drivers/iommu/s390-iommu.c

   1 /*
   2  * IOMMU API for s390 PCI devices
   3  *
   4  * Copyright IBM Corp. 2015
   5  * Author(s): Gerald Schaefer <gerald.schaefer@de.ibm.com>
   6  */
   7
   8 #include <linux/pci.h>
   9 #include <linux/iommu.h>
  10 #include <linux/iommu-helper.h>
  11 #include <linux/sizes.h>
  12 #include <asm/pci_dma.h>
  13
  14 /*
  15  * Physically contiguous memory regions can be mapped with 4 KiB alignment,
  16  * we allow all page sizes that are an order of 4KiB (no special large page
  17  * support so far).
  18  */
  19 #define S390_IOMMU_PGSIZES      (~0xFFFUL)
  20
  21 static const struct iommu_ops s390_iommu_ops;
  22
  23 struct s390_domain {
  24         struct iommu_domain     domain;
  25         struct list_head        devices;
  26         unsigned long           *dma_table;
  27         spinlock_t              dma_table_lock;
  28         spinlock_t              list_lock;
  29 };
  30
  31 struct s390_domain_device {
  32         struct list_head        list;
  33         struct zpci_dev         *zdev;
  34 };
  35
  36 static struct s390_domain *to_s390_domain(struct iommu_domain *dom)
  37 {
  38         return container_of(dom, struct s390_domain, domain);
  39 }
  40
  41 static bool s390_iommu_capable(enum iommu_cap cap)
  42 {
  43         switch (cap) {
  44         case IOMMU_CAP_CACHE_COHERENCY:
  45                 return true;
  46         case IOMMU_CAP_INTR_REMAP:
  47                 return true;
  48         default:
  49                 return false;
  50         }
  51 }
  52
  53 static struct iommu_domain *s390_domain_alloc(unsigned domain_type)
  54 {
  55         struct s390_domain *s390_domain;
  56
  57         if (domain_type != IOMMU_DOMAIN_UNMANAGED)
  58                 return NULL;
  59
  60         s390_domain = kzalloc(sizeof(*s390_domain), GFP_KERNEL);
  61         if (!s390_domain)
  62                 return NULL;
  63
  64         s390_domain->dma_table = dma_alloc_cpu_table();
  65         if (!s390_domain->dma_table) {
  66                 kfree(s390_domain);
  67                 return NULL;
  68         }
  69
  70         spin_lock_init(&s390_domain->dma_table_lock);
  71         spin_lock_init(&s390_domain->list_lock);
  72         INIT_LIST_HEAD(&s390_domain->devices);
  73
  74         return &s390_domain->domain;
  75 }
  76
  77 static void s390_domain_free(struct iommu_domain *domain)
  78 {
  79         struct s390_domain *s390_domain = to_s390_domain(domain);
  80
  81         dma_cleanup_tables(s390_domain->dma_table);
  82         kfree(s390_domain);
  83 }
  84
  85 static int s390_iommu_attach_device(struct iommu_domain *domain,
  86                                     struct device *dev)
  87 {
  88         struct s390_domain *s390_domain = to_s390_domain(domain);
  89         struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
  90         struct s390_domain_device *domain_device;
  91         unsigned long flags;
  92         int rc;
  93
  94         if (!zdev)
  95                 return -ENODEV;
  96
  97         domain_device = kzalloc(sizeof(*domain_device), GFP_KERNEL);
  98         if (!domain_device)
  99                 return -ENOMEM;
 100
 101         if (zdev->dma_table)
 102                 zpci_dma_exit_device(zdev);
 103
 104         zdev->dma_table = s390_domain->dma_table;
 105         rc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma,
 106                                 (u64) zdev->dma_table);
 107         if (rc)
 108                 goto out_restore;
 109
 110         spin_lock_irqsave(&s390_domain->list_lock, flags);
 111         /* First device defines the DMA range limits */
 112         if (list_empty(&s390_domain->devices)) {
 113                 domain->geometry.aperture_start = zdev->start_dma;
 114                 domain->geometry.aperture_end = zdev->end_dma;
 115                 domain->geometry.force_aperture = true;
 116         /* Allow only devices with identical DMA range limits */
 117         } else if (domain->geometry.aperture_start != zdev->start_dma ||
 118                    domain->geometry.aperture_end != zdev->end_dma) {
 119                 rc = -EINVAL;
 120                 spin_unlock_irqrestore(&s390_domain->list_lock, flags);
 121                 goto out_restore;
 122         }
 123         domain_device->zdev = zdev;
 124         zdev->s390_domain = s390_domain;
 125         list_add(&domain_device->list, &s390_domain->devices);
 126         spin_unlock_irqrestore(&s390_domain->list_lock, flags);
 127
 128         return 0;
 129
 130 out_restore:
 131         zpci_dma_init_device(zdev);
 132         kfree(domain_device);
 133
 134         return rc;
 135 }
 136
