drivers/pci/controller/vmd.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Volume Management Device driver
   4  * Copyright (c) 2015, Intel Corporation.
   5  */
   6
   7 #include <linux/device.h>
   8 #include <linux/interrupt.h>
   9 #include <linux/irq.h>
  10 #include <linux/kernel.h>
  11 #include <linux/module.h>
  12 #include <linux/msi.h>
  13 #include <linux/pci.h>
  14 #include <linux/srcu.h>
  15 #include <linux/rculist.h>
  16 #include <linux/rcupdate.h>
  17
  18 #include <asm/irqdomain.h>
  19 #include <asm/device.h>
  20 #include <asm/msi.h>
  21 #include <asm/msidef.h>
  22
  23 #define VMD_CFGBAR      0
  24 #define VMD_MEMBAR1     2
  25 #define VMD_MEMBAR2     4
  26
  27 #define PCI_REG_VMCAP           0x40
  28 #define BUS_RESTRICT_CAP(vmcap) (vmcap & 0x1)
  29 #define PCI_REG_VMCONFIG        0x44
  30 #define BUS_RESTRICT_CFG(vmcfg) ((vmcfg >> 8) & 0x3)
  31 #define PCI_REG_VMLOCK          0x70
  32 #define MB2_SHADOW_EN(vmlock)   (vmlock & 0x2)
  33
  34 #define MB2_SHADOW_OFFSET       0x2000
  35 #define MB2_SHADOW_SIZE         16
  36
  37 enum vmd_features {
  38         /*
  39          * Device may contain registers which hint the physical location of the
  40          * membars, in order to allow proper address translation during
  41          * resource assignment to enable guest virtualization
  42          */
  43         VMD_FEAT_HAS_MEMBAR_SHADOW              = (1 << 0),
  44
  45         /*
  46          * Device may provide root port configuration information which limits
  47          * bus numbering
  48          */
  49         VMD_FEAT_HAS_BUS_RESTRICTIONS           = (1 << 1),
  50
  51         /*
  52          * Device contains physical location shadow registers in
  53          * vendor-specific capability space
  54          */
  55         VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP        = (1 << 2),
  56 };
  57
  58 /*
  59  * Lock for manipulating VMD IRQ lists.
  60  */
  61 static DEFINE_RAW_SPINLOCK(list_lock);
  62
  63 /**
  64  * struct vmd_irq - private data to map driver IRQ to the VMD shared vector
  65  * @node:       list item for parent traversal.
  66  * @irq:        back pointer to parent.
  67  * @enabled:    true if driver enabled IRQ
  68  * @virq:       the virtual IRQ value provided to the requesting driver.
  69  *
  70  * Every MSI/MSI-X IRQ requested for a device in a VMD domain will be mapped to
  71  * a VMD IRQ using this structure.
  72  */
  73 struct vmd_irq {
  74         struct list_head        node;
  75         struct vmd_irq_list     *irq;
  76         bool                    enabled;
  77         unsigned int            virq;
  78 };
  79
  80 /**
  81  * struct vmd_irq_list - list of driver requested IRQs mapping to a VMD vector
  82  * @irq_list:   the list of irq's the VMD one demuxes to.
  83  * @srcu:       SRCU struct for local synchronization.
  84  * @count:      number of child IRQs assigned to this vector; used to track
  85  *              sharing.
