drivers/char/agp/amd-k7-agp.c

   1 /*
   2  * AMD K7 AGPGART routines.
   3  */
   4
   5 #include <linux/module.h>
   6 #include <linux/pci.h>
   7 #include <linux/init.h>
   8 #include <linux/agp_backend.h>
   9 #include <linux/gfp.h>
  10 #include <linux/page-flags.h>
  11 #include <linux/mm.h>
  12 #include "agp.h"
  13
  14 #define AMD_MMBASE      0x14
  15 #define AMD_APSIZE      0xac
  16 #define AMD_MODECNTL    0xb0
  17 #define AMD_MODECNTL2   0xb2
  18 #define AMD_GARTENABLE  0x02    /* In mmio region (16-bit register) */
  19 #define AMD_ATTBASE     0x04    /* In mmio region (32-bit register) */
  20 #define AMD_TLBFLUSH    0x0c    /* In mmio region (32-bit register) */
  21 #define AMD_CACHEENTRY  0x10    /* In mmio region (32-bit register) */
  22
  23 static struct pci_device_id agp_amdk7_pci_table[];
  24
  25 struct amd_page_map {
  26         unsigned long *real;
  27         unsigned long __iomem *remapped;
  28 };
  29
  30 static struct _amd_irongate_private {
  31         volatile u8 __iomem *registers;
  32         struct amd_page_map **gatt_pages;
  33         int num_tables;
  34 } amd_irongate_private;
  35
  36 static int amd_create_page_map(struct amd_page_map *page_map)
  37 {
  38         int i;
  39
  40         page_map->real = (unsigned long *) __get_free_page(GFP_KERNEL);
  41         if (page_map->real == NULL)
  42                 return -ENOMEM;
  43
  44         SetPageReserved(virt_to_page(page_map->real));
  45         global_cache_flush();
  46         page_map->remapped = ioremap_nocache(virt_to_gart(page_map->real),
  47                                             PAGE_SIZE);
  48         if (page_map->remapped == NULL) {
  49                 ClearPageReserved(virt_to_page(page_map->real));
  50                 free_page((unsigned long) page_map->real);
  51                 page_map->real = NULL;
  52                 return -ENOMEM;
  53         }
  54         global_cache_flush();
  55
  56         for (i = 0; i < PAGE_SIZE / sizeof(unsigned long); i++) {
  57                 writel(agp_bridge->scratch_page, page_map->remapped+i);
  58                 readl(page_map->remapped+i);    /* PCI Posting. */
  59         }
  60
  61         return 0;
  62 }
  63
  64 static void amd_free_page_map(struct amd_page_map *page_map)
  65 {
  66         iounmap(page_map->remapped);
  67         ClearPageReserved(virt_to_page(page_map->real));
  68         free_page((unsigned long) page_map->real);
  69 }
  70
  71 static void amd_free_gatt_pages(void)
  72 {
  73         int i;
  74         struct amd_page_map **tables;
  75         struct amd_page_map *entry;
  76
  77         tables = amd_irongate_private.gatt_pages;
  78         for (i = 0; i < amd_irongate_private.num_tables; i++) {
  79                 entry = tables[i];
  80                 if (entry != NULL) {
  81                         if (entry->real != NULL)
  82                                 amd_free_page_map(entry);
  83                         kfree(entry);
  84                 }
  85         }
  86         kfree(tables);
  87         amd_irongate_private.gatt_pages = NULL;
  88 }
  89
  90 static int amd_create_gatt_pages(int nr_tables)
  91 {
  92         struct amd_page_map **tables;
  93         struct amd_page_map *entry;
  94         int retval = 0;
  95         int i;
  96
  97         tables = kzalloc((nr_tables + 1) * sizeof(struct amd_page_map *),GFP_KERNEL);
  98         if (tables == NULL)
  99                 return -ENOMEM;
 100
 101         for (i = 0; i < nr_tables; i++) {
 102                 entry = kzalloc(sizeof(struct amd_page_map), GFP_KERNEL);
