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