  86  */
  87 struct vmd_irq_list {
  88         struct list_head        irq_list;
  89         struct srcu_struct      srcu;
  90         unsigned int            count;
  91 };
  92
  93 struct vmd_dev {
  94         struct pci_dev          *dev;
  95
  96         spinlock_t              cfg_lock;
  97         char __iomem            *cfgbar;
  98
  99         int msix_count;
 100         struct vmd_irq_list     *irqs;
 101
 102         struct pci_sysdata      sysdata;
 103         struct resource         resources[3];
 104         struct irq_domain       *irq_domain;
 105         struct pci_bus          *bus;
 106         u8                      busn_start;
 107 };
 108
 109 static inline struct vmd_dev *vmd_from_bus(struct pci_bus *bus)
 110 {
 111         return container_of(bus->sysdata, struct vmd_dev, sysdata);
 112 }
 113
 114 static inline unsigned int index_from_irqs(struct vmd_dev *vmd,
 115                                            struct vmd_irq_list *irqs)
 116 {
 117         return irqs - vmd->irqs;
 118 }
 119
 120 /*
 121  * Drivers managing a device in a VMD domain allocate their own IRQs as before,
 122  * but the MSI entry for the hardware it's driving will be programmed with a
 123  * destination ID for the VMD MSI-X table.  The VMD muxes interrupts in its
 124  * domain into one of its own, and the VMD driver de-muxes these for the
 125  * handlers sharing that VMD IRQ.  The vmd irq_domain provides the operations
 126  * and irq_chip to set this up.
 127  */
 128 static void vmd_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
 129 {
 130         struct vmd_irq *vmdirq = data->chip_data;
 131         struct vmd_irq_list *irq = vmdirq->irq;
 132         struct vmd_dev *vmd = irq_data_get_irq_handler_data(data);
 133
 134         msg->address_hi = MSI_ADDR_BASE_HI;
 135         msg->address_lo = MSI_ADDR_BASE_LO |
 136                           MSI_ADDR_DEST_ID(index_from_irqs(vmd, irq));
 137         msg->data = 0;
 138 }
 139
 140 /*
 141  * We rely on MSI_FLAG_USE_DEF_CHIP_OPS to set the IRQ mask/unmask ops.
 142  */
 143 static void vmd_irq_enable(struct irq_data *data)
 144 {
 145         struct vmd_irq *vmdirq = data->chip_data;
 146         unsigned long flags;
 147
 148         raw_spin_lock_irqsave(&list_lock, flags);
 149         WARN_ON(vmdirq->enabled);
 150         list_add_tail_rcu(&vmdirq->node, &vmdirq->irq->irq_list);
 151         vmdirq->enabled = true;
 152         raw_spin_unlock_irqrestore(&list_lock, flags);
 153
 154         data->chip->irq_unmask(data);
 155 }
 156
 157 static void vmd_irq_disable(struct irq_data *data)
 158 {
 159         struct vmd_irq *vmdirq = data->chip_data;
 160         unsigned long flags;
 161
 162         data->chip->irq_mask(data);
 163
 164         raw_spin_lock_irqsave(&list_lock, flags);
 165         if (vmdirq->enabled) {
 166                 list_del_rcu(&vmdirq->node);
 167                 vmdirq->enabled = false;
 168         }
 169         raw_spin_unlock_irqrestore(&list_lock, flags);
 170 }
 171
 172 /*
 173  * XXX: Stubbed until we develop acceptable way to not create conflicts with
 174  * other devices sharing the same vector.
 175  */
 176 static int vmd_irq_set_affinity(struct irq_data *data,
 177                                 const struct cpumask *dest, bool force)
 178 {
 179         return -EINVAL;
 180 }
 181
 182 static struct irq_chip vmd_msi_controller = {
 183         .name                   = "VMD-MSI",
 184         .irq_enable             = vmd_irq_enable,
 185         .irq_disable            = vmd_irq_disable,
 186         .irq_compose_msi_msg    = vmd_compose_msi_msg,
 187         .irq_set_affinity       = vmd_irq_set_affinity,
 188 };
 189
 190 static irq_hw_number_t vmd_get_hwirq(struct msi_domain_info *info,
 191                                      msi_alloc_info_t *arg)
 192 {
 193         return 0;
 194 }
 195
 196 /*
 197  * XXX: We can be even smarter selecting the best IRQ once we solve the
 198  * affinity problem.
 199  */
 200 static struct vmd_irq_list *vmd_next_irq(struct vmd_dev *vmd, struct msi_desc *desc)
 201 {
 202         int i, best = 1;
 203         unsigned long flags;
 204
 205         if (vmd->msix_count == 1)
 206                 return &vmd->irqs[0];
 207
 208         /*
 209          * White list for fast-interrupt handlers. All others will share the
 210          * "slow" interrupt vector.
