd->dev->dev);
 406         struct dma_map_ops *dest = &vmd->dma_ops;
 407         struct dma_domain *domain = &vmd->dma_domain;
 408
 409         domain->domain_nr = vmd->sysdata.domain;
 410         domain->dma_ops = dest;
 411
 412         if (!source)
 413                 return;
 414         ASSIGN_VMD_DMA_OPS(source, dest, alloc);
 415         ASSIGN_VMD_DMA_OPS(source, dest, free);
 416         ASSIGN_VMD_DMA_OPS(source, dest, mmap);
 417         ASSIGN_VMD_DMA_OPS(source, dest, get_sgtable);
 418         ASSIGN_VMD_DMA_OPS(source, dest, map_page);
 419         ASSIGN_VMD_DMA_OPS(source, dest, unmap_page);
 420         ASSIGN_VMD_DMA_OPS(source, dest, map_sg);
 421         ASSIGN_VMD_DMA_OPS(source, dest, unmap_sg);
 422         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_cpu);
 423         ASSIGN_VMD_DMA_OPS(source, dest, sync_single_for_device);
 424         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_cpu);
 425         ASSIGN_VMD_DMA_OPS(source, dest, sync_sg_for_device);
 426         ASSIGN_VMD_DMA_OPS(source, dest, mapping_error);
 427         ASSIGN_VMD_DMA_OPS(source, dest, dma_supported);
 428 #ifdef ARCH_HAS_DMA_GET_REQUIRED_MASK
 429         ASSIGN_VMD_DMA_OPS(source, dest, get_required_mask);
 430 #endif
 431         add_dma_domain(domain);
 432 }
 433 #undef ASSIGN_VMD_DMA_OPS
 434 #else
 435 static void vmd_teardown_dma_ops(struct vmd_dev *vmd) {}
 436 static void vmd_setup_dma_ops(struct vmd_dev *vmd) {}
 437 #endif
 438
 439 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 440                                   unsigned int devfn, int reg, int len)
 441 {
 442         char __iomem *addr = vmd->cfgbar +
 443                              (bus->number << 20) + (devfn << 12) + reg;
 444
 445         if ((addr - vmd->cfgbar) + len >=
 446             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 447                 return NULL;
 448
 449         return addr;
 450 }
 451
 452 /*
 453  * CPU may deadlock if config space is not serialized on some versions of this
 454  * hardware, so all config space access is done under a spinlock.
 455  */
 456 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 457                         int len, u32 *value)
 458 {
 459         struct vmd_dev *vmd = vmd_from_bus(bus);
 460         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 461         unsigned long flags;
 462         int ret = 0;
 463
 464         if (!addr)
 465                 return -EFAULT;
 466
 467         spin_lock_irqsave(&vmd->cfg_lock, flags);
 468         switch (len) {
 469         case 1:
 470                 *value = readb(addr);
 471                 break;
 472         case 2:
 473                 *value = readw(addr);
 474                 break;
 475         case 4:
 476                 *value = readl(addr);
 477                 break;
 478         default:
 479                 ret = -EINVAL;
 480                 break;
 481         }
 482         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 483         return ret;
 484 }
 485
 486 /*
 487  * VMD h/w converts non-posted config writes to posted memory writes. The
 488  * read-back in this function forces the completion so it returns only after
 489  * the config space was written, as expected.
 490  */
 491 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 492                          int len, u32 value)
 493 {
 494         struct vmd_dev *vmd = vmd_from_bus(bus);
 495         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 496         unsigned long flags;
 497         int ret = 0;
 498
 499         if (!addr)
 500                 return -EFAULT;
 501
 502         spin_lock_irqsave(&vmd->cfg_lock, flags);
 503         switch (len) {
 504         case 1:
 505                 writeb(value, addr);
 506                 readb(addr);
 507                 break;
 508         case 2:
 509                 writew(value, addr);
 510                 readw(addr);
 511                 break;
 512         case 4:
 513                 writel(value, addr);
 514                 readl(addr);
 515                 break;
 516         default:
 517                 ret = -EINVAL;
 518                 break;
 519         }
 520         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 521         return ret;
 522 }
 523
 524 static struct pci_ops vmd_ops = {
 525         .read           = vmd_pci_read,
 526         .write          = vmd_pci_write,
 527 };
 528
 529 static void vmd_attach_resources(struct vmd_dev *vmd)
 530 {
 531         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 532         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 533 }
 534
 535 static void vmd_detach_resources(struct vmd_dev *vmd)
 536 {
 537         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 538         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 539 }
 540
 541 /*
 542  * VMD domains start at 0x1000 to not clash with ACPI _SEG domains.
