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 static int s390_iommu_strict; 20 21 static int zpci_refresh_global(struct zpci_dev *zdev) 22 { 23 return zpci_refresh_trans((u64) zdev->fh << 32, zdev->start_dma, 24 zdev->iommu_pages * PAGE_SIZE); 25 } 26 27 unsigned long *dma_alloc_cpu_table(void) 28 { 29 unsigned long *table, *entry; 30 31 table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC); 32 if (!table) 33 return NULL; 34 35 for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++) 36 *entry = ZPCI_TABLE_INVALID; 37 return table; 38 } 39 40 static void dma_free_cpu_table(void *table) 41 { 42 kmem_cache_free(dma_region_table_cache, table); 43 } 44 45 static unsigned long *dma_alloc_page_table(void) 46 { 47 unsigned long *table, *entry; 48 49 table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC); 50 if (!table) 51 return NULL; 52 53 for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++) 54 *entry = ZPCI_PTE_INVALID; 55 return table; 56 } 57 58 static void dma_free_page_table(void *table) 59 { 60 kmem_cache_free(dma_page_table_cache, table); 61 } 62 63 static unsigned long *dma_get_seg_table_origin(unsigned long *entry) 64 { 65 unsigned long *sto; 66 67 if (reg_entry_isvalid(*entry)) 68 sto = get_rt_sto(*entry); 69 else { 70 sto = dma_alloc_cpu_table(); 71 if (!sto) 72 return NULL; 73 74 set_rt_sto(entry, sto); 75 validate_rt_entry(entry); 76 entry_clr_protected(entry); 77 } 78 return sto; 79 } 80 81 static unsigned long *dma_get_page_table_origin(unsigned long *entry) 82 { 83 unsigned long *pto; 84 85 if (reg_entry_isvalid(*entry)) 86 pto = get_st_pto(*entry); 87 else { 88 pto = dma_alloc_page_table(); 89 if (!pto) 90 return NULL; 91 set_st_pto(entry, pto); 92 validate_st_entry(entry); 93 entry_clr_protected(entry); 94 } 95 return pto; 96 } 97 98 unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr) 99 { 100 unsigned long *sto, *pto; 101 unsigned int rtx, sx, px; 102 103 rtx = calc_rtx(dma_addr); 104 sto = dma_get_seg_table_origin(&rto[rtx]); 105 if (!sto) 106 return NULL; 107 108 sx = calc_sx(dma_addr); 109 pto = dma_get_page_table_origin(&sto[sx]); 110 if (!pto) 111 return NULL; 112 113 px = calc_px(dma_addr); 114 return &pto[px]; 115 } 116 117 void dma_update_cpu_trans(unsigned long *entry, void *page_addr, int flags) 118 { 119 if (flags & ZPCI_PTE_INVALID) { 120 invalidate_pt_entry(entry); 121 } else { 122 set_pt_pfaa(entry, page_addr); 123 validate_pt_entry(entry); 124 } 125 126 if (flags & ZPCI_TABLE_PROTECTED) 127 entry_set_protected(entry); 128 else 129 entry_clr_protected(entry); 130 } 131 132 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa, 133 dma_addr_t dma_addr, size_t size, int flags) 134 { 135 unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; 136 u8 *page_addr = (u8 *) (pa & PAGE_MASK); 137 dma_addr_t start_dma_addr = dma_addr; 138 unsigned long irq_flags; 139 unsigned long *entry; 140 int i, rc = 0; 141 142 if (!nr_pages) 143 return -EINVAL; 144 145 spin_lock_irqsave(&zdev->dma_table_lock, irq_flags); 146 if (!zdev->dma_table) { 147 rc = -EINVAL; 148 goto no_refresh; 149 } 150 151 for (i = 0; i < nr_pages; i++) { 152 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr); 153 if (!entry) { 154 rc = -ENOMEM; 155 goto undo_cpu_trans; 156 } 157 dma_update_cpu_trans(entry, page_addr, flags); 158 page_addr += PAGE_SIZE; 159 dma_addr += PAGE_SIZE; 160 } 161 162 /* 163 * With zdev->tlb_refresh == 0, rpcit is not required to establish new 164 * translations when previously invalid translation-table entries are 165 * validated. With lazy unmap, it also is skipped for previously valid 166 * entries, but a global rpcit is then required before any address can 167 * be re-used, i.e. after each iommu bitmap wrap-around. 