 211          */
 212         switch (msi_desc_to_pci_dev(desc)->class) {
 213         case PCI_CLASS_STORAGE_EXPRESS:
 214                 break;
 215         default:
 216                 return &vmd->irqs[0];
 217         }
 218
 219         raw_spin_lock_irqsave(&list_lock, flags);
 220         for (i = 1; i < vmd->msix_count; i++)
 221                 if (vmd->irqs[i].count < vmd->irqs[best].count)
 222                         best = i;
 223         vmd->irqs[best].count++;
 224         raw_spin_unlock_irqrestore(&list_lock, flags);
 225
 226         return &vmd->irqs[best];
 227 }
 228
 229 static int vmd_msi_init(struct irq_domain *domain, struct msi_domain_info *info,
 230                         unsigned int virq, irq_hw_number_t hwirq,
 231                         msi_alloc_info_t *arg)
 232 {
 233         struct msi_desc *desc = arg->desc;
 234         struct vmd_dev *vmd = vmd_from_bus(msi_desc_to_pci_dev(desc)->bus);
 235         struct vmd_irq *vmdirq = kzalloc(sizeof(*vmdirq), GFP_KERNEL);
 236         unsigned int index, vector;
 237
 238         if (!vmdirq)
 239                 return -ENOMEM;
 240
 241         INIT_LIST_HEAD(&vmdirq->node);
 242         vmdirq->irq = vmd_next_irq(vmd, desc);
 243         vmdirq->virq = virq;
 244         index = index_from_irqs(vmd, vmdirq->irq);
 245         vector = pci_irq_vector(vmd->dev, index);
 246
 247         irq_domain_set_info(domain, virq, vector, info->chip, vmdirq,
 248                             handle_untracked_irq, vmd, NULL);
 249         return 0;
 250 }
 251
 252 static void vmd_msi_free(struct irq_domain *domain,
 253                         struct msi_domain_info *info, unsigned int virq)
 254 {
 255         struct vmd_irq *vmdirq = irq_get_chip_data(virq);
 256         unsigned long flags;
 257
 258         synchronize_srcu(&vmdirq->irq->srcu);
 259
 260         /* XXX: Potential optimization to rebalance */
 261         raw_spin_lock_irqsave(&list_lock, flags);
 262         vmdirq->irq->count--;
 263         raw_spin_unlock_irqrestore(&list_lock, flags);
 264
 265         kfree(vmdirq);
 266 }
 267
 268 static int vmd_msi_prepare(struct irq_domain *domain, struct device *dev,
 269                            int nvec, msi_alloc_info_t *arg)
 270 {
 271         struct pci_dev *pdev = to_pci_dev(dev);
 272         struct vmd_dev *vmd = vmd_from_bus(pdev->bus);
 273
 274         if (nvec > vmd->msix_count)
 275                 return vmd->msix_count;
 276
 277         memset(arg, 0, sizeof(*arg));
 278         return 0;
 279 }
 280
 281 static void vmd_set_desc(msi_alloc_info_t *arg, struct msi_desc *desc)
 282 {
 283         arg->desc = desc;
 284 }
 285
 286 static struct msi_domain_ops vmd_msi_domain_ops = {
 287         .get_hwirq      = vmd_get_hwirq,
 288         .msi_init       = vmd_msi_init,
 289         .msi_free       = vmd_msi_free,
 290         .msi_prepare    = vmd_msi_prepare,
 291         .set_desc       = vmd_set_desc,
 292 };
 293
 294 static struct msi_domain_info vmd_msi_domain_info = {
 295         .flags          = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
 296                           MSI_FLAG_PCI_MSIX,
 297         .ops            = &vmd_msi_domain_ops,
 298         .chip           = &vmd_msi_controller,
 299 };
 300
 301 static char __iomem *vmd_cfg_addr(struct vmd_dev *vmd, struct pci_bus *bus,
 302                                   unsigned int devfn, int reg, int len)
 303 {
 304         char __iomem *addr = vmd->cfgbar +
 305                              ((bus->number - vmd->busn_start) << 20) +
 306                              (devfn << 12) + reg;
 307
 308         if ((addr - vmd->cfgbar) + len >=
 309             resource_size(&vmd->dev->resource[VMD_CFGBAR]))
 310                 return NULL;
 311
 312         return addr;
 313 }
 314
 315 /*
 316  * CPU may deadlock if config space is not serialized on some versions of this
 317  * hardware, so all config space access is done under a spinlock.
