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/pci.h>
  14 #include <asm/pci_dma.h>
  15
  16 static enum zpci_ioat_dtype zpci_ioat_dt = ZPCI_IOTA_RTTO;
  17
  18 static struct kmem_cache *dma_region_table_cache;
  19 static struct kmem_cache *dma_page_table_cache;
  20
  21 static unsigned long *dma_alloc_cpu_table(void)
  22 {
  23         unsigned long *table, *entry;
  24
  25         table = kmem_cache_alloc(dma_region_table_cache, GFP_ATOMIC);
  26         if (!table)
  27                 return NULL;
  28
  29         for (entry = table; entry < table + ZPCI_TABLE_ENTRIES; entry++)
  30                 *entry = ZPCI_TABLE_INVALID | ZPCI_TABLE_PROTECTED;
  31         return table;
  32 }
  33
  34 static void dma_free_cpu_table(void *table)
  35 {
  36         kmem_cache_free(dma_region_table_cache, table);
  37 }
  38
  39 static unsigned long *dma_alloc_page_table(void)
  40 {
  41         unsigned long *table, *entry;
  42
  43         table = kmem_cache_alloc(dma_page_table_cache, GFP_ATOMIC);
  44         if (!table)
  45                 return NULL;
  46
  47         for (entry = table; entry < table + ZPCI_PT_ENTRIES; entry++)
  48                 *entry = ZPCI_PTE_INVALID | ZPCI_TABLE_PROTECTED;
  49         return table;
  50 }
  51
  52 static void dma_free_page_table(void *table)
  53 {
  54         kmem_cache_free(dma_page_table_cache, table);
  55 }
  56
  57 static unsigned long *dma_get_seg_table_origin(unsigned long *entry)
  58 {
  59         unsigned long *sto;
  60
  61         if (reg_entry_isvalid(*entry))
  62                 sto = get_rt_sto(*entry);
  63         else {
  64                 sto = dma_alloc_cpu_table();
  65                 if (!sto)
  66                         return NULL;
  67
  68                 set_rt_sto(entry, sto);
  69                 validate_rt_entry(entry);
  70                 entry_clr_protected(entry);
  71         }
  72         return sto;
  73 }
  74
  75 static unsigned long *dma_get_page_table_origin(unsigned long *entry)
  76 {
  77         unsigned long *pto;
  78
  79         if (reg_entry_isvalid(*entry))
  80                 pto = get_st_pto(*entry);
  81         else {
  82                 pto = dma_alloc_page_table();
  83                 if (!pto)
  84                         return NULL;
  85                 set_st_pto(entry, pto);
  86                 validate_st_entry(entry);
  87                 entry_clr_protected(entry);
  88         }
  89         return pto;
  90 }
  91
  92 static unsigned long *dma_walk_cpu_trans(unsigned long *rto, dma_addr_t dma_addr)
  93 {
  94         unsigned long *sto, *pto;
  95         unsigned int rtx, sx, px;
  96
  97         rtx = calc_rtx(dma_addr);
  98         sto = dma_get_seg_table_origin(&rto[rtx]);
  99         if (!sto)
 100                 return NULL;
 101
 102         sx = calc_sx(dma_addr);
 103         pto = dma_get_page_table_origin(&sto[sx]);
 104         if (!pto)
 105                 return NULL;
 106
 107         px = calc_px(dma_addr);
 108         return &pto[px];
 109 }
 110
 111 static void dma_update_cpu_trans(struct zpci_dev *zdev, void *page_addr,
 112                                  dma_addr_t dma_addr, int flags)
 113 {
 114         unsigned long *entry;
 115
 116         entry = dma_walk_cpu_trans(zdev->dma_table, dma_addr);
 117         if (!entry) {
 118                 WARN_ON_ONCE(1);
 119                 return;
 120         }
 121
 122         if (flags & ZPCI_PTE_INVALID) {
 123                 invalidate_pt_entry(entry);
 124                 return;
 125         } else {
 126                 set_pt_pfaa(entry, page_addr);
 127                 validate_pt_entry(entry);
 128         }
 129
 130         if (flags & ZPCI_TABLE_PROTECTED)
 131                 entry_set_protected(entry);
 132         else
 133                 entry_clr_protected(entry);
 134 }
 135
 136 static int dma_update_trans(struct zpci_dev *zdev, unsigned long pa,
 137                             dma_addr_t dma_addr, size_t size, int flags)
 138 {
 139         unsigned int nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
 140         u8 *page_addr = (u8 *) (pa & PAGE_MASK);
 141         dma_addr_t start_dma_addr = dma_addr;
 142         unsigned long irq_flags;
 143         int i, rc = 0;
 144
 145         if (!nr_pages)
 146                 return -EINVAL;
 147
 148         spin_lock_irqsave(&zdev->dma_table_lock, irq_flags);
 149         if (!zdev->dma_table) {
 150                 dev_err(&zdev->pdev->dev, "Missing DMA table\n");
 151                 goto no_refresh;
 152         }
 153
 154         for (i = 0; i < nr_pages; i++) {
 155                 dma_update_cpu_trans(zdev, page_addr, dma_addr, flags);
 156                 page_addr += PAGE_SIZE;
 157                 dma_addr += PAGE_SIZE;
 158         }
 159
 160         /*
 161          * rpcit is not required to establish new translations when previously
 162          * invalid translation-table entries are validated, however it is
 163          * required when altering previously valid entries.