168 */ 169 if (!zdev->tlb_refresh && 170 (!s390_iommu_strict || 171 ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID))) 172 goto no_refresh; 173 174 rc = zpci_refresh_trans((u64) zdev->fh << 32, start_dma_addr, 175 nr_pages * PAGE_SIZE); 176 undo_cpu_trans: 177 if (rc && ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)) { 178 flags = ZPCI_PTE_INVALID; 179 while (i-- > 0) { 180 page_addr -= PAGE_SIZE; 181 dma_addr -= PAGE_SIZE; 182 entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr); 183 if (!entry) 184 break; 185 dma_update_cpu_trans(entry, page_addr, flags); 186 } 187 } 188 189 no_refresh: 190 spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags); 191 return rc; 192 } 193 194 void dma_free_seg_table(unsigned long entry) 195 { 196 unsigned long *sto = get_rt_sto(entry); 197 int sx; 198 199 for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++) 200 if (reg_entry_isvalid(sto[sx])) 201 dma_free_page_table(get_st_pto(sto[sx])); 202 203 dma_free_cpu_table(sto); 204 } 205 206 void dma_cleanup_tables(unsigned long *table) 207 { 208 int rtx; 209 210 if (!table) 211 return; 212 213 for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++) 214 if (reg_entry_isvalid(table[rtx])) 215 dma_free_seg_table(table[rtx]); 216 217 dma_free_cpu_table(table); 218 } 219 220 static unsigned long __dma_alloc_iommu(struct device *dev, 221 unsigned long start, int size) 222 { 223 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 224 unsigned long boundary_size; 225 226 boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1, 227 PAGE_SIZE) >> PAGE_SHIFT; 228 return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages, 229 start, size, 0, boundary_size, 0); 230 } 231 232 static unsigned long dma_alloc_iommu(struct device *dev, int size) 233 { 234 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 235 unsigned long offset, flags; 236 int wrap = 0; 237 238 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags); 239 offset = __dma_alloc_iommu(dev, zdev->next_bit, size); 240 if (offset == -1) { 241 /* wrap-around */ 242 offset = __dma_alloc_iommu(dev, 0, size); 243 wrap = 1; 244 } 245 246 if (offset != -1) { 247 zdev->next_bit = offset + size; 248 if (!zdev->tlb_refresh && !s390_iommu_strict && wrap) 249 /* global flush after wrap-around with lazy unmap */ 250 zpci_refresh_global(zdev); 251 } 252 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags); 253 return offset; 254 } 255 256 static void dma_free_iommu(struct device *dev, unsigned long offset, int size) 257 { 258 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 259 unsigned long flags; 260 261 spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags); 262 if (!zdev->iommu_bitmap) 263 goto out; 264 bitmap_clear(zdev->iommu_bitmap, offset, size); 265 /* 266 * Lazy flush for unmap: need to move next_bit to avoid address re-use 267 * until wrap-around. 268 */ 269 if (!s390_iommu_strict && offset >= zdev->next_bit) 270 zdev->next_bit = offset + size; 271 out: 272 spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags); 273 } 274 275 static inline void zpci_err_dma(unsigned long rc, unsigned long addr) 276 { 277 struct { 278 unsigned long rc; 279 unsigned long addr; 280 } __packed data = {rc, addr}; 281 282 zpci_err_hex(&data, sizeof(data)); 283 } 284 285 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page, 286 unsigned long offset, size_t size, 287 enum dma_data_direction direction, 288 struct dma_attrs *attrs) 289 { 290 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 291 unsigned long nr_pages, iommu_page_index; 292 unsigned long pa = page_to_phys(page) + offset; 293 int flags = ZPCI_PTE_VALID; 294 dma_addr_t dma_addr; 295 int ret; 296 297 /* This rounds up number of pages based on size and offset */ 298 nr_pages = iommu_num_pages(pa, size, PAGE_SIZE); 299 iommu_page_index = dma_alloc_iommu(dev, nr_pages); 300 if (iommu_page_index == -1) { 301 ret = -ENOSPC; 302 goto out_err; 303 } 304 305 /* Use rounded up size */ 306 size = nr_pages * PAGE_SIZE; 307 308 dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE; 309 if (dma_addr + size > zdev->end_dma) { 310 ret = -ERANGE; 311 goto out_free; 312 } 313 314 if (direction == DMA_NONE || direction == DMA_TO_DEVICE) 315 flags |= ZPCI_TABLE_PROTECTED; 316 317 ret = dma_update_trans(zdev, pa, dma_addr, size, flags); 318 if (ret) 319 goto out_free; 320 321 atomic64_add(nr_pages, &zdev->mapped_pages); 322 return dma_addr + (offset & ~PAGE_MASK); 323 324 out_free: 325 dma_free_iommu(dev, iommu_page_index, nr_pages); 326 out_err: 327 zpci_err("map error:\n"); 328 zpci_err_dma(ret, pa); 329 return DMA_ERROR_CODE; 330 } 331 332 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr, 333 size_t size, enum dma_data_direction direction, 334 struct dma_attrs *attrs) 335 { 336 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 337 unsigned long iommu_page_index; 338 int npages, ret; 339 340 npages = iommu_num_pages(dma_addr, size, PAGE_SIZE); 341 dma_addr = dma_addr & PAGE_MASK; 342 ret = dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE, 343 ZPCI_PTE_INVALID); 344 if (ret) { 345 zpci_err("unmap error:\n"); 346 zpci_err_dma(ret, dma_addr); 347 return; 348 } 349 350 atomic64_add(npages, &zdev->unmapped_pages); 351 iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT; 352 dma_free_iommu(dev, iommu_page_index, npages); 353 } 354 355 static void *s390_dma_alloc(struct device *dev, size_t size, 356 dma_addr_t *dma_handle, gfp_t flag, 357 struct dma_attrs *attrs) 358 { 359 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 360 struct page *page; 361 unsigned long pa; 362 dma_addr_t map; 363 364 size = PAGE_ALIGN(size); 365 page = alloc_pages(flag, get_order(size)); 366 if (!page) 367 return NULL; 368 369 pa = page_to_phys(page); 370 memset((void *) pa, 0, size); 371 372 map = s390_dma_map_pages(dev, page, 0, size, DMA_BIDIRECTIONAL, NULL); 373 if (dma_mapping_error(dev, map)) { 374 free_pages(pa, get_order(size)); 375 return NULL; 376 } 377 378 atomic64_add(size / PAGE_SIZE, &zdev->allocated_pages); 379 if (dma_handle) 380 *dma_handle = map; 381 return (void *) pa; 382 } 383 384 static void s390_dma_free(struct device *dev, size_t size, 385 void *pa, dma_addr_t dma_handle, 386 struct dma_attrs *attrs) 387 { 388 struct zpci_dev *zdev = to_zpci(to_pci_dev(dev)); 389 390 size = PAGE_ALIGN(size); 391 atomic64_sub(size / PAGE_SIZE, &zdev->allocated_pages); 392 s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, NULL); 393 free_pages((unsigned long) pa, get_order(size)); 394 } 395 396 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg, 397 int nr_elements, enum dma_data_direction dir, 398 struct dma_attrs *attrs) 399 { 400 int mapped_elements = 0; 401 struct scatterlist *s; 402 int i; 403 404 for_each_sg(sg, s, nr_elements, i) { 405 struct page *page = sg_page(s); 406 s->dma_address = s390_dma_map_pages(dev, page, s->offset, 407 s->length, dir, NULL); 408 if (!dma_mapping_error(dev, s->dma_address)) { 409 s->dma_length = s->length; 410 mapped_elements++; 411 } else 412 goto unmap; 413 } 414 out: 415 return mapped_elements; 416 417 unmap: 418 for_each_sg(sg, s, mapped_elements, i) { 419 if (s->dma_address) 420 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, 421 dir, NULL); 422 s->dma_address = 0; 423 s->dma_length = 0; 424 } 425 mapped_elements = 0; 426 goto out; 427 } 428 429 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg, 430 int nr_elements, enum dma_data_direction dir, 431 struct dma_attrs *attrs) 432 { 433 struct scatterlist *s; 434 int i; 435 436 for_each_sg(sg, s, nr_elements, i) { 437 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL); 438 s->dma_address = 0; 439 s->dma_length = 0; 440 } 441 } 442 443 int zpci_dma_init_device(struct zpci_dev *zdev) 444 { 445 int rc; 446 447 /* 448 * At this point, if the device is part of an IOMMU domain, this would 449 * be a strong hint towards a bug in the IOMMU API (common) code and/or 450 * simultaneous access via IOMMU and DMA API. So let's issue a warning. 