 543  */
 544 static int vmd_find_free_domain(void)
 545 {
 546         int domain = 0xffff;
 547         struct pci_bus *bus = NULL;
 548
 549         while ((bus = pci_find_next_bus(bus)) != NULL)
 550                 domain = max_t(int, domain, pci_domain_nr(bus));
 551         return domain + 1;
 552 }
 553
 554 static int vmd_enable_domain(struct vmd_dev *vmd)
 555 {
 556         struct pci_sysdata *sd = &vmd->sysdata;
 557         struct resource *res;
 558         u32 upper_bits;
 559         unsigned long flags;
 560         LIST_HEAD(resources);
 561
 562         res = &vmd->dev->resource[VMD_CFGBAR];
 563         vmd->resources[0] = (struct resource) {
 564                 .name  = "VMD CFGBAR",
 565                 .start = 0,
 566                 .end   = (resource_size(res) >> 20) - 1,
 567                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 568         };
 569
 570         /*
 571          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 572          * put 32-bit resources in the window.
 573          *
 574          * There's no hardware reason why a 64-bit window *couldn't*
 575          * contain a 32-bit resource, but pbus_size_mem() computes the
 576          * bridge window size assuming a 64-bit window will contain no
 577          * 32-bit resources.  __pci_assign_resource() enforces that
 578          * artificial restriction to make sure everything will fit.
 579          *
 580          * The only way we could use a 64-bit non-prefechable MEMBAR is
 581          * if its address is <4GB so that we can convert it to a 32-bit
 582          * resource.  To be visible to the host OS, all VMD endpoints must
 583          * be initially configured by platform BIOS, which includes setting
 584          * up these resources.  We can assume the device is configured
 585          * according to the platform needs.
 586          */
 587         res = &vmd->dev->resource[VMD_MEMBAR1];
 588         upper_bits = upper_32_bits(res->end);
 589         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 590         if (!upper_bits)
 591                 flags &= ~IORESOURCE_MEM_64;
 592         vmd->resources[1] = (struct resource) {
 593                 .name  = "VMD MEMBAR1",
 594                 .start = res->start,
 595                 .end   = res->end,
 596                 .flags = flags,
 597                 .parent = res,
 598         };
 599
 600         res = &vmd->dev->resource[VMD_MEMBAR2];
 601         upper_bits = upper_32_bits(res->end);
 602         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 603         if (!upper_bits)
 604                 flags &= ~IORESOURCE_MEM_64;
 605         vmd->resources[2] = (struct resource) {
 606                 .name  = "VMD MEMBAR2",
 607                 .start = res->start + 0x2000,
 608                 .end   = res->end,
 609                 .flags = flags,
 610                 .parent = res,
 611         };
 612
 613         sd->vmd_domain = true;
 614         sd->domain = vmd_find_free_domain();
 615         if (sd->domain < 0)
 616                 return sd->domain;
 617
 618         sd->node = pcibus_to_node(vmd->dev->bus);
 619
 620         vmd->irq_domain = pci_msi_create_irq_domain(NULL, &vmd_msi_domain_info,
 621                                                     x86_vector_domain);
 622         if (!vmd->irq_domain)
 623                 return -ENODEV;
 624
 625         pci_add_resource(&resources, &vmd->resources[0]);
 626         pci_add_resource(&resources, &vmd->resources[1]);
 627         pci_add_resource(&resources, &vmd->resources[2]);
 628         vmd->bus = pci_create_root_bus(&vmd->dev->dev, 0, &vmd_ops, sd,
 629                                        &resources);
 630         if (!vmd->bus) {
 631                 pci_free_resource_list(&resources);
 632                 irq_domain_remove(vmd->irq_domain);
 633                 return -ENODEV;
 634         }
 635
 636         vmd_attach_resources(vmd);
 637         vmd_setup_dma_ops(vmd);
 638         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 639         pci_rescan_bus(vmd->bus);
 640
 641         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 642                                "domain"), "Can't create symlink to domain\n");
 643         return 0;
 644 }
 645
 646 static irqreturn_t vmd_irq(int irq, void *data)
 647 {
 648         struct vmd_irq_list *irqs = data;
