1 /* 2 * Copyright IBM Corp. 2012 3 * 4 * Author(s): 5 * Jan Glauber <jang@linux.vnet.ibm.com> 6 */ 7 8 #include <linux/kernel.h> 9 #include <linux/slab.h> 10 #include <linux/export.h> 11 #include <linux/iommu-helper.h> 12 #include <linux/dma-mapping.h> 13 #include <linux/vmalloc.h> 14 #include <linux/pci.h> 15 #include <asm/pci_dma.h> 16 17 static struct kmem_cache *dma_region_table_cache; 18 static struct kmem_cache *dma_page_table_cache; 19 20 static unsigned long *dma_alloc_cpu_table(void) 21 { 22 unsigned long *table, *entry; 23 24 table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC); 25 if (!table) 26 return NULL; 27 28 for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++) 29 *entry = ZPCI_TABLE_INVALID | ZPCI_TABLE_PROTECTED; 30 return table; 31 } 32 33 static void dma_free_cpu_table(void *table) 34 { 35 kmem_cache_free(dma_region_table_cache, table); 36 } 37 38 static unsigned long *dma_alloc_page_table(void) 39 { 40 unsigned long *table, *entry; 41 42 table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC); 43 if (!table) 44 return NULL; 45 46 for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++) 47 *entry = ZPCI_PTE_INVALID | ZPCI_TABLE_PROTECTED; 48 return table; 49 } 50 51 static void dma_free_page_table(void *table) 52 { 53 kmem_cache_free(dma_page_table_cache, table); 54 } 55 56 static unsigned long *dma_get_seg_table_origin(unsigned long *entry) 57 { 58 unsigned long *sto; 59 60 if (reg_entry_isvalid(*entry)) 61 sto = get_rt_sto(*entry); 62 else { 63 sto = dma_alloc_cpu_table(); 64 if (!sto) 65 return NULL; 66 67 set_rt_sto(entry, sto); 68 validate_rt_entry(entry); 69 entry_clr_protected(entry); 70 } 71 return sto; 72 } 73 74 static unsigned long *dma_get_page_table_origin(unsigned long *entry) 75 { 76 unsigned long *pto; 77 78 if (reg_entry_isvalid(*entry)) 79 pto = get_st_pto(*entry); 80 else { 81 pto = dma_alloc_page_table(); 82 if (!pto) 83 return NULL; 84 set_st_pto(entry, pto); 85 validate_st_entry(entry); 86 entry_clr_protected(entry); 87 } 88 return pto; 89 } 90 91 static unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr) 92 { 93 unsigned long *sto, *pto; 94 unsigned int rtx, sx, px; 95 96 rtx = calc_rtx(dma_addr); 97 sto = dma_get_seg_table_origin(&rto[rtx]); 98 if (!sto) 99 return NULL; 100 101 sx = calc_sx(dma_addr); 102 pto = dma_get_page_table_origin(&sto[sx]); 103 if (!pto) 104 return NULL; 105 106 px = calc_px(dma_addr); 107 return &pto[px]; 108 } 109 110 static void dma_update_cpu_trans(struct zpci_dev *zdev, void *page_addr, 111 dma_addr_t dma_addr, int flags) 112 { 113 unsigned long *entry; 114 115 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr); 116 if (!entry) { 117 WARN_ON_ONCE(1); 118 return; 119 } 120 121 if (flags & ZPCI_PTE_INVALID) { 122 invalidate_pt_entry(entry); 123 return; 124 } else { 125 set_pt_pfaa(entry, page_addr); 126 validate_pt_entry(entry); 127 } 128 129 if (flags & ZPCI_TABLE_PROTECTED) 130 entry_set_protected(entry); 131 else 132 entry_clr_protected(entry); 133 } 134 135 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa, 136 dma_addr_t dma_addr, size_t size, int flags) 137 { 138 unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; 139 u8 *page_addr = (u8 *) (pa & PAGE_MASK); 140 dma_addr_t start_dma_addr = dma_addr; 141 unsigned long irq_flags; 142 int i, rc = 0; 143 144 if (!nr_pages) 145 return -EINVAL; 146 147 spin_lock_irqsave(&zdev->dma_table_lock, irq_flags); 148 if (!zdev->dma_table) 149 goto no_refresh; 150 151 for (i = 0; i < nr_pages; i++) { 152 dma_update_cpu_trans(zdev, page_addr, dma_addr, flags); 153 page_addr += PAGE_SIZE; 154 dma_addr += PAGE_SIZE; 155 } 156 157 /* 158 * rpcit is not required to establish new translations when previously 159 * invalid translation-table entries are validated, however it is 160 * required when altering previously valid entries. 