 103                 if (entry == NULL) {
 104                         while (i > 0) {
 105                                 kfree(tables[i-1]);
 106                                 i--;
 107                         }
 108                         kfree(tables);
 109                         retval = -ENOMEM;
 110                         break;
 111                 }
 112                 tables[i] = entry;
 113                 retval = amd_create_page_map(entry);
 114                 if (retval != 0)
 115                         break;
 116         }
 117         amd_irongate_private.num_tables = nr_tables;
 118         amd_irongate_private.gatt_pages = tables;
 119
 120         if (retval != 0)
 121                 amd_free_gatt_pages();
 122
 123         return retval;
 124 }
 125
 126 /* Since we don't need contiguous memory we just try
 127  * to get the gatt table once
 128  */
 129
 130 #define GET_PAGE_DIR_OFF(addr) (addr >> 22)
 131 #define GET_PAGE_DIR_IDX(addr) (GET_PAGE_DIR_OFF(addr) - \
 132         GET_PAGE_DIR_OFF(agp_bridge->gart_bus_addr))
 133 #define GET_GATT_OFF(addr) ((addr & 0x003ff000) >> 12)
 134 #define GET_GATT(addr) (amd_irongate_private.gatt_pages[\
 135         GET_PAGE_DIR_IDX(addr)]->remapped)
 136
 137 static int amd_create_gatt_table(struct agp_bridge_data *bridge)
 138 {
 139         struct aper_size_info_lvl2 *value;
 140         struct amd_page_map page_dir;
 141         unsigned long addr;
 142         int retval;
 143         u32 temp;
 144         int i;
 145
 146         value = A_SIZE_LVL2(agp_bridge->current_size);
 147         retval = amd_create_page_map(&page_dir);
 148         if (retval != 0)
 149                 return retval;
 150
 151         retval = amd_create_gatt_pages(value->num_entries / 1024);
 152         if (retval != 0) {
 153                 amd_free_page_map(&page_dir);
 154                 return retval;
 155         }
 156
 157         agp_bridge->gatt_table_real = (u32 *)page_dir.real;
 158         agp_bridge->gatt_table = (u32 __iomem *)page_dir.remapped;
 159         agp_bridge->gatt_bus_addr = virt_to_gart(page_dir.real);
 160
 161         /* Get the address for the gart region.
 162          * This is a bus address even on the alpha, b/c its
 163          * used to program the agp master not the cpu
 164          */
 165
 166         pci_read_config_dword(agp_bridge->dev, AGP_APBASE, &temp);
 167         addr = (temp & PCI_BASE_ADDRESS_MEM_MASK);
 168         agp_bridge->gart_bus_addr = addr;
 169
 170         /* Calculate the agp offset */
 171         for (i = 0; i < value->num_entries / 1024; i++, addr += 0x00400000) {
 172                 writel(virt_to_gart(amd_irongate_private.gatt_pages[i]->real) | 1,
 173                         page_dir.remapped+GET_PAGE_DIR_OFF(addr));
 174                 readl(page_dir.remapped+GET_PAGE_DIR_OFF(addr));        /* PCI Posting. */
 175         }
 176
 177         return 0;
 178 }
 179
 180 static int amd_free_gatt_table(struct agp_bridge_data *bridge)
 181 {
 182         struct amd_page_map page_dir;
 183
 184         page_dir.real = (unsigned long *)agp_bridge->gatt_table_real;
 185         page_dir.remapped = (unsigned long __iomem *)agp_bridge->gatt_table;
 186
 187         amd_free_gatt_pages();
 188         amd_free_page_map(&page_dir);
 189         return 0;
 190 }
 191
 192 static int amd_irongate_fetch_size(void)
 193 {
 194         int i;
 195         u32 temp;
 196         struct aper_size_info_lvl2 *values;
 197
 198         pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
 199         temp = (temp & 0x0000000e);
