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 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 void dma_update_cpu_trans(unsigned long *dma_table, void *page_addr,
 118                           dma_addr_t dma_addr, int flags)
 119 {
 120         unsigned long *entry;
 121
 122         entry = dma_walk_cpu_trans(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->dma_table, page_addr, dma_addr,
 160                                      flags);
 161                 page_addr += PAGE_SIZE;
 162                 dma_addr += PAGE_SIZE;
 163         }
 164
 165         /*
 166          * With zdev->tlb_refresh == 0, rpcit is not required to establish new
 167          * translations when previously invalid translation-table entries are
 168          * validated. With lazy unmap, it also is skipped for previously valid
 169          * entries, but a global rpcit is then required before any address can
 170          * be re-used, i.e. after each iommu bitmap wrap-around.
 171          */
 172         if (!zdev->tlb_refresh &&
 173                         (!s390_iommu_strict ||
 174                         ((flags & ZPCI_PTE_VALID_MASK) == ZPCI_PTE_VALID)))
 175                 goto no_refresh;
 176
 177         rc = zpci_refresh_trans((u64) zdev->fh << 32, start_dma_addr,
 178                                 nr_pages * PAGE_SIZE);
 179
 180 no_refresh:
 181         spin_unlock_irqrestore(&zdev->dma_table_lock, irq_flags);
 182         return rc;
 183 }
 184
 185 void dma_free_seg_table(unsigned long entry)
 186 {
 187         unsigned long *sto = get_rt_sto(entry);
 188         int sx;
 189
 190         for (sx = 0; sx < ZPCI_TABLE_ENTRIES; sx++)
 191                 if (reg_entry_isvalid(sto[sx]))
 192                         dma_free_page_table(get_st_pto(sto[sx]));
 193
 194         dma_free_cpu_table(sto);
 195 }
 196
 197 void dma_cleanup_tables(unsigned long *table)
 198 {
 199         int rtx;
 200
 201         if (!table)
 202                 return;
 203
 204         for (rtx = 0; rtx < ZPCI_TABLE_ENTRIES; rtx++)
 205                 if (reg_entry_isvalid(table[rtx]))
 206                         dma_free_seg_table(table[rtx]);
 207
 208         dma_free_cpu_table(table);
 209 }
 210
 211 static unsigned long __dma_alloc_iommu(struct zpci_dev *zdev,
 212                                        unsigned long start, int size)
 213 {
 214         unsigned long boundary_size;
 215
 216         boundary_size = ALIGN(dma_get_seg_boundary(&zdev->pdev->dev) + 1,
 217                               PAGE_SIZE) >> PAGE_SHIFT;
 218         return iommu_area_alloc(zdev->iommu_bitmap, zdev->iommu_pages,
 219                                 start, size, 0, boundary_size, 0);
 220 }
 221
 222 static unsigned long dma_alloc_iommu(struct zpci_dev *zdev, int size)
 223 {
 224         unsigned long offset, flags;
 225         int wrap = 0;
 226
 227         spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
 228         offset = __dma_alloc_iommu(zdev, zdev->next_bit, size);
 229         if (offset == -1) {
 230                 /* wrap-around */
 231                 offset = __dma_alloc_iommu(zdev, 0, size);
 232                 wrap = 1;
 233         }
 234
 235         if (offset != -1) {
 236                 zdev->next_bit = offset + size;
 237                 if (!zdev->tlb_refresh && !s390_iommu_strict && wrap)
 238                         /* global flush after wrap-around with lazy unmap */
 239                         zpci_refresh_global(zdev);
 240         }
 241         spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
 242         return offset;
 243 }
 244
 245 static void dma_free_iommu(struct zpci_dev *zdev, unsigned long offset, int size)
 246 {
 247         unsigned long flags;
 248
 249         spin_lock_irqsave(&zdev->iommu_bitmap_lock, flags);
 250         if (!zdev->iommu_bitmap)
 251                 goto out;
 252         bitmap_clear(zdev->iommu_bitmap, offset, size);
 253         /*
 254          * Lazy flush for unmap: need to move next_bit to avoid address re-use
 255          * until wrap-around.
