1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * IOMMU API for s390 PCI devices 4 * 5 * Copyright IBM Corp. 2015 6 * Author(s): Gerald Schaefer <gerald.schaefer@de.ibm.com> 7 */ 8 9 #include <linux/pci.h> 10 #include <linux/iommu.h> 11 #include <linux/iommu-helper.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 static const struct iommu_ops s390_iommu_ops; 23 24 struct s390_domain { 25 struct iommu_domain domain; 26 struct list_head devices; 27 unsigned long *dma_table; 28 spinlock_t dma_table_lock; 29 spinlock_t list_lock; 30 }; 31 32 struct s390_domain_device { 33 struct list_head list; 34 struct zpci_dev *zdev; 35 }; 36 37 static struct s390_domain *to_s390_domain(struct iommu_domain *dom) 38 { 39 return container_of(dom, struct s390_domain, domain); 40 } 41 42 static bool s390_iommu_capable(enum iommu_cap cap) 43 { 44 switch (cap) { 45 case IOMMU_CAP_CACHE_COHERENCY: 46 return true; 47 case IOMMU_CAP_INTR_REMAP: 48 return true; 49 default: 50 return false; 51 } 52 } 53 54 static struct iommu_domain *s390_domain_alloc(unsigned domain_type) 55 { 56 struct s390_domain *s390_domain; 57 58 if (domain_type != IOMMU_DOMAIN_UNMANAGED) 59 return NULL; 60 61 s390_domain = kzalloc(sizeof(*s390_domain), GFP_KERNEL); 62 if (!s390_domain) 63 return NULL; 64 65 s390_domain->dma_table = dma_alloc_cpu_table(); 66 if (!s390_domain->dma_table) { 67 kfree(s390_domain); 68 return NULL; 69 } 70 71 spin_lock_init(&s390_domain->dma_table_lock); 72 spin_lock_init(&s390_domain->list_lock); 73 INIT_LIST_HEAD(&s390_domain->devices); 74 75 return &s390_domain->domain; 76 } 77 78 static void s390_domain_free(struct iommu_domain *domain) 79 { 80 struct s390_domain *s390_domain = to_s390_domain(domain); 81 82 dma_cleanup_tables(s390_domain->dma_table); 83 kfree(s390_domain); 84 } 85 86 static int s390_iommu_attach_device(struct iommu_domain *domain, 87 struct device *dev) 88 { 89 struct s390_domain *s390_domain = to_s390_domain(domain); 90 struct zpci_dev *zdev = to_zpci_dev(dev); 91 struct s390_domain_device *domain_device; 92 unsigned long flags; 93 int cc, rc; 94 95 if (!zdev) 96 return -ENODEV; 97 98 domain_device = kzalloc(sizeof(*domain_device), GFP_KERNEL); 99 if (!domain_device) 100 return -ENOMEM; 101 102 if (zdev->dma_table && !zdev->s390_domain) { 103 cc = zpci_dma_exit_device(zdev); 104 if (cc) { 105 rc = -EIO; 106 goto out_free; 107 } 108 } 109 110 if (zdev->s390_domain) 111 zpci_unregister_ioat(zdev, 0); 112 113 zdev->dma_table = s390_domain->dma_table; 114 cc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma, 115 virt_to_phys(zdev->dma_table)); 116 if (cc) { 117 rc = -EIO; 118 goto out_restore; 119 } 120 121 spin_lock_irqsave(&s390_domain->list_lock, flags); 122 /* First device defines the DMA range limits */ 123 if (list_empty(&s390_domain->devices)) { 124 domain->geometry.aperture_start = zdev->start_dma; 125 domain->geometry.aperture_end = zdev->end_dma; 126 domain->geometry.force_aperture = true; 127 /* Allow only devices with identical DMA range limits */ 128 } else if (domain->geometry.aperture_start != zdev->start_dma || 129 domain->geometry.aperture_end != zdev->end_dma) { 130 rc = -EINVAL; 131 spin_unlock_irqrestore(&s390_domain->list_lock, flags); 132 goto out_restore; 133 } 134 domain_device->zdev = zdev; 135 zdev->s390_domain = s390_domain; 136 list_add(&domain_device->list, &s390_domain->devices); 137 spin_unlock_irqrestore(&s390_domain->list_lock, flags); 138 139 return 0; 140 141 out_restore: 142 if (!zdev->s390_domain) { 143 