arch/s390/pci/pci_dma.c

   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 static 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 | ZPCI_TABLE_PROTECTED;
  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 | ZPCI_TABLE_PROTECTED;
  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 static 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 static void dma_update_cpu_trans(struct zpci_dev *zdev, void *page_addr,
 118                                  dma_addr_t dma_addr, int flags)
 119 {
 120         unsigned long *entry;
 121
 122         entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
 123         if (!entry) {
 124                 WARN_ON_ONCE(1);
 125                 return;
 126         }
 127
 128         if (flags & ZPCI_PTE_INVALID) {
 129                 invalidate_pt_entry(entry);
 130                 return;
 131         } else {
 132                 set_pt_pfaa(entry, page_addr);
 133                 validate_pt_entry(entry);
 134         }
 135
 136         if (flags & ZPCI_TABLE_PROTECTED)
 137                 entry_set_protected(entry);
 138         else
 139                 entry_clr_protected(entry);
 140 }
 141
 142 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
 143                             dma_addr_t dma_addr, size_t size, int flags)
 144 {
 145         unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
 146         u8 *page_addr = (u8 *) (pa & PAGE_MASK);
 147         dma_addr_t start_dma_addr = dma_addr;
 148         unsigned long irq_flags;
 149         int i, rc = 0;
 150
 151         if (!nr_pages)
 152                 return -EINVAL;
 153
 154         spin_lock_irqsave(&zdev->dma_table_lock, irq_flags);
 155         if (!zdev->dma_table)
 156                 goto no_refresh;
 157
 158         for (i = 0; i < nr_pages; i++) {
 159                 dma_update_cpu_trans(zdev, page_addr, dma_addr, flags);
 160                 page_addr += PAGE_SIZE;
 161                 dma_addr += PAGE_SIZE;
 162         }
 163
 164         /*
 165          * With zdev->tlb_refresh == 0, rpcit is not required to establish new
 166          * translations when previously invalid translation-table entries are
 167          * validated. With lazy unmap, it also is skipped for previously valid
 168          * entries, but a global rpcit is then required before any address can
 169          * be re-used, i.e. after each iommu bitmap wrap-around.
 170          */
 171         if (!zdev->tlb_refresh &&
 172                         (!s390_iommu_strict ||
 173                         ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)))
 174                 goto no_refresh;
 175
 176         rc = zpci_refresh_trans((u64) zdev->fh << 32, start_dma_addr,
 177                                 nr_pages * PAGE_SIZE);
 178
 179 no_refresh:
 180         spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
 181         return rc;
 182 }
 183
 184 static void dma_free_seg_table(unsigned long entry)
 185 {
 186         unsigned long *sto = get_rt_sto(entry);
 187         int sx;
 188
 189         for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
 190                 if (reg_entry_isvalid(sto[sx]))
 191                         dma_free_page_table(get_st_pto(sto[sx]));
 192
 193         dma_free_cpu_table(sto);
 194 }
 195
 196 static void dma_cleanup_tables(struct zpci_dev *zdev)
 197 {
 198         unsigned long *table;
 199         int rtx;
 200
 201         if (!zdev || !zdev->dma_table)
 202                 return;
 203
 204         table = zdev->dma_table;
 205         for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
 206                 if (reg_entry_isvalid(table[rtx]))
 207                         dma_free_seg_table(table[rtx]);
 208
 209         dma_free_cpu_table(table);
 210         zdev->dma_table = NULL;
 211 }
 212
 213 static unsigned long __dma_alloc_iommu(struct zpci_dev *zdev,
 214                                        unsigned long start, int size)
 215 {
 216         unsigned long boundary_size;
 217
 218         boundary_size = ALIGN(dma_get_seg_boundary(&zdev->pdev->dev) + 1,
 219                               PAGE_SIZE) >> PAGE_SHIFT;
 220         return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages,
 221                                 start, size, 0, boundary_size, 0);
 222 }
 223
 224 static unsigned long dma_alloc_iommu(struct zpci_dev *zdev, int size)
 225 {
 226         unsigned long offset, flags;
 227         int wrap = 0;
 228
 229         spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
 230         offset = __dma_alloc_iommu(zdev, zdev->next_bit, size);
 231         if (offset == -1) {
 232                 /* wrap-around */
 233                 offset = __dma_alloc_iommu(zdev, 0, size);
 234                 wrap = 1;
 235         }
 236
 237         if (offset != -1) {
 238                 zdev->next_bit = offset + size;
 239                 if (!zdev->tlb_refresh && !s390_iommu_strict && wrap)
 240                         /* global flush after wrap-around with lazy unmap */
 241                         zpci_refresh_global(zdev);
 242         }
 243         spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
 244         return offset;
 245 }
 246
 247 static void dma_free_iommu(struct zpci_dev *zdev, unsigned long offset, int size)
 248 {
 249         unsigned long flags;
 250
 251         spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
 252         if (!zdev->iommu_bitmap)
 253                 goto out;
 254         bitmap_clear(zdev->iommu_bitmap, offset, size);
 255         /*
 256          * Lazy flush for unmap: need to move next_bit to avoid address re-use
 257          * until wrap-around.