451 */ 452 WARN_ON(zdev->s390_domain); 453 454 spin_lock_init(&zdev->iommu_bitmap_lock); 455 spin_lock_init(&zdev->dma_table_lock); 456 457 zdev->dma_table = dma_alloc_cpu_table(); 458 if (!zdev->dma_table) { 459 rc = -ENOMEM; 460 goto out_clean; 461 } 462 463 /* 464 * Restrict the iommu bitmap size to the minimum of the following: 465 * - main memory size 466 * - 3-level pagetable address limit minus start_dma offset 467 * - DMA address range allowed by the hardware (clp query pci fn) 468 * 469 * Also set zdev->end_dma to the actual end address of the usable 470 * range, instead of the theoretical maximum as reported by hardware. 471 */ 472 zdev->iommu_size = min3((u64) high_memory, 473 ZPCI_TABLE_SIZE_RT - zdev->start_dma, 474 zdev->end_dma - zdev->start_dma + 1); 475 zdev->end_dma = zdev->start_dma + zdev->iommu_size - 1; 476 zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT; 477 zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8); 478 if (!zdev->iommu_bitmap) { 479 rc = -ENOMEM; 480 goto out_reg; 481 } 482 483 rc = zpci_register_ioat(zdev, 0, zdev->start_dma, zdev->end_dma, 484 (u64) zdev->dma_table); 485 if (rc) 486 goto out_reg; 487 return 0; 488 489 out_reg: 490 dma_free_cpu_table(zdev->dma_table); 491 out_clean: 492 return rc; 493 } 494 495 void zpci_dma_exit_device(struct zpci_dev *zdev) 496 { 497 /* 498 * At this point, if the device is part of an IOMMU domain, this would 499 * be a strong hint towards a bug in the IOMMU API (common) code and/or 500 * simultaneous access via IOMMU and DMA API. So let's issue a warning. 501 */ 502 WARN_ON(zdev->s390_domain); 503 504 zpci_unregister_ioat(zdev, 0); 505 dma_cleanup_tables(zdev->dma_table); 506 zdev->dma_table = NULL; 507 vfree(zdev->iommu_bitmap); 508 zdev->iommu_bitmap = NULL; 509 zdev->next_bit = 0; 510 } 511 512 static int __init dma_alloc_cpu_table_caches(void) 513 { 514 dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables", 515 ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN, 516 0, NULL); 517 if (!dma_region_table_cache) 518 return -ENOMEM; 519 520 dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables", 521 ZPCI_PT_SIZE, ZPCI_PT_ALIGN, 522 0, NULL); 523 if (!dma_page_table_cache) { 524 kmem_cache_destroy(dma_region_table_cache); 525 return -ENOMEM; 526 } 527 return 0; 528 } 529 530 int __init zpci_dma_init(void) 531 { 532 return dma_alloc_cpu_table_caches(); 533 } 534 535 void zpci_dma_exit(void) 536 { 537 kmem_cache_destroy(dma_page_table_cache); 538 kmem_cache_destroy(dma_region_table_cache); 539 } 540 541 #define PREALLOC_DMA_DEBUG_ENTRIES (1 << 16) 542 543 static int __init dma_debug_do_init(void) 544 { 545 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES); 546 return 0; 547 } 548 fs_initcall(dma_debug_do_init); 549 550 struct dma_map_ops s390_pci_dma_ops = { 551 .alloc = s390_dma_alloc, 552 .free = s390_dma_free, 553 .map_sg = s390_dma_map_sg, 554 .unmap_sg = s390_dma_unmap_sg, 555 .map_page = s390_dma_map_pages, 556 .unmap_page = s390_dma_unmap_pages, 557 /* if we support direct DMA this must be conditional */ 558 .is_phys = 0, 559 /* dma_supported is unconditionally true without a callback */ 560 }; 561 EXPORT_SYMBOL_GPL(s390_pci_dma_ops); 562 563 static int __init s390_iommu_setup(char *str) 564 { 565 if (!strncmp(str, "strict", 6)) 566 s390_iommu_strict = 1; 567 return 0; 568 } 569 570 __setup("s390_iommu=", s390_iommu_setup); 571