 318  */
 319 static int vmd_pci_read(struct pci_bus *bus, unsigned int devfn, int reg,
 320                         int len, u32 *value)
 321 {
 322         struct vmd_dev *vmd = vmd_from_bus(bus);
 323         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 324         unsigned long flags;
 325         int ret = 0;
 326
 327         if (!addr)
 328                 return -EFAULT;
 329
 330         spin_lock_irqsave(&vmd->cfg_lock, flags);
 331         switch (len) {
 332         case 1:
 333                 *value = readb(addr);
 334                 break;
 335         case 2:
 336                 *value = readw(addr);
 337                 break;
 338         case 4:
 339                 *value = readl(addr);
 340                 break;
 341         default:
 342                 ret = -EINVAL;
 343                 break;
 344         }
 345         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 346         return ret;
 347 }
 348
 349 /*
 350  * VMD h/w converts non-posted config writes to posted memory writes. The
 351  * read-back in this function forces the completion so it returns only after
 352  * the config space was written, as expected.
 353  */
 354 static int vmd_pci_write(struct pci_bus *bus, unsigned int devfn, int reg,
 355                          int len, u32 value)
 356 {
 357         struct vmd_dev *vmd = vmd_from_bus(bus);
 358         char __iomem *addr = vmd_cfg_addr(vmd, bus, devfn, reg, len);
 359         unsigned long flags;
 360         int ret = 0;
 361
 362         if (!addr)
 363                 return -EFAULT;
 364
 365         spin_lock_irqsave(&vmd->cfg_lock, flags);
 366         switch (len) {
 367         case 1:
 368                 writeb(value, addr);
 369                 readb(addr);
 370                 break;
 371         case 2:
 372                 writew(value, addr);
 373                 readw(addr);
 374                 break;
 375         case 4:
 376                 writel(value, addr);
 377                 readl(addr);
 378                 break;
 379         default:
 380                 ret = -EINVAL;
 381                 break;
 382         }
 383         spin_unlock_irqrestore(&vmd->cfg_lock, flags);
 384         return ret;
 385 }
 386
 387 static struct pci_ops vmd_ops = {
 388         .read           = vmd_pci_read,
 389         .write          = vmd_pci_write,
 390 };
 391
 392 static void vmd_attach_resources(struct vmd_dev *vmd)
 393 {
 394         vmd->dev->resource[VMD_MEMBAR1].child = &vmd->resources[1];
 395         vmd->dev->resource[VMD_MEMBAR2].child = &vmd->resources[2];
 396 }
 397
 398 static void vmd_detach_resources(struct vmd_dev *vmd)
 399 {
 400         vmd->dev->resource[VMD_MEMBAR1].child = NULL;
 401         vmd->dev->resource[VMD_MEMBAR2].child = NULL;
 402 }
 403
 404 /*
 405  * VMD domains start at 0x10000 to not clash with ACPI _SEG domains.
 406  * Per ACPI r6.0, sec 6.5.6,  _SEG returns an integer, of which the lower
 407  * 16 bits are the PCI Segment Group (domain) number.  Other bits are
 408  * currently reserved.