 200         values = A_SIZE_LVL2(agp_bridge->driver->aperture_sizes);
 201         for (i = 0; i < agp_bridge->driver->num_aperture_sizes; i++) {
 202                 if (temp == values[i].size_value) {
 203                         agp_bridge->previous_size =
 204                             agp_bridge->current_size = (void *) (values + i);
 205
 206                         agp_bridge->aperture_size_idx = i;
 207                         return values[i].size;
 208                 }
 209         }
 210
 211         return 0;
 212 }
 213
 214 static int amd_irongate_configure(void)
 215 {
 216         struct aper_size_info_lvl2 *current_size;
 217         u32 temp;
 218         u16 enable_reg;
 219
 220         current_size = A_SIZE_LVL2(agp_bridge->current_size);
 221
 222         /* Get the memory mapped registers */
 223         pci_read_config_dword(agp_bridge->dev, AMD_MMBASE, &temp);
 224         temp = (temp & PCI_BASE_ADDRESS_MEM_MASK);
 225         amd_irongate_private.registers = (volatile u8 __iomem *) ioremap(temp, 4096);
 226
 227         /* Write out the address of the gatt table */
 228         writel(agp_bridge->gatt_bus_addr, amd_irongate_private.registers+AMD_ATTBASE);
 229         readl(amd_irongate_private.registers+AMD_ATTBASE);      /* PCI Posting. */
 230
 231         /* Write the Sync register */
 232         pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL, 0x80);
 233
 234         /* Set indexing mode */
 235         pci_write_config_byte(agp_bridge->dev, AMD_MODECNTL2, 0x00);
 236
 237         /* Write the enable register */
 238         enable_reg = readw(amd_irongate_private.registers+AMD_GARTENABLE);
 239         enable_reg = (enable_reg | 0x0004);
 240         writew(enable_reg, amd_irongate_private.registers+AMD_GARTENABLE);
 241         readw(amd_irongate_private.registers+AMD_GARTENABLE);   /* PCI Posting. */
 242
 243         /* Write out the size register */
 244         pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
 245         temp = (((temp & ~(0x0000000e)) | current_size->size_value) | 1);
 246         pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
 247
 248         /* Flush the tlb */
 249         writel(1, amd_irongate_private.registers+AMD_TLBFLUSH);
 250         readl(amd_irongate_private.registers+AMD_TLBFLUSH);     /* PCI Posting.*/
 251         return 0;
 252 }
 253
 254 static void amd_irongate_cleanup(void)
 255 {
 256         struct aper_size_info_lvl2 *previous_size;
 257         u32 temp;
 258         u16 enable_reg;
 259
 260         previous_size = A_SIZE_LVL2(agp_bridge->previous_size);
 261
 262         enable_reg = readw(amd_irongate_private.registers+AMD_GARTENABLE);
 263         enable_reg = (enable_reg & ~(0x0004));
 264         writew(enable_reg, amd_irongate_private.registers+AMD_GARTENABLE);
 265         readw(amd_irongate_private.registers+AMD_GARTENABLE);   /* PCI Posting. */
 266
 267         /* Write back the previous size and disable gart translation */
 268         pci_read_config_dword(agp_bridge->dev, AMD_APSIZE, &temp);
 269         temp = ((temp & ~(0x0000000f)) | previous_size->size_value);
 270         pci_write_config_dword(agp_bridge->dev, AMD_APSIZE, temp);
 271         iounmap((void __iomem *) amd_irongate_private.registers);
 272 }
 273
 274 /*
 275  * This routine could be implemented by taking the addresses
 276  * written to the GATT, and flushing them individually.  However
 277  * currently it just flushes the whole table.  Which is probably
 278  * more efficent, since agp_memory blocks can be a large number of
 279  * entries.