 649         struct vmd_irq *vmdirq;
 650         int idx;
 651
 652         idx = srcu_read_lock(&irqs->srcu);
 653         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 654                 generic_handle_irq(vmdirq->virq);
 655         srcu_read_unlock(&irqs->srcu, idx);
 656
 657         return IRQ_HANDLED;
 658 }
 659
 660 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 661 {
 662         struct vmd_dev *vmd;
 663         int i, err;
 664
 665         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 666                 return -ENOMEM;
 667
 668         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 669         if (!vmd)
 670                 return -ENOMEM;
 671
 672         vmd->dev = dev;
 673         err = pcim_enable_device(dev);
 674         if (err < 0)
 675                 return err;
 676
 677         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 678         if (!vmd->cfgbar)
 679                 return -ENOMEM;
 680
 681         pci_set_master(dev);
 682         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 683             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 684                 return -ENODEV;
 685
 686         vmd->msix_count = pci_msix_vec_count(dev);
 687         if (vmd->msix_count < 0)
 688                 return -ENODEV;
 689
 690         vmd->msix_count = pci_alloc_irq_vectors(dev, 1, vmd->msix_count,
 691                                         PCI_IRQ_MSIX | PCI_IRQ_AFFINITY);
 692         if (vmd->msix_count < 0)
 693                 return vmd->msix_count;
 694
 695         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 696                                  GFP_KERNEL);
 697         if (!vmd->irqs)
 698                 return -ENOMEM;
 699
 700         for (i = 0; i < vmd->msix_count; i++) {
 701                 err = init_srcu_struct(&vmd->irqs[i].srcu);
 702                 if (err)
 703                         return err;
 704
 705                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 706                 err = devm_request_irq(&dev->dev, pci_irq_vector(dev, i),
 707                                        vmd_irq, IRQF_NO_THREAD,
 708                                        "vmd", &vmd->irqs[i]);
 709                 if (err)
 710                         return err;
 711         }
 712
 713         spin_lock_init(&vmd->cfg_lock);
 714         pci_set_drvdata(dev, vmd);
 715         err = vmd_enable_domain(vmd);
 716         if (err)
 717                 return err;
 718
 719         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 720                  vmd->sysdata.domain);
 721         return 0;
 722 }
 723
 724 static void vmd_cleanup_srcu(struct vmd_dev *vmd)
 725 {
 726         int i;
 727
 728         for (i = 0; i < vmd->msix_count; i++)
 729                 cleanup_srcu_struct(&vmd->irqs[i].srcu);
 730 }
 731
 732 static void vmd_remove(struct pci_dev *dev)
 733 {
 734         struct vmd_dev *vmd = pci_get_drvdata(dev);
 735
 736         vmd_detach_resources(vmd);
 737         vmd_cleanup_srcu(vmd);
 738         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 739         pci_stop_root_bus(vmd->bus);
 740         pci_remove_root_bus(vmd->bus);
 741         vmd_teardown_dma_ops(vmd);
 742         irq_domain_remove(vmd->irq_domain);
 743 }
 744
 745 #ifdef CONFIG_PM_SLEEP
 746 static int vmd_suspend(struct device *dev)
 747 {
 748         struct pci_dev *pdev = to_pci_dev(dev);
 749
 750         pci_save_state(pdev);
 751         return 0;
 752 }
 753
 754 static int vmd_resume(struct device *dev)
 755 {
 756         struct pci_dev *pdev = to_pci_dev(dev);
 757
 758         pci_restore_state(pdev);
 759         return 0;
 760 }
 761 #endif
 762 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 763
 764 static const struct pci_device_id vmd_ids[] = {
 765         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x201d),},
 766         {0,}
 767 };
 768 MODULE_DEVICE_TABLE(pci, vmd_ids);
 769
 770 static struct pci_driver vmd_drv = {
 771         .name           = "vmd",
 772         .id_table       = vmd_ids,
 773         .probe          = vmd_probe,
 774         .remove         = vmd_remove,
 775         .driver         = {
 776                 .pm     = &vmd_dev_pm_ops,
 777         },
 778 };
 779 module_pci_driver(vmd_drv);
 780
 781 MODULE_AUTHOR("Intel Corporation");
 782 MODULE_LICENSE("GPL v2");
 783 MODULE_VERSION("0.6");