 409  */
 410 static int vmd_find_free_domain(void)
 411 {
 412         int domain = 0xffff;
 413         struct pci_bus *bus = NULL;
 414
 415         while ((bus = pci_find_next_bus(bus)) != NULL)
 416                 domain = max_t(int, domain, pci_domain_nr(bus));
 417         return domain + 1;
 418 }
 419
 420 static int vmd_enable_domain(struct vmd_dev *vmd, unsigned long features)
 421 {
 422         struct pci_sysdata *sd = &vmd->sysdata;
 423         struct fwnode_handle *fn;
 424         struct resource *res;
 425         u32 upper_bits;
 426         unsigned long flags;
 427         LIST_HEAD(resources);
 428         resource_size_t offset[2] = {0};
 429         resource_size_t membar2_offset = 0x2000;
 430         struct pci_bus *child;
 431
 432         /*
 433          * Shadow registers may exist in certain VMD device ids which allow
 434          * guests to correctly assign host physical addresses to the root ports
 435          * and child devices. These registers will either return the host value
 436          * or 0, depending on an enable bit in the VMD device.
 437          */
 438         if (features & VMD_FEAT_HAS_MEMBAR_SHADOW) {
 439                 u32 vmlock;
 440                 int ret;
 441
 442                 membar2_offset = MB2_SHADOW_OFFSET + MB2_SHADOW_SIZE;
 443                 ret = pci_read_config_dword(vmd->dev, PCI_REG_VMLOCK, &vmlock);
 444                 if (ret || vmlock == ~0)
 445                         return -ENODEV;
 446
 447                 if (MB2_SHADOW_EN(vmlock)) {
 448                         void __iomem *membar2;
 449
 450                         membar2 = pci_iomap(vmd->dev, VMD_MEMBAR2, 0);
 451                         if (!membar2)
 452                                 return -ENOMEM;
 453                         offset[0] = vmd->dev->resource[VMD_MEMBAR1].start -
 454                                         (readq(membar2 + MB2_SHADOW_OFFSET) &
 455                                          PCI_BASE_ADDRESS_MEM_MASK);
 456                         offset[1] = vmd->dev->resource[VMD_MEMBAR2].start -
 457                                         (readq(membar2 + MB2_SHADOW_OFFSET + 8) &
 458                                          PCI_BASE_ADDRESS_MEM_MASK);
 459                         pci_iounmap(vmd->dev, membar2);
 460                 }
 461         }
 462
 463         if (features & VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP) {
 464                 int pos = pci_find_capability(vmd->dev, PCI_CAP_ID_VNDR);
 465                 u32 reg, regu;
 466
 467                 pci_read_config_dword(vmd->dev, pos + 4, &reg);
 468
 469                 /* "SHDW" */
 470                 if (pos && reg == 0x53484457) {
 471                         pci_read_config_dword(vmd->dev, pos + 8, &reg);
 472                         pci_read_config_dword(vmd->dev, pos + 12, &regu);
 473                         offset[0] = vmd->dev->resource[VMD_MEMBAR1].start -
 474                                         (((u64) regu << 32 | reg) &
 475                                          PCI_BASE_ADDRESS_MEM_MASK);
 476
 477                         pci_read_config_dword(vmd->dev, pos + 16, &reg);
 478                         pci_read_config_dword(vmd->dev, pos + 20, &regu);
 479                         offset[1] = vmd->dev->resource[VMD_MEMBAR2].start -
 480                                         (((u64) regu << 32 | reg) &
 481                                          PCI_BASE_ADDRESS_MEM_MASK);
 482                 }
 483         }
 484
 485         /*
 486          * Certain VMD devices may have a root port configuration option which
 487          * limits the bus range to between 0-127, 128-255, or 224-255
 488          */
 489         if (features & VMD_FEAT_HAS_BUS_RESTRICTIONS) {
 490                 u16 reg16;
 491
 492                 pci_read_config_word(vmd->dev, PCI_REG_VMCAP, &reg16);
 493                 if (BUS_RESTRICT_CAP(reg16)) {
 494                         pci_read_config_word(vmd->dev, PCI_REG_VMCONFIG,
 495                                              &reg16);
 496
 497                         switch (BUS_RESTRICT_CFG(reg16)) {
 498                         case 1:
 499                                 vmd->busn_start = 128;
 500                                 break;
 501                         case 2:
 502                                 vmd->busn_start = 224;
 503                                 break;
 504                         case 3:
 505                                 pci_err(vmd->dev, "Unknown Bus Offset Setting\n");
 506                                 return -ENODEV;
 507                         default:
 508                                 break;
 509                         }
 510                 }
 511         }
 512
 513         res = &vmd->dev->resource[VMD_CFGBAR];
 514         vmd->resources[0] = (struct resource) {
 515                 .name  = "VMD CFGBAR",
 516                 .start = vmd->busn_start,
 517                 .end   = vmd->busn_start + (resource_size(res) >> 20) - 1,
 518                 .flags = IORESOURCE_BUS | IORESOURCE_PCI_FIXED,
 519         };
 520
 521         /*
 522          * If the window is below 4GB, clear IORESOURCE_MEM_64 so we can
 523          * put 32-bit resources in the window.