 280  */
 281
 282 static void amd_irongate_tlbflush(struct agp_memory *temp)
 283 {
 284         writel(1, amd_irongate_private.registers+AMD_TLBFLUSH);
 285         readl(amd_irongate_private.registers+AMD_TLBFLUSH);     /* PCI Posting. */
 286 }
 287
 288 static int amd_insert_memory(struct agp_memory *mem, off_t pg_start, int type)
 289 {
 290         int i, j, num_entries;
 291         unsigned long __iomem *cur_gatt;
 292         unsigned long addr;
 293
 294         num_entries = A_SIZE_LVL2(agp_bridge->current_size)->num_entries;
 295
 296         if (type != 0 || mem->type != 0)
 297                 return -EINVAL;
 298
 299         if ((pg_start + mem->page_count) > num_entries)
 300                 return -EINVAL;
 301
 302         j = pg_start;
 303         while (j < (pg_start + mem->page_count)) {
 304                 addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
 305                 cur_gatt = GET_GATT(addr);
 306                 if (!PGE_EMPTY(agp_bridge, readl(cur_gatt+GET_GATT_OFF(addr))))
 307                         return -EBUSY;
 308                 j++;
 309         }
 310
 311         if (mem->is_flushed == FALSE) {
 312                 global_cache_flush();
 313                 mem->is_flushed = TRUE;
 314         }
 315
 316         for (i = 0, j = pg_start; i < mem->page_count; i++, j++) {
 317                 addr = (j * PAGE_SIZE) + agp_bridge->gart_bus_addr;
 318                 cur_gatt = GET_GATT(addr);
 319                 writel(agp_generic_mask_memory(agp_bridge,
 320                         mem->memory[i], mem->type), cur_gatt+GET_GATT_OFF(addr));
 321                 readl(cur_gatt+GET_GATT_OFF(addr));     /* PCI Posting. */
 322         }
 323         amd_irongate_tlbflush(mem);
 324         return 0;
 325 }
 326
 327 static int amd_remove_memory(struct agp_memory *mem, off_t pg_start, int type)
 328 {
 329         int i;
 330         unsigned long __iomem *cur_gatt;
 331         unsigned long addr;
 332
 333         if (type != 0 || mem->type != 0)
 334                 return -EINVAL;
 335
 336         for (i = pg_start; i < (mem->page_count + pg_start); i++) {
 337                 addr = (i * PAGE_SIZE) + agp_bridge->gart_bus_addr;
 338                 cur_gatt = GET_GATT(addr);
 339                 writel(agp_bridge->scratch_page, cur_gatt+GET_GATT_OFF(addr));
 340                 readl(cur_gatt+GET_GATT_OFF(addr));     /* PCI Posting. */
 341         }
 342
 343         amd_irongate_tlbflush(mem);
 344         return 0;
 345 }
 346
 347 static struct aper_size_info_lvl2 amd_irongate_sizes[7] =
 348 {
 349         {2048, 524288, 0x0000000c},
 350         {1024, 262144, 0x0000000a},
 351         {512, 131072, 0x00000008},
 352         {256, 65536, 0x00000006},
 353         {128, 32768, 0x00000004},
 354         {64, 16384, 0x00000002},
 355         {32, 8192, 0x00000000}
 356 };
 357
 358 static struct gatt_mask amd_irongate_masks[] =
 359 {
 360         {.mask = 1, .type = 0}
 361 };
 362
 363 static struct agp_bridge_driver amd_irongate_driver = {
 364         .owner                  = THIS_MODULE,
 365         .aperture_sizes         = amd_irongate_sizes,
 366         .size_type              = LVL2_APER_SIZE,
 367         .num_aperture_sizes     = 7,
 368         .configure              = amd_irongate_configure,
 369         .fetch_size             = amd_irongate_fetch_size,
 370         .cleanup                = amd_irongate_cleanup,
 371         .tlb_flush              = amd_irongate_tlbflush,
 372         .mask_memory            = agp_generic_mask_memory,
 373         .masks                  = amd_irongate_masks,
 374         .agp_enable             = agp_generic_enable,
 375         .cache_flush            = global_cache_flush,