161 */ 162 if (!zdev->tlb_refresh && 163 ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) 164 /* 165 * TODO: also need to check that the old entry is indeed INVALID 166 * and not only for one page but for the whole range... 167 * -> now we WARN_ON in that case but with lazy unmap that 168 * needs to be redone! 169 */ 170 goto no_refresh; 171 172 rc = zpci_refresh_trans((u64) zdev->fh << 32, start_dma_addr, 173 nr_pages * PAGE_SIZE); 174 175 no_refresh: 176 spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags); 177 return rc; 178 } 179 180 static void dma_free_seg_table(unsigned long entry) 181 { 182 unsigned long *sto = get_rt_sto(entry); 183 int sx; 184 185 for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++) 186 if (reg_entry_isvalid(sto[sx])) 187 dma_free_page_table(get_st_pto(sto[sx])); 188 189 dma_free_cpu_table(sto); 190 } 191 192 static void dma_cleanup_tables(struct zpci_dev *zdev) 193 { 194 unsigned long *table; 195 int rtx; 196 197 if (!zdev || !zdev->dma_table) 198 return; 199 200 table = zdev->dma_table; 201 for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++) 202 if (reg_entry_isvalid(table[rtx])) 203 dma_free_seg_table(table[rtx]); 204 205 dma_free_cpu_table(table); 206 zdev->dma_table = NULL; 207 } 208 209 static unsigned long __dma_alloc_iommu(struct zpci_dev *zdev, 210 unsigned long start, int size) 211 { 212 unsigned long boundary_size; 213 214 boundary_size = ALIGN(dma_get_seg_boundary(&zdev->pdev->dev) + 1, 215 PAGE_SIZE) >> PAGE_SHIFT; 216 return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages, 217 start, size, 0, boundary_size, 0); 218 } 219 220 static unsigned long dma_alloc_iommu(struct zpci_dev *zdev, int size) 221 { 222 unsigned long offset, flags; 223 224 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags); 225 offset = __dma_alloc_iommu(zdev, zdev->next_bit, size); 226 if (offset == -1) 227 offset = __dma_alloc_iommu(zdev, 0, size); 228 229 if (offset != -1) { 230 zdev->next_bit = offset + size; 231 if (zdev->next_bit >= zdev->iommu_pages) 232 zdev->next_bit = 0; 233 } 234 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags); 235 return offset; 236 } 237 238 static void dma_free_iommu(struct zpci_dev *zdev, unsigned long offset, int size) 239 { 240 unsigned long flags; 241 242 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags); 243 if (!zdev->iommu_bitmap) 244 goto out; 245 bitmap_clear(zdev->iommu_bitmap, offset, size); 246 if (offset >= zdev->next_bit) 247 zdev->next_bit = offset + size; 248 out: 249 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags); 250 } 251 252 int dma_set_mask(struct device *dev, u64 mask) 253 { 254 if (!dev->dma_mask || !dma_supported(dev, mask)) 255 return -EIO; 256 257 *dev->dma_mask = mask; 258 return 0; 259 } 260 EXPORT_SYMBOL_GPL(dma_set_mask); 261 262 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page, 263 unsigned long offset, size_t