 164          */
 165         if (!zdev->tlb_refresh &&
 166             ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID))
 167                 /*
 168                  * TODO: also need to check that the old entry is indeed INVALID
 169                  * and not only for one page but for the whole range...
 170                  * -> now we WARN_ON in that case but with lazy unmap that
 171                  * needs to be redone!
 172                  */
 173                 goto no_refresh;
 174         rc = rpcit_instr((u64) zdev->fh << 32, start_dma_addr,
 175                           nr_pages * PAGE_SIZE);
 176
 177 no_refresh:
 178         spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
 179         return rc;
 180 }
 181
 182 static void dma_free_seg_table(unsigned long entry)
 183 {
 184         unsigned long *sto = get_rt_sto(entry);
 185         int sx;
 186
 187         for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
 188                 if (reg_entry_isvalid(sto[sx]))
 189                         dma_free_page_table(get_st_pto(sto[sx]));
 190
 191         dma_free_cpu_table(sto);
 192 }
 193
 194 static void dma_cleanup_tables(struct zpci_dev *zdev)
 195 {
 196         unsigned long *table;
 197         int rtx;
 198
 199         if (!zdev || !zdev->dma_table)
 200                 return;
 201
 202         table = zdev->dma_table;
 203         for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
 204                 if (reg_entry_isvalid(table[rtx]))
 205                         dma_free_seg_table(table[rtx]);
 206
 207         dma_free_cpu_table(table);
 208         zdev->dma_table = NULL;
 209 }
 210
 211 static unsigned long __dma_alloc_iommu(struct zpci_dev *zdev, unsigned long start,
 212                                    int size)
 213 {
 214         unsigned long boundary_size = 0x1000000;
 215
 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(container_of(dev, struct 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         WARN_ON_ONCE(offset > PAGE_SIZE);
 274
 275         /* This rounds up number of pages based on size and offset */
 276         nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
 277         iommu_page_index = dma_alloc_iommu(zdev, nr_pages);
 278         if (iommu_page_index == -1)
 279                 goto out_err;
 280
 281         /* Use rounded up size */
 282         size = nr_pages * PAGE_SIZE;
 283
 284         dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE;
 285         if (dma_addr + size > zdev->end_dma) {
 286                 dev_err(dev, "(dma_addr: 0x%16.16LX + size: 0x%16.16lx) > end_dma: 0x%16.16Lx\n",
 287                          dma_addr, size, zdev->end_dma);
 288                 goto out_free;
 289         }
 290
 291         if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
 292                 flags |= ZPCI_TABLE_PROTECTED;
 293
 294         if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) {
 295                 atomic64_add(nr_pages, (atomic64_t *) &zdev->fmb->mapped_pages);
 296                 return dma_addr + offset;
 297         }
 298
 299 out_free:
 300         dma_free_iommu(zdev, iommu_page_index, nr_pages);
 301 out_err:
 302         dev_err(dev, "Failed to map addr: %lx\n", pa);
 303         return DMA_ERROR_CODE;
 304 }
 305
 306 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
 307                                  size_t size, enum dma_data_direction direction,
 308                                  struct dma_attrs *attrs)
 309 {
 310         struct zpci_dev *zdev = get_zdev(container_of(dev, struct pci_dev, dev));
 311         unsigned long iommu_page_index;
 312         int npages;
 313
 314         npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
 315         dma_addr = dma_addr & PAGE_MASK;
 316         if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
 317                              ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID))
 318                 dev_err(dev, "Failed to unmap addr: %Lx\n", dma_addr);
 319
 320         atomic64_add(npages, (atomic64_t *) &zdev->fmb->unmapped_pages);
 321         iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
 322         dma_free_iommu(zdev, iommu_page_index, npages);
 323 }
 324
 325 static void *s390_dma_alloc(struct device *dev, size_t size,
 326                             dma_addr_t *dma_handle, gfp_t flag,
 327                             struct dma_attrs *attrs)
 328 {
 329         struct zpci_dev *zdev = get_zdev(container_of(dev, struct pci_dev, dev));
 330         struct page *page;
 331         unsigned long pa;
 332         dma_addr_t map;
 333
 334         size = PAGE_ALIGN(size);
 335         page = alloc_pages(flag, get_order(size));
 336         if (!page)
 337                 return NULL;
 338
 339         atomic64_add(size / PAGE_SIZE, (atomic64_t *) &zdev->fmb->allocated_pages);
 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         if (dma_handle)
 351                 *dma_handle = map;
 352         return (void *) pa;
 353 }
 354
 355 static void s390_dma_free(struct device *dev, size_t size,
 356                           void *pa, dma_addr_t dma_handle,
 357                           struct dma_attrs *attrs)
 358 {
 359         s390_dma_unmap_pages(dev, dma_handle, PAGE_ALIGN(size),