 258          */
 259         if (!s390_iommu_strict && offset >= zdev->next_bit)
 260                 zdev->next_bit = offset + size;
 261 out:
 262         spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
 263 }
 264
 265 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page,
 266                                      unsigned long offset, size_t size,
 267                                      enum dma_data_direction direction,
 268                                      struct dma_attrs *attrs)
 269 {
 270         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 271         unsigned long nr_pages, iommu_page_index;
 272         unsigned long pa = page_to_phys(page) + offset;
 273         int flags = ZPCI_PTE_VALID;
 274         dma_addr_t dma_addr;
 275
 276         /* This rounds up number of pages based on size and offset */
 277         nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
 278         iommu_page_index = dma_alloc_iommu(zdev, nr_pages);
 279         if (iommu_page_index == -1)
 280                 goto out_err;
 281
 282         /* Use rounded up size */
 283         size = nr_pages * PAGE_SIZE;
 284
 285         dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE;
 286         if (dma_addr + size > zdev->end_dma)
 287                 goto out_free;
 288
 289         if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
 290                 flags |= ZPCI_TABLE_PROTECTED;
 291
 292         if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) {
 293                 atomic64_add(nr_pages, &zdev->mapped_pages);
 294                 return dma_addr + (offset & ~PAGE_MASK);
 295         }
 296
 297 out_free:
 298         dma_free_iommu(zdev, iommu_page_index, nr_pages);
 299 out_err:
 300         zpci_err("map error:\n");
 301         zpci_err_hex(&pa, sizeof(pa));
 302         return DMA_ERROR_CODE;
 303 }
 304
 305 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
 306                                  size_t size, enum dma_data_direction direction,
 307                                  struct dma_attrs *attrs)
 308 {
 309         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 310         unsigned long iommu_page_index;
 311         int npages;
 312
 313         npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
 314         dma_addr = dma_addr & PAGE_MASK;
 315         if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
 316                              ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID)) {
 317                 zpci_err("unmap error:\n");
 318                 zpci_err_hex(&dma_addr, sizeof(dma_addr));
 319         }
 320
 321         atomic64_add(npages, &zdev->unmapped_pages);
 322         iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
 323         dma_free_iommu(zdev, iommu_page_index, npages);
 324 }
 325
 326 static void *s390_dma_alloc(struct device *dev, size_t size,
 327                             dma_addr_t *dma_handle, gfp_t flag,
 328                             struct dma_attrs *attrs)
 329 {
 330         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 331         struct page *page;
 332         unsigned long pa;
 333         dma_addr_t map;
 334
 335         size = PAGE_ALIGN(size);
 336         page = alloc_pages(flag, get_order(size));
 337         if (!page)
 338                 return NULL;
 339
 340         pa = page_to_phys(page);
 341         memset((void *) pa, 0, size);
 342
 343         map = s390_dma_map_pages(dev, page, pa % PAGE_SIZE,
 344                                  size, DMA_BIDIRECTIONAL, NULL);
 345         if (dma_mapping_error(dev, map)) {
 346                 free_pages(pa, get_order(size));
 347                 return NULL;
 348         }
 349
 350         atomic64_add(size / PAGE_SIZE, &zdev->allocated_pages);
 351         if (dma_handle)
 352                 *dma_handle = map;
 353         return (void *) pa;
 354 }
 355
 356 static void s390_dma_free(struct device *dev, size_t size,
 357                           void *pa, dma_addr_t dma_handle,
 358                           struct dma_attrs *attrs)
 359 {
 360         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 361
 362         size = PAGE_ALIGN(size);
 363         atomic64_sub(size / PAGE_SIZE, &zdev->allocated_pages);
 364         s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, NULL);
 365         free_pages((unsigned long) pa, get_order(size));
 366 }
 367
 368 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
 369                            int nr_elements, enum dma_data_direction dir,
 370                            struct dma_attrs *attrs)
 371 {
 372         int mapped_elements = 0;
 373         struct scatterlist *s;
 374         int i;
 375
 376         for_each_sg(sg, s, nr_elements, i) {
 377                 struct page *page = sg_page(s);
 378                 s->dma_address = s390_dma_map_pages(dev, page, s->offset,