size, 264 enum dma_data_direction direction, 265 struct dma_attrs *attrs) 266 { 267 struct zpci_dev *zdev = get_zdev(to_pci_dev(dev)); 268 unsigned long nr_pages, iommu_page_index; 269 unsigned long pa = page_to_phys(page) + offset; 270 int flags = ZPCI_PTE_VALID; 271 dma_addr_t dma_addr; 272 273 /* This rounds up number of pages based on size and offset */ 274 nr_pages = iommu_num_pages(pa, size, PAGE_SIZE); 275 iommu_page_index = dma_alloc_iommu(zdev, nr_pages); 276 if (iommu_page_index == -1) 277 goto out_err; 278 279 /* Use rounded up size */ 280 size = nr_pages * PAGE_SIZE; 281 282 dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE; 283 if (dma_addr + size > zdev->end_dma) 284 goto out_free; 285 286 if (direction == DMA_NONE || direction == DMA_TO_DEVICE) 287 flags |= ZPCI_TABLE_PROTECTED; 288 289 if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) { 290 atomic64_add(nr_pages, &zdev->fmb->mapped_pages); 291 return dma_addr + (offset & ~PAGE_MASK); 292 } 293 294 out_free: 295 dma_free_iommu(zdev, iommu_page_index, nr_pages); 296 out_err: 297 zpci_err("map error:\n"); 298 zpci_err_hex(&pa, sizeof(pa)); 299 return DMA_ERROR_CODE; 300 } 301 302 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr, 303 size_t size, enum dma_data_direction direction, 304 struct dma_attrs *attrs) 305 { 306 struct zpci_dev *zdev = get_zdev(to_pci_dev(dev)); 307 unsigned long iommu_page_index; 308 int npages; 309 310 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE); 311 dma_addr = dma_addr & PAGE_MASK; 312 if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE, 313 ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID)) { 314 zpci_err("unmap error:\n"); 315 zpci_err_hex(&dma_addr, sizeof(dma_addr)); 316 } 317 318 atomic64_add(npages, &zdev->fmb->unmapped_pages); 319 iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT; 320 dma_free_iommu(zdev, iommu_page_index, npages); 321 } 322 323 static void *s390_dma_alloc(struct device *dev, size_t size, 324 dma_addr_t *dma_handle, gfp_t flag, 325 struct dma_attrs *attrs) 326 { 327 struct zpci_dev *zdev = get_zdev(to_pci_dev(dev)); 328 struct page *page; 329 unsigned long pa; 330 dma_addr_t map; 331 332 size = PAGE_ALIGN(size); 333 page = alloc_pages(flag, get_order(size)); 334 if (!page) 335 return NULL; 336 337 pa = page_to_phys(page); 338 memset((void *) pa, 0, size); 339 340 map = s390_dma_map_pages(dev, page, pa % PAGE_SIZE, 341 size, DMA_BIDIRECTIONAL, NULL); 342 if (dma_mapping_error(dev, map)) { 343 free_pages(pa, get_order(size)); 344 return NULL; 345 } 346 347 atomic64_add(size / PAGE_SIZE, &zdev->fmb->allocated_pages); 348 if (dma_handle) 349 *dma_handle = map; 350 return (void *) pa; 351 } 352 353 static void s390_dma_free(struct device *dev, size_t size, 354 void *pa, dma_addr_t dma_handle, 355 struct dma_attrs *attrs) 356 { 357 struct zpci_dev *zdev = get_zdev(to_pci_dev(dev)); 358 359 size = PAGE_ALIGN(size); 360 atomic64_sub(size / PAGE_SIZE, &zdev->fmb->allocated_pages); 361 s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, NULL); 362 free_pages((unsigned long) pa, get_order(size)); 363 } 364 365 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg, 366 int nr_elements, enum dma_data_direction dir, 367 struct dma_attrs *attrs) 368 { 369 int mapped_elements = 0; 370 struct scatterlist *s; 371 int i; 372 373 for_each_sg(sg, s, nr_elements, i) { 