 376         .create_gatt_table      = amd_create_gatt_table,
 377         .free_gatt_table        = amd_free_gatt_table,
 378         .insert_memory          = amd_insert_memory,
 379         .remove_memory          = amd_remove_memory,
 380         .alloc_by_type          = agp_generic_alloc_by_type,
 381         .free_by_type           = agp_generic_free_by_type,
 382         .agp_alloc_page         = agp_generic_alloc_page,
 383         .agp_destroy_page       = agp_generic_destroy_page,
 384         .agp_type_to_mask_type  = agp_generic_type_to_mask_type,
 385 };
 386
 387 static struct agp_device_ids amd_agp_device_ids[] __devinitdata =
 388 {
 389         {
 390                 .device_id      = PCI_DEVICE_ID_AMD_FE_GATE_7006,
 391                 .chipset_name   = "Irongate",
 392         },
 393         {
 394                 .device_id      = PCI_DEVICE_ID_AMD_FE_GATE_700E,
 395                 .chipset_name   = "761",
 396         },
 397         {
 398                 .device_id      = PCI_DEVICE_ID_AMD_FE_GATE_700C,
 399                 .chipset_name   = "760MP",
 400         },
 401         { }, /* dummy final entry, always present */
 402 };
 403
 404 static int __devinit agp_amdk7_probe(struct pci_dev *pdev,
 405                                      const struct pci_device_id *ent)
 406 {
 407         struct agp_bridge_data *bridge;
 408         u8 cap_ptr;
 409         int j;
 410
 411         cap_ptr = pci_find_capability(pdev, PCI_CAP_ID_AGP);
 412         if (!cap_ptr)
 413                 return -ENODEV;
 414
 415         j = ent - agp_amdk7_pci_table;
 416         printk(KERN_INFO PFX "Detected AMD %s chipset\n",
 417                amd_agp_device_ids[j].chipset_name);
 418
 419         bridge = agp_alloc_bridge();
 420         if (!bridge)
 421                 return -ENOMEM;
 422
 423         bridge->driver = &amd_irongate_driver;
 424         bridge->dev_private_data = &amd_irongate_private,
 425         bridge->dev = pdev;
 426         bridge->capndx = cap_ptr;
 427
 428         /* 751 Errata (22564_B-1.PDF)
 429            erratum 20: strobe glitch with Nvidia NV10 GeForce cards.
 430            system controller may experience noise due to strong drive strengths
 431          */
 432         if (agp_bridge->dev->device == PCI_DEVICE_ID_AMD_FE_GATE_7006) {
 433                 u8 cap_ptr=0;
 434                 struct pci_dev *gfxcard=NULL;
 435                 while (!cap_ptr) {
 436                         gfxcard = pci_get_class(PCI_CLASS_DISPLAY_VGA<<8, gfxcard);
 437                         if (!gfxcard) {
 438                                 printk (KERN_INFO PFX "Couldn't find an AGP VGA controller.\n");
 439                                 return -ENODEV;
 440                         }
 441                         cap_ptr = pci_find_capability(gfxcard, PCI_CAP_ID_AGP);
 442                         if (!cap_ptr) {
 443                                 pci_dev_put(gfxcard);
 444                                 continue;
 445                         }
 446                 }
 447
 448                 /* With so many variants of NVidia cards, it's simpler just
 449                    to blacklist them all, and then whitelist them as needed
 450                    (if necessary at all). */
 451                 if (gfxcard->vendor == PCI_VENDOR_ID_NVIDIA) {
 452                         agp_bridge->flags |= AGP_ERRATA_1X;
 453                         printk (KERN_INFO PFX "AMD 751 chipset with NVidia GeForce detected. Forcing to 1X due to errata.\n");
 454                 }
 455                 pci_dev_put(gfxcard);
 456         }
 457
 458         /* 761 Errata (23613_F.pdf)
 459          * Revisions B0/B1 were a disaster.