 256          */
 257         if (!s390_iommu_strict && offset >= zdev->next_bit)
 258                 zdev->next_bit = offset + size;
 259 out:
 260         spin_unlock_irqrestore(&zdev->iommu_bitmap_lock, flags);
 261 }
 262
 263 static dma_addr_t s390_dma_map_pages(struct device *dev, struct page *page,
 264                                      unsigned long offset, size_t size,
 265                                      enum dma_data_direction direction,
 266                                      struct dma_attrs *attrs)
 267 {
 268         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 269         unsigned long nr_pages, iommu_page_index;
 270         unsigned long pa = page_to_phys(page) + offset;
 271         int flags = ZPCI_PTE_VALID;
 272         dma_addr_t dma_addr;
 273
 274         /* This rounds up number of pages based on size and offset */
 275         nr_pages = iommu_num_pages(pa, size, PAGE_SIZE);
 276         iommu_page_index = dma_alloc_iommu(zdev, nr_pages);
 277         if (iommu_page_index == -1)
 278                 goto out_err;
 279
 280         /* Use rounded up size */
 281         size = nr_pages * PAGE_SIZE;
 282
 283         dma_addr = zdev->start_dma + iommu_page_index * PAGE_SIZE;
 284         if (dma_addr + size > zdev->end_dma)
 285                 goto out_free;
 286
 287         if (direction == DMA_NONE || direction == DMA_TO_DEVICE)
 288                 flags |= ZPCI_TABLE_PROTECTED;
 289
 290         if (!dma_update_trans(zdev, pa, dma_addr, size, flags)) {
 291                 atomic64_add(nr_pages, &zdev->mapped_pages);
 292                 return dma_addr + (offset & ~PAGE_MASK);
 293         }
 294
 295 out_free:
 296         dma_free_iommu(zdev, iommu_page_index, nr_pages);
 297 out_err:
 298         zpci_err("map error:\n");
 299         zpci_err_hex(&pa, sizeof(pa));
 300         return DMA_ERROR_CODE;
 301 }
 302
 303 static void s390_dma_unmap_pages(struct device *dev, dma_addr_t dma_addr,
 304                                  size_t size, enum dma_data_direction direction,
 305                                  struct dma_attrs *attrs)
 306 {
 307         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 308         unsigned long iommu_page_index;
 309         int npages;
 310
 311         npages = iommu_num_pages(dma_addr, size, PAGE_SIZE);
 312         dma_addr = dma_addr & PAGE_MASK;
 313         if (dma_update_trans(zdev, 0, dma_addr, npages * PAGE_SIZE,
 314                              ZPCI_TABLE_PROTECTED | ZPCI_PTE_INVALID)) {
 315                 zpci_err("unmap error:\n");
 316                 zpci_err_hex(&dma_addr, sizeof(dma_addr));
 317         }
 318
 319         atomic64_add(npages, &zdev->unmapped_pages);
 320         iommu_page_index = (dma_addr - zdev->start_dma) >> PAGE_SHIFT;
 321         dma_free_iommu(zdev, iommu_page_index, npages);
 322 }
 323
 324 static void *s390_dma_alloc(struct device *dev, size_t size,
 325                             dma_addr_t *dma_handle, gfp_t flag,
 326                             struct dma_attrs *attrs)
 327 {
 328         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 329         struct page *page;
 330         unsigned long pa;
 331         dma_addr_t map;
 332
 333         size = PAGE_ALIGN(size);
 334         page = alloc_pages(flag, get_order(size));
 335         if (!page)
 336                 return NULL;
 337
 338         pa = page_to_phys(page);
 339         memset((void *) pa, 0, size);
 340
 341         map = s390_dma_map_pages(dev, page, pa % PAGE_SIZE,
 342                                  size, DMA_BIDIRECTIONAL, NULL);
 343         if (dma_mapping_error(dev, map)) {
 344                 free_pages(pa, get_order(size));
 345                 return NULL;
 346         }
 347
 348         atomic64_add(size / PAGE_SIZE, &zdev->allocated_pages);
 349         if (dma_handle)
 350                 *dma_handle = map;
 351         return (void *) pa;
 352 }
 353
 354 static void s390_dma_free(struct device *dev, size_t size,
 355                           void *pa, dma_addr_t dma_handle,
 356                           struct dma_attrs *attrs)
 357 {
 358         struct zpci_dev *zdev = to_zpci(to_pci_dev(dev));
 359
 360         size = PAGE_ALIGN(size);
 361         atomic64_sub(size / PAGE_SIZE, &zdev->allocated_pages);
 362         s390_dma_unmap_pages(dev, dma_handle, size, DMA_BIDIRECTIONAL, NULL);
 363         free_pages((unsigned long) pa, get_order(size));
 364 }
 365
 366 static int s390_dma_map_sg(struct device *dev, struct scatterlist *sg,
 367                            int nr_elements, enum dma_data_direction dir,
 368                            struct dma_attrs *attrs)
 369 {
 370         int mapped_elements = 0;
 371         struct scatterlist *s;
 372         int i;
 373
 374         for_each_sg(sg, s, nr_elements, i) {
 375                 struct page *page = sg_page(s);
 376                 s->dma_address = s390_dma_map_pages(dev, page, s->offset,
 377                                                     s->length, dir, NULL);
 378                 if (!dma_mapping_error(dev, s->dma_address)) {
 379                         s->dma_length = s->length;
 380                         mapped_elements++;
 381                 } else
 382                         goto unmap;
 383         }
 384 out:
 385         return mapped_elements;
 386
 387 unmap:
 388         for_each_sg(sg, s, mapped_elements, i) {
 389                 if (s->dma_address)
 390                         s390_dma_unmap_pages(dev, s->dma_address, s->dma_length,
 391                                              dir, NULL);
 392                 s->dma_address = 0;
 393                 s->dma_length = 0;
 394         }
 395         mapped_elements = 0;
 396         goto out;
 397 }
 398
 399 static void s390_dma_unmap_sg(struct device *dev, struct scatterlist *sg,
 400                               int nr_elements, enum dma_data_direction dir,
 401                               struct dma_attrs *attrs)
 402 {
 403         struct scatterlist *s;
 404         int i;
 405
 406         for_each_sg(sg, s, nr_elements, i) {
 407                 s390_dma_unmap_pages(dev, s->dma_address, s->dma_length, dir, NULL);
 408                 s->dma_address = 0;
 409                 s->dma_length = 0;
 410         }
 411 }
 412
 413 int zpci_dma_init_device(struct zpci_dev *zdev)
 414 {
 415         int rc;
 416
 417         /*
 418          * At this point, if the device is part of an IOMMU domain, this would
 419          * be a strong hint towards a bug in the IOMMU API (common) code and/or
 420          * simultaneous access via IOMMU and DMA API. So let's issue a warning.