 137 static void s390_iommu_detach_device(struct iommu_domain *domain,
 138                                      struct device *dev)
 139 {
 140         struct s390_domain *s390_domain = to_s390_domain(domain);
 141         struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
 142         struct s390_domain_device *domain_device, *tmp;
 143         unsigned long flags;
 144         int found = 0;
 145
 146         if (!zdev)
 147                 return;
 148
 149         spin_lock_irqsave(&s390_domain->list_lock, flags);
 150         list_for_each_entry_safe(domain_device, tmp, &s390_domain->devices,
 151                                  list) {
 152                 if (domain_device->zdev == zdev) {
 153                         list_del(&domain_device->list);
 154                         kfree(domain_device);
 155                         found = 1;
 156                         break;
 157                 }
 158         }
 159         spin_unlock_irqrestore(&s390_domain->list_lock, flags);
 160
 161         if (found) {
 162                 zdev->s390_domain = NULL;
 163                 zpci_unregister_ioat(zdev, 0);
 164                 zpci_dma_init_device(zdev);
 165         }
 166 }
 167
 168 static int s390_iommu_add_device(struct device *dev)
 169 {
 170         struct iommu_group *group = iommu_group_get_for_dev(dev);
 171         struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
 172
 173         if (IS_ERR(group))
 174                 return PTR_ERR(group);
 175
 176         iommu_group_put(group);
 177         iommu_device_link(&zdev->iommu_dev, dev);
 178
 179         return 0;
 180 }
 181
 182 static void s390_iommu_remove_device(struct device *dev)
 183 {
 184         struct zpci_dev *zdev = to_pci_dev(dev)->sysdata;
 185         struct iommu_domain *domain;
 186
 187         /*
 188          * This is a workaround for a scenario where the IOMMU API common code
 189          * "forgets" to call the detach_dev callback: After binding a device
 190          * to vfio-pci and completing the VFIO_SET_IOMMU ioctl (which triggers
 191          * the attach_dev), removing the device via
 192          * "echo 1 > /sys/bus/pci/devices/.../remove" won't trigger detach_dev,
 193          * only remove_device will be called via the BUS_NOTIFY_REMOVED_DEVICE
 194          * notifier.
 195          *
 196          * So let's call detach_dev from here if it hasn't been called before.
 197          */
 198         if (zdev && zdev->s390_domain) {
 199                 domain = iommu_get_domain_for_dev(dev);
 200                 if (domain)
 201                         s390_iommu_detach_device(domain, dev);
 202         }
 203
 204         iommu_device_unlink(&zdev->iommu_dev, dev);
 205         iommu_group_remove_device(dev);
 206 }
 207
 208 static int s390_iommu_update_trans(struct s390_domain *s390_domain,
 209                                    unsigned long pa, dma_addr_t dma_addr,
 210                                    size_t size, int flags)
 211 {
 212         struct s390_domain_device *domain_device;
 213         u8 *page_addr = (u8 *) (pa & PAGE_MASK);
 214         dma_addr_t start_dma_addr = dma_addr;
 215         unsigned long irq_flags, nr_pages, i;
 216         unsigned long *entry;
 217         int rc = 0;
 218
 219         if (dma_addr < s390_domain->domain.geometry.aperture_start ||
 220             dma_addr + size > s390_domain->domain.geometry.aperture_end)
 221                 return -EINVAL;
 222
 223         nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
 224         if (!nr_pages)
 225                 return 0;
 226
 227         spin_lock_irqsave(&s390_domain->dma_table_lock, irq_flags);
 228         for (i = 0; i < nr_pages; i++) {
 229                 entry = dma_walk_cpu_trans(s390_domain->dma_table, dma_addr);
 230                 if (!entry) {
 231                         rc = -ENOMEM;
 232                         goto undo_cpu_trans;
 233                 }
 234                 dma_update_cpu_trans(entry, page_addr, flags);
 235                 page_addr += PAGE_SIZE;
 236                 dma_addr += PAGE_SIZE;
 237         }
 238
 239         spin_lock(&s390_domain->list_lock);
 240         list_for_each_entry(domain_device, &s390_domain->devices, list) {
 241                 rc = zpci_refresh_trans((u64) domain_device->zdev->fh << 32,
 242                                         start_dma_addr, nr_pages * PAGE_SIZE);
 243                 if (rc)
 244                         break;
 245         }
 246         spin_unlock(&s390_domain->list_lock);
 247
 248 undo_cpu_trans:
 249         if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) {
 250                 flags = ZPCI_PTE_INVALID;
 251                 while (i-- > 0) {