 360                              DMA_BIDIRECTIONAL, NULL);
 361         free_pages((unsigned long) pa, get_order(size));
 362 }
 363
 364 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
 365                            int nr_elements, enum dma_data_direction dir,
 366                            struct dma_attrs *attrs)
 367 {
 368         int mapped_elements = 0;
 369         struct scatterlist *s;
 370         int i;
 371
 372         for_each_sg(sg, s, nr_elements, i) {
 373                 struct page *page = sg_page(s);
 374                 s->dma_address = s390_dma_map_pages(dev, page, s->offset,
 375                                                     s->length, dir, NULL);
 376                 if (!dma_mapping_error(dev, s->dma_address)) {
 377                         s->dma_length = s->length;
 378                         mapped_elements++;
 379                 } else
 380                         goto unmap;
 381         }
 382 out:
 383         return mapped_elements;
 384
 385 unmap:
 386         for_each_sg(sg, s, mapped_elements, i) {
 387                 if (s->dma_address)
 388                         s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
 389                                              dir, NULL);
 390                 s->dma_address = 0;
 391                 s->dma_length = 0;
 392         }
 393         mapped_elements = 0;
 394         goto out;
 395 }
 396
 397 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
 398                               int nr_elements, enum dma_data_direction dir,
 399                               struct dma_attrs *attrs)
 400 {
 401         struct scatterlist *s;
 402         int i;
 403
 404         for_each_sg(sg, s, nr_elements, i) {
 405                 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL);
 406                 s->dma_address = 0;
 407                 s->dma_length = 0;
 408         }
 409 }
 410
 411 int zpci_dma_init_device(struct zpci_dev *zdev)
 412 {
 413         unsigned int bitmap_order;
 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         bitmap_order = get_order(zdev->iommu_pages / 8);
 428         pr_info("iommu_size: 0x%lx  iommu_pages: 0x%lx  bitmap_order: %i\n",
 429                  zdev->iommu_size, zdev->iommu_pages, bitmap_order);
 430
 431         zdev->iommu_bitmap = (void *) __get_free_pages(GFP_KERNEL | __GFP_ZERO,
 432                                                        bitmap_order);
 433         if (!zdev->iommu_bitmap) {
 434                 rc = -ENOMEM;
 435                 goto out_reg;
 436         }
 437
 438         rc = zpci_register_ioat(zdev,
 439                                 0,
 440                                 zdev->start_dma + PAGE_OFFSET,
 441                                 zdev->start_dma + zdev->iommu_size - 1,
 442                                 (u64) zdev->dma_table);
 443         if (rc)
 444                 goto out_reg;
 445         return 0;
 446
 447 out_reg:
 448         dma_free_cpu_table(zdev->dma_table);
 449 out_clean:
 450         return rc;
 451 }
 452
 453 void zpci_dma_exit_device(struct zpci_dev *zdev)
 454 {
 455         zpci_unregister_ioat(zdev, 0);
 456         dma_cleanup_tables(zdev);
 457         free_pages((unsigned long) zdev->iommu_bitmap,
 458                    get_order(zdev->iommu_pages / 8));
 459         zdev->iommu_bitmap = NULL;
 460         zdev->next_bit = 0;
 461 }
 462
 463 static int __init dma_alloc_cpu_table_caches(void)
 464 {
 465         dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
 466                                         ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
 467                                         0, NULL);
 468         if (!dma_region_table_cache)
 469                 return -ENOMEM;
 470
 471         dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
 472                                         ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
 473                                         0, NULL);
 474         if (!dma_page_table_cache) {
 475                 kmem_cache_destroy(dma_region_table_cache);
 476                 return -ENOMEM;
 477         }
 478         return 0;
 479 }
 480
 481 int __init zpci_dma_init(void)
 482 {
 483         return dma_alloc_cpu_table_caches();
 484 }
 485
 486 void zpci_dma_exit(void)
 487 {
 488         kmem_cache_destroy(dma_page_table_cache);
 489         kmem_cache_destroy(dma_region_table_cache);
 490 }
 491
 492 #define PREALLOC_DMA_DEBUG_ENTRIES      (1 << 16)
 493
 494 static int __init dma_debug_do_init(void)
 495 {
 496         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 497         return 0;
 498 }
 499 fs_initcall(dma_debug_do_init);
 500
 501 struct dma_map_ops s390_dma_ops = {
 502         .alloc          = s390_dma_alloc,
 503         .free           = s390_dma_free,
 504         .map_sg         = s390_dma_map_sg,
 505         .unmap_sg       = s390_dma_unmap_sg,
 506         .map_page       = s390_dma_map_pages,
 507         .unmap_page     = s390_dma_unmap_pages,
 508         /* if we support direct DMA this must be conditional */
 509         .is_phys        = 0,
 510         /* dma_supported is unconditionally true without a callback */
 511 };
 512 EXPORT_SYMBOL_GPL(s390_dma_ops);