 421          */
 422         WARN_ON(zdev->s390_domain);
 423
 424         spin_lock_init(&zdev->iommu_bitmap_lock);
 425         spin_lock_init(&zdev->dma_table_lock);
 426
 427         zdev->dma_table = dma_alloc_cpu_table();
 428         if (!zdev->dma_table) {
 429                 rc = -ENOMEM;
 430                 goto out_clean;
 431         }
 432
 433         zdev->iommu_size = (unsigned long) high_memory - PAGE_OFFSET;
 434         zdev->iommu_pages = zdev->iommu_size >> PAGE_SHIFT;
 435         zdev->iommu_bitmap = vzalloc(zdev->iommu_pages / 8);
 436         if (!zdev->iommu_bitmap) {
 437                 rc = -ENOMEM;
 438                 goto out_reg;
 439         }
 440
 441         rc = zpci_register_ioat(zdev,
 442                                 0,
 443                                 zdev->start_dma + PAGE_OFFSET,
 444                                 zdev->start_dma + zdev->iommu_size - 1,
 445                                 (u64) zdev->dma_table);
 446         if (rc)
 447                 goto out_reg;
 448         return 0;
 449
 450 out_reg:
 451         dma_free_cpu_table(zdev->dma_table);
 452 out_clean:
 453         return rc;
 454 }
 455
 456 void zpci_dma_exit_device(struct zpci_dev *zdev)
 457 {
 458         /*
 459          * At this point, if the device is part of an IOMMU domain, this would
 460          * be a strong hint towards a bug in the IOMMU API (common) code and/or
 461          * simultaneous access via IOMMU and DMA API. So let's issue a warning.
 462          */
 463         WARN_ON(zdev->s390_domain);
 464
 465         zpci_unregister_ioat(zdev, 0);
 466         dma_cleanup_tables(zdev->dma_table);
 467         zdev->dma_table = NULL;
 468         vfree(zdev->iommu_bitmap);
 469         zdev->iommu_bitmap = NULL;
 470         zdev->next_bit = 0;
 471 }
 472
 473 static int __init dma_alloc_cpu_table_caches(void)
 474 {
 475         dma_region_table_cache = kmem_cache_create("PCI_DMA_region_tables",
 476                                         ZPCI_TABLE_SIZE, ZPCI_TABLE_ALIGN,
 477                                         0, NULL);
 478         if (!dma_region_table_cache)
 479                 return -ENOMEM;
 480
 481         dma_page_table_cache = kmem_cache_create("PCI_DMA_page_tables",
 482                                         ZPCI_PT_SIZE, ZPCI_PT_ALIGN,
 483                                         0, NULL);
 484         if (!dma_page_table_cache) {
 485                 kmem_cache_destroy(dma_region_table_cache);
 486                 return -ENOMEM;
 487         }
 488         return 0;
 489 }
 490
 491 int __init zpci_dma_init(void)
 492 {
 493         return dma_alloc_cpu_table_caches();
 494 }
 495
 496 void zpci_dma_exit(void)
 497 {
 498         kmem_cache_destroy(dma_page_table_cache);
 499         kmem_cache_destroy(dma_region_table_cache);
 500 }
 501
 502 #define PREALLOC_DMA_DEBUG_ENTRIES      (1 << 16)
 503
 504 static int __init dma_debug_do_init(void)
 505 {
 506         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 507         return 0;
 508 }
 509 fs_initcall(dma_debug_do_init);
 510
 511 struct dma_map_ops s390_dma_ops = {
 512         .alloc          = s390_dma_alloc,
 513         .free           = s390_dma_free,
 514         .map_sg         = s390_dma_map_sg,
 515         .unmap_sg       = s390_dma_unmap_sg,
 516         .map_page       = s390_dma_map_pages,
 517         .unmap_page     = s390_dma_unmap_pages,
 518         /* if we support direct DMA this must be conditional */
 519         .is_phys        = 0,
 520         /* dma_supported is unconditionally true without a callback */
 521 };
 522 EXPORT_SYMBOL_GPL(s390_dma_ops);
 523
 524 static int __init s390_iommu_setup(char *str)
 525 {
 526         if (!strncmp(str, "strict", 6))
 527                 s390_iommu_strict = 1;
 528         return 0;
 529 }
 530
 531 __setup("s390_iommu=", s390_iommu_setup);