 252                         page_addr -= PAGE_SIZE;
 253                         dma_addr -= PAGE_SIZE;
 254                         entry = dma_walk_cpu_trans(s390_domain->dma_table,
 255                                                    dma_addr);
 256                         if (!entry)
 257                                 break;
 258                         dma_update_cpu_trans(entry, page_addr, flags);
 259                 }
 260         }
 261         spin_unlock_irqrestore(&s390_domain->dma_table_lock, irq_flags);
 262
 263         return rc;
 264 }
 265
 266 static int s390_iommu_map(struct iommu_domain *domain, unsigned long iova,
 267                           phys_addr_t paddr, size_t size, int prot)
 268 {
 269         struct s390_domain *s390_domain = to_s390_domain(domain);
 270         int flags = ZPCI_PTE_VALID, rc = 0;
 271
 272         if (!(prot & IOMMU_READ))
 273                 return -EINVAL;
 274
 275         if (!(prot & IOMMU_WRITE))
 276                 flags |= ZPCI_TABLE_PROTECTED;
 277
 278         rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
 279                                      size, flags);
 280
 281         return rc;
 282 }
 283
 284 static phys_addr_t s390_iommu_iova_to_phys(struct iommu_domain *domain,
 285                                            dma_addr_t iova)
 286 {
 287         struct s390_domain *s390_domain = to_s390_domain(domain);
 288         unsigned long *sto, *pto, *rto, flags;
 289         unsigned int rtx, sx, px;
 290         phys_addr_t phys = 0;
 291
 292         if (iova < domain->geometry.aperture_start ||
 293             iova > domain->geometry.aperture_end)
 294                 return 0;
 295
 296         rtx = calc_rtx(iova);
 297         sx = calc_sx(iova);
 298         px = calc_px(iova);
 299         rto = s390_domain->dma_table;
 300
 301         spin_lock_irqsave(&s390_domain->dma_table_lock, flags);
 302         if (rto && reg_entry_isvalid(rto[rtx])) {
 303                 sto = get_rt_sto(rto[rtx]);
 304                 if (sto && reg_entry_isvalid(sto[sx])) {
 305                         pto = get_st_pto(sto[sx]);
 306                         if (pto && pt_entry_isvalid(pto[px]))
 307                                 phys = pto[px] & ZPCI_PTE_ADDR_MASK;
 308                 }
 309         }
 310         spin_unlock_irqrestore(&s390_domain->dma_table_lock, flags);
 311
 312         return phys;
 313 }
 314
 315 static size_t s390_iommu_unmap(struct iommu_domain *domain,
 316                                unsigned long iova, size_t size)
 317 {
 318         struct s390_domain *s390_domain = to_s390_domain(domain);
 319         int flags = ZPCI_PTE_INVALID;
 320         phys_addr_t paddr;
 321         int rc;
 322
 323         paddr = s390_iommu_iova_to_phys(domain, iova);
 324         if (!paddr)
 325                 return 0;
 326
 327         rc = s390_iommu_update_trans(s390_domain, (unsigned long) paddr, iova,
 328                                      size, flags);
 329         if (rc)
 330                 return 0;
 331
 332         return size;
 333 }
 334
 335 int zpci_init_iommu(struct zpci_dev *zdev)
 336 {
 337         int rc = 0;
 338
 339         rc = iommu_device_sysfs_add(&zdev->iommu_dev, NULL, NULL,
 340                                     "s390-iommu.%08x", zdev->fid);
 341         if (rc)
 342                 goto out_err;
 343
 344         iommu_device_set_ops(&zdev->iommu_dev, &s390_iommu_ops);
 345
 346         rc = iommu_device_register(&zdev->iommu_dev);
 347         if (rc)
 348                 goto out_sysfs;
 349
 350         return 0;
 351
 352 out_sysfs:
 353         iommu_device_sysfs_remove(&zdev->iommu_dev);
 354
 355 out_err:
 356         return rc;
 357 }
 358
 359 void zpci_destroy_iommu(struct zpci_dev *zdev)
 360 {
 361         iommu_device_unregister(&zdev->iommu_dev);
 362         iommu_device_sysfs_remove(&zdev->iommu_dev);
 363 }
 364
 365 static const struct iommu_ops s390_iommu_ops = {
 366         .capable = s390_iommu_capable,
 367         .domain_alloc = s390_domain_alloc,
 368         .domain_free = s390_domain_free,
 369         .attach_dev = s390_iommu_attach_device,
 370         .detach_dev = s390_iommu_detach_device,
 371         .map = s390_iommu_map,
 372         .unmap = s390_iommu_unmap,
 373         .iova_to_phys = s390_iommu_iova_to_phys,
 374         .add_device = s390_iommu_add_device,
 375         .remove_device = s390_iommu_remove_device,
 376         .device_group = generic_device_group,
 377         .pgsize_bitmap = S390_IOMMU_PGSIZES,
 378 };
 379
 380 static int __init s390_iommu_init(void)
 381 {
 382         return bus_set_iommu(&pci_bus_type, &s390_iommu_ops);
 383 }
 384 subsys_initcall(s390_iommu_init);