 460          * erratum 44: SYSCLK/AGPCLK skew causes 2X failures -- Force mode to 1X
 461          * erratum 45: Timing problem prevents fast writes -- Disable fast write.
 462          * erratum 46: Setup violation on AGP SBA pins - Disable side band addressing.
 463          * With this lot disabled, we should prevent lockups. */
 464         if (agp_bridge->dev->device == PCI_DEVICE_ID_AMD_FE_GATE_700E) {
 465                 u8 revision=0;
 466                 pci_read_config_byte(pdev, PCI_REVISION_ID, &revision);
 467                 if (revision == 0x10 || revision == 0x11) {
 468                         agp_bridge->flags = AGP_ERRATA_FASTWRITES;
 469                         agp_bridge->flags |= AGP_ERRATA_SBA;
 470                         agp_bridge->flags |= AGP_ERRATA_1X;
 471                         printk (KERN_INFO PFX "AMD 761 chipset with errata detected - disabling AGP fast writes & SBA and forcing to 1X.\n");
 472                 }
 473         }
 474
 475         /* Fill in the mode register */
 476         pci_read_config_dword(pdev,
 477                         bridge->capndx+PCI_AGP_STATUS,
 478                         &bridge->mode);
 479
 480         pci_set_drvdata(pdev, bridge);
 481         return agp_add_bridge(bridge);
 482 }
 483
 484 static void __devexit agp_amdk7_remove(struct pci_dev *pdev)
 485 {
 486         struct agp_bridge_data *bridge = pci_get_drvdata(pdev);
 487
 488         agp_remove_bridge(bridge);
 489         agp_put_bridge(bridge);
 490 }
 491
 492 /* must be the same order as name table above */
 493 static struct pci_device_id agp_amdk7_pci_table[] = {
 494         {
 495         .class          = (PCI_CLASS_BRIDGE_HOST << 8),
 496         .class_mask     = ~0,
 497         .vendor         = PCI_VENDOR_ID_AMD,
 498         .device         = PCI_DEVICE_ID_AMD_FE_GATE_7006,
 499         .subvendor      = PCI_ANY_ID,
 500         .subdevice      = PCI_ANY_ID,
 501         },
 502         {
 503         .class          = (PCI_CLASS_BRIDGE_HOST << 8),
 504         .class_mask     = ~0,
 505         .vendor         = PCI_VENDOR_ID_AMD,
 506         .device         = PCI_DEVICE_ID_AMD_FE_GATE_700E,
 507         .subvendor      = PCI_ANY_ID,
 508         .subdevice      = PCI_ANY_ID,
 509         },
 510         {
 511         .class          = (PCI_CLASS_BRIDGE_HOST << 8),
 512         .class_mask     = ~0,
 513         .vendor         = PCI_VENDOR_ID_AMD,
 514         .device         = PCI_DEVICE_ID_AMD_FE_GATE_700C,
 515         .subvendor      = PCI_ANY_ID,
 516         .subdevice      = PCI_ANY_ID,
 517         },
 518         { }
 519 };
 520
 521 MODULE_DEVICE_TABLE(pci, agp_amdk7_pci_table);
 522
 523 static struct pci_driver agp_amdk7_pci_driver = {
 524         .name           = "agpgart-amdk7",
 525         .id_table       = agp_amdk7_pci_table,
 526         .probe          = agp_amdk7_probe,
 527         .remove         = agp_amdk7_remove,
 528 };
 529
 530 static int __init agp_amdk7_init(void)
 531 {
 532         if (agp_off)
 533                 return -EINVAL;
 534         return pci_register_driver(&agp_amdk7_pci_driver);
 535 }
 536
 537 static void __exit agp_amdk7_cleanup(void)
 538 {
 539         pci_unregister_driver(&agp_amdk7_pci_driver);
 540 }
 541
 542 module_init(agp_amdk7_init);
 543 module_exit(agp_amdk7_cleanup);
 544
 545 MODULE_LICENSE("GPL and additional rights");