 524          *
 525          * There's no hardware reason why a 64-bit window *couldn't*
 526          * contain a 32-bit resource, but pbus_size_mem() computes the
 527          * bridge window size assuming a 64-bit window will contain no
 528          * 32-bit resources.  __pci_assign_resource() enforces that
 529          * artificial restriction to make sure everything will fit.
 530          *
 531          * The only way we could use a 64-bit non-prefetchable MEMBAR is
 532          * if its address is <4GB so that we can convert it to a 32-bit
 533          * resource.  To be visible to the host OS, all VMD endpoints must
 534          * be initially configured by platform BIOS, which includes setting
 535          * up these resources.  We can assume the device is configured
 536          * according to the platform needs.
 537          */
 538         res = &vmd->dev->resource[VMD_MEMBAR1];
 539         upper_bits = upper_32_bits(res->end);
 540         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 541         if (!upper_bits)
 542                 flags &= ~IORESOURCE_MEM_64;
 543         vmd->resources[1] = (struct resource) {
 544                 .name  = "VMD MEMBAR1",
 545                 .start = res->start,
 546                 .end   = res->end,
 547                 .flags = flags,
 548                 .parent = res,
 549         };
 550
 551         res = &vmd->dev->resource[VMD_MEMBAR2];
 552         upper_bits = upper_32_bits(res->end);
 553         flags = res->flags & ~IORESOURCE_SIZEALIGN;
 554         if (!upper_bits)
 555                 flags &= ~IORESOURCE_MEM_64;
 556         vmd->resources[2] = (struct resource) {
 557                 .name  = "VMD MEMBAR2",
 558                 .start = res->start + membar2_offset,
 559                 .end   = res->end,
 560                 .flags = flags,
 561                 .parent = res,
 562         };
 563
 564         sd->vmd_dev = vmd->dev;
 565         sd->domain = vmd_find_free_domain();
 566         if (sd->domain < 0)
 567                 return sd->domain;
 568
 569         sd->node = pcibus_to_node(vmd->dev->bus);
 570
 571         fn = irq_domain_alloc_named_id_fwnode("VMD-MSI", vmd->sysdata.domain);
 572         if (!fn)
 573                 return -ENODEV;
 574
 575         vmd->irq_domain = pci_msi_create_irq_domain(fn, &vmd_msi_domain_info,
 576                                                     NULL);
 577
 578         if (!vmd->irq_domain) {
 579                 irq_domain_free_fwnode(fn);
 580                 return -ENODEV;
 581         }
 582
 583         /*
 584          * Override the irq domain bus token so the domain can be distinguished
 585          * from a regular PCI/MSI domain.