zpci_dma_init_device(zdev); 144 } else { 145 zdev->dma_table = zdev->s390_domain->dma_table; 146 zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma, 147 virt_to_phys(zdev->dma_table)); 148 } 149 out_free: 150 kfree(domain_device); 151 152 return rc; 153 } 154 155 static void s390_iommu_detach_device(struct iommu_domain *domain, 156 struct device *dev) 157 { 158 struct s390_domain *s390_domain = to_s390_domain(domain); 159 struct zpci_dev *zdev = to_zpci_dev(dev); 160 struct s390_domain_device *domain_device, *tmp; 161 unsigned long flags; 162 int found = 0; 163 164 if (!zdev) 165 return; 166 167 spin_lock_irqsave(&s390_domain->list_lock, flags); 168 list_for_each_entry_safe(domain_device, tmp, &s390_domain->devices, 169 list) { 170 if (domain_device->zdev == zdev) { 171 list_del(&domain_device->list); 172 kfree(domain_device); 173 found = 1; 174 break; 175 } 176 } 177 spin_unlock_irqrestore(&s390_domain->list_lock, flags); 178 179 if (found && (zdev->s390_domain == s390_domain)) { 180 zdev->s390_domain = NULL; 181 zpci_unregister_ioat(zdev, 0); 182 zpci_dma_init_device(zdev); 183 } 184 } 185 186 static struct iommu_device *s390_iommu_probe_device(struct device *dev) 187 { 188 struct zpci_dev *zdev = to_zpci_dev(dev); 189 190 return &zdev->iommu_dev; 191 } 192 193 static void s390_iommu_release_device(struct device *dev) 194 { 195 struct zpci_dev *zdev = to_zpci_dev(dev); 196 struct iommu_domain *domain; 197 198 /* 199 * This is a workaround for a scenario where the IOMMU API common code 200 * "forgets" to call the detach_dev callback: After binding a device 201 * to vfio-pci and completing the VFIO_SET_IOMMU ioctl (which triggers 202 * the attach_dev), removing the device via 203 * "echo 1 > /sys/bus/pci/devices/.../remove" won't trigger detach_dev, 204 * only release_device will be called via the BUS_NOTIFY_REMOVED_DEVICE 205 * notifier. 206 * 207 * So let's call detach_dev from here if it hasn't been called before. 208 */ 209 if (zdev && zdev->s390_domain) { 210 domain = iommu_get_domain_for_dev(dev); 211 if (domain) 212 s390_iommu_detach_device(domain, dev); 213 } 214 } 215 216 static int s390_iommu_update_trans(struct s390_domain *s390_domain, 217 phys_addr_t pa, dma_addr_t dma_addr, 218 size_t size, int flags) 219 { 220 struct s390_domain_device *domain_device; 221 phys_addr_t page_addr = pa & PAGE_MASK; 222 dma_addr_t start_dma_addr = dma_addr; 223 unsigned long irq_flags, nr_pages, i; 224 unsigned long *entry; 225 int rc = 0; 226 227 if (dma_addr < s390_domain->domain.geometry.aperture_start || 228 dma_addr + size > s390_domain->domain.geometry.aperture_end) 229 return -EINVAL; 230 231 nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; 232 if (!nr_pages) 233 return 0; 234 235 spin_lock_irqsave(&s390_domain->dma_table_lock, irq_flags); 236 for (i = 0; i < nr_pages; i++) { 237 entry = dma_walk_cpu_trans(s390_domain->dma_table, dma_addr); 238 if (!entry) { 239 rc = -ENOMEM; 240 goto undo_cpu_trans; 241 } 242 dma_update_cpu_trans(entry, page_addr, flags); 243 page_addr += PAGE_SIZE; 244 dma_addr += PAGE_SIZE; 245 } 246 247 spin_lock(&s390_domain->list_lock); 248 list_for_each_entry(domain_device, &s390_domain->devices, list) { 249 rc = zpci_refresh_trans((u64) domain_device->zdev->fh << 32, 250 start_dma_addr, nr_pages * PAGE_SIZE); 251 if (rc) 252 break; 253 } 254 spin_unlock(&s390_domain->list_lock); 255 256 undo_cpu_trans: 257 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) { 258 flags = ZPCI_PTE_INVALID; 259 while (i-- > 0) { 260 page_addr -= PAGE_SIZE; 261 dma_addr -= PAGE_SIZE; 262 entry = dma_walk_cpu_trans(s390_domain->dma_table, 263 dma_addr); 264 if (!entry) 265 break; 266 dma_update_cpu_trans(entry, page_addr, flags); 267 } 268 } 269 spin_unlock_irqrestore(&s390_domain->dma_table_lock, irq_flags); 270 271 return rc; 272 } 273 274 static int s390_iommu_map(struct iommu_domain *domain, unsigned long iova, 275 phys_addr_t paddr, size_t size, int prot, gfp_t gfp) 276 { 277 struct s390_domain *s390_domain = to_s390_domain(domain); 278 int flags = ZPCI_PTE_VALID, rc = 0; 279 280 if (!(prot & IOMMU_READ)) 281 return -EINVAL; 282 283 if (!(prot & IOMMU_WRITE)) 284 flags |= ZPCI_TABLE_PROTECTED; 285 286 rc = s390_iommu_update_trans(s390_domain, paddr, iova, 287 size, flags); 288 289 return rc; 290 } 291 292 static phys_addr_t s390_iommu_iova_to_phys(struct iommu_domain *domain, 293 dma_addr_t iova) 294 { 295 struct s390_domain *s390_domain = to_s390_domain(domain); 296 unsigned long *sto, *pto, *rto, flags; 297 unsigned int rtx, sx, px; 298 phys_addr_t phys = 0; 299 300 if (iova < domain->geometry.aperture_start || 301 iova > domain->geometry.aperture_end) 302 return 0; 303 304 rtx = calc_rtx(iova); 305 sx = calc_sx(iova); 306 px = calc_px(iova); 307 rto = s390_domain->dma_table; 308 309 spin_lock_irqsave(&s390_domain->dma_table_lock, flags); 310 if (rto && reg_entry_isvalid(rto[rtx])) { 311 sto = get_rt_sto(rto[rtx]); 312 if (sto && reg_entry_isvalid(sto[sx])) { 313 pto = get_st_pto(sto[sx]); 314 if (pto && pt_entry_isvalid(pto[px])) 315 phys = pto[px] & ZPCI_PTE_ADDR_MASK; 316 } 317 } 318 spin_unlock_irqrestore(&s390_domain->dma_table_lock, flags); 319 320 return phys; 321 } 322 323 static size_t s390_iommu_unmap(struct iommu_domain *domain, 324 unsigned long iova, size_t size, 325 struct iommu_iotlb_gather *gather) 326 { 327 struct s390_domain *s390_domain = to_s390_domain(domain); 328 int flags = ZPCI_PTE_INVALID; 329 phys_addr_t paddr; 330 int rc; 331 332 paddr = s390_iommu_iova_to_phys(domain, iova); 333 if (!paddr) 334 return 0; 335 336 rc = s390_iommu_update_trans(s390_domain, paddr, iova, 337 size, flags); 338 if (rc) 339 return 0; 340 341 return size; 342 } 343 344 int zpci_init_iommu(struct zpci_dev *zdev) 345 { 346 int rc = 0; 347 348 rc = iommu_device_sysfs_add(&zdev->iommu_dev, NULL, NULL, 349 "s390-iommu.%08x", zdev->fid); 350 if (rc) 351 goto out_err; 352 353 rc = iommu_device_register(&zdev->iommu_dev, &s390_iommu_ops, NULL); 354 if (rc) 355 goto out_sysfs; 356 357 return 0; 358 359 out_sysfs: 360 iommu_device_sysfs_remove(&zdev->iommu_dev); 361 362 out_err: 363 return rc; 364 } 365 366 void zpci_destroy_iommu(struct zpci_dev *zdev) 367 { 368 iommu_device_unregister(&zdev->iommu_dev); 369 iommu_device_sysfs_remove(&zdev->iommu_dev); 370 } 371 372 static const struct iommu_ops s390_iommu_ops = { 373 .capable = s390_iommu_capable, 374 .domain_alloc = s390_domain_alloc, 375 .probe_device = s390_iommu_probe_device, 376 .release_device = s390_iommu_release_device, 377 .device_group = generic_device_group, 378 .pgsize_bitmap = S390_IOMMU_PGSIZES, 379 .default_domain_ops = &(const struct iommu_domain_ops) { 380 .attach_dev = s390_iommu_attach_device, 381 .detach_dev = s390_iommu_detach_device, 382 .map = s390_iommu_map, 383 .unmap = s390_iommu_unmap, 384 .iova_to_phys = s390_iommu_iova_to_phys, 385 .free = s390_domain_free, 386 } 387 }; 388 389 static int __init s390_iommu_init(void) 390 { 391 return bus_set_iommu(&pci_bus_type, &s390_iommu_ops); 392 } 393 subsys_initcall(s390_iommu_init); 394