 379                                                     s->length, dir, NULL);
 380                 if (!dma_mapping_error(dev, s->dma_address)) {
 381                         s->dma_length = s->length;
 382                         mapped_elements++;
 383                 } else
 384                         goto unmap;
 385         }
 386 out:
 387         return mapped_elements;
 388
 389 unmap:
 390         for_each_sg(sg, s, mapped_elements, i) {
 391                 if (s->dma_address)
 392                         s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
 393                                              dir, NULL);
 394                 s->dma_address = 0;
 395                 s->dma_length = 0;
 396         }
 397         mapped_elements = 0;
 398         goto out;
 399 }
 400
 401 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
 402                               int nr_elements, enum dma_data_direction dir,
 403                               struct dma_attrs *attrs)
 404 {
 405         struct scatterlist *s;
 406         int i;
 407
 408         for_each_sg(sg, s, nr_elements, i) {
 409                 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL);
 410                 s->dma_address = 0;
 411                 s->dma_length = 0;
 412         }
 413 }
 414
 415 int zpci_dma_init_device(struct zpci_dev *zdev)
 416 {
 417         int rc;
 418
 419         spin_lock_init(&zdev->iommu_bitmap_lock);
 420         spin_lock_init(&zdev->dma_table_lock);
 421
 422         zdev->dma_table = dma_alloc_cpu_table();
 423         if (!zdev->dma_table) {
 424                 rc = -ENOMEM;
 425                 goto out_clean;
 426         }
 427
 428         zdev->iommu_size = (unsigned long) high_memory - PAGE_OFFSET;
 429         zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
 430         zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8);
 431         if (!zdev->iommu_bitmap) {
 432                 rc = -ENOMEM;
 433                 goto out_reg;
 434         }
 435
 436         rc = zpci_register_ioat(zdev,
 437                                 0,
 438                                 zdev->start_dma + PAGE_OFFSET,
 439                                 zdev->start_dma + zdev->iommu_size - 1,
 440                                 (u64) zdev->dma_table);
 441         if (rc)
 442                 goto out_reg;
 443         return 0;
 444
 445 out_reg:
 446         dma_free_cpu_table(zdev->dma_table);
 447 out_clean:
 448         return rc;
 449 }
 450
 451 void zpci_dma_exit_device(struct zpci_dev *zdev)
 452 {
 453         zpci_unregister_ioat(zdev, 0);
 454         dma_cleanup_tables(zdev);
 455         vfree(zdev->iommu_bitmap);
 456         zdev->iommu_bitmap = NULL;
 457         zdev->next_bit = 0;
 458 }
 459
 460 static int __init dma_alloc_cpu_table_caches(void)
 461 {
 462         dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
 463                                         ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
 464                                         0, NULL);
 465         if (!dma_region_table_cache)
 466                 return -ENOMEM;
 467
 468         dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
 469                                         ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
 470                                         0, NULL);
 471         if (!dma_page_table_cache) {
 472                 kmem_cache_destroy(dma_region_table_cache);
 473                 return -ENOMEM;
 474         }
 475         return 0;
 476 }
 477
 478 int __init zpci_dma_init(void)
 479 {
 480         return dma_alloc_cpu_table_caches();
 481 }
 482
 483 void zpci_dma_exit(void)
 484 {
 485         kmem_cache_destroy(dma_page_table_cache);
 486         kmem_cache_destroy(dma_region_table_cache);
 487 }
 488
 489 #define PREALLOC_DMA_DEBUG_ENTRIES      (1 << 16)
 490
 491 static int __init dma_debug_do_init(void)
 492 {
 493         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 494         return 0;
 495 }
 496 fs_initcall(dma_debug_do_init);
 497
 498 struct dma_map_ops s390_dma_ops = {
 499         .alloc          = s390_dma_alloc,
 500         .free           = s390_dma_free,
 501         .map_sg         = s390_dma_map_sg,
 502         .unmap_sg       = s390_dma_unmap_sg,
 503         .map_page       = s390_dma_map_pages,
 504         .unmap_page     = s390_dma_unmap_pages,
 505         /* if we support direct DMA this must be conditional */
 506         .is_phys        = 0,
 507         /* dma_supported is unconditionally true without a callback */
 508 };
 509 EXPORT_SYMBOL_GPL(s390_dma_ops);
 510
 511 static int __init s390_iommu_setup(char *str)
 512 {
 513         if (!strncmp(str, "strict", 6))
 514                 s390_iommu_strict = 1;
 515         return 0;
 516 }
 517
 518 __setup("s390_iommu=", s390_iommu_setup);