 586          */
 587         irq_domain_update_bus_token(vmd->irq_domain, DOMAIN_BUS_VMD_MSI);
 588
 589         pci_add_resource(&resources, &vmd->resources[0]);
 590         pci_add_resource_offset(&resources, &vmd->resources[1], offset[0]);
 591         pci_add_resource_offset(&resources, &vmd->resources[2], offset[1]);
 592
 593         vmd->bus = pci_create_root_bus(&vmd->dev->dev, vmd->busn_start,
 594                                        &vmd_ops, sd, &resources);
 595         if (!vmd->bus) {
 596                 pci_free_resource_list(&resources);
 597                 irq_domain_remove(vmd->irq_domain);
 598                 irq_domain_free_fwnode(fn);
 599                 return -ENODEV;
 600         }
 601
 602         vmd_attach_resources(vmd);
 603         dev_set_msi_domain(&vmd->bus->dev, vmd->irq_domain);
 604
 605         pci_scan_child_bus(vmd->bus);
 606         pci_assign_unassigned_bus_resources(vmd->bus);
 607
 608         /*
 609          * VMD root buses are virtual and don't return true on pci_is_pcie()
 610          * and will fail pcie_bus_configure_settings() early. It can instead be
 611          * run on each of the real root ports.
 612          */
 613         list_for_each_entry(child, &vmd->bus->children, node)
 614                 pcie_bus_configure_settings(child);
 615
 616         pci_bus_add_devices(vmd->bus);
 617
 618         WARN(sysfs_create_link(&vmd->dev->dev.kobj, &vmd->bus->dev.kobj,
 619                                "domain"), "Can't create symlink to domain\n");
 620         return 0;
 621 }
 622
 623 static irqreturn_t vmd_irq(int irq, void *data)
 624 {
 625         struct vmd_irq_list *irqs = data;
 626         struct vmd_irq *vmdirq;
 627         int idx;
 628
 629         idx = srcu_read_lock(&irqs->srcu);
 630         list_for_each_entry_rcu(vmdirq, &irqs->irq_list, node)
 631                 generic_handle_irq(vmdirq->virq);
 632         srcu_read_unlock(&irqs->srcu, idx);
 633
 634         return IRQ_HANDLED;
 635 }
 636
 637 static int vmd_probe(struct pci_dev *dev, const struct pci_device_id *id)
 638 {
 639         struct vmd_dev *vmd;
 640         int i, err;
 641
 642         if (resource_size(&dev->resource[VMD_CFGBAR]) < (1 << 20))
 643                 return -ENOMEM;
 644
 645         vmd = devm_kzalloc(&dev->dev, sizeof(*vmd), GFP_KERNEL);
 646         if (!vmd)
 647                 return -ENOMEM;
 648
 649         vmd->dev = dev;
 650         err = pcim_enable_device(dev);
 651         if (err < 0)
 652                 return err;
 653
 654         vmd->cfgbar = pcim_iomap(dev, VMD_CFGBAR, 0);
 655         if (!vmd->cfgbar)
 656                 return -ENOMEM;
 657
 658         pci_set_master(dev);
 659         if (dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(64)) &&
 660             dma_set_mask_and_coherent(&dev->dev, DMA_BIT_MASK(32)))
 661                 return -ENODEV;
 662
 663         vmd->msix_count = pci_msix_vec_count(dev);
 664         if (vmd->msix_count < 0)
 665                 return -ENODEV;
 666
 667         vmd->msix_count = pci_alloc_irq_vectors(dev, 1, vmd->msix_count,
 668                                         PCI_IRQ_MSIX);
 669         if (vmd->msix_count < 0)
 670                 return vmd->msix_count;
 671
 672         vmd->irqs = devm_kcalloc(&dev->dev, vmd->msix_count, sizeof(*vmd->irqs),
 673                                  GFP_KERNEL);
 674         if (!vmd->irqs)
 675                 return -ENOMEM;
 676
 677         for (i = 0; i < vmd->msix_count; i++) {
 678                 err = init_srcu_struct(&vmd->irqs[i].srcu);
 679                 if (err)
 680                         return err;
 681
 682                 INIT_LIST_HEAD(&vmd->irqs[i].irq_list);
 683                 err = devm_request_irq(&dev->dev, pci_irq_vector(dev, i),
 684                                        vmd_irq, IRQF_NO_THREAD,
 685                                        "vmd", &vmd->irqs[i]);