374 struct page *page = sg_page(s); 375 s->dma_address = s390_dma_map_pages(dev, page, s->offset, 376 s->length, dir, NULL); 377 if (!dma_mapping_error(dev, s->dma_address)) { 378 s->dma_length = s->length; 379 mapped_elements++; 380 } else 381 goto unmap; 382 } 383 out: 384 return mapped_elements; 385 386 unmap: 387 for_each_sg(sg, s, mapped_elements, i) { 388 if (s->dma_address) 389 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, 390 dir, NULL); 391 s->dma_address = 0; 392 s->dma_length = 0; 393 } 394 mapped_elements = 0; 395 goto out; 396 } 397 398 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg, 399 int nr_elements, enum dma_data_direction dir, 400 struct dma_attrs *attrs) 401 { 402 struct scatterlist *s; 403 int i; 404 405 for_each_sg(sg, s, nr_elements, i) { 406 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL); 407 s->dma_address = 0; 408 s->dma_length = 0; 409 } 410 } 411 412 int zpci_dma_init_device(struct zpci_dev *zdev) 413 { 414 int rc; 415 416 spin_lock_init(&zdev->iommu_bitmap_lock); 417 spin_lock_init(&zdev->dma_table_lock); 418 419 zdev->dma_table = dma_alloc_cpu_table(); 420 if (!zdev->dma_table) { 421 rc = -ENOMEM; 422 goto out_clean; 423 } 424 425 zdev->iommu_size = (unsigned long) high_memory - PAGE_OFFSET; 426 zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT; 427 zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8); 428 if (!zdev->iommu_bitmap) { 429 rc = -ENOMEM; 430 goto out_reg; 431 } 432 433 rc = zpci_register_ioat(zdev, 434 0, 435 zdev->start_dma + PAGE_OFFSET, 436 zdev->start_dma + zdev->iommu_size - 1, 437 (u64) zdev->dma_table); 438 if (rc) 439 goto out_reg; 440 return 0; 441 442 out_reg: 443 dma_free_cpu_table(zdev->dma_table); 444 out_clean: 445 return rc; 446 } 447 448 void zpci_dma_exit_device(struct zpci_dev *zdev) 449 { 450 zpci_unregister_ioat(zdev, 0); 451 dma_cleanup_tables(zdev); 452 vfree(zdev->iommu_bitmap); 453 zdev->iommu_bitmap = NULL; 454 zdev->next_bit = 0; 455 } 456 457 static int __init dma_alloc_cpu_table_caches(void) 458 { 459 dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables", 460 ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN, 461 0, NULL); 462 if (!dma_region_table_cache) 463 return -ENOMEM; 464 465 dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables", 466 ZPCI_PT_SIZE, ZPCI_PT_ALIGN, 467 0, NULL); 468 if (!dma_page_table_cache) { 469 kmem_cache_destroy(dma_region_table_cache); 470 return -ENOMEM; 471 } 472 return 0; 473 } 474 475 int __init zpci_dma_init(void) 476 { 477 return dma_alloc_cpu_table_caches(); 478 } 479 480 void zpci_dma_exit(void) 481 { 482 kmem_cache_destroy(dma_page_table_cache); 483 kmem_cache_destroy(dma_region_table_cache); 484 } 485 486 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16) 487 488 static int __init dma_debug_do_init(void) 489 { 490 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES); 491 return 0; 492 } 493 fs_initcall(dma_debug_do_init); 494 495 struct dma_map_ops s390_dma_ops = { 496 .alloc = s390_dma_alloc, 497 .free = s390_dma_free, 498 .map_sg = s390_dma_map_sg, 499 .unmap_sg = s390_dma_unmap_sg, 500 .map_page = s390_dma_map_pages, 501 .unmap_page = s390_dma_unmap_pages, 502 /* if we support direct DMA this must be conditional */ 503 .is_phys = 0, 504 /* dma_supported is unconditionally true without a callback */ 505 }; 506 EXPORT_SYMBOL_GPL(s390_dma_ops); 507