 686                 if (err)
 687                         return err;
 688         }
 689
 690         spin_lock_init(&vmd->cfg_lock);
 691         pci_set_drvdata(dev, vmd);
 692         err = vmd_enable_domain(vmd, (unsigned long) id->driver_data);
 693         if (err)
 694                 return err;
 695
 696         dev_info(&vmd->dev->dev, "Bound to PCI domain %04x\n",
 697                  vmd->sysdata.domain);
 698         return 0;
 699 }
 700
 701 static void vmd_cleanup_srcu(struct vmd_dev *vmd)
 702 {
 703         int i;
 704
 705         for (i = 0; i < vmd->msix_count; i++)
 706                 cleanup_srcu_struct(&vmd->irqs[i].srcu);
 707 }
 708
 709 static void vmd_remove(struct pci_dev *dev)
 710 {
 711         struct vmd_dev *vmd = pci_get_drvdata(dev);
 712         struct fwnode_handle *fn = vmd->irq_domain->fwnode;
 713
 714         sysfs_remove_link(&vmd->dev->dev.kobj, "domain");
 715         pci_stop_root_bus(vmd->bus);
 716         pci_remove_root_bus(vmd->bus);
 717         vmd_cleanup_srcu(vmd);
 718         vmd_detach_resources(vmd);
 719         irq_domain_remove(vmd->irq_domain);
 720         irq_domain_free_fwnode(fn);
 721 }
 722
 723 #ifdef CONFIG_PM_SLEEP
 724 static int vmd_suspend(struct device *dev)
 725 {
 726         struct pci_dev *pdev = to_pci_dev(dev);
 727         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 728         int i;
 729
 730         for (i = 0; i < vmd->msix_count; i++)
 731                 devm_free_irq(dev, pci_irq_vector(pdev, i), &vmd->irqs[i]);
 732
 733         pci_save_state(pdev);
 734         return 0;
 735 }
 736
 737 static int vmd_resume(struct device *dev)
 738 {
 739         struct pci_dev *pdev = to_pci_dev(dev);
 740         struct vmd_dev *vmd = pci_get_drvdata(pdev);
 741         int err, i;
 742
 743         for (i = 0; i < vmd->msix_count; i++) {
 744                 err = devm_request_irq(dev, pci_irq_vector(pdev, i),
 745                                        vmd_irq, IRQF_NO_THREAD,
 746                                        "vmd", &vmd->irqs[i]);
 747                 if (err)
 748                         return err;
 749         }
 750
 751         pci_restore_state(pdev);
 752         return 0;
 753 }
 754 #endif
 755 static SIMPLE_DEV_PM_OPS(vmd_dev_pm_ops, vmd_suspend, vmd_resume);
 756
 757 static const struct pci_device_id vmd_ids[] = {
 758         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_201D),
 759                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP,},
 760         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_28C0),
 761                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW |
 762                                 VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 763         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x467f),
 764                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 765                                 VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 766         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x4c3d),
 767                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 768                                 VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 769         {PCI_DEVICE(PCI_VENDOR_ID_INTEL, PCI_DEVICE_ID_INTEL_VMD_9A0B),
 770                 .driver_data = VMD_FEAT_HAS_MEMBAR_SHADOW_VSCAP |
 771                                 VMD_FEAT_HAS_BUS_RESTRICTIONS,},
 772         {0,}
 773 };
 774 MODULE_DEVICE_TABLE(pci, vmd_ids);
 775
 776 static struct pci_driver vmd_drv = {
 777         .name           = "vmd",
 778         .id_table       = vmd_ids,
 779         .probe          = vmd_probe,
 780         .remove         = vmd_remove,
 781         .driver         = {
 782                 .pm     = &vmd_dev_pm_ops,
 783         },
 784 };
 785 module_pci_driver(vmd_drv);
 786
 787 MODULE_AUTHOR("Intel Corporation");
 788 MODULE_LICENSE("GPL v2");
 789